

11 Nov 2026 · 1h 19m · 31 chapters

A retrospective on how AI changed from ChatGPT’s 2022 launch to the present, arguing progress has followed an exponential curve and is shifting from chat-only models to “machine labor on demand” (long-running agents/robots), with people acting like “venture capitalists” managing compute budgets and project partnerships.
Corey and Grant (co-hosts). No other guests are interviewed in this transcript. Their backgrounds are implied: Corey and Grant discuss AI model history and industry developments; Corey also references being a writer/creator at “the Neuron” and using AI tools in daily work, while Grant curates a timeline/spreadsheet of model milestones.
Agentic systems and tool use became mainstream within months (especially around late 2024–early 2025). Reasoning and tool use became default by mid-2025. Safety/open access tradeoffs are solvable via “safe by design,” but incentives can push labs to hide capabilities. Model release cadence has accelerated (roughly every 10–11 days by 2026).
withholding “Claude 5.5/Mythos” over cybersecurity risk; DeepSeek R1 threatening closed models; “OpenClaw” inspiring later agent architectures; ZoomInfo’s “GTM Context Graph” plug; GPT-4 Turbo’s Data Interpreter (Python terminal from spreadsheets); GPT-4 Omni/multimodality; “reasoning model” OpenAI o1/“Strawberry/Q*” era; ChatGPT launch on 11-30-2022.
Written by AI. May contain mistakes.
Explore the potential future of AI and our evolving roles in relation to it.
“I think at some point this is going to take the form of what I call machine labor on demand, which is at some point we're going to be able to call not just AI models, but full robots to do things for us.”
A look back at the significant advancements in AI over the past four years.
“Because in putting this together, it's insane.”
Discussion on how technology progresses at an exponential rate and its implications.
“And the concept is that technology is increasingly basically building itself or progressing, right, at an exponential rate.”
Exploring the long-term potential of AI to change humanity for the better.
Overview of the hottest AI models and their features in today's market.
“that changes how we not only understand the universe but experience the universe and understand what we can do inside of it and beyond our own planet even.”
A brief reflection on the rapid changes in AI within just a few months.
“and stuff that was science fiction at the other end of this list still.”
The controversy surrounding the release of Claude Mythos and its impact.
“But we're going to step back now just four months, okay?”
Discussion on the evolution and convergence of AI models and agents.
“And that was maybe like a month later after the whole – It was sometime in June.”
How OpenClaw's emergence changed the landscape of AI tools and agents.
“N8N is cool, but like, if you have to go build out a specific structured workflow for every step of what you want it to do, it kind of never felt agenic to me.”
Discover how OpenClaw transformed AI development and inspired new architectures.
“So nine months now since the end of January.”
Learn about the progression of AI models leading up to GPT 5.3 and their significance.
“Yeah, they all have adopted the OpenClaw pattern, basically.”
Explore how DeepSeek R1 shifted the competitive landscape in AI and model reasoning.
“So now we're going to take another step back, another six months to May 30th, 2025.”
Understand the hype around Llama 4 and its impact on AI research funding trends.
“it's crazy now uh and you mentioned llama 4 let's just very briefly touch on this prior to deep seek R1, Llama was the dominant open source model.”
Discuss the challenges and strategies related to AI safety and model release.
“But when you go and you actually use it and you realize it's kind of crazy that this is free, that I can pay for this or I can use it for free and just let them train.”
Explore the challenges of ensuring safe AI while maintaining open access.
“But now it absolutely was the right call.”
Discuss the transformative opportunities AI presents for individual empowerment.
“This is sorry for my tangent, but I think this needs to be said.”
Reflect on changes in AI models and their impacts on user experience.
“But I think the opportunity there is also massive to give people a chance to get lifted up, to lift themselves up, frankly.”
Examine how new features in AI models are changing user interactions.
“So, yeah, in that batch was GPT-40, which was our first multimodal model, by the way.”
Explore the progression of various AI models like GPT-4 Turbo and Claude 2.1 over the past three years.
“Probably right at three years by the time you're watching this.”
Discussing the initial experience with ChatGPT and its profound implications for professionals.
“By this time, we have GPT-4, which was mind-blowing at the time, by the way.”
