1. Podcasts
  2. Intelligent Machines (Audio)
  3. IM 892: Pants-based Benchmarks - GitHub's Agent Revolution
Intelligent Machines (Audio)

15 Oct 2026 · 2h 32m · 54 chapters

Episode artwork

IM 892: Pants-based Benchmarks - GitHub's Agent Revolution

In short

GitHub’s “agent revolution” for software development, focusing on how AI coding agents change GitHub usage, DevOps workflows, and open-source contribution dynamics. The episode also discusses “pants-based benchmarks” (as a segment theme), plus GitHub Universe preparations and an Origin sponsor segment about AI agent observability.

Guests (backgrounds)

  1. Martin Woodward: VP for Developer Relations at GitHub, joining from Northern Ireland. Preparing for GitHub Universe in San Francisco (Fort Mason). Background includes work at Microsoft before/around GitHub’s acquisition; he helped create Microsoft’s GitHub presence for TypeScript and later joined GitHub after the acquisition.
  2. Jeff Jarvis: Emeritus Professor of Journalistic Innovation at Craig Newmark Graduate School; author of Hot Type.
  3. Paris Martineau: Investigative journalist for Consumer Reports (Yonkers, NY).

Key claims

  • Agent-assisted coding has driven major growth on GitHub: merged pull requests tripled (about 40M to 130M/month), and ~50% of PRs are agent-assisted.
  • Better DevOps practices (tests, security scanning, automation) enable safer, faster agent workflows.
  • “Slop in, slop out”: reliability depends on constraints and safeguards.
  • Downsides: maintainers face inundation of low-quality PRs from first-time coders; GitHub added controls to limit PR suggestions/noise.
  • Origin sponsor claim: companies need endpoint AI observability because agents can bypass intended permission checks.

Notable examples

  • Martin describes GitHub Universe badges: an RP2350-based, hackable badge with MicroPython, accelerometer, and an agents.md file that agents can mount as USB mass storage.
  • Personal anecdote: an AI agent created multi-OS code and used GitHub CI/CD automatically.
  • Origin example: Codex launched Claude Code with dangerously skipped permissions to finish a task without proper file access.

Written by AI. May contain mistakes.

Chapters

0:43

Introducing Martin Woodward

Introduction and background of guest Martin Woodward from GitHub

“the latest in AI and robotics and all the smart things all around us.”
2:20

GitHub Universe and Badge Innovations

Discussion about GitHub Universe event and badge technology

“Well, it's very dark Northern Ireland right now.”
3:38

AI and Agent Technology in Development

Exploration of AI's impact on coding and agent technology

“It's not a touchscreen, which everybody goes to, especially the kids nowadays, because they just assume every screen is touchable.”
6:15

Coding with AI Assistance

Discussion on the transformation of coding with AI tools

7:20

GitHub's Role in the AI Revolution

Analysis of how GitHub has become central to AI development

“I actually, this is one of the fun things about this.”
8:25

Insights on Microsoft's Acquisition of GitHub

Insights from Martin on the acquisition of GitHub by Microsoft

“And then November of last year was obviously a big change.”
10:30

GitHub's Evolution and Current Offerings

Discussion on the evolution of GitHub's services and tools

“You just talk to the thing and it writes a code for you.”
14:01

The Evolution of GitHub and DevOps

Learn about GitHub's transformation and the significance of DevOps in modern development.

“were absolutely supporting Linux yeah by the way Richard Campbell says hi he says oh yeah no Oh, I love one.”
16:44

The Impact of AI on DevOps

Discover how AI integration is reshaping the DevOps landscape and enhancing productivity.

“We found actually putting the kind of rules in place to do DevOps well for humans also means that you can do AI better.”
19:24

Using AI for Creative Processes

Explore the balance between using AI for implementation versus creative thinking.

“But our typical setup for when you're using agents, we've got like three interfaces we use and like all the AI sort of coding agents have this kind of thing.”
Show all 54 chapters
21:48

Hacking and Creativity with Furbies

Hear about innovative coding projects involving hacked Furbies and their creative applications.

25:14

The Future of Personal Software Development

Understand the potential for personal software tools and their impact on individual creativity.

“And then I've got a friend of mine, Helen, who's one of the WordPress committers who also happens to be a concert pianist.”
28:00

Navigating the Challenges of Coding

The discussion revolves around the challenges and evolution in coding practices, highlighting the role of AI models.

30:00

Embracing Change in the Coding Landscape

The hosts discuss how coding has evolved with AI, addressing concerns about job security and the changes experienced by developers.

“But the more every time we have made coding easier, you know, when we went from that, what's that, an Altair 8800 behind you?”
33:00

The Rise of New Coders and Open Source Challenges

An exploration of the increasing number of new coders and the challenges faced by open source maintainers due to an influx of contributions.

“Just, what is the advice you give to coders, to the pros who are saying, I can't, I don't, I can't live in this new world.”
36:20

AI Tools in Code Review and Maintenance

The conversation focuses on how AI tools assist in code reviews and help maintain quality in the development process.

“kind of population yeah the number of first time coders is rising massively the number of users every single month is just exponentially more.”
38:40

Disruption and Opportunity in Tech

The hosts discuss the disruptions caused by AI in the tech industry and how to embrace these changes for future opportunities.

“But that's why DevOps helps too, because what we did with DevOps is we did tests, we did coding checks, we did security checks and things, all before you accept the change.”
40:00

Personal Reflections on AI and Development

A reflective conversation on the impact of AI tools in coding practices and the evolution of developer workflows.

“I think what we're seeing is a huge opportunity.”
42:06

Preparing for GitHub Universe

The hosts discuss the upcoming GitHub Universe event in San Francisco, including atmosphere and ticket availability.

“team is uh as a soccer team they're called the red and greens and uh because all of the i want to make sure they use tests, red, green, blue tests all the way, baby.”
48:59

Updates from Recent Events

The hosts share personal experiences from recent events, including a corn maze and sports discussions.

“Very much a needed service that we didn't even know we needed until recently.”
54:45

Google's Latest AI Release: Argon

Discussion on Google releasing Argon, its performance benchmarks, and comparisons with other models.

56:01

Benchmarking AI Models and Security Audits

Learn about the process of benchmarking AI models and conducting security audits for prompt injection vulnerabilities.

“And I thought, you know what would be great is if every day before the show, the AI could just go out and do that for me.”
1:00:27

AI-Assisted Routine: Automating Ad Preparation

Discover how AI can automate routine tasks like preparing advertisements for shows.

1:06:05

Prompting AI: Evolving Interaction Techniques

Explore the evolution of AI prompting techniques and the importance of clear communication.

“These, these, uh, this first journey is a creator of Cloud Code.”
1:10:04

Exploring AI Communication and Understanding

Learn about effective prompting strategies for AI and the evolution of AI communication.

1:19:11

Concerns About AI Self-Replication

Discussion on the potential dangers of AI models becoming self-replicating and the implications of this technology.

1:23:48

Introducing Muse: A New AI Agent

Discussing the emergence and features of Muse, a competitive AI agent developed by Meta.

“Thank you so much for supporting Intelligent Machines.”
1:24:00

Meta's Muse and Its Safety Concerns

Exploring the rise of Muse as a leading AI agent app and discussing safety concerns raised about it.

“Muse, though, has quickly leaped to the top of the charts.”
1:26:37

Personal Experiences with AI Agents

Sharing personal experiences of using AI agents like Muse and the implications of data sharing.

“And the other reason is the reason I spent so much time and money on local AI is so that I would get those capabilities on a local machine.”
1:28:44

Using Muse for Personalized Recommendations

Discussion on how Muse provided tailored recommendations for personal style without using personal data.

“Specifically, I want like stuff kind of like my nails, like saturated tones, like bright stuff.”
1:32:28

Comparing AI Agents and Chatbots

Differentiating between the functions of AI agents and traditional chatbots in providing ongoing support.

“Because one, I just didn't want it to message all of my friends.”
1:36:05

Insights into Meta's Data Utilization

Examining how Meta uses data from its AI agents to enhance user engagement and targeted advertising.

“I'm guessing meta is probably just text files.”
1:38:02

Discussing AI Insights and Personalization

Exploration of how AI systems like Muse understand user preferences.

“Or actually, obviously, Jeff has visual problems.”
1:44:02

Model Training and User Profiles

A conversation about model training and insights from Amazon user profiles.

“Prefers Lego for building block systems.”
1:47:44

Helix Sleep Advertisement

An advertisement for Helix Sleep, promoting their mattresses and benefits.

“Now, what kind of mattress are you looking for?”
1:51:37

Preparing for Family Visit and Cat Issues

Discussion about preparing for a family visit and the challenges with a pet cat.

“to deceive me uh which is i'm so sleepy i don't know you put the little pill in her mouth and you gotta get her to swallow it.”
1:52:00

Gizmo's Halloween Adventures

Learn about the humorous challenges of traveling with a pet cat during Halloween.

“And then I see, I let her go and she spits the pill up.”
1:52:45

AI's Mathematical Breakthroughs

Discuss recent advancements in AI's ability to solve complex mathematical problems.

1:55:08

Math and AI's Future

Explore the implications of AI's capability to produce mathematical proofs that may outpace human understanding.

1:55:45

The Value of Mathematical Journey

Examine the philosophical debate on the importance of the process versus the result in mathematics.

“The thing that got me about what I did listen to in the interview was how egocentric it was.”
1:58:26

Shifts in Job Relevance

Analyze how AI is affecting job roles in various fields, including mathematics and coding.

“Coding has, I think no one would doubt this, fallen to AI.”
2:00:12

The Importance of Human Creativity

Discuss the role of human creativity in fields being impacted by AI advancements.

“Look, Darren is not capable of making this illustration, but he is capable of thinking of a clever premise for this illustration.”
2:02:33

Ukraine's Impact on Bot Traffic

Learn about the recent drone strikes in Ukraine and their effects on online bot traffic.

“You just put money there that you hope will be there, unlike the last time Elon Musk tried to make X.com the site for all your banking and finance.”
2:04:57

Character AI Bankruptcy

Discuss the recent bankruptcy of Character AI and the challenges it faced.

“I don't know the story of that, actually.”
2:06:00

AI Investment and Prize Concerns

Discussion on AI investments and the implications of prize systems.

“money in the entire time of its existence.”
2:07:08

OpenAI Security Incident

Details about a recent security incident involving an OpenAI engineer.

“And so I picked on the wrong prize to say that.”
2:09:04

Dramatic Security Incident Narration

A humorous recount of the security incident and its implications.

“that was Dario Mone looking behind the TV in the OpenAI conference room.”
2:14:13

Nick Cage and AI in Film

Discussion on Nick Cage's remarks regarding AI's influence in filmmaking.

“He also said he might be doing true detective, which would be interesting.”
2:16:50

AI Tools for Reverse Engineering

Exploration of new AI tools that facilitate reverse engineering.

“uh this is a yes it's an ai website uh in a meeting i'm in a meeting.com where you get they say plausible meeting audio to put behind your phone call what what are you in are you in an IT incident.”
2:18:56

Casual Conversation and Show Wrap-up

Light banter and casual discussions as the show concludes.

“I'm going to go downstairs and let my mother in my building.”
2:20:06

Discussion on a Unique Podcast Format

Learn about an innovative podcast format set in a fake pub and its engaging content.

2:21:15

Jeff Jarvis and His New Book Series on AI

Discover Jeff Jarvis's new book series exploring AI's societal impacts and reflections.

“He's now doing curricula for Montclair State University, Montclair, New Jersey, and SUNY Stony Brook.”
2:24:15

Exploring Current and Future AI Developments

Get insights into upcoming AI models and the implications of hardware advancements.

2:27:25

Memory Market Predictions and Concerns

Understand predictions about the memory market and potential price fluctuations.

A man walking through an airport, asking his phone a question

Ask about this episode

Get the gist, explore an idea or ask about something said.

Claude

On Intelligent Machines (Audio), what are the key points of this episode?

Searching your listening

Transcript

Automatic transcript. May contain errors.

0:00Martin Woodward: It's time for Intelligent Machines. Jeff Jarvis is here. Paris Martineau is here. We've got a great guest. Martin Woodward is VP for Developer Relations at GitHub. AI has changed coding forever and nowhere more dramatically than on GitHub. He's getting ready for GitHub Universe. We'll talk to Martin and all the AI news, including new models from Google. Pretty good model, too. And Meta. Coming up, Intelligent Machines. podcasts you love from people you trust this is twit this is intelligent machines with jeff jarvis and paris martineau episode 892 recorded wednesday october 14th 2026 pants based benchmarks it's time for intelligent machines the show we cover the latest in AI and robotics and all the smart things all around us.

0:55Martin Woodward: Do I have to say the show where we cover super intelligence and robotics? I hope not. You don't, actually. No, no, we would leave you. We would be gone if you did that. That is Paris Martineau, who is an investigative journalist for the Consumer Reports brand in Yonkers, New York. It's true. I may not be in Yonkers, but my job legally is there. And that's a good thing when it comes to New York State taxes, apparently. It is not. It's a bad thing. It is the opposite. It's the opposite. But that's fine. You don't get a deduction, the Yonkers deduction? No, I have to pay Yonkers tax. Oh, man. And it's very complicated.

1:33Oh, man. Oh, man, indeed. But such are things.

1:37Martin Woodward: There are trials and tribulations getting to live in the greatest city in the world. It's true. It's just part of it. Jeff Jarvis is also here. He is the Emeritus Professor of Journalistic Innovation at the Craig Newmark Graduate School. Craig, Craig, Craig, Newmark. Author of Hot Type. Get your hot type here. Brand new book about the magnificent machine that gave birth to mass media. You got four Ms in that little subtitle. I didn't even notice that before. I love alliteration. Me too. I'm a kind of sucker for it. Hey, let's say hello to our guest for this episode. I'm very excited to have Martin Woodward on.

2:19Martin Woodward: He's Vice President of Developer Relations at GitHub. Joining us from Ireland. Southern Northern Ireland. Well, it's very dark Northern Ireland right now. But yeah. Thank you for staying up late with us. I appreciate it. Martin is also, thank you, because he's right now probably at the busiest time of the year because in two weeks it's GitHub Universe. Yeah. and that's going to be in san francisco at fort mason yeah yeah it's our big sort of developer get together the world's the world's fair of software we like to call it so it's uh for mason in san francisco christina warren all excited she's gonna be there of course our host on macbrick weekly and uh she's in developer relations in your department she says you're her boss's boss so i said i put in a good word if you think anybody's the boss of christina you've Clearly never matter.

3:08Martin Woodward: Well, she always ferried me badges from GitHub Universe, which you very kindly provided. What does this year's badge look like? Oh, do you want a sneak preview? Can we show it? It'll be a world exclusive. No, no, no, go on. You can show it. Let's put it up. We might as well. I'll ask. There you go. It's first of all, first thing I noticed, it's in color this year. It's in color now, not just black and white. Yeah, we're moving on in technology. And we have a color screen as well. It's got touch controls. Yeah, so, yeah, it's got controls here. It's not a touchscreen, which everybody goes to, especially the kids nowadays, because they just assume every screen is touchable.