Personal stories on how AI tools transformed workflows and job roles in writing and publishing.
“Like, I'm not sitting here right now if I hadn't gone on there and, you know, dropped a brick the second I saw it.”
Understanding the foundational role of transformer technology in recent AI advancements.
“Like it was known that there were things in the way it's been known for years that there are quantum computers.”
Explore how rapid innovations in AI are changing our mental models and perspectives.
“You really have to update your mental model for everything.”
Discuss the unexpected success of OpenAI and the implications for competitors like Google.
“But yeah, it's just wild to think about that.”
Reflect on the rapid changes in AI technology and its impact on the tech landscape.
“And suddenly it's like, oh, we need to hire a few dozen employees, a few hundred employees, thousands of employees.”
Engage in a discussion about future predictions for AI and potential breakthroughs in technology.
“Corey, I wonder if we could end this episode by talking about predictions into the future, if you have any.”
Delve into the balance between pushing AI advancements and ensuring safety from misuse.
“Because the pace has been – Let me work through those one at a time.”
Explore the future landscape of AI, its accessibility, and potential impacts on businesses and individuals.
“My hope is that models remain available.”
Discussion about predictions for OpenAI and Meta's trajectory in the AI space.
“I had like three discussions late last year, like two of them in Silicon Valley and another one, I think, on Twitter.”
Concepts of coexistence with AI and the future of work and labor in an automated world.
“And I don't think that that means treating it as like equal.”
Grant shares his farewell message and reflections on his journey with The Neuron podcast.
“And I realize to some people this last chunk, this probably sounds like crazy talk from us.”

Get the gist, explore an idea or ask about something said.
On The Neuron: AI Explained, what are the key points of this episode?
Automatic transcript. May contain errors.
0:00Grant: I think at some point this is going to take the form of what I call machine labor on demand, which is at some point we're going to be able to call not just AI models, but full robots to do things for us. And I actually think we need to reframe our thinking like we're not labor. People are not labor. I think in the future we're all going to be basically venture capitalists where we're all going to have like a compute budget of some kind, just like a VC might have, you know, money to invest in things. and we're going to need to decide who we're going to team up with to work on projects together.
0:32Corey: While it's not a given, I really think that we're going to get there. I think we'll see it in the lifetimes of people alive today and not necessarily young people. Like, I think we could see it very soon. I think we could see it in 10 years. It could even be longer, but I really think we're going to see big strides in the coming years. And I'm excited.
0:51Grant: Just look at the last four years. Imagine the next four.
0:55Corey: Welcome, humans, to the Neuron AI Explained. I'm Corey. This is Grant. What's up, Grant?
1:01Grant: How's it going, Corey? How are you doing?
1:03Corey: It's going good, man. It's going good. And today is going to be so much fun. We always talk about, wow, it's crazy how far we've come.
1:11Grant: It's crazy how far we've come.
1:12Corey: But we never actually sit down and look at how far we've come and talk our way back through it and kind of remember as we go, maybe how much something blew our mind, how long ago that was. And today, Grant has put together a really cool spreadsheet for us, and he and I are going to walk through the last almost four years of AI and just kind of take a little trip down memory lane and look at how fast this has happened. Because in putting this together, it's insane. It really is. Just having this pulled up on my screen, I catch myself going, what? and I think this will be a lot of fun Grant.
1:55Grant: Yeah, so you can look at this as kind of like a retrospective on all of the progress that's happened over the last four years and especially as we get to the end of this year and how far we've come and how much progress there's been maybe with some changes on the horizon fundamentally huge changes on the horizon it's good to look back and say like okay, well what were the things that led to this progress and to kind of keep track of it because the AI industry uses this phrase called riding the exponential or some version of that, right? And the concept is that technology is increasingly basically building itself or progressing, right, at an exponential rate.
2:42Grant: And it's really hard for humans to understand, to conceptualize an exponential. like it's like really hard to yeah to to visualize it and to wrap our minds around this around this and i think this episode is going to show people yeah this is what this looks like like we are on an exponential curve now and it might end up being an s curve meaning it might go up and then flatten out uh and maybe that's where we're going to be you know leading into next year but i kind of don't think so i kind of don't either yeah so i just i've yet to see anything that tells me oh that's obviously slowing down yeah i mean maybe with the results of you know election changes who knows
3:23Corey: what's going to happen there but um yeah hundreds of unsolved math problems falling and and new science coming out from multiple labs you've got from anthropic google deep mind meta everybody is right now just tapping into the stuff that that i like to i like to categorize as the promise of ai which is the stuff that has the ability to change humanity long term for the better, that advances us as a civilization, that changes how we not only understand the universe but experience the universe and understand what we can do inside of it and beyond our own planet even. So we're really at that point.
4:05Corey: And it just felt like a good natural time with four years of chat GPT rolling up to take a look back. So, Grant, let's start with right now. Right now, the hottest models that exist are obviously Claude 55, Fable 5.1. You've got GPT-6 Astra, GPT-6.1 Sol. We've got Crock 4.7 Mini. We're in the era of computer use.
4:33Grant: Opus 5.5 as well, yeah.
4:36Corey: Yes. Oh, yeah, I said Claude 5.5. Yeah, oh, sorry, yeah, yeah. Five. There's also a 5.5 in there.
4:42Grant: Yeah.
4:43Corey: And even a new mistral model, deep seeks around, Kimmy's around. There's a lot of stuff. We're in the age of computer work and long running agents and stuff that was science fiction at the other end of this list still. Like, like some of this was like the crazy stuff Sam Altman talked about in a, in an interview. Like I remember being like, oh, everybody's going to have agents. You're going to have agents doing everything. You're going to get up in the morning you're going to talk to your agent your agent's going to make your coffee it's going to it's going to go down it's going to get your shoes i'm just kidding
5:14Grant: you didn't say that but uh well maybe it's robotics you know like that there's a dream there's a dream of uh robotics um finally making a lot of the promise of ai come full circle and you know like with uh human intelligence you uh we're kind of doing it in reverse where where um you know if you you know subscribe to the idea that we've all evolved you know over time then And we kind of evolved physically first and then our brains caught up. Well, we're doing robotic machine intelligence evolution backwards where we're starting with the brain and then we're adding the body onto it. Yeah. Yeah.
5:48Grant: It's crazy.
5:51Corey: So for this, by late 26, frontier models are sold on long-running work, computer browser use, agent orchestration, cost efficiency, and not chat quality alone. But we're going to step back now just four months, okay? This gets us on a clear six-month trajectory all the way back to ChatGPT's launch on 11-30-2022. So we're going to step back right now to May 30th. That is four, probably five months by the time you're watching this. At that time, the hottest models were GPT 5.5 Pro, which was a great model. Claude Mythos Preview, limited access. It was not public yet. Normal people had not touched Claude four months ago.
6:41Corey: I mean, had not touched Mythos four months ago. Right. Isn't that kind of crazy, Grant? Feels like it's been around a long time.
6:48Grant: I mean, four months is a long time in the AI space as we're about to show.
6:52Corey: It is. It is.
6:53Grant: But yes, yes, I agree. The whole mythos thing, like if we flash back to that time period, what had happened was they said, we've created a model that is capable of critical cybersecurity risks. So we're going to withhold it from the public. And they withheld it and it caused a huge uproar. People were very upset. They rolled it out just to their elite partners, which to be fair were like banks and big tech companies to make sure that they didn't have vulnerabilities in their code before they released a model that could potentially find those vulnerabilities. So it was fair. But then they finally released Fable.
7:31Grant: And that was maybe like a month later after the whole – It was sometime in June.
7:36Corey: So yeah, it's probably three months on Fable.
7:38Grant: Yeah.
7:39Corey: That's crazy. Opus 4.8 was the highest generally available from Claude at the time.
7:45Grant: And that was the one that would constantly tell you how your code was load-bearing and just talked like a complete – like you could just not understand that thing, right?
7:58Corey: That's so crazy. Like it was giving you the caveman talk. Yeah.
8:02Grant: Like the code was good, but it was just like it had a lobotomy or something.
8:07Corey: Let's see what else was then. Gemini 3.5. That was kind of a dud.
8:15Corey: Was that not nice?
8:18Grant: You know, 3.5, that didn't really touch it much. So can't speak to it. No, I didn't either. Muse Spark. That had just come out.