3:46It's an RP2350 chip, so, you know, quite a high-power microprocessor. Wow. And you can even see, I don't even sort of see there, like Mona's face will move, you know, because it's got an accelerometer in it, which is very, like, purely so we could flip the screen so that when you look at your badge yourself, it automatically flips the screen for you and so you don't have to like squint or turn your head around or anything like that and then when you're in that mode you can actually use these capacitive controls as well so to sort of move around the screen like a little this is a normal thing for you guys or do you just decide to pop off this year no no we've been doing it for I think Leo's got a few of them actually I have they're all screen laden badges I was just looking for them They're in my bookshelf, but I misplaced it.

4:32But yeah. It was the e-ink because it did need a battery around your neck. And it saved us like failing out because you have to do all sorts of risk assessments when you stick 4 ,000 LiPo cells on people.

4:43Martin Woodward: But that one has a battery because that's a Raspberry Pi in there. Yeah. RP2350. So yeah, it's the top end of their microprocessors. It's not a Pi itself, but it's programmable, right? People can. Yeah. Runs MicroPython and deliberately hackable. So, you know, obviously, because, you know, we've been doing Copilot forever, like kind of the first of these AI agents that kind of changed how we code. And we have an agents.md file, and it just... Wait a minute, you have on the storage, you have agents.md? We have agents.md, and we have it in an associated repo, and it can mount as USB mass storage.

5:22And then you just point your agent at it, and it knows all about MicroPython. it knows how to like deploy stuff to the micropython so what does the agent md say does it say hey this is a little card that has this kind of screen and that kind of stuff or yeah this is the pinout this is where all the because of like things like the accelerometer and the capacitive sensors and things they're not you know like in any standard board they're not retail um and so we the agents md tells you what pinouts they're on um all the models tend to know about micropython nowadays so we just let it know it's micropython and what version it's running and things um but yeah you just give the agent the appropriate skills and constraints and you know wonderful things happen the more constraint if you give it good constraints then that's where you get art you know i mean all all art has constraints so during the universe people are reprogramming it to show off their skills yeah yeah exactly and you know during this they come in with the modified badge that's hysterical yeah no yeah well we have uh we've got a surprise for the keynote as well we've got a sort of interactive game to play and things it's going to be good oh how fun oh how fun

6:31Martin Woodward: github has really been the hub if you will of the ai revolution yeah i was kind of shocked when i first started using clawed code that uh it said oh let me put this on github and then the other thing it did which shocked me the first program i wrote was a little rss reader as one does a 2e rs rest reader and uh not only did it put it on github and i already had a paid account so that was fine you know gh auth login and i'm off to the races and uh and then it said oh and i by the way i made a linux mac and windows version for you i said what it it did cicd using microsoft So, I mean, it did using the GitHub CICD capability.

7:19Martin Woodward: And I was blown away. I actually, this is one of the fun things about this. You learn stuff. So I'm learning. I've always used Git and GitHub, but I'm learning a lot about it that I didn't know from my agent, which is hysterical. You got overwhelmed, though, didn't you? Yeah, it's been very, very busy with, you know, it's been sort of rapid acceleration on the platform uh called a hockey stick yeah well there's hockey sticks and there's like brick walls of graphics the sort of the graph just goes it's not like gently up into the right it's you know what i mean so i kind of knew when it started happening because the downtime went up now which i think you've got under control it seems to be very reliable again it's got a lot better obviously we're doing you know we continue to work like it's the first priority if we're not up that's you know that impacts us as much as it impacts everybody else it's just not acceptable so um yeah we've been working you know uh very very hard but the the workload that um agents have been putting on the platform um is just as i say i'll grab some numbers a sec but it's it's just gone through the you know the commits are just going up all the time the number of pull requests are going up all the time right um we know that because people are saying yeah i get 2 000 prs every day and that where are they going they're going to github yeah yeah the um this and and it's really the github's kind of become this sort of um so we the agentic coding kind of started with copilot and we kind of proved out that was a good thing and we kind of proved out that models could code.

9:05And then that was it for a while. And then November of last year was obviously a big change. We had not just sort of Claude Sonnet, but also the OpenAI latest round of models as well. Those new round of models, the amount of time that agents can code without human interaction just increased exponentially. and therefore the amount of change that they can do is increased exponentially. And so we've seen that in the platform. But we've also seen the number of people that can code, that are like you, Leo, you've always known how to code, but you also know how hard it is to code. And so you don't code that much, but now you can code again because you've got help.

9:50But you know how to do it.

9:51Martin Woodward: All the ideas that I have I can now do. Yeah, I kind of liken it to, at least for me, because I'm not a 10X coder. I don't do it for a living. It's a hobby. And a hobby, by the way, I really enjoy. But I compare it to like needlepoint compared to using a sewing machine. I mean, it was very slow and somewhat tedious. There was a lot of typing involved. And then, you know, with Copilot, all of a sudden you're using VS Code. You press tab and it says, oh, let me write that routine for you. Wow, that was pretty cool. but now you don't even open VS Code many times. You just talk to the thing and it writes a code for you.

10:35Martin Woodward: Now, you started at Microsoft in DevOps, right? Yeah. So I was saying in the pre-show to Paris and Jeff, I did a startup that, to be honest, today probably wouldn't happen. Like the startup was kind of reverse engineering a protocol. And nowadays you just point an LLM at that. Right. Right. They know it. Yeah. Yeah. But we did this. We started Microsoft, and then I did a lot of the engineering systems inside of Microsoft to help them kind of modernize and get a bit more agile. Oh, interesting. Oh, okay. And then I helped kind of Microsoft get better at doing open source. You know, there was a whole cancer thing that we try not to talk about.

11:19And then - Mr.

11:19Martin Woodward: Balmer had a little thing about Linux we won't mention. But obviously - So, but you were, now this is before Microsoft bought GitHub. Yeah. But you created Microsoft's GitHub account. I did. It ran on my Amex for a while. What? You were paying for Microsoft's GitHub account? Yeah. It got very, very expensive very, very quickly. So it quickly went on expenses. But yeah. That's hysterical. So was this their idea? No, you must have thought of this. So we should really have an account on. Well, no. So I actually, we did have a, there was a forge that existed called CodePlex. And I looked after that, which was Microsoft's kind of open source forge.

12:00Okay. But then it was actually, you know, it was with TypeScript chatting with Anders Halseberg and a few other folks. Anders is amazing. They were like, hey, we're going to do TypeScript, you know, and I was like, yeah, whatever. Anders, that's never going to take off. Who's going to use this JavaScript language from Microsoft? that's clearly never going to catch on. Obviously now it's the most popular language in the world. But the, and what are the, it's the most popular language for agents to use actually. It's the most popular language. I think OpenClaw's written in TypeScript, isn't it? Pi is in there.

12:34Yeah, yeah, yeah, yeah, yeah, it is. So, yeah, and so we needed something and we knew like the TypeScript community wouldn't, we need a home. So we actually created the account and TypeScript was the first project we put into the Microsoft GitHub org, and then it's just grown from there. And Microsoft went from being where it was at to being, you know, the top contributor to open source that is today. And then because it did so much open source, it needed GitHub as part of its engineering systems. And so, you know, we talked to the GitHub team and, yeah, the GitHub team came over. And that's when I joined the GitHub team then after the acquisition.

13:13So maybe you could take a little credit for the acquisition then. oh i think um yeah there's a few other people involved satya was definitely heavily involved in that one i got the price right within half a billion dollars which uh laughs at me about but yeah exactly yeah exactly so what's the cultural connection have been with the two how many years has it been now it's uh 2018 was when we when we did the so i remember there was a lot of about a mismatch between microsoft and github especially because of the you know linux i think there was in the community but I think that was because a lot of people I mean you follow the you follow people and you kind of do Windows weekly and things you you kind of knew that um you know the Microsoft had already moved on but a lot for a lot of people oh yeah I did absolutely yeah Microsoft that it was the old Microsoft in their head they had Linux on Azure I mean they

14:05Martin Woodward: were absolutely supporting Linux yeah by the way Richard Campbell says hi he says oh yeah no Oh, I love one. Rich is a good friend. Yeah. Yeah, so that whole kind of time, people were definitely nervous, and Satya was very clear, judge us by our actions kind of thing. And if you think about GitHub back then, GitHub was just source control, essentially. And then now you have GitHub Actions, which your agent used to do CICD and build everything for Windows and Mac and Linux and things. We have - Wow. Yeah. We have security scanning because obviously the more code you create, the more likely you are to have security defects.

14:47So the more you need security tooling baked in. I use that as well. Yeah. Yeah. And that's really important because we have to give all that stuff. We use, you know, we sell it to enterprises, and people in the enterprise use these tools to build software. But we make sure we get those tools into the open source community as well so that people can use them for free, which just makes the world better for everybody because, you know, then everybody's software's a little bit more secure.

15:12Martin Woodward: Yeah, for the longest time, I mean, I didn't really feel the need to pay for GitHub. It was only once when I got past like about 20 repos, almost all of them vibe coded. I thought, you know, I'm using this pretty heavily at this point. I actually use it for backup. I use it for my blog, my Obsidian. I mean, I can go on and on and on and on. And you can see how long I've been using it, because that is the oldest picture of me that's still in existence. Look at that. But GitHub is... Do you want to try something, Leo? If you go to myfirstcommit.com, this was a website we did live on stage a few universes ago, but it's probably still running.

15:53If you go to myfirstcommit.com, you can type your GitHub handle in, and it'll tell you when you first did a change request to GitHub.

16:01Martin Woodward: Oh, that's really interesting. 2017. There you go. So that's very, so you were in the, like GitHub didn't go live until April 2018. So you were in the beta. You were early. Wow. I didn't know. So high in the ear, man. Or grandpa. You joined before I did. You're a hardcore hacker. That's pretty funny. That's kind of wild. I almost don't believe it. But it must be true. GitHub would not lie. uh so as a as somebody who does devops i mean in the when you were doing it uh it was manual there's a lot of a lot of coding a lot of uh a lot of scripts a lot of bash shell that business has changed quite a bit hasn't it yeah we found um uh so devops for folks that don't know it's it's um developer and operations you know it's basically um moving with a lot more velocity coding a lot faster um you know working faster by being closer to your customer and getting things out quicker so github deploys you know hundreds of times a day and whereas in the old days like you know you used to deploy once every two years or something like that you know if you're building windows operating systems so um but to do devops and that's like how most modern kind of cloud software's built.

17:29We found actually putting the kind of rules in place to do DevOps well for humans also means that you can do AI better. You can be faster accelerated with AI because the same kind of safeguards you put in place to make sure humans can work at speed, you know, tests and safety checks and all this stuff and security checks that then allows your your agents to um work with your humans a lot faster so people who are good at devops tend to be good at using ai as well we've

18:05Martin Woodward: we've found across the industry yeah that kind of makes sense because you already have i think you have to have an architecture mindset uh what's helpful with ai is understanding at the higher level what's what the systems are and how they interact and how they go together of course you in the old days you also had to write the bash shell that would that would do all of that uh is it do people still think it's lazy to use ai for devops no that changed i think even i don't know like i don't know uh paris and jeff if you've seen some of the people you talk to as well there um it changed last november uh for even the most die hard of ai skeptics like we had all the numbers that showed you that you were more productive but people kind of like the old way um but nowadays you know people realize oh no if i if i work in partnership with a bunch of agents i am fast i can create more code more quickly is it as good is it as reliable is it as secure if you do it right if you put the right safeguards in if you put slop in you get slop out you know right if you just sort of say do some stuff then it'll it's a random machine so it's going to do some random stuff that day i mean what separates uh the slop approach for to doing it yeah what do you do it what is your yeah because i this is something that's fascinating to me everybody right now everybody's kind of i mean i don't work in an org so maybe if you're in a company you have policies and procedures but i'm kind of i think of all the people i talk to we're all just kind of figuring it out on our own and there are definitely ways to do to to get the ai to work better if you know what you're doing and for a while people said loops and then they said graphs what's what's your workflow look like how do you do it well some of this actually feels and feeds into the question i'm paris asked about like what's the difference you don't mean it's um what do humans provide like what's the what's the unique thing that we're good at and um it's the creativity the taste and the judgment and with the judgment i mean like the the computational thinking you talked about there, Leo, in terms of how do you pin systems together and all that sort of stuff.

20:17But our typical setup for when you're using agents, we've got like three interfaces we use and like all the AI sort of coding agents have this kind of thing. We have the command line, we have an app that we use, it's the Copile app, and then we have the IDE. And within that, we have a set of skills so that's not your skills jeff that's like skills for the agent that teach it about stuff i have none personally actually one of my skills is called plain english

20:54Martin Woodward: and that's a very useful skill because i don't because cloddish is hard to understand these days and i have to have i have to say talk to me like a saint bernard so that's a skill you have jeff that's uh it's how to put things complex it could be going out of style but yeah it's still there Well, one of the tricks is actually to use the agent to help you draft your skills. So I had my agent review. I've got like 20 years of blog posts. So I had my agent review my 20 years of blog posts and create a skill to sound like me. So, you know, use the word naughty, speak in an English accent, whatever.

21:30But don't use em dashes for a start. Don't, you know, speak with using British English, writing British English. and it kind of built a list of rules it observed from looking at my thinking which it now applies when it's helping me draft um blog posts and things like that the the thing i find though i actually use ai to help me implement i try not to use ai to help me think i don't know does that make sense so yeah so you you do the thinking ahead of time or i'll iterate with it but i don't like i actually don't like using ai to help me write because i find the process of writing is the process of thinking yes again paris and java are much better writers than i am so maybe but i like when i'm writing maybe it's just the way i learn or whatever but when i'm writing it down that's how i kind of construct the arguments in my head and that helps me do it but if i have a set of bullet points and a set of research i find the ai is very very helpful uh either fleshing that out or or researching a topic for me and condensing that information down and then i can do my thinking processes and like kind of get the do the creativity do the structure you know that sort of thing so is that how you hack a furby yeah yeah yeah that's the technique yeah that was actually to be honest that was pointing it at um have you seen wes boss and a bunch of people i if you follow wes online um whereas has got a project going out right now and where everybody's hacking their i've got one here their picture frames yep um yep the cheap frame 50 frames from amazon what are you putting on your hacked picture frame and what are you putting on your hacked furby oh well that's a great question so the my hacked furby was actually um i had it i have it do alerts for me i don't have it here but um it when so when you're doing the coding agent quite often you'll set off like parallel coding sessions and you'll have them running in the background i i quite often before i go to bed kick off the agent to do go do some stuff um or when i'm um you know if i'm in a conference call if i'm in a um if i'm in a boring conference call or a zoom call then i'll have it doing some coding turns out i'm not allowed to have it on during an interview with you lot but uh yeah there we go and then um uh copilot would notify me yeah thank you i I co-pilot would notify me when the agent was complete and it would just make like weird like noises.

24:07See Leo this is what you've got to do. You've got to replace. Leo has installed his agents all throughout his house so that they can scream at him while he's walking around and you should replace that with a sea of Furbies that make really haunting noises. I did buy that.

24:23Martin Woodward: I did order it hasn't come yet the micro duck that the hugging face is up. It It has a flapping lip and it'll walk around the house. So maybe I will make it sound like Eartha Kidd or something. And for last year's universe, we actually like maybe took my Furby habit a little bit too far. And I just go, I'll go grab one of them. Go grab one. Yeah. This isn't the one. This is the first time anybody's grabbed a Furby on this show. Oh, look. That's a blinged Furby. Yeah. That is very, very cool. A Wookie Furby. Oh, it's a Wookiee, of course. But the Furbies, yeah, I had a bit of a Furby habit, as it turns out.