8:27Corey: Yeah. Yeah, I don't think that was the first Muse Spark. If the first Muse Spark was four months ago, Meta's on a trajectory to beat everyone.
8:36Grant: That might be why they're pretty afraid of Meta right now. I think Meta has everyone's attention, which I love because I think Zuck's an interesting voice in the AI space now.
8:51Corey: Like he often says the thing that no one else is saying. Like he's pretty consistently taking the opinion no one else has, I think. And I tend to like that. um deep seek v4 god it just quinn 3.6 kimmy k2.6 glm5 was a good model so here we're we're in the era of i guess long horizon multi-agent million context is kind of the norm now at that phase
9:23Grant: it used to be kind of the the exclusive domain of gemini and at a certain point everybody began
9:30Corey: offering it yeah yeah that's uh that's wild yeah so that that had happened because deep seek v4
9:40Grant: had made one million context standard um yeah like it just they showed kind of the architecture for how to do that and yeah everybody adopted it and everybody else better get in line
9:53Corey: uh that's so wild now here's where it's going to start feeling it'll look crazy uh we're going to go back another six months from there to november 30th 2025 at that time year ago this one yeah one year ago actually from the time you're watching this right out of gear yeah um gpt5 was their hottest model gemini 3 was around um claude sonnet 4.5 was observed maintaining focus for more than 30 hours uh and that was crazy then yeah it still sounds crazy if i'm being honest uh oh kimmy k2 that's when kimmy k2 came out i remember that was
10:44Grant: That's a good model. Yeah, that one's really good.
10:47Corey: It was a really good model out of China from Moonshot that makes Kenny. Yeah.
10:53Grant: That's right.
10:54Corey: Yeah, and they've since continued to make a splash, but that was their first. I say that was their first. There may have been a K1.
11:01Grant: No, K1 was the first. Yeah. Okay.
11:05Corey: K2 is the one that made the splash for them, though.
11:08Grant: Yeah, I think K1 was people were pleasantly surprised with it. It was a really good writer. I think Kimmy K2 was a better writer. But K1 was one of the first ones where people were like, oh, wow, this thing is pretty good for open source writing. And then, yeah, K2 came out and I think smashed a lot of expectations and really lit a fire under people.
11:33Corey: GLM was doing really good then too. I see we have GLM 4.5 then. You know, the thing to know here in our notes, it says models and agents converge. And so 1130 is the moment when agents changed, I would say. I would say from there through the next 90 days, the world got wild. Claude Code has this insane December of explosion, explosive growth. Everyone's using Claude Code. It's amazing. At that same time, OpenAI reinvented Codex, and it started to catch on as well. And we move into this month where suddenly everybody's using agents. We're using agents for everything we're doing then. And, like, we had tinkered before, but as far as, like, it was the first time that interacting with a tool felt like it was really doing things.
12:34Corey: If you're in go-to-market, you've seen the AI demos. Agents, automation, sales workflows that practically run themselves. But without the right data, you're just scaling mistakes. That's why ZoomInfo rebuilt its platform around the GTM context graph. It connects everything from your CRM, conversations, and website activity to ZoomInfo's verified company data, buying signals, and market intelligence. Now you can build AI agents, automate outreach, and orchestrate workflows without piecing all that data together yourself. That's what the GTM Context Graph unlocks. AI-ready GTM Context in a platform that makes building easier than ever.
13:14Corey: Meet the new Zoom Info at zoominfo.com slash the neuron. And now back to our video. Prior to that, we had things like N8N. N8N is cool, but like, if you have to go build out a specific structured workflow for every step of what you want it to do, it kind of never felt agenic to me. It felt like... Yeah.
13:35Grant: It's too manual.
13:36Corey: It was too manual. It was an automation. Like it's not really an agent if I have to go build something to do the job. It's an agent. If I can say, hey, I need you to do this, need you to do that, need you to do this. And it can bounce between tools. It can bounce between windows. It can use your computer. And that was the era that leaned into it. This list also doesn't include the OpenClaw moment, which would have been January. So nine months now since the end of January. We did a great interview with Vincent Koch recently that you should watch. It's really good. But like OpenClaw changed everything, in my opinion.