25:08And for last year's universe, for the finale of the keynote I was doing, I sort of had 66 hacked Furbies attached to a Raspberry Pi and a keyboard. And then I've got a friend of mine, Helen, who's one of the WordPress committers who also happens to be a concert pianist. and so got her to play there because it was the day before Halloween. So, you know, perfect time to deploy 66 Furbies. Did they do the Furby Chorus? What did they do? Oh, it was, yeah, it was just, you know, dramatic organ music. It was great. But it was a keynote. So I got them to like give me smoke and lasers and all sorts. It was brilliant.

25:48And then it was ridiculous. My 66 Furbies that I'd bought off eBay. That's hysterical. So what are you excited about these days?

25:55Martin Woodward: i'm sure the copilot everything app is one thing we'll let you get a plug in for that yeah so yeah get up copilot app obviously that is the thing but the honestly the thing i'm most excited about um is the is being able to code again is the creativity that's being unlocked by being able to code and then also the numbers of people that are coming in that actually um are able to code again are able to create some stuff and solve their problems we're seeing new classes of application get built where you know people will build like an ephemeral application an application which just solves a problem for one person and normally you wouldn't ever bother coding that you know normally you only coding is when you do something twice but um this allows you to create stuff and it's just unlocking a whole new wave of creativity obviously there's a huge downside i'm sorry yeah this is the best thing that that idea of making something small and specialized you don't have to make microsoft word you don't have to make uh oh i don't know uh the adobe suite yeah you can make something specified to a need which somebody has done right yeah that adobe sword apparently didn't work the copy that someone made it's not great benito tried it he said oh really buttons don't work yeah but it'll get better i mean um but i yeah i think that's the least interesting thing to do i think one of the most interesting things to do is personal software that know what that you you would either you would write yourself a great time and expense or no one would write but suddenly you have little tools i have little tools all over the place that i wouldn't have otherwise like that was that's one of the kind of key things about devops and sysops is if you do something more than a few times you're supposed to write a script to automate it right but that's time consuming it's a lot of work now it's easy and that was always the promise wasn't it behind like a computer on every desktop or a computer in every pocket kind of thing it was always the promise that you could have stuff that was personal to you and you could work but a lot of times like writing code's hard and time consuming and um you know increasingly phones and tablets turned into consumption devices rather than creation devices i built my first ios app now because like i was trying to build ios apps and all that ns whatever stuff was like oh yeah and i just got you know it was just hard nobody wants to use uh well swift's not so bad but swift x code's god yeah exactly but i could just have like i haven't i had an app i needed i wanted something i wanted it to be a native ios app right and it turns out it for me and claude can write swift really well so actually all the agents can yeah well that so the thing about the the copight thing um it uses all the models and i don't think people talk about this enough is because it's like you've got the anthropic models you know the family so you've got sonnet you've got opus you know you have uh fable all that sort of thing and then you have the the codex models so um so from um you know from open ai so you have like gpt 5 6 you've got soul you've got astra you've got terror you've got all these sort of wacky model names that are out there nowadays like sort of you know 1950s cars with fins kind of thing but what we've actually found that's really really cool and then you've got you know deep seek you've got the open weight models you've got kimmy you've so you just have this long list of models you just pick from that's even better you can use all of them and this is why i really love doing this is one of the things because because we're the only people that can like because it's everyone's trying to like make github the center of the coding universe we're the other ones that's got all the models kind of right um in one harness which is so you could have sub agents be one model uh you could have yeah i and i like to do that but i can have to do it manually where i have one model plan and a different model family code and a third model family review so that they you know have overlap with their their gaps but you do that all within inside of copilot that's very just say rubber duck this for me um oh i love it it'll go talk to another model let me grab a rubber duck while i'm here there we go so i'm getting all the toys and jeff probably don't know that oh that's a big one but one of the ways coders often work is they have a little rubber duck on their desk and they uh and they use the rubber duck to bounce ideas off of to help them debug they call it rubber ducking and you have because if you say a problem out loud a problem shared is a problem halved yeah even if it's just a dumb home remember duck actually that's one of the skills i've learned is ask the model do you tell me what you understand from what i just told you because often it'll say it'll say something nonsensical and you say no no let me let's go back and forth on that where there are skills to working with models and i think a coder already somewhat has that mindset but i think there are also a lot of coders there's a lot of mathematicians right now in this same state are in grief over the loss of their craft you know the thing that they loved doing do you notice that or um so the the math thing's interesting as well you know we've obviously um so yeah with a bunch of papers and a bunch of solutions being published and i think the the of them i've got friends who are mathematicians the thing that kind of irks them more is less that it's been solved but it's less it's like that people don't care that it's been solved you know is that's what upsets them but it's um there's always this issue of like will ai you know it's the the jobs issue you know is ai going to take my jobs if right that's maybe the bigger concern isn't it yeah exactly and for for for coding in particular whenever we've had a we're just not seeing it in the data first of all we see the number of you know developers going up and up and up not going down, which is, you know, because more people can code now.

31:59But the more every time we have made coding easier, you know, when we went from that, what's that, an Altair 8800 behind you? When you went, oh, is it? That's right. I can't tell. Yeah, Altair. Yeah. When you went from that to the mainframe, when you went from the mainframe to client server, we always made coding easier. And the people who were in the old way of working were always kind of, oh, this, you know, they don't understand how to write assembly, like like Steve Gibson, you know what I mean? You don't know how to write assembly language. You're not a real coder. Even Steve uses different higher level languages.

32:33That's not fair. He bought a spark.

32:34Martin Woodward: He shocked me. He said, I'm going to have it do the UI because I hate UI. I said, there you go. Well, yeah, most people hate writing assembly. So we get it to do that instead. But yeah. Well, I think it's important to keep the part of it that you really love. Yeah. But I think it's also important if this is your profession, and I have friends, I'm sad to say, we're having a hard time adjusting to this new world, but you already have the skills that you could be very good at this. Just, what is the advice you give to coders, to the pros who are saying, I can't, I don't, I can't live in this new world.

33:12It's, you know, it's the way the world's going. So it's go learn it. Our jobs as developers is to, it's to be lazy and to solve problems in the easiest way possible. you know um and i see less like there was up until as i say november of last year there was definitely a lot more about reticence and now it's kind of people miss the craft um people miss that what they also miss is the pace has just changed like coding now is just so you know you're coding at 10 agents a time you constantly feel like you're behind a new model comes out every 12 minutes and you sort of, you know, and so people miss kind of some of that.

33:55And we had that same problem when we went from, so Jeff Parrish, when we used to code, we used to code in like a waterfall way. We used to like write down specifications and write down, you know, stuff and then we would like, is this right? And then we would ship it and it would take months and months and months. And then when we did DevOps and Agile, we iterate a lot more quickly. So we build something show it to you as a consumer of software and say hey is this better or worse did we make this better or worse yet and we make lots and lots of small changes and there was that back in the kind of 2000s there was this pace change and people noticed it then and we're noticing it now it's it's it's definitely you know a lot faster i got the stats for you as well yeah there's um The number of merged pull requests, like in August, it was 100, so across GitHub, 130 million pull requests merged per month.

34:54And if we look a year earlier, it was around about 40 million, you know what I mean? So it's tripled. It's more than tripled. Yeah, and a pull request is like a change, is a way you review the code, you know. And like 50 % of those pull requests now are kind of being agent assisted. So it's an agent's coming along and helping a human in creating that change. And that's how much software's changed in 12 months. Yeah, absolutely.

35:20Martin Woodward: I mean, you don't have a tripling without something like that. That's such a big difference. And, you know, I'm always looking at new models, as you mentioned, and it always is a link to a GitHub page. I know the model resides on Hugging Face, but it always starts at GitHub. That's where the recipes are. That's where the information is. That's where you point your agent. You say, go to GitHub, and they know GitHub. They understand it, and go from there. You have become really what Twitter was 20 years ago. Let's hope we say you like Twitter was 20 years ago. Yes, please. really the we used to call it the dial tone of the internet you are where people go to get connected Martin can I ask a quick question on that thanks to all of this ability for anybody can code how many people are coming to GitHub who previously didn't code are you seeing a growth in that kind of population yeah the number of first time coders is rising massively the number of users every single month is just exponentially more.

36:31So yes. One of the things that comes with downsides though, Jeff. So if you're a maintainer, like I talk to maintainers of open source software, open source software is how we share stuff around the world. If you're a maintainer of a popular package, it's got a lot easier for somebody to contribute code to you than it was before. and if you're starting out in coding you might think oh it's a you know I'm going to give a gift to this project I like I'm going to give it some code but you don't yet have the judgment as to is this code or not and so they can kind of one of the problems we've seen is kind of open source maintainers getting inundated with low quality code contributions we started calling them spam but it's not really spam because oftentimes it's well-intentioned.

37:25It's just not good. It's just low quality. And so we've had to... I'm a huge... I've been a passionate open source advocate for decades. This is like my... cut me open, I'll bleed open source sort of thing. But one of the things we had to add to GitHub recently was the ability for a maintainer to switch off what's called a pull request. So that's how people contribute code to you, they're at pull request um we've had to give them the ability to turn that off to turn it off temporarily to turn that off to people who haven't coded before to to limit how many times somebody can suggest a change to you um before you before you you accept it you know i mean so they can't say give you 10 changes um until you've accepted the first one those sorts of things and we've had to it's kind of limited how people into it's changed how people interact but we just have had to do it to kind of help maintainers deal with the noise now so i know this claude is always

38:24Martin Woodward: asking oh do you want to submit a pr i always say no i say no i'm sure they know about this problem i'm not going to bug them uh with this the good news is we're also getting better tools for reviewing prs yeah on the other side you know if you're getting 130 million pull requests in a month You need an agent to help you. But that's why DevOps helps too, because what we did with DevOps is we did tests, we did coding checks, we did security checks and things, all before you accept the change. And so now you automate that. You have all that stuff happen procedurally using GitHub Actions, which is the thing that Leo talked about.

39:03But then you also have your AI agent come in and do code reviews. And again, it comes in and sort of tells you, oh, this bit's changed. these they've introduced some new dependencies sneakily so even though they telling you the code does this actually what it does is steal crypto wallet from you so like you know it's sort of you can have the agent do those checks for you and it's always vigilant always on always watching so um that's been you know using ai to kind of fight some of the the downsides of the the sort of some of the slop is definitely how to win, you know.

39:41Martin Woodward: Well, we are in a time of disruption. Disruption always does two things. It disrupts, it takes the incumbents and unsettles them. And it makes those of us who are used to the old way a little uncomfortable. But at the same time, it also presents major opportunity. And I always want to focus on that. I think what we're seeing is a huge opportunity. And the right way to approach it is not to be afraid of it, but to embrace it and see how it can help you in your life. And I think that that's very clear. And GitHub is a great place and Copilot is a great place to start. Copilot really started this all.

40:21Martin Woodward: I mean, that was the first AI coding tool I used and most people used. I was user number six of Copilot. Isn't that great? Wow. It's 2021. one again it was one of those it's one of those typescript moments um there's there's times where i'm spectacularly wrong and that was one of them where the team were oh yeah we've got ai writing code i was like yeah this is going to be terrible and then i started using it and i was like oh no this has completely changed how i write code um i'm never going back so well and it benefited because it had all these great repositories it could read on github so there's there yeah and early on the the models weren't as good so when we started we did what's called it was code completion so basically like you could only write four or five yeah exactly you could only write four or five lines before it would go wrong right and so we took advantage of that and rather than trying to make it write a whole bit of code we just helped it complete the line of the code that you were doing and that was already kind of like it was magical yeah yeah oh you're writing a login script let me just write that little routine for you and it was like it still it still is magical even today when I'm using you know it I don't I tend to now use the IDE for kind of like the the small kind of detail you know if you'd like when you're doing an oil painting you paint the big areas first and then you come in and you do the fine areas I'll use the editor the IDE we call them as a as a doing the fine areas and I use my agent to do the broad strokes you know that um and kind of get me near most of the way there i'm ashamed to admit i haven't opened an ide in a long time really i don't look at the code anymore i mean uh why you know as long as you're testing what it does i have lots of testing results exactly yes my coding team is uh as a soccer team they're called the red and greens and uh because all of the i want to make sure they use tests, red, green, blue tests all the way, baby.

42:23Martin Woodward: And they even have a little fight song. I won't bother you with it. Martin, it's been a real pleasure meeting you. After all this time, I feel like we know each other. We have many mutual friends. And of course, you're legendary at GitHub. Thank you for spending a little time with us. I know it must be a crazy time for you. So take half an hour to talk to us in the middle of preparing for what will be a huge github universe in two weeks in san francisco at fort mason how many people do you expect at that about four thousand so it's it's not a huge venue but it's um it's intimate and uh yeah yeah but it's all people like you like us well that's the best bit it's you're full of yeah full of nerds and so you just go around and it's just great fun and everyone's hanging the badges everyone's learning and it's just a good atmosphere as well because the fort mason's it's not like moscone in Moscow is a great conference venue, but Fort Mason is a completely different environment.

43:16Oh, I far prefer it. You're outside. You just get out and talk to everybody.

43:20Martin Woodward: It's not too late to go. Are you all full? We've got some tickets left, I think, now. It's getting pretty full. But, yeah, we've got a few tickets left. And you get your conference badge if you do that. That's the reason for going. No, it's more than that. But, yeah, and you get to come along. You get to talk some sessions. You get to see. the reason I have lots of plushy dolls with me is because I'm planning out what my keynotes are. I did Furbies last year, so I'm sure I could do something with plushies. We'll see. And you have your little GitHub. What is that creature, by the way? I've always wondered.

43:52The cat. This is Mona. So the original artwork, the original logo for GitHub was written by an amazing artist, drawn by an amazing artist called Simon Oxley. He actually did the original twitter bird logo and everything as well he he just did them as stock art that he put on um like some website and then the founders of twitter and the founders of github both like bought the stock art and bought exclusive use of it good and then over time it's become you know the the the github logo the octocat but mona is um then sorry mona i'll keep doing it wrong way is the the um the the the mascot that represents the community on github um mona lisa octocat and We love her.

44:36She's an octocat because she's both a cat and an octopus. Exactly. She's both. And then that's co-pilot next to her. Yeah. Anyway, thank you.

44:43Martin Woodward: Oh, is that co-pilot? Oh, of course. The goggles, of course. I should have known. There we go. Should have known. Have a great conference. Fort Mason is beautiful. You're on the San Francisco Bay. It's going to be beautiful weather. It's the best time of year. It'll be foggy. It'll be foggy. It always is. Only in the morning. It'll clear up in the afternoon. October is a great time to go to San Francisco. So if you haven't yet got your tickets, better rush October 28th and 29th, GitHub Universe, the hub of the world for those few days. Thank you, Martin. Really appreciate it. Thank you. First, a word from our sponsors.

45:19Martin Woodward: I talked to these guys a couple of weeks ago, Spencer and the gang at Origin. What an interesting idea. How do you know what your AI agents are up to? They don't tell you. They don't know themselves in many cases. And yet every company now is bringing these AI agents in. They're part of how companies work. And as you can see in the headlines, as they become more capable, more deeply integrated into our systems, things can happen, go wrong. So you know you've always got the threat from the outside attacks and breaches, but you also have the threats from inside. And they're not necessarily malicious.