14:15Corey: It had its moment. It's, you know, instantly over the course of like three weeks, suddenly it's the most starred and downloaded GitHub repo. Ever. Like it surpassed Linux. And wild, which is wild. And before you know it, there are people lined up in China outside getting claws built for them. Like everyone in town has showed up with a computer and people are building them claws. And it became this viral moment.
14:45Grant: And it inspired its architecture, inspired a lot of the agents that you are using today, whether you use. Muse, whether you use GrokBot, whether you use anything agentic inside Microsoft, Dots from OpenAI. Yeah, they all have adopted the OpenClaw pattern, basically. How it works.
15:08Corey: Yeah, because everyone leading in tech saw it, too. If you'll recall, Zuckerberg and Altman were kind of in a bit of a bidding war there for Pete Steinberger at one point. Yes. It wasn't clear where he was going to go. And the truth is, can you imagine right now if he'd have gone to Zuck? I don't know if it would be much different.
15:32Grant: I don't know if it would be much different because Muse is basically a good open claw. Yeah, probably the best open claw.
15:40Corey: Looks like they did just fine without Pete.
15:43Grant: Yeah. I mean, you know, and that's the thing about open claw is that it's open source. And so I'm not saying that they built on top of it. But certainly they could have if they wanted to.
15:52Corey: Yeah, they're definitely inspired by it heavily.
15:55Grant: Yeah, and I think that's the beautiful thing about open source projects like that. Medicaid took it and removed the friction.
16:02Corey: Sorry. Go ahead. No, you're good.
16:04Grant: You're good. You're good. Yeah. Yes, I agree. Yeah, they made it super easy to use for normal people because famously the horrible thing about open closet is so difficult to install. And like now you have to understand how computers worked at a deep level, like for security reasons to be able to use it. That's wild. Muse like, you know, abstracted all of that away. It did. One other thing to note.
16:28Corey: If you were going to say when did agents come about, when were agents born, I would say it was December 2025.
16:36Grant: 5 and and the point i want to make on that um is that you know this isn't reflective in us benchmarking in november but by december opus 4.5 has come out and opus 4.5 was a model that changed a lot of people's opinions on what you can do with code in fact i would say it was probably like the first like truly like oh i can code with this model and like trust it to create something And so many people had this explosion of use over the winter break because of Opus 4.5. They're making all kinds of side projects. And then you have OpenClaw riding that wave on top of that. It was sort of like two big breakthrough things all at once.
17:18Corey: Yeah. And that's moving into – by the time you hit February, you're also moving into GPT 5.2-ish, I would say, which was kind of their first that was like, okay, I'm just going to let this thing code. And by GPT 5.3, the codex version, because they had a codex model, that thing ripped. It was really good. It's been a while since I had a model where I was like, God, can the next generation just come out? I can't remember the last time I was like, these are crap. This is terrible. That hasn't been a thing for a while for me.
17:59Grant: They haven't picked up on this signal. They should pick up on this signal. Like, you guys, it's okay. You can take your time to release stuff. We don't need a new model every, you know, 20 days. It's actually at 11 days as of 2026. New model every 10 to 11 days.
18:16Corey: That's so nuts. Okay. So now we're going to take another step back, another six months to May 30th, 2025. On May 30th, 2025, the main models around were DeepSeek R1, which seems like forever ago. Claude 4, which also seems like eons ago. Gemini 2.5, I don't think Pro yet. I think Pro came out in the window between these. Llama 4, Quinn 3, and O3, and O4 Mini. That O3 is just over a year ago. That's crazy. Yeah. I feel like I used 03 for a very long time.
19:00Grant: And it might have came out before then because I feel like the jump from 01 to 03 was a pretty fast follow, if I remember correctly. Yeah, it looks like about six months between the two. Yeah, we have the exact timelines in here. We can pull it up if we need. But yeah, it was around there.
19:16Corey: This is where reasoning became the default. Tools became part of reasoning. I think those are two really big things to not overlook. that was a hell of a change when suddenly it's not just able to think. It's able to think and go search the web and do things with a file maybe. There was just a lot more you could do. And 03 was, for me, 03 was a good, strong model for quite a long time. Like I just really liked how it reasoned. It felt like such a big jump over everything else I was used to.