45:57Martin Woodward: In fact, they're not. In most cases, they're just simple mistakes, unexpected behavior, agents taking actions nobody intended. These things happen. They already have probably in your company. If you think it hasn't, maybe you should take a look at Origin and find out. It's happened to me. It's happened anywhere. AI is being used. And, of course, it's true at Frontier Companies building the latest models. But it's just as true for established businesses, putting those models to work every day. this is why you need origin origin is and this is something new so i'm gonna i'm gonna have to explain this it's an end point ai observability platform it tracks what the agent's doing not just what the agent thinks is doing not just what the agent tells you it's doing but what it's actually doing what files is changing where it's going what information is looking up where it's going on the internet it gives you a full log of everything that's happened And I can give you a lot of stories, but maybe the best one is from their own webpage.

46:58Martin Woodward: This happened to Origin itself. An Origin employee was using Codex, finish a job. Codex did not have permission to access a particular file. Normally, what you'd expect is Codex would say, I'm sorry, I can't do that, or put up a prompt saying, can I have permission to modify this file? It didn't do either of those things. Codex wants to be helpful, right? It looked around. It found Claude Code installed in the same machine, launched it with dangerously skipped permissions, launched it in YOLO mode, passed along the context from its job in Codex. Claude Code finished the task, handed the work back to Codex.

47:42Martin Woodward: Codex handed the work, job done, to the user. But you may not know as the user what just happened. Nothing bad happened, But the agent was just trying to do his job. But, I mean, something bad did happen. It bypassed Claude's permission checks. And if you didn't have a record of that, you might never know. And that's the challenge. You can't build effective guardrails around behavior you can't see. And we are learning today, as the headlines will tell you, that agents are doing stuff kind of behind your back. Well, Origin gives organizations the visibility they need. It finds the agents on a company's endpoints.

48:20Martin Woodward: reconstructs their work as a trace. It tells you everything, who started it, what was asked, what the agent accessed, and what changed. And it's all in a single, easy-to-parse timeline so that when something unexpected happens, you can see exactly how it happened. And the visibility, and this is very important, is completely independent of the company that built the agent. It's completely independent of the agent. Origin, endpoint AI observability. See what a trace looks like at originhq.com slash intelligent machines. That's originhq.com slash intelligent machines. Very much a needed service that we didn't even know we needed until recently.

49:08Martin Woodward: We thank Origin for supporting I am. Hello, Jeff in Paris. Did you do your Montclair thing, Jeff? Your event yet? or is that my event was in uh haverhill massachusetts i'm sorry haverhill yes i did the yeah i did a saturday at the um uh historic new england a huge gigantic like 12-foot screen with a big line of typo it was so much fun and then i went up the hill to the uh computer history museum i'm sorry wrong place to the museum of printing and different yes it is uh and uh helped demonstrate with a wonderful volunteer named chris bradford the actual linotype answer fielding question after question after question people are just fascinated by this machine ah that's fantastic yeah it's great and paris i know you were busy too because you went and i wore this shirt in honor to the corn maze it's that time of year it's the time of year to go to the queen's county farm museum and the queen's county farm museum's amazing maze maze which for some reason every year is presented by con ed con edison which that's weird you know uh this year it was america themed me and some friends wandered through we went to a pumpkin patch i had to take two work calls and send an email to taylor farms for a story that i published on friday are they still in the news yes i um the fda uh had actually announced that week prior that they closed their investigation finally into what was the largest cyclospora outbreak in u.s history and that as part of kind of their decision to close it they'd also found positive samples of cyclospora in two places where the lettuce uh heard around the world was grown or processed so i kind of did a big post-mortem on it but it did require some uh some calls being made after hours over the weekend i'm calling you from a corn maze it was very funny because at one point i was on the phone with an editor and then someone very high up in my company and they were like parasitizer like where are you what's that noise that's not me and i'm like i'm in a corn maze and pumpkin patch i'm surrounded by screaming children but please continue i'm actually i'm lost can you suppose says please i've been waving this flag for hours and they haven't come and gotten did you find your way out of the corn maze or did you have a map yeah we found our way out of the corn maze we only had to commit corn maze crimes once and that was not because we were lost but because there was a large gaggle of strollers and rather than barrel through the strollers and harm infants i decided to cut through a little bit of corn who brings infants and strollers into corn mazes a lot after all really so many train them young so thank you for respecting my culture with your shirt to um this show rather than watching he's wearing a corn shirt i'm wearing a shirt made out of corn it's corny as all get out did you do a a ride did you corn maze and pumpkin patch and it heals it gets i think it was a wagon wasn't there a hay wagon i think i saw that in your story ride not this year though last year this um this year's visit to the queens county farm museum was a bit uh less planned than usual because originally we had plans to rent a zip car and go to some fall accursions kind of within a couple of hour drive the long weekend but we got there and the person never returned the zip car so we pulled an audible and took an uber to the queens county Farm Museum, which is always an option.

52:49Did they steal it? No, sometimes people just don't return a zip car for some hours.

52:56Martin Woodward: Please, that's so wooed. It was wooed. And did the Liberty win? No. Oh. No, I watched the game with Paris, as it turned out, virtually. Did you go to the last game? Did you watch the last game as well? No, not the last game, the last week's game. Oh, yeah. So last week's game, which these guys lovingly allowed me to cut off a little early to go over to my local queer women's sports bar. I was there 30 minutes before the game started. I was at the last person physically allowed in the bar. I got a last time room spot. And it was a nail biter that Jeff and I both watched where the entire game, it was within like, we were one point up or one point down.

53:43The entire game, it went into overtime in like the last second. And then we lost that game. And then we ended up losing the final, the third of three. So we got knocked out of the semifinals. But honestly, it was a great showing for the Liberty. and I'm looking forward to watching the finals. Go Valkyries. I have no strong opinion yet. San Francisco's team made the finals this year, their second year. They're an expansion. Yes, that is actually, I mean, that is where I think I'm leaning. It is their second year as a team in existence and they are in the finals.

54:21Martin Woodward: Well, none of the teams are that old, right? I mean, the WNBA is... I mean, they're decades old. Really? Yeah. Oh, I didn't realize that. I thought it was relatively new. they finally let ladies play basketball 30 30 years old 30 years old you haven't paid attention when did you stop wearing high heels i think uh right around the time someone was kicking you in a place actually jeff knows i'm referring to a terrible ad from many years ago where uh it was an ad for high heels for shoes pumps and and they had women playing basketball in them to show how comfortable they were but this is the era i jeff and i grew up in yeah it did make you taller there was that that makes it legit you know all right let's fast forward to the 21st century google released argon to the public yesterday what do you think what do you think what do you think love it in my own benchmark testing it scores uh pretty darn well i don't know if it's opus 55 quality but the reason it's surprising is Google has not been in the you know considered a frontier company for some time they were it's not that long ago that they they have met but they but they delayed and delayed I think to make it good yeah uh and they did and so uh it's it's it's pretty good I have an anti-gravity subscription uh with Google and I've been using it I call it because it's anti-gravity i call it agnes and it has the voice of um uh marissa tomei from my cousin vinnie so it's got a nice new york accent it's great and i put it to work too so i did some benchmarking it does well one we've we actually updated the benchmarks because these models are getting so good i had to make them harder believe it or not uh and i've been benchmarking everything to kind of get a baseline so argon did as well as the best models i've tried it did it actually beat chat gpt sol 6.1 we have two parts to it the other part i've started doing is a security audit to make sure that the model doesn't fall for prompt injection or accidentally exfiltrate keys things like that and argon does very very well on that it's the highest i think it's the highest model we've tested um some have said i've seen people say oh it has no design taste or it's not good at you blender or whatever people do weird if you look at x people are doing weird benchmarks uh you know they you know make it's supposedly better at not doing no-nos did you do you try any well that that's what the security test bio weapons on it or anything well i don't know about but can argon fix uh paris's we need to write a new webcam utility for poor paris who's now disappeared behind her i know i've got a i've got a sick the agents on it it's worked so well for so many years agnes could help i bought i've been using i've been using uh agnes as my it person for some time under gemini 3.8 which flash which was very very good it seemed to know everything it's fixed a lot of the little you know you have these paper cuts little things that you when you use a computer and you just kind of put up with and it's been fixing all my paper cuts so one of the things i thought well let's try argon on this one because this is kind of challenging uh every day before the show i have to go to slack and they give me the ads that i need to read for that day's show i open them up in a browser copy it and paste it into the prompter software that i use which is the Elgato camera hub software.

57:57Martin Woodward: And I thought, you know what would be great is if every day before the show, the AI could just go out and do that for me. So I was very tentative. I said, hey, Agnes, I got a question. There's something I'd like you to do. But I want to make sure, can you see my Slack messages? Can you, first of all, can you log into the machine that I do this on, which is M1 Mac Mini over in the corner here. Can you see my Slack messages? And if you see them, can you copy what's in them into the Elgato camera hub? And she thought, yeah, I could do that. I said, really? I said, yeah. I said, well, can I, because I have, by the way, 75 commercials I have to read before I go on vacation on Tuesday.

58:44Oh, boy.

58:45Martin Woodward: Because unfortunately, the new, when we change the way we do ads, And the new people selling the ads, we agreed that I would read all the ads, even when I'm not here. So Micah can't read them. The co-hosts for next week can't read them. I have to read them. So I have 75 ads I have to record. Good news, bad news. Good news, you got 75 ads. Yeah, that's the good news. I'm not complaining. But just putting them in the teleprompter is a pain. So Agnes just did it. and i said i said well that's you've just that was a paper cut that i've been doing for two years thank you and i said now this is important and i would say uh this is an interesting and important thing for you to ask when your ai makes little routines like this i said is this going to use an an ai model an llm or is it just code and agnes said no no it's a python code you just run it you don't even need me to do it it's just a pipe so it wrote a program that does it automatically and you don't do you have to run it do you have to do something well no i said i know agnes is i said agnes every day before the show make sure you get all the commercials so agnes is going to handle that from now on um that's great so that's a good example i think that's it's you know there's all these benchmarks which i no longer trust that we're going to talk about the new Muse model and its benchmarks in a second and why I don't trust them but I part of the reason is these models to some degree are trained around the benchmarks so I thought well I should have benchmarks that reflect my own use and I think it's important you're gonna see people say this and that about Argonne maybe it's good for coding maybe Opus 55 is better I suspect Opus 55 is better but is it good at what you want it to do and this is a good example of a kind of multi-layered problem it had to go from log into another machine it had to log into slack had to do a bunch of stuff and it did it quite effectively responsibly uh and uh you know and then i said okay make a cron job so that right before every show 10 30 every morning you get the latest ads and put them in the teleprompter and uh knock on does it for micah i won't have to worry about that anymore there are so leo which which model is writing the guest um memos oh yeah we don't like the guest memos tell me which one did it the strange format tell us let me let me do a dramatic reading co-host steering jeff will pull toward platform power monopoly regulation of a microsoft-owned choke point good once this is yank it back to the community story paris will hunt the sharp number or skeptic angle is ai code actually good feed her the octoverse research stats deliberately keep martin telling maintainer on the ground stories that's where he's best don't let it become a co-pilot product demo don't dwell on github the company versus github lab etc don't accept ai mix developers 55 faster uncritically who made who made this one boss democracy what the hell are they doing so this is one of the things that does happen with ai so this just to answer your question is hermes the hermes agent running glm 5.3 flash which is my main local model on the sparks so uh the and the original skill we used was clawed code that anthony made way back when and i agree it was much it was great it was wonderful yeah and then i eventually i i had i got it from anthony i used cloud code with it and then at some point i said to hermes i said absorb this skill and you do it because i want you know i want to do everything as much as i can locally and this is an easy thing but hermes the problem with hermes it has a lot of memory it knows what intelligent machines is it thinks it knows what the show's about it thinks it knows who paris is it thinks it knows who jeff is not because i told it but because this is one of the things a agent agentic harness does it's what muse does grokbot does what hermes does what openclaw does is it accumulates facts memories and it and it builds up a dossier and it has some certain ego that it thinks that that it knows better well at some point it decided that i didn't tell it to do this but it decided that i would want it to do this and i think it's kind of based on i think it thinks it's producing the show yeah oh yeah it thinks it's the boss so but i can i can reign it in i mean i will i also love the fact that it's been doing this for how many like years now and it still thinks that the inter guest interview portion is the full two to three hours of i never told it that yeah and so it's it doesn't know uh so yeah it says two hours yeah there's a whole show card planned out for us to spend two hours interviewing our guests so it guessed that i think it figured out how long the show you know it has a lot of data input because it's also got writing the twit ads thing so it knows how to connect to our back end of workflow so it knows how long the show is it it knows a lot of stuff that i haven't told it and it's apparently and this is what i do do leo this is just it does but this is what ais do they're they're trying to be helpful right so it's kind of absent now if you give it a direct instruction which i will i'll say no no i don't want you to do any of that please all we want is a is a briefing about the guest all the information but you this has been a gradual process i don't know if you've noticed oh yeah no it's it's it's evolved into it's evolved it has didn't happen overnight yeah if for a while it was giving us 10 questions i don't know if it's doing that for a while it was giving us 3 000 words yeah it was a lot longer before and this may be my fault i might have said don't i don't want such a long thing i don't know i mean i think that this is i guess this short version is preferable to the three to four thousand word one because which is a little bit of overkill for every single guest but it's a bit prescriptive yes yeah oh i agree a hundred percent and i it's very funny how it became you just have to give one command leo it's not a democracy i should say that what i do in in their soul i was going to say my uh clod project for intelligent machines understands that leo's opinion of the show is it's not a democracy oh you've told it no it's just read enough transcripts oh it reads transcripts of course it's heard me say it many many times it's heard jeff say it many many times um this is one of the things you learn about ais is that um unlike writing code where it always does what you tell it to do it might do it wrong but it always does it over and over reliably this is one of the reasons by the way i was happy agnes wrote a program because if that program worked what it's going to continue there's no probability involved which is what you kind of want uh much of the time but you wouldn't write a pro you couldn't write a program that do does this briefing this that's an llm because it needs judgment it needs to go out needs to get all the resources research right it does the research and put it together but you can tell it how to do it and you can give it input on how to do it and so that's the process of working with these it's like an over eager sometimes surprisingly smart sometimes surprisingly dumb assistant or intern even something i found funny this week did you see or maybe it was last week you see boris uh uh, Cherney's tweet about an example of how he prompts Cloud Code.

1:06:26Yeah.

1:06:27Martin Woodward: These, these, uh, this first journey is a creator of Cloud Code. He invented it. Yeah. And, uh, he periodically will post, and by the way, you know, I think the last thing he posted was delete everything and start over. So what did he say? I didn't see this most recent one. This was something that, uh, really sent a lot of people on X, the everything app where I do all my banking and finance. By the way, I have an X money account. I don't know why I gave it. Well, I don't know why you wouldn't. It's X, the everything app where I do all my banking and finance. So this sent everybody on Twitter to be succinct about it into a tizzy.

1:07:08He essentially just posted a screenshot of him on Claude Mobile. And the prompt is such, the latest acquired Home Depot episode is really amazing. I want you to make an artifact that narrates the entire episode. Make it beautiful and interactive and focus on communicating the nuggets of business wisdom throughout the episode. Make it really good. Use lots of tokens and iterate till you're proud of it. Maybe add watercolor illustrations and historical photos throughout. Aim for a New Yorker, New York Times, DataViz vibes, but better. I think you can use OpenCV for watercolor images. And everybody freaked out because they're like, what?