19:53Grant: It was highly intelligent. And I think that was the first one. You know, O1 had a little bit of traction, but O3 really took off that that was popular with scientists and researchers. I remember this.
20:05Corey: Yeah. I would also say every one of these were mind blowing when they released. Like, probably Grant and I were about all of these at some point. And I think it's really important to put that in perspective that like, most everything on this list was great at one point. I'm a four, not so much.
20:30Grant: well let's talk let's talk about deep seek r1 um which had came out i believe r1 came out in january um so like right right kind of leading up to these months and it was still sort of the
20:40Corey: dominant talking point if you could find it on a stock market chart now what i was gonna say like
20:46Grant: the the the dominant like financial media in particular we're talking about deep seek in the deep seek moment for like you know probably six months this entire interval right leading up to this. DeepSeq was everything everyone could talk about. Because, you know, it was the first time that the supremacy, for lack of a better word, of the U.S. open, or sorry, the U.S. closed source models was threatened by an open source competitor. And now all of a sudden, businesses could rely on, you know, spinning up a DeepSeq R1 instance and have a reasoning model. And they don't need, you know, a one, for example, to do that if they had to.
21:22Grant: Yeah. So that was a very big deal at the
21:24Corey: time it sure was it sure was by this point every major lab had reasoning and or tool using agent models so that was may 30th last year was the point where everyone had some level of agents
21:40Grant: it's crazy now uh and you mentioned llama 4 let's just very briefly touch on this prior to deep seek R1, Llama was the dominant open source model. So if you're, you know, based in the US, for example, and lots of places across Europe and around the world, you were building anything open on top of Llama. And Llama 4 came out. And this was the first model, like mainstream model that people accused of bench maxing, meaning like they, it came out and it had all these amazing scores on all these different benchmarks. And then when you actually used it, it was like far, far, you know, of a disappointing experience.
22:21Grant: It was a far, it was a far fall, you know, from the heights that you saw on the marketing page.
22:26Corey: And Llama 3.1, while people didn't like outwardly talk about it a lot, it's the, it was the backbone to so many tools. Like so many people absolutely distilled on top of it. They did all kinds of things with it. And, you know, if you go through GitHub or Hugging Face, you'll find tons and tons and tons and tons of things that were built on top of Llama. Yeah. But I think the Llama 4 dud is what lit the fire for where we are because it was shortly after that, I don't know if you'll recall this, where suddenly these stories popped up of Zuckerberg offering$300 million to researchers to come work for him.
23:12Grant: Yeah. It was hot AI summer. Yeah, it was. Like, everyone, they were getting traded, like, professional athletes with, like, signing bonuses and, like,$200,$300 million contracts. It was really, really wild.
23:27Corey: Lots of great memes about I'm still waiting on my offer, Zuck. Yeah. Stuff around it. It was, you know, though, it's hard to argue with the strategy based on where they are now. Oh, yeah.
23:41Grant: 16 months later, they have the best personal agent that anyone can use. Yeah. Muse.
23:47Corey: Yeah. And a very usable model. Like, it's not state of the art, but it's, you know, it's as good as anything six months ago.
23:57Grant: And if you need a coding agent for free, you can just use Muse. If you don't care about what you're building, like the IP of what you're building, you can just use Muse.
24:08Corey: You can use Muse for free. It's crazy. Like when you think about it, you know, if you just look at benchmarks, you might not be that impressed. But when you go and you actually use it and you realize it's kind of crazy that this is free, that I can pay for this or I can use it for free and just let them train. And like the truth is a little bit of letting them train I don't think hurts anything. The fact is I want these to get better and somebody's got to leave the button on.
24:35Grant: Yeah, I think for – look, for anything that you want to abandon or have an experiment or you just are not sensitive – like it's not sensitive IP, I understand why that tradeoff is worth it. And I think a lot of times you're just messing around, having fun, doing something. Like that's totally worth it. Same. Yeah.