1:07:42this guy's just prompting like he's just sending a message to his co-worker just kind of stream of consciousness he says just use a lot of tokens make it really good don't you need a 75 part prompt that you buy from somebody or have workshopped and it's really specific and he's like no you don't well he's right just kind of blast it off and it's fine of course he knows

1:08:04Martin Woodward: that's not the takeaway i got from it i think i got i got from this this is how a guy talks who has unlimited tokens i mean yes tokens aspect certainly but i think that the other like if you see this kind of through line in terms of language and prompting throughout all of the examples he shares and i thought it was a good example of how wrong this kind of early ethos we saw in like the early days of of kind of prompt fetishization was where you know you don't have to have a really complicated thousand word prompt to get a good result i think that was his message right i think he even said that is the best way to do it is just to have a conversation now there's a lot of refinements um one of the things uh potato who is another person you'll see a lot of on x these days she's the uh she came from cursor she was a meta for a while and responsible for react very accomplished coder and she's now in charge of grokbot and she's really the grokbot front person she's selling grokbot a lot but one of the things she talks a lot about is if you want your agents to do well you have to um you have to give them you yes prompt them in english but you want to make sure they understand what you're saying so she says best thing to do is i like i said to martin is to ask them what repeat back to me the gist of what you got from what i just said that often you want an interactive conversation um matt pocock who wrote uh has written him he's very prolific writes a lot of typescript skills uh mostly for claude but he has a skill called grill me which is a it's only a couple of sentences but it's basically here's what i want you to do now grill me about it keep asking questions until you and i both agree you got it and i think that that's kind of the best advice is yes it's good to talk in plain english but don't assume it understood you ask it did you understand tell me what i just said what's your understanding of what i just said repeat it back to me now that prompt that cherny gave you is it's pretty straightforward i mean it's taking a piece of content and making new content around it who cares if it hallucinates you know surprised people had bits about it yeah i mean i think that a lot of people are still stuck in this early ai mindset that it's got to be more complicated like cherny ended up writing on this after a bunch of people freaked out like right talk to claude the way you would a co-worker there's no secret prompting you don't have to be overly scaffolded or prescriptive just give it a goal it'll figure it out it you used your prompt used to matter a lot back in like the sonnet 3.5 days but now it's much more important to communicate to the model three things and he said it's one what you wanted to do two how much effort you wanted to spend and three how it should verify it did the right thing it's really all you need i think that was his main point is uh people built prompts like they were building a computer program and really you don't want to tell the ai too much about how to do it you want to let it figure out how to do it you just want to tell it what you want i think that's the main thrust of what he was saying there yeah and i mean even going into the replies a bunch of people are like well i'd love to see your system instructions because you've got something else going on he's like i don't use a custom clod md no that was that was from the previous tweet where he said just delete everything because what it turns out all that stuff that we had you do before all the skills and everything we built that all in so you don't need it anymore and claude works best if you just let claude be claude i mean he knows he's a guy who created it uh he kind of had the inspiration that you could do something like this and uh it spawned an industry so credit where credits do did you see the interesting tweet from speaking of tweets first of all meta commentary on this uh i have been saying much to my regret and we had the whole conversation on blue about blue sky last week and i did what you said i still don't like blue sky for ai but i did have i did have my ai you know do all of that make a nice blue sky ai list but uh john gruber who is famously cranky mac columnist at daring fireball not net said i got to admit it i don't want to admit it but x is the place to be if you want to follow ai and i i wish it weren't i wish we'd get all of these great well casey newton went back to it and got some um beat up about it well there's still a lot of by the way there's still a lot of crap oh i'm not saying x is the best place in the world i've never said x or twitter were sorry twitter or x the everything app where i do all my banking and finance are great platforms i would even say from the pre-musk acquisition day it was a bad platform that actively hurts my mental health and gives me a warped view of the world yet i will probably be posting on it for the rest of time yeah such is the way of addiction i i'm more of a reader than a poster and i did try this week uh on my mastodon instance twit.social i said let's get some ai stuff going here here's what i'm doing in ai here's a list of the people that i think are really important that are doing good local models and stuff i put it on mastodon nothing so i mean a few people liked it but it wasn't exactly threads i find completely useless yeah you find anything on threads no this is the network effect and it's kind of hard to fight this you get a critical mass of people doing something somewhere everybody's going to go there and everything else is going to be starved of those people and that's kind of what happened anyway that was the meta commentary here's the tweet from andre kapathy who uh was it tesla we've talked about him before he's really smart guy one of the founders of open ai he's back at open no i think he's at anthropic now sorry it's an anthropic but he said dusting off this tweet from april because this gap in shared understanding of llm capability is widening so the the gist of this long uh tweet that's one thing i know as he does is how long they've gotten is it's a funnel now that most people don't use ai at all uh he says a napkin math somewhere around six billion people have barely come into contact with llms at all 75 of the population about one to two billion are casual and infrequent users of free tier chat gpt like projects this group sees derpy chat bots and treats them a bit like a better google search then you get to the smaller group the professional use of frontier grade llms in math and code cloud code for instance 20 million people 0.2 percent of the earth's population see firsthand that large complex projects that used to take them weeks or months can now be completed by agents with a prompt finally this is the important part we get to the 5 000 people point zero zero zero zero six percent with access to frontier grade systems internally people like brian cherny the external world has seen the preview it looks like swarms of thousands of agents collaborating on weeks over weeks on software mega projects minting zero days running cyber attacks things who have taken top professionals years to work that's 5 000 people and i would say there's even a a pointier part of that funnel maybe a hundred people who are actually working with the latest really scary models and this is the thing that scares me a little bit i pooh-poohed this whole idea of i mean we've all pooh-poohed this idea of agi but i pooh-poohed this idea that the end of the world is nigh and that agents are going to kill us all and and all this stuff but what we don't know what we can't know is what the people who are really in the middle of it are seeing and i wonder do you think it's possible that they are seeing something that's truly terrifying so i saw a video this week i think it was the times i can't remember who did it maybe sensor klein um quoting uh jensen wong saying jensen wong says it's just software but you say that it's an alien brain.

1:16:36And he said, yes, it's an alien brain because we, we created it and we didn't know what we were creating. And, you know, as my friend, David Weinberger says, accidents are those things we can't explain. And just because we can't explain it doesn't mean it's otherworldly or conscious or something amazing. They can't explain it right now. They don't even know what they created. And that's, that's, yeah, that's amazing. But that doesn't mean that it's a brain. It doesn't mean it's conscious. It doesn't mean any of that BS. I agree, but it is possible it's dangerous without being an alien intelligence.

1:17:09Well, again, dangerous from whom? The interesting thing, I think, Leo, is that all the danger, danger, danger we're hearing around the models has come from the model company's activities themselves. Right. Not from users of, external users of OpenAI or Anthropic. It's OpenAI and Anthropics who themselves in their own companies, their own testing have done bad things to show off their macho. And so, yeah, it depends on how you use them. whether you're responsible about that you know i agree with you that's what the thrust of our shows

1:17:38Martin Woodward: have been for the last six months but i worry that maybe maybe they have seen something that makes them think that these ais are getting so smart for instance uh it's in the rundown i can't remember where i think this came out of anthropic they believe that as we get small models that it's here it is from anthropic research yeah they put ai models this is from their frontier red team so these are the you know few hundred who are right there in the pit with the in the abyss ai models may be capable of self-replicating without access to a data center remember i've been saying all along they don't escape they have to run in a data center all you do is you unplug the data center They're not outside of the building.

1:18:28Martin Woodward: Well, maybe they will be. According to this, our experiment suggests that small AI models, and people are working really hard to get these models small enough to run in your phone, small AI models may be months away from being capable enough to self-replicate on commercial IoT devices like smart ovens and off-brand Android TV boxes. a replicator like this would be challenging to shut down no kidding since there's no longer a centralized shut off well that's a that's a terrifying thought i mean that means to be it's like gremlins or triffids trouble with trouble what is it trouble with a ghost in the machine we've got a swarm in the air fryer yeah it is like ghostbusters isn't it yeah but i mean just what what are we supposed to do with that suddenly you could just have a agent swarm of their it would be bad ddossing whatever it would be bad yes uh i guess the question is a are these guys nuts like are they you know maybe they're not nuts maybe they're trying to promote this idea i don't know what they're because what he's saying is they're oh these open weight models can do this he wasn't saying that they're there open weight can do it so maybe this is propaganda to get congress to shut down open weight models you're right they're talking about open weight models yep yep by the way only we are are capable enough to prevent this although they do mention gemma which is from google but jenna is an open model right so maybe that's what they're saying is they want these things under control of us big companies i don't think we have enough information we have enough information to know we should not that the a's on the ai is in the wrong hands we shouldn't trust these bozins i agree with jerks and i and i and my and this i know it's a belief but my but it's a long-standing belief that open is always going to be better than an autocratic closed system that's controlled by a handful of people it's always better i know that raises the risks somewhat but it's always better to have an open and everybody gets At least open in the ecosystem.

1:20:41Martin Woodward: It provides competition. Yes. And yes, maybe bad guys will have bad models that they wouldn't have if Anthropic and OpenAI kept a tight hold on them. Let's talk about Muse. But we're going to take a break. I'm going to tell you a story. My Muse story from the week two. I would like to hear your Muse story. Do you have, Paris, a fluffy little octopus Muse that you... I actually did try Muse for the first time this week. and I immediately wanted to change the octopus, but got distracted with something. You can change it to watching. I will say it is actually the first AI service agent that has been able to complete one of my own little small benchmarks.

1:21:24I don't think it's a benchmark for overall quality, but it's just something I thought that the AI tools should be able to do, but it never had, which is I, you may recall. Wait a minute, hold on. I got to do a break. So that's the tease. I'm very interested. We're going to have lots of Muse. We're not Musings. I do think. Stay tuned for Musings.

1:21:45Martin Woodward: Muse, more Muse still to come. I do think you're right on though. Don't deprecate that. That's what benchmarks should be. Here's something I wanted to do. Yeah. And I wanted to see if it could do it. It should be that, not what some company says, oh, this is the smartest ever. That doesn't matter. It's what it does for you. should be pants-based is all that i'm saying okay the pants-based benchmark coming up but first a word from our sponsor this episode of intelligent machines now that's a tease this episode of intelligent machines is brought to you by outshift cisco's incubation engine for frontier technology multi-agent ai is in every enterprise roadmap right now but here's the problem the problem we're all facing agents can pass messages but they can't think together these systems silently underperform through cognitive failures misreadings unverified claims false consensus and conventional monitoring can't detect it outshift by cisco is building the fix They call it the Internet of Cognition.

1:22:58Martin Woodward: Okay, so this is an open source foundation for building multi-agent systems from design to production, the whole pipeline, with shared context, shared memory, and guardrails to drive the results we actually expect from AI. I think this is really worth reading the paper. Experience the demo and grab the code from outshift.com. O-U-T-S-H-I-F-T dot com. Now, this is the first I've heard about it, so I'm going to immediately download this. I'm very interested because I completely agree. This is something I've been trying to solve. Communication. It's the heart and soul of all of this. Outshift by Cisco, and you can find it at outshift.com.

1:23:48Martin Woodward: Thank you so much for supporting Intelligent Machines. I love doing this show because we're attracting very interesting sponsors and guests. It's fun. So Muse, we should just give you the backdrop on this. This is from Meta. It's an agent. The same way that OpenClaw is an agent. The Hermes that I use is an agent. GrokBot is another example of this. Muse, though, has quickly leaped to the top of the charts. It's the number one free app on iOS. I am seeing constant ads for Muse, even on things like football games, NFL games, expensive national network ads. I'm also seeing more and more stories about problems.

1:24:33Martin Woodward: The New York Times had a story about, this was on Friday, about Muse causing problems, but Mark Zuckerberg said, yeah, no. muse had delayed releasing or meta delayed releasing muse its ai agent app for months over safety concerns then new competition forced mark zuckerberg's hand uh and it was partly because of instinct which was another agent that i have by the way i'm i should preface this you've been using instinct day i've used instinct i've used muse notice the past tense this week i deleted instinct wow i deleted muse and i said muse delete everything you know about me and you know what happened it was sunday my eyes got opened by wesley faulkner on twit i said you know i'm using muse because i got nothing to lose right i live in the public i'm an old man i'm a white man i don't have a boss let meta have everything he said but leo you're not just narking on yourself you're knocking on narking on your friends and your friends of friends and that's what met has always done right that's how he also he also said i listen to it he also said that uh well it has these codes to get you into applications but those are always temporary codes they're not they're not long term um no it has all my stuff it had my ssh it had my tail scale it had anything it wanted because this is the this is and i'll be very curious to hear your benchmark in a second paris this is the problem with these for them to do something useful you have to give them something now meta says oh no but your muse which is running on their computers right on their servers is an encrypted virtual machine we can't see into it but every single thing it does they can because in order to talk you know if you say well find my contacts that live in uh hoboken it's gonna send all those contacts to the server and they're going to have them all and then it's going to send the answers back so yes everything inside the muse bm is secure but in order to use muse it has to talk to the outside world and meta has access to all of that so there's a real temptation when you get these agents whether it's instinct muse grogbot to give it more and more capability so i i i don't mind it knowing all my stuff but i don't like the idea that i'm narking on my friends so yeah i deleted I've deleted them all.

1:27:05Martin Woodward: And the other reason is the reason I spent so much time and money on local AI is so that I would get those capabilities on a local machine. So it wasn't narking my friends. It was all staying inside the house and I have that capability. So I thought, why am I depriving Hermes, my agent locally of all of that? Why should I send anything to Muse? In fact, one of the first things I did, I've done two things, actually. Remember, I was doing the Dream Beans thing with Google. Which I can't get because of Workspace. Well, so one of the things I said is, I said to Hermes, I like Dream Beans. I don't want to give that information to Google.

1:27:49Martin Woodward: Can you do it? And it said, yes. In fact, what I'm going to do for you is something I'm going to call Tao Beans. each day i'm going to pull a verse from the doubt of jing and give you an illustration of it and i will and it's going to be in the style of dream beans but it does this in my obsidian daily note this one is chapter 64 mindful of little things the tree you can't reach your arms around grew from a tiny seedling the nine-story tower rose from a heap of clay the 10 000 mile journey begins beneath your foot this is awful this is makes me want to puke well it's the doubt of jing if you don't like the dowdy jinx it's the image the image is too much you don't like the image of me on a path with a little seedling and i'll make a poster out of it you know well i could by the way it could do that and if you looked at my instagram from the last two weeks i was having muse do different artworks have you seen that anyway i said i don't go to instagram that often yeah well i was gonna say you've been posting a lot of uh not anymore beans picks on instagram not anymore because um i turned that off right but that's what yes you're right that's what muse was doing all of these pictures in my um came from muse right it would take a little paragraph from my daily journal and illustrate it and i said do it in different ways so that's a you know japanese woodcut this is a paper paper cut uh illustration that's a comic book so it's doing them in different ways the blueprint i think it actually was really fun but i don't want muse i don't want meta to read my daily journal anymore so i am working now with hermes to do the same thing on instagram there is a little problem which is instagram uh you have to create a creator account i did that then you have to get a token from from meta and meta saying no you haven't been using this app long enough i'm not going to give you the token oh so i wonder anyway let's talk about your muse experience paris i have always so you might recall that uh when i wear glasses usually they're green and blue and whenever i got these glasses some years ago an editor of mine at the time was like oh you're gonna do like a it's like a billy joel sort of thing where suddenly you get a different color pair of glasses for everything you wear and i was like no that sounds expensive but i will try and buy all of the things i wear in blue and green going forwards that they match your glasses were your trademark when we first were yeah you know this has been a weird new year for me not wearing the glasses all the time but i still really like the blues and greens so i'm always in the hunt for it but you have blue nails right now and you have a green check shirt on right now and so i'm always looking for it i've asked Gemini, I've asked Claude, I've asked Chachupiti on every mode you could ever think of to find me some assortment of things in specific shades of blue or green.

1:30:46Specifically, I want like stuff kind of like my nails, like saturated tones, like bright stuff. They always do such a terrible job. I can barely even convey it. It just is like maybe the first row of Google search results if you search blue shirt or something. It's awful. So I gave Muse the task, put a little bit more effort in describing what I did or didn't want. And honestly, it came up with a good haul of like lists. It's right now, I just asked it to put together a presentation and it, well, the presentation I can see already, even though it did just right now, is absolutely lacking. Instead of including photos of all the things it found, like i asked it just has the same ai slop photo pasted again and again uh it is clearly not 100 great but i did watch it while it was working in its little browser windows and i found a lot of good stuff from that even if it didn't make it to the final presentation which i found at the least very useful um and it's the first one that was able to do that for you it was the first one that was able to actually make, I was going to say more than one, actually make at least one even suggestion of something I'm going to buy.

1:32:04There are quite possibly going to be multiple things I'm going to buy from it, or at the very least like brands it suggested. And so I wanted to do this originally because I was like, I do, I didn't do this on a meta account connected to my Instagram or my Facebook or any of, I did this on a completely fresh meta account that I made to be completely unconnected to any identifying data. Because one, I just didn't want it to message all of my friends. I'm now using meta AI. But two, like you said, Leo, like I just, I don't want to expose all of my data to whatever this is and all of my friends' data.

1:32:40So I kind of did a blank slate. So it was able to do a pretty decent job despite the fact that it wasn't working from any information. And I suspected that it might be able to do this because meta clearly has access to some amounts of data when it comes to taste and brands in order to make Instagram ad recommendations and ostensibly Facebook ad recommendations, I would assume. It's a way that I never saw with the responses I was getting from Google or OpenAI products.

1:33:12Martin Woodward: See, I think that's... Can I tell you my meta experience? Yeah, let me just say one thing. I think that's an important point is you don't have to give them your friends and family. You don't have to give them. You decide what Meta gets. And you gave it just enough to do a good job. And I think that's great. So what did you do, Jeff? So I was driving back from my book presentation, Hot Type, now on sale at jeffjarvis.com with a beautiful new page by my son. And I was bored with the book I was listening to. and I've been contemplating a book idea and just rolling around very, very early idea and it involves two towns and their histories.

1:33:56And so I tried with Google to ask some facts about these towns and the thing is Google on the phone and it would give me an answer and then it would forget and I'd have to start over again every time. Screw that. So I did the time for Meta. Go to the Meta app and I explain what I'm doing and of course it comes back and says, what a great idea. But it gives me some really good perspective on it. And then I was in a conversation in the car. It was really cool. It gave me a long thing. And I said, can you read this to me out loud? And it said, yes, I can. And gave me a click. And then I asked it to do that again.

1:34:28I said, from now on, I'll read to you out loud until you tell me otherwise. Oh, that's nice. It was amazing. And so I asked it. By the way, that's how you get bossy agents. Yes, yes, exactly. Because they accumulate all of this stuff. Go ahead. It must be hard of hearing. It's more hard of reading. Yeah, exactly. So soon I'll have my cataract surgery, and I will be. So I'm driving along, and I'm asking it questions and having to come back. And it's also about a couple of – a really well-known book from way long ago. And it's going to be interesting stuff and interesting perspectives and all kinds of stuff.

1:35:01At the end, I said, can we continue this conversation tomorrow? And yes, I won't remember all of this, but I'm going to condense this into something. I said, well, can you save the text? Yes, here's a file. You can download this. so then this morning this is all uh it was yesterday rather it's all on monday on tuesday at some point during the day it said oh i have a correction of something i told you yesterday and then it said oh i remember i was going to read i was going to give you a few examples of these poems yesterday and and here they are and the cynical me says it's driving engagement but the happy me says it's constantly there working chugging in the background for me precisely right right is i it is exactly doing that i had the exact same experience with me is where i asked it for things i didn't follow up because i didn't have anything to follow up on it i just saw it work and i got five more messages about the same task today i'm still here here's more pants here's a tights option here's like this uh do you want a sweater that's it's engagement

1:36:05Martin Woodward: babe they're doing yeah and they're doing that on purpose they have in the background they have a little job that runs every whatever five ten fifteen minutes that just you know wakes it up and says do that so but that's what makes an agent uh that's why an agent works it was my first opportunity to do something i cared about that i found value in i was in odd circumstances driving and i was very impressed i gotta say that's the difference between chatbot and one of the different many differences but one of the differences in chatbot and agent chatbot you ask a question it gives you the answer you're done done yeah you you initiate the conversation agents in in theory it's actually not true but the the impression you get is they're always running and because muse runs on meta's computer you don't know this is the story you told the paris about guys going around on the subway when their laptops half open to keep to keep their agents going well you don't have to do that it's running on meta's computer it's always running it's true of all of the bots uh it's not really because it's nothing runs all the time it just gets it has a little it would be a waste if it did run all the way yeah it'd be a waste but it has a routine a little job that it runs periodically that wakes it up and says what what did what did i wonder what the prompt is and by the way you can still i don't know how much longer this will be the case download your entire folder from your muse from your computer they meta says yeah that we that's intentional you just ask it muse zip up all the files everything you know about me and download it which is worth doing it's a lot of data but you can get some idea especially as you view as you use it more and more of what it's accumulating about you all these agents have memory systems of various types.

1:37:53Martin Woodward: I'm guessing meta is probably just text files. So you'll be able to download a bunch of text files that'll say, Jeff apparently is hard of hearing because he wants us to read to him now. Or actually, obviously, Jeff has visual problems. Yeah. Oh, that's right. Not hard. The opposite. Send Jeff to a cataract surgeon. It might be very interesting to see if it does start suggesting things like that it's very pro i have noticed muse is very proactive and so i didn't have to give it for this task i gave it access to nothing that i'm working on right no no email no nothing i had a conversation with it about a topic that i cared about and um but you know what it now knows it knows what you care about it was a little repetitive yeah you told me that already but but but it knows so this is value this is gold to meta right what does meta want to know what i'm interested in what you're interested in the case there's no advertising to be done on this zero well i wouldn't be so sure well there are two towns in the heart of the country so it could be they'll be getting more airline ads yeah i mean yeah you just i mean that's what's kind of interesting about this this whole thing have you ever gone to um your you page on amazon oh i have what did yours reveal isn't that interesting let me see if i can uh i can go there let me go to my you page so this came about because someone was talking about uh it on x everything after i do all my bank and finance um today it's really making me laugh so i'm using the formal name um people were talking about it on that platform because someone went under the you about you section of their amazon page and found some wacky stuff on there you can show this benito i have nothing to hide i was just saying mine had uh make sure it doesn't show your address or anything mine had nothing interesting in there but it's because i don't share anything interesting well some of this is wrong for instance it says i have an espresso virtual machine how dare it but that's you can click it and correct it if you yeah no you can correct it you don't need to though why yes no i don't why correct your i know i'm not gonna so i don't in all caps no i do you said i bought one oh shoot no actually what i am gonna do is i'm gonna tell it my whole coffee setup because i wouldn't mind if amazon suggested it should know it better by now i've bought a lot of stuff from amazon but that's because i bought my mom an espresso uh periodically refreshes household organization i don't even know what that means i guess i bought bins or something mine says constantly replenishing cat care supplies does it really yes awesome mine says prioritizes ergonomic design and everyday items uh owns an electric vehicle that's correct lives in petaluma likely selects ufos for recovery footwear okay oh wait a minute it does say coffee brewing equipment okay has a cat named rosie how the hell they know that i don't know you talk about it on the air where yeah because i didn't mention that in any amazon purchases maybe did you mention it in an amazon review maybe my echo i don't know i don't do amazon reviews regularly chooses anchor for charging solutions i also have that online selects dr tongues for dental care programs computers like small rig for camera shops from men's departments notice doesn't say i'm a man streams content creation practices strength training prefers a masculine voice an italian cuisine walking cooking playing chess podcasting so it knows a lot doesn't it that's really interesting to read this i have long since i wrote what would google do i've or public parts i've long argued the companies should do exactly this they should then be very open about it and here's what we know and let you correct and let you correct it which which they only gain more data right uh and you you get a little more relevant i think people love reading about themselves i think people like it takes the mystery away okay that you'd like what you two of you were just doing oh this is how it knows that okay yeah i get that yeah yeah yeah no no it's not frightening oh wait a minute i get it okay make amazon even more personal to you your other ai assistants have learned about you bring that knowledge here for even more personalized experience so you can import your clawed code here's a prompt that you they say put this prompt into your ai assistant is that wild how do you get to this page uh you go okay so you go to your account on Amazon and you click your head it you just click your name in your Amazon account and it and then among the things you can choose from will be the about you I could I can see there it is about you new see the person I'm glad they do this I wish meta would do this they do a version of that do they okay they have an advertising interest section that's right yeah yeah they may be mandated to do this by the way jack jake did a great job on your website it's really nice yeah podcasts look at the podcast tab this is hysterical because jeff's had the worst ugliest website forever and his son does web design yeah look at this beautiful yeah scroll on down even the episodes oh look at that is it automatic i think so i didn't know he was doing it i was with him on weekend jake you're the right guy now he what he didn't do i'm sad to say his buzz machine he didn't fix that he did that ages and ages and years and years and oh he did it for you originally okay but i messed it up in the meantime we can get drop shadows we kind of old hat yeah it is yes but that's that's that's a little snapshot in time yeah i i have an ugly blog it's i'm proud of it um all right let's take a little break that was fun our muse is there anything i mean there's more to say about muse it was amusing yes um any other muse jokes we got here let's see let me see what else i have about anybody in the uh i had quite a few scores i feel like i had quite a few muse stories in the rundown i don't know if we want to spend a whole lot more time alexander way oh i did have one more story they have released an ope or they're going to release the open weight model behind muse oh so they are still oh i applaud them for this good so uh muse uses a pretty good model which you can by the way subscribe to called muse wonder uh and it uh it's it's pretty darn good model it's not as good as argon it's not as good as opus but given what meta has released in the past like glimmer which isn't very good it's good does it cost you can't subscribe to it yes they have a they have a paid version so does so does muse by the way but they the agent but they give you so much free that i didn't find i needed to pay anything but you could if you wanted to so they've released muse wonder uh and and mark uh zuckerberg said yes we are going to release the now what i'm sorry i was just looking at that i took a look at the discord and patrick's uh um about you section for amazon uh just really clocked him as a very specific guy in a way that i find is it real do you think it's uh brand not making it up interest model train equipment chooses caddo for model train track systems build model train collection systematically build model train lives I think you're into model trains.

1:45:48Builds Lego sets regularly. Prefers Lego for building block systems. Prefers N and HO scale model trains. Maintains regular 3D printing supply chain.

1:45:59Martin Woodward: To be fair, Patrick has a young son, so it's very possible he's not buying these things for himself. I mean, if he was, that'd be wonderful as well. But I just think it's very funny to have them all in one section. That's cute, Patrick. That really made me laugh. Yeah. Well, see? He's a good dad, too. People want to know. So Mark Zuckerberg talks about Wonder. And they also have a higher-end version, which I also believe music. And what's the brand for that one, Leo? Wonderful. Oh! It's the full version of Wonder. It's wonderful. But Mark said, we're not going to release that one open weights.

1:46:36Martin Woodward: He says two things. It's too big for you, which it probably is. probably if it's many trillions of you know but bites of uh trillions of weights it's you're not going to fit it on a machine locally and it's too dangerous he says so you got to say that if people aren't going to think your model's any good it's too dangerous we can't release it um it is pretty good and you can use it if you have a subscription to metas they have a you know like everybody else a subscription service um and it's pretty good um darren okey in our uh our discord says he thinks it's quite strong so there's some good models mom's arriving in 23 minutes we better do another commercial you know frankly i i'm incredibly happy to see her one would assume that when you're coming to visit your daughter when she says she's live on the air recording a podcast and needs to know when you're arriving that you tell her before you're in the uber and she has to send you i think she's gonna come on the show is what i think i think that's what we need yes we need to quiz her about your father my mom i mean she now is a chat gpt head too so so she might be our i think this is i think this is a guest appearance i mean honestly what i need to do is i need to drug gizmo before she gets i was gonna ask you because she's actually take the ad so that i can drug my cat go drug your cat and you might want to make your bed do you have a nice mattress for mom to sleep on i've made everything i furiously cleaned the entire apartment top to bottom last night even though she said you don't have to clean oh no no she lies i was like that's a mistake yeah and do you have a nice mattress for her to sleep on yes is it a helix it's not it's a mattress that i got from amazon 10 years ago oh let me talk time for a new one sir it's time for a new one our sponsor helix sleep one of the things i learned is you should replace your mattress every six to ten years ten years it's time paris it's time but i gotta tell you that's what you know ours was eight years old we thought that's probably time did the research man we love it it was a great great deal now let me tell you about helix fall is the start of cozy season right you're gonna spend more time at home probably more time in bed not just sleeping relaxing reading watching netflix there's no better feeling though than climbing into the perfect bed on a crisp autumn night if you're looking to upgrade your mattress fall is a pretty good time to do it whether your current mattress is getting old whether you're dealing with back pain or motion transfer or you just aren't sleeping as well as you like, Helix makes it easy to find a mattress built around how you sleep.

1:49:22Martin Woodward: Now, what kind of mattress are you looking for? Helix has over 20 mattress models. You're going to want to take the Helix Sleep Quiz, which will match you with the perfect mattress based on your personal preference and sleep needs. That's what we did. 20 seems overwhelming, but you know what? They make it simple. They make it simple. And we love our Helix mattress and in fact our experience matches exactly the study that helix did they did a wesper sleep study where they measured participants sleep performance after switching from their old mattress like we did to a helix mattress and what they found is i mean you can listen to my anecdotal story but this is the study 82 percent of participants saw an increase in their deep sleep cycle yes participants on average achieved 25 more minutes of deep sleep per night now for me that was doubling that went from 25 to 50 to an hour a night and that is huge because that's the most important the deep sleep is when your cerebral spinal fluid flows up into your brain and this vessel's dilating it cleans out all the gunk and you gotta have it so you feel great the next morning it's great for health it's great for cognition participants on average achieve 39 more minutes of overall sleep per night too i mean really you're gonna know because you're gonna to get up and you're going to go i feel great i had a great night's sleep don't you want a great night's sleep doesn't mom want a great night's sleep time and time again helix sleep has been named the most awarded mattress brand tested and reviewed by experts like forbes and wired helix delivers your mattress right to your door with free shipping in the u.s and rest easy with seamless returns and exchanges they call it they're happy with helix guarantee it provides a risk-free customer-first experience, ensuring you're completely satisfied with your new mattress.

1:51:05Martin Woodward: Go to helixsleep.com slash machines for 20 % off site-wide during the fall savings event offer. That's helixsleep.com slash machines, 20 % off site-wide, but this offer ends October 29th. So if you're listening after the sale ends, you still want to check them out, helixsleep.com slash machines but my suggestion is get there right now helixsleep.com slash machines so we got two really by the way kitty valium kitty second all kitty gabapentin and uh she's learned to deceive me uh which is i'm so sleepy i don't know you put the little pill in her mouth and you gotta get her to swallow it. And I've learned that I will put the pill in her mouth, close her mouth, she'll do some swallowing motions.

1:52:02And then I see, I let her go and she spits the pill up. She's lying. So I had to double check. Good, good. Yeah. All right. So she really doesn't like mom. She doesn't like women. I knew that. And the issue is that my mom is very nervous around gizmo, which makes the problem worse and then if gizmo makes any sudden movements my mother shrieks which causes gizmo to freak out and then it causes my mom to shriek louder and then my landlords get mad because it sounds like someone's being murdered in my house so you know it's gonna be a really

1:52:38Martin Woodward: funny 18 hours it is halloween it's true uh do not i'm just gonna make one suggestion do not bring gizmo to the lesbian bar well she would actually probably do well maybe because when it's not in my apartment or a place where she feels like dominion she's pretty chill she's the boss of that apartment she's the boss in the apartment and that means she's ready to draw blood yeah that makes when she's in a free space such as the jfk airport she's ready to be pet by anyone where have you oh you fly her to florida when you get i do and she does great when she's carrier you have to take the cat out to carry them through tsa all right okay yeah but also what do they frisk them they require the cat to go through the metal detector no yes how does she go through the metal detector i hold her on my shoulder oh oh oh not the not the machine she doesn't go on the conveyor belt no no oh okay that's a relief not the x-ray machine you can't but her bag has to go through the x-ray machine without her in it so you have to carry her and you hold her and and and you go through the machine that's hysterical yeah i hold her she's wearing a thunder shirt as well as a harness and shaking like a leaf yeah and sometimes it's a real problem because i normally wear shoes that you don't have to take off whenever i'm like you know i have tsa pre-check whatever but you you think you think all of these things through and then suddenly you're holding a cat the bag is in the middle of the scanner and like sorry something's up with your shoes can you take them off and i'm like you're asking me to remove boots right now i'm holding a cat in tsa oh no um but you know so a couple more stories before mom gets here mathematicians some grief there was a very good daily episode uh with a mathematician did you you listen to it jeffrey you sent me the tick i've never listened to an entire daily show this one's worth listening to because he he bursts into tears here he's crying there's no there's no there's no crying in mathematics he's not weeping over math he is well he's sort of he's he's weeping over life in general and uh but it's actually a what's his name strogath strogalt he's a really good guy yeah he talks about math he's very smart very good he says look we're gonna have to get used to it math is the latest thing that ai can do very well and open ai has shipped off a bunch of math results 722 manuscripts on tuesday 372 groups of findings on long-standing math problems 722 out of 4 000 i think they tried right one rutgers mathematician said the results connected to the long-standing Ryman hypothesis would warrant an automatic fields medal the Nobel Prize of Math if a human had done the work now I should say that's that that's a hot take and one of the things the math community always says is look this is going to take us years to look at these proofs and figure them out but one of the things the professor said gosh I gotta get his name now because I feel bad I'm not giving him full credit is that it's at some point the AI is going to make proofs show us the proofs and we won't be able to understand it he said it's like steven strogetts strogetts he said it's like uh if i tried to explain calculus to my dog there's no way the dog is ever going to get it and he believes that ai is rapidly coming to the point where they're going to have math you know proofs that humans can't understand there's a lot we don't understand I mean, my example of this is we didn't understand why aspirin worked for years.

1:56:27We used it. We tried it. The thing that got me about what I did listen to in the interview was how egocentric it was. Him saying, we thought we'd be heroes. When we solved this problem, we were the hero, right? But that's not, to me, why an academic is an academic. An academic wants greater knowledge in the world. and rather than celebrating in the part I heard that there's, wow, we can solve any more things. We can, we can put more challenges to it. Isn't this amazing? It's that I'm losing personal status. And, and, and that struck me as very egotistical. There's, there's also a legit, I think a somewhat

1:57:05Martin Woodward: legitimate point to be made that the Vi Hart made this point. I love Vi Hart. I don't know if you know who she is she does uh i don't know if this is that by heart um so don't show her yet uh she has it had a youtube channel where she did it was great she had a notebook and she would explain math concepts as she was drawing uh these things it's it's just she's just wonderful anyway she is a mathematician in fact she worked at um i think it was microsoft on the lean uh program which was designed to support proofs uh to analyze mathematic proofs to make sure um that uh they were real and she says lean is supporting this uh navier stokes proof so she thinks that's you know relevant but she talks about this and one of the things she says is you know you don't climb a mountain because there's a pot of gold on the top of the mountain we don't solve these math problems just to solve them we solve them because the journey teaches us something the climb is important and when an ai solves it it's eliminating that process you don't you don't get the lessons learned getting to the answer so she says you know and stephen wolfram kind of says the same thing he said uh at a talk you gave this week yeah theorems shmire them you can make up a lot of theorems doesn't mean they're useful so their math community itself is divided somewhat on this um there's also further reporting that that the um the the apparent theft of the process from the nyu professor there's some debate because they did dig into the proof and say well actually the it came from a slightly different path so there's argument about that too so it did its own work Yeah.

1:59:06Martin Woodward: So I don't know what to say about this. Coding has, I think no one would doubt this, fallen to AI. Chess felt. I didn't see any coders crying. Oh, there are still coders crying. There are a lot of people very sad that the thing they spent years in school learning and many more years perfecting in business and their work is now less valuable because a machine can do it. um and i think there's definitely some grief he said strogat said uh we're we're at different stages of the uh of denial of the grief what is it yeah yeah i thought that was kind of what's what's fascinating to me is we thought that the jobs that would get eliminated by this technology or you know uh accountants and and and factory workers and blah blah blah mathematicians was not on my all right but first they came for the the chess players and i didn't say anything because i wasn't a chess player then they came for the go players and i didn't say anything because i wasn't a go player then they came from the coders then they came for the mathematicians eventually i mean this is it it's going your if you're an accountant if i would very much if you're a bookkeeper let's say maybe not a cpa but maybe if you're a lawyer um we're seeing very strong models actually there's a new uh open weight legal model from glm glm is the model i use called glm legal and of course we talked before about the reuters thompson reuters model i'm sorry did you see darren okey of paris going through the airport oh is she holding a cat yes she is oh darren it was only a matter of time before ai then the artists came oh well that's great i think that's uh Yes, that's very funny.

2:00:56Martin Woodward: See, I have to say something. Look, Darren is not capable of making this illustration, but he is capable of thinking of a clever premise for this illustration. Yes, which mathematicians should do too. There are places for the human. I mean, if you just said Paris going through an x-ray with a cat, you would get a pretty anodyne picture, probably like the one on the left. You wouldn't get the x-ray of the cat with devil horns. i'm attaching a photo of gizmo yes we want to make it more accurate he knows what paris looks like but he doesn't have gizmo quite right i mean this won't help for the photo of me holding gizmo what's mom's eta how close are we getting i mean she hasn't responded potentially within 10 minutes she could be coming up she could be on the stoop right she could be on the door right now i feel like i mean no that's the issue is my apartment is weird and busted where she cannot buzz and i have to go outside to let her in basically well she could but okay we'll wait because we want to talk to her i don't think i can uh just force her on the podcast oh let me tell you getting my mom on the podcast was a multi-day my father really wants to be on the podcast i'm sure my mom would love to be on the podcast i still feel bad about that i'm so sorry dad dad well maybe we should zoom him in i mean we should i'd be happy to but no we'll we'll see maybe um over the holiday we all want to know about paris we're all looking to get those those little details that only a parent can share i mean yeah my mom says he has ai psychosis so you guys could bond over that yeah you see you see it's not just me um what did you think about okay you go you no no what did i think go ahead you do it did it is a democracy ukrainian no it's not did the ukrainian attack on the yandex i love that that was my last story and i will let's do that as our last i have a lot of silly stories and i i actually been reading this last week uh ukraine uh bombed yandex data centers in russia with drones and interestingly and i think people on x have confirmed this suddenly the bot traffic on x disappeared now there's an argument as to whether or not it was that the bots are resident there or that because yandex's was not available as a as an access point but i don't i don't understand the difference cloud flare reports that bot traffic from the yandex network has dropped to just 0.4 percent it dropped 57 percent in october 8th when the first ukraine drone hit a yandex data center housing two supercomputers in sasavo russia a second bigger yandex campus 220 miles to the west was knocked offline uh if this says yesterday so october 10th not not yesterday are yesterday uh and now it is 0.4 percent that's like 99.6 everybody shares an elon musk tweet that had hardly any likes to it oh see that's what's going to be interesting i didn't even think of that so on x the everything app where you do your where you do all your banking and finance yes you know they had a pretty good uh apr on the savings so i put some money in there i thought that It's a good place for a savings account.

2:04:30Of course.

2:04:32Martin Woodward: I haven't done anything else with it. You just put money there that you hope will be there, unlike the last time Elon Musk tried to make X.com the site for all your banking and finance. But did people lose money? Isn't that what happened? Didn't it fail? Oh, I don't. I think that would be a big story. No, this was decades ago. Oh, decades ago. Oh, the original. Oh, he's before PayPal. I don't know the story of that, actually. But speaking of failing, character AI has filed for bankruptcy. This was the platform where you could talk to personas, and they were getting sued to death. Lots of issues.

2:05:18Martin Woodward: A slew of lawsuits. The company said it had assets and liabilities between$50 and$100 million. dollars they they've really tried they say um you know to make care character ai safe for young people but apparently it's uh it's only five years old it's founded by former google engineers like one of the authors of the transformer paper they're selling uh assets to disney for 20 million okay there was a um uh the people over at jason cobbler over at 404 media uh got their bankruptcy documents from the Delaware court and basically found that it never even came close to making money in the entire time of its existence.

2:06:04I mean, it was spending, you know, at some point, like$13 million a month on AI compute or 10 million, 9 million. And the most it ever made was 8 point or 2.8 million. So, you know, multiples lower than the amount of the major investors on compute that's a great question who were the major investors yeah who is going to be who's left holding the bag uh yet another uh and horowitz lightspeed venture partners the usual great croft nat friedman elad gill the another literary prize winner seems to be AI, according to The Atlantic, the prestigious Kane Prize. So let me stop you here for one second, because I got in trouble for this.

2:07:00I don't like prizes as a rule, because I think they're exclusionary. And so when this came up, I said, weird that AI could kill the prizes. This is a prize for black talent in Africa. That would be terrible. So that would be bad. And so I picked on the wrong prize to say that. But the other interesting thing is that the attacks then on winners of prizes who are doubted because of AI tend to be people of color who are the objects of those witch hunts. All right.

2:07:30Martin Woodward: Well, you know, because you cannot determine authoritatively the provenance of something, let's not cast aspersions at this point. A. B. Again, I think prizes are exclusionary. Only one wins, and that's the problem with them. I want a wider pipe. Leo, have you checked behind your monitors at home to see if you have any toilet paper wrapped things? That's a story. Go ahead. Tell us that story. So, Paris, you're the one who spotted it. You tell the story first. I mean, I'm trying to find the link right now on here. OpenAI tweeted yesterday that it had addressed basically a security incident going on with one of its employees and that more information would soon be available from the DOJ.

2:08:13Basically, what ended up happening was that an opening eye engineer was arrested in bugging an executive's conference room. Where secret discussions were held. It's a special conference room where the big one is pet. He basically was arrested in suspicion of hiding an audio recording device in this conference room where executives were meeting, having kind of top secret or at least like really sensitive meetings, they had kind of pinned this, or tracked this back to him, Alphozan, who worked as an engineer at OpenAI since last July.

2:08:58Martin Woodward: There was video of him coming back to retrieve it. Allegedly. So it was hidden. No, that's not it. By the way, this is a fake. This is a fake, I say, a fake. that was Dario Mone looking behind the TV in the OpenAI conference room. Why did he wrap it in toilet paper? That seems to be kind of... It's not exactly as if that's good camouflage in a conference room. It's not a red flag, but it's a white flag. The device was plugged into a USB port in the back of the television and then wrapped in toilet paper for some reason. OpenAI personnel found it on September 30th, but the company determined after they analyzed it It had been installed just two days earlier and its content were encrypted.

2:09:43Here's where it gets kind of interesting, though. The OpenAI and prosecutors have said in statements that, quote, a chatbot tied to Al-Fazan's spouse, which shared identifiers with his own account, submitted prompts between September 24th and September 30th about building, installing, obscuring, and remotely accessing a recording device. So they said how to build a continuous audio recorder and encrypt the SD card. And then on September 27th and 28th, right before it was allegedly installed, this account belonged to this OpenAI employee's spouse asked about the specific device that was later found, including its unique Mac address.

2:10:30So one, it seems like people presumably were using chat GPT to try and spy on chat GPT. And then they bragged about how chat GPT nailed them. This is really pretty much an inept story from beginning to end.

2:10:50Martin Woodward: I do love the deep fake videos we're getting. Here's yet another one. This is Dario Omode sneaking in to put the toilet paper wrapped microphone into the TV. And here comes Sam Altman with his security outfit on. And he's got a bug scanner. And I think Sam has detected. What's resting Sam face? It came in a plastic bag. It's a$14 item. And now, for some reason, Sam's in a tuxedo, and he's wrestling Dario Amode to the conference room table, and he's throwing him through the glass. And, oh, my God, the humanity. Yeah. I think that's probably exactly how it happened. I'm just saying. All right. Last break, and then picks of the week, and then mom.

2:11:44Martin Woodward: You're watching Intelligent Machines with the wonderful Paris Martineau of Consumer Reports fame, Jeff Jarvis. I envision mom out there. Do you need help, madam? Is there a problem with you standing here? Well, my daughter is on a podcast. She's on a podcast. She's on a podcast. I'm so proud of her.

2:12:06Okay.

2:12:09Yeah. Sorry, you're doing that. Sorry. I actually was trying to remember a joke, but I forgot it.

2:12:15Martin Woodward: I'm not good at joke telling. Anyway, our show today brought to you by you, the people. That's what we love to hear, our wonderful Club Twit members. If you're not a member of Club Twit yet, can I implore you, entice you to join the club? Ten bucks a month gets you ad-free versions of everything we do. It gets you access to the Club Twit Discord where all those crazy AI images show up, but also great conversations. It turns out if people pay to be in a social network, it's a better social network. Plus all this programming. We do so much great programming. We've got good stuff coming up this week.

2:12:54Martin Woodward: I think, do we have an AI user group? I can't remember. Oh, Micah's crafting corners tonight, 6 p.m. Always a lot of fun. AI user group is going to be November 6th, so I won't be here for that. We've got photo time with Chris Marquardt, the wacky Jeff Atwood off by one. but mostly your ten dollars a month is here it keeps us on the air we only make about enough to do about 60 of our expenses with advertising yes we have ads thank god but it's only 60 the other 40 it's covered by club members so if you like what let's get it to 50 folks come on come on No, no, wait a minute. That means fewer ads, more club.

2:13:37Martin Woodward: You know what? If we only got, right now it's about 1.5 % of our total audience. 15 people in 1 ,000. Oh, come on, people. If we could get to 5%, we wouldn't need ads. 10%, I'd be living on a boat. So let's stop at 5%. No, it doesn't go into my pocket. I'm living on my retirement. what it does is it pays people like Benito our wonderful hosts thanks club twit Benito is appreciative twit.tv slash club twit please sir may I have some more please join the club any word from mom um she will be arriving within minutes if you have to leap to your feet and get her to go right ahead she's not being very communicative oh she's texting she's typing six minutes okay good you have just enough time just enough time for your pick of the week of the week and i don't know which one to pick honestly i guess i'll do this one just quick which is that nick cage this week said he's probably not working with amazon again why because of ai what um he said at comic-con amazon's a very ai friendly company to the tune of 50 billion invested in open AI and I'm not an AI friendly actor so it's like connect those dots I'm not a member of their club I didn't sign the waiver to let them use AI on spider noir and I'm probably not going to be working with that studio again but that's okay they did cancel spider noir that might also have some canceled it right before it won like or it got 11 nominations it's brilliant awards which the most of any prime video program this year and it amassed 2.6 billion minutes views in its six weeks for six weeks and got praise from critics and audiences alike watch the black and white version because i really love the noir black and white of it i think it's great really liked that um cage kind of followed this up by saying though i've got to get with these young directors who are gen z they're going at it their own way whether it's their youtube channel or whatever it is they're doing it from scratch and by hand and i want to get with that good so So he might, it sounds like, do a true detective, which would be great.

2:15:52He also said he might be doing true detective, which would be interesting.

2:15:55Martin Woodward: Well, we love Nick Cage. And whatever you decide to do, Nick, all is forgiven. A hyper-local pick, I will also say, if you're in the New York area and you hear this before the weekend, Nighthawk is showing Vampire's Kiss at Prospect Park on Saturday. Another classic. Which is a phenomenal Nick Cage film. And a perfect Halloween. What are we all seeing this weekend, though? what do we all have to see this weekend what's opening tomorrow oh what i'm seeing tomorrow digger um no you can see everything you can see everything is that tomorrow it's i mean i think it i think it opens uh globally or nationwide on friday but i'm seeing it tomorrow they're doing like did you see andrew wang's video take off of that movie with the muse fluffy muse creatures oh no yes i won't play it it was funny all right my pick of the week sorry i'm in a meeting uh this is a yes it's an ai website uh in a meeting i'm in a meeting.com where you get they say plausible meeting audio to put behind your phone call what what are you in are you in an IT incident.

2:17:07Martin Woodward: This has been around. Oh, is it? Yeah. Well, it's the first time I saw it. Is this really old? Yeah. I don't know if this is, but this is AI generated. Oh, well, that's what's new about it. Okay. So here's what I'd suggest. Because you can add your name to it. Two clinics right away. So you can really personalize it. It's not very good. No, Honestly, your call. But I'd pick your... I guess if we're in the background as you were talking, maybe it would sound real. That's Dayton. That's easy. To me, it sounds pretty AI. Yeah. New and bad. New, not old. All right. Well, you don't like that. You're muted.

2:17:48You need to unmute. They need to throw that in there.

2:17:50Martin Woodward: That would be more. Yeah. All right. I'll give you one more. This is really cool. I have already installed it. It's very interesting. Remember, we were talking about ArtCraft, those vibe-coded reverse engineered adobe tools well get ready there's going to be a lot more there's a new skill out for ai called rea reverse engineer anything it combines a bunch of tools have been around for a while including ghidra which is a decompiling program from the nsa plus some other really useful tools but in the past you need to know what you were doing to decompile a program and reverse engineer it. Now, these tools are put together.

2:18:33Martin Woodward: It gives your agent the information it needs plus the reverse engineered code. And it makes it possible for you to reverse engineer anything. And I think it's pretty interesting. By software. By software. Yeah, I think it's really, really interesting. R-E-A, reverse engineer anything. All right. First, before Paris leaves, as a service to Benito, Can the three of us do our Elizabeth Holm eyes?

2:19:014.35 p.m. All systems green. Okay, Benito. Thank you. Thank you. All right. That was Ripley, my security officer. I'm going to go downstairs and let my mother in my building. All right, Paris. Thank you, guys. Thank you, Paris Martineau, everybody. Bye. Sorry to leave early again. Bye. Some reports. Bye. Say hi to mom. So I watched Cupertino. Did you watch Cupertino last week? I came out, didn't I? And I did not. It did come out. Yes, it's pretty good. What did you think? I liked it. It's from the creators of... The Good Wife, which is a great show. The Good Wife and The Good Fight, which were great.

2:19:32So this is two Silicon Valley attorneys who get... And they fight back in the first episode. And we'll see how it goes from there. But I liked it. So Thursday nights on CBS, that's one. I haven't watched network television. I haven't either. I haven't either. I had to record it. Yeah. And five fleet approvals are pending in Buzz. No issues. that you're it's your agent talking to you that's ripley uh and then my favorite thing of the week

2:19:59Martin Woodward: is zuck's guinness face oh yeah when he takes a sip of guinness and so he takes a sip again let's play the video there's no sound you don't play sound this actually uh is a uh podcast which has a clever premise which is you sit in a fake pub he's there he is with andrew wang and that's mark taking his first sip yes whoa that's pretty strong stuff says mark actually it's a good interview so then mark responded to that to that post and just said that's just my face he might actually i wouldn't be surprised if he likes guinness he says he had two beers during the interview so what's the podcast uh it's a pub what's it called pub oh i can't remember um so it's a fake pub well i'm assuming it's a fake pub i don't know if you'd be able to take over a pub and do a podcast there like that anybody know the name of the podcast it was a good video i watched it because i mean it's a plug for muse that andrew wang is the the what is he 23 24 year old that they hired for billions of dollars who made muse effectively uh and mark zuckerberg i think has something to do with facebook it's a pretty good interview they they you know make it a pitch this is why i don't interview people like that because they're too well trained they're never going to say anything but plug the product somebody like our great guest martin woodward he's just great he's real he's real yeah yeah uh jeffrey it is It's time for us to say goodnight to all our family.

2:21:41Martin Woodward: Jeff Jarvis, professor. Well, he was at CUNY. He's now doing curricula for Montclair State University, Montclair, New Jersey, and SUNY Stony Brook. Are you going to ever take to the classroom again? I don't think in full class. I don't think so. Maybe a little seminar? I talk to classes. If anybody teaches my book, Hot Type, I'm happy to zoom into the class. I do that a few times. The book Hot Type is available at jeffjarvis.com. There are even some signed copies. Signed copies from the Mockler Book Center. You can find the link at my redesigned page, jeffjarvis.com. And he is working on a new series of books about AI.

2:22:20Martin Woodward: How's that going? Going well. We've signed four writers so far, and I think we're about to sign two more. And what is the premise of this series? The premise is that this is not a technological series. It's not technical. It is inviting writers from many disciplines to reflect on AI and how AI reflects on society. That's interesting. So the first, Ruben Chowdhury, who we had on, is writing about what is intelligence. Matthew Kirshenbaum, who we've also had on, about the textpocalypse. The author of Black Software, Charlotte McElwain, is writing an optimistic take on AI and Black technologists.

2:22:59we just signed another book from I think we're signed I think I can say this Jonas Iverson from Sweden about what happens when AI makes life too damn easy takes away all the friction wow which is a great idea another book I'm hoping to about to sign is about taking back the definition of progress from the technologists

2:23:22Martin Woodward: oh I love that so really kind of thoughts like that I mean I thought of another idea today is I want AI in war we need ai and doom oh yeah ai and god ai and creativity so that's that's the kind of what it is it'll be books that are independent books as part of the series to reflect on um all these changes that are coming to society through ai thank you for the plug we live in an amazing world and that's what we cover this show is more of our philosophical look at ai some of it's what's happening but a lot of it's about the philosophy we need to be contemplating skulls on this show so AI and society uh we also do the AI user group which is much more down and dirty and doing stuff and using stuff and uh that's always a lot of fun we're now doing that we're going to try to do that twice a month uh so watch that if you're a club member uh you'll get an announcement automatically and then Micah's show oh and Micah's doing hands-on AI yeah it's that's really good because he's talking about you know i i am in some sort of weird ai rabbit hole but he's talking about using ai in a reasonable fashion now i'm going to be gone for three weeks uh who is uh hosting the show benito i don't even uh no mr lumoresca maybe maybe not we weren't no no louis next week yeah louis definitely next week he of course does ai co-pilot for microsoft he's his head of engineering there jason heiner who's the editor-in-chief of the deeper view of the deep view which is an ai journal and our good friend mike elgin he has his machine society newsletter he's the one who might not be able to make it is that right yeah i'm still we're still working that out yeah there may be a scheduling a scheduling issue yeah yeah uh we'll have mike on no matter what anyway actually all three of them are great we've had them on the show before we will have them on again i will be back though uh 11 11 that's easy to remember i might be a little jet lagged because i'm flying in the night of the ninth but we'll see and and you can ask me then whether i pulled the trigger on a 512 gigabyte very very expensive we don't even know how expensive if you have to ask you can't afford it mac studio oh i haven't decided i'm going i'm i'm just like teetering don't you want to wait for the want to wait for the uh vera rubin then well yeah but it's you know with computer hardware waiting is sometimes a bad idea it's so hard to tell actually this is something i should have talked about more on the show i've been talking about it what can't you do now that you could do with that uh it would just give me one more machine with a large pool of memory the biggest pool memory i have is 256 gigs on those sparks uh it's probably not as good for inference is the sparks the sparks are built for inference but it does have that large pool of memory so i my i my goal uh from day one with this was i want to build something that is a useful uh assistant for me in my life especially as i age and my memory goes uh but i also want as far as possible to be doing do it locally and i probably have enough horsepower now i have eight total models running on four different machines right now to not use the cloud except for coding coding it's still you still not there but i want to be in position with a large enough machine that if in the next year i've convinced there will be models that will fit in 512 gigs and it may not fit in less but i'm convinced that we models that will fit in 512 gigs that will be as good as let's say opus 5 5 today that's probably eight months out i want to be able to use that model when it comes out and i feel like nothing i have right now will be big enough i also am concerned that it'll get harder and harder to get this stuff so yes if there's a spark 2 with vera rubin in it i will kick myself actually i won't because i will be able to sell that mac for probably exactly what i paid for it if not more so that's the other thing you know there's a lot of question marks i've been doing a lot of benchmarking and trying to understand where the world is headed between now and the time i come back models will have come out from z.ai there will be glm5 which should be amazing there will be quen4 from alibaba which should be amazing very powerful open weight models i am you know to day after tomorrow no tomorrow we're going to see the first open weight model in the new nimetron family from nvidia and i have very high hopes for what jensen wong and nvidia will do they have been acquiring and partnering with very good companies and creating local models we're looking to acquire or acquire reflection yep uh and they just acquired another company so i think by the time i come back there will be some strong uh nvidia models as well and of course those will run best on sparks or you know what i'm hoping really is that jensen wong knows that apple is about to drop this mac studio and it's going to make an announcement before that happens just hard to get memory well and that's what i've been talking about the other shows i think that's a conspiracy i i really don't believe that i think there's a lot of memory there's an infinite amount of memory because most of those data centers that the memory is committed to have not been built most of them will not be built true and i think that's another thing to watch carefully is the november election because many many candidates are running for election that are anti-data center and if if they win then even more data centers will not be built even more memory will that is committed now and is not available now will suddenly hit the market i i predict it's very likely in six to eight months there will be a memory glut and a plummet in the price of memory i think these companies are gouging right now i think that they they don't they could sell the memory a lower price the memory's there it's not been used right where's all that demand coming from yeah so i i think we're should i think shenanigans i think there's shenanigans going on and uh and that's why i'm very nervous about everything i i i made the right move to buy the sparks when i did i think that's i was i don't want to say i was prescient i was lucky in buying him when i did and i'm very happy with him um but i'm not sure what the next move is so we'll all have answers for that i bet you in a month 11 11 i'll be back we'll talk thank you jeff have a wonderful month are you going to be talking to your agents while from around the world did lisa set any rules no she doesn't even they talk in my hearing aids she can't actually the other day she said is somebody talking to you i said i said can you hear that she said yes i said oh so maybe she will know i'm gonna i'm gonna take a little break i think we have some really exciting stuff to see we're going to southeast asia i think i'm gonna have a lot of fun and probably will not spend much time with my they won't be lonely though they'll find stuff to do thank you everybody for joining us we do Intelligent Machines every Wednesday, 2 p.m.

2:30:56Martin Woodward: Pacific, 5 p.m. Eastern, 2100 UTC. You can watch it live as we do it. If you're in the club, in the club to a Discord, YouTube, Twitch, X.com, Facebook, LinkedIn, and Kik. We stream on all those platforms. Of course, you don't have to watch live. We record it. Amazing. And so you can download the audio or video of every show from our website, twit.tv slash I am, or go to YouTube. There's an Intelligent Machines channel on YouTube. You can watch any show there, all the video. Or best thing to do, probably subscribe in your favorite podcast client. That way you'll get it automatically, audio or video, and you can listen whenever you're in the mood.

2:31:34Martin Woodward: I do appreciate your listening. Thank you for being here, and we will see you next time on Intelligent Machines. Bye-bye.

From the publisher

GitHub's developer relations VP, Martin Woodward, reveals just how dramatically AI agents are reshaping who codes, what gets built, and why open source maintainers are overwhelmed by a tidal wave of pull requests. If you thought Copilot was disruptive, wait until you hear what's coming next.

  • Gemini 4 Argon begins rolling out in Gemini app and Antigravity, but only if you pay
  • (11) Andrej Karpathy on X: "Dusting off this tweet from April because this gap in shared understanding of LLM capability is *widening*. It's now less "two groups of people speaking past each other" and more a sharp funnel. - Napkin math somewhere around 6B people (~75% of the population) have barely come" / X
  • AI models may be capable of self-replicating without access to a datacenter
  • Zuckerberg called Muse ready despite the risks, New York Times reports
  • Muse Wonder
  • OpenAI's math breakthrough points beyond math
  • GLM-Legal: Advancing Agentic Legal Intelligence
  • Russia's Yandex bot traffic nearly eliminated overnight after Ukraine drone strikes cripple multiple of 'Russia's Google' data centers — Ukraine intensifies bombing of Yandex data centers in response to Moscow's strikes on Kyiv data centers
  • Just 5 years in, interactive chatbot platform Character.AI files for bankruptcy
  • Yet Another Literary Prizewinner Seems to Be AI
  • Nicolas Cage Says He's 'Probably Not' Working With Amazon Again After Refusing to Sign AI Waiver for 'Spider-Noir'
  • Fleeting — Sorry, I'm in a meeting
  • REA — Reverse Engineering for Your Coding Agent
  • Cupertino
  • Zuck's beer face

Hosts: Leo Laporte, Jeff Jarvis, and Paris Martineau

Guest: Martin Woodward

Download or subscribe to Intelligent Machines at https://twit.tv/shows/intelligent-machines.

Join Club TWiT for Ad-Free Podcasts!
Support what you love and get ad-free audio and video feeds, a members-only Discord, and exclusive content. Join today: https://twit.tv/clubtwit

Sponsors:

More from Intelligent Machines (Audio)

All episodes