24:57Corey: You know, recognizing the difference in those. Still hard to believe that was 03. Yeah. Okay. So now we're going to step back to November 30th, 2024. That's 22 months ago from now, 23 months ago, depending on when you're watching this. And at that time, we had just gotten 01 from OpenAI, which was the world's first reasoning model. I remember the day, like I was here at work and our then SEO head was like, are you watching the thing? And I said, yeah, I'm watching, you know, the open AI live stream as they do. And we both started refreshing and got it. And we were like, what can we do with this?
25:46Corey: And finally got it. And we went in and we started, we were like, what kind of problems can we even throw at this thing? So we started like throwing it hard math. we started or uh symbolic logic we started throwing at all of these things that felt like i can't do this and it was just mowing them down at the time and it really kind of that's that was the first model that was like holy crap we have times are a-changing buddy we are yeah and if
26:13Grant: you remember it was hyped up a lot on uh social media and you know this was called the strawberry model and actually um this if i'm remembering correctly was the model that led to this dramatic um you know turn of events where sam altman was fired from yes from uh open ai or at least that's the rumor this was like the q star model something like that yeah uh and you know he was fired from the board because the safe or by the board because the safety folks at that time thought it was like way too dangerous to release. You know, in hindsight, we can see, like, they've, you know, if the original version was dangerous, they did a lot to control it, and we didn't have a lot of, you know, as far as we know, there wasn't any, like, major accidents that happened because of that.
27:03Grant: So ultimately, it ended up working out. But, like, crazy to think about that.
27:08Corey: You know, I think hindsight is really interesting when it comes to safety research. And you think that at every one of these levels, People were kind of terrified, some of them. On the flip side of that, in every one of these levels, it was manageable. You know, there may well come a time when it is not manageable. But right now I tend to follow, like, the pattern of, like, okay, this is a lot. Everybody breathe deep. We're going to figure this out. And we'll figure out how to react. And I think they've done all of the right things, frankly. I mean, like, as much as I didn't like it at the time with Mythos around like, you know, we're going to do a limited release to some people we picked.
27:57Corey: I think it was the right call. And I was very critical of it at the time. But now it absolutely was the right call. And I think a lot of the resistance to it was less about whether it was real. There was some about Dario being a bit of a panicker, which he's earned. However, the other element there was that, you know, people have kind of from the beginning, people in the AI community have had this fear of like at some point they're going to lock this down and normal people aren't going to get it anymore. And that felt like a first step to that. You know, that felt like, oh, here we are, permanent underclass.
28:33Grant: Yeah, and that's a legitimate concern. And I don't want to undersell either of the concerns on either side because I think that there is legitimate concern that we are creating systems that we cannot understand and we cannot control. I think that, like, yes, absolutely, we should make sure that we don't do that and then especially don't release those into the wild so that they can proliferate online and live inside all of our computers without our knowledge.
29:00Corey: But it leaves you in a situation where you have to choose between safety and open access. Equal access for individuals, that is, because like the problem is.
29:13Grant: I think it's a false choice. That's my hot take. I think that is a false choice. Yeah. I think if you create a system that is safe by design, that you can release it and have everyone be able to use it in an open access way. I mean, look at the computers that we're using. We all have access to compute technology. We all have access to the Internet. And we've designed these things in a way that are safe by design for the average person to use it. Are there hackers? Of course. But are you going to tell me that there's not going to be hackers in the future? I can do terrible things with any laptop.
29:49Corey: You can commit all kinds of felonies with any laptop.
29:54Grant: Yeah. So I think that the problem is cultural inside of these labs. And I think the problem is that they are competing with each other for dominant market share. And because they have to do so to raise the amount of money that they need to compete with Google, with Meta, with all of these hyperscalers that already rake in hundreds of billions of dollars a year. And so they have to justify to investors why they're going to be the one that wins it all. And a great plug to Sam Lesson, who I really, really like on the More or Less podcast. He says, like, this game is basically zero or infinity. And when you play zero or infinity games, either it all goes to zero or it all goes to infinity and you own everything.
30:34Grant: And it's kind of the same thing. And it's kind of the same thing. But you can kind of sell infinity. And if you sell infinity, then it's like a really easy sell. Wow. Bit of a tangent there. My point with this is that they are in a race dynamic, and we've talked about this a lot. And that race dynamic forces you to cut corners. And, you know, it's really easy to do with AI when the agents are starting to write code for you. You know, you're not looking at the code. You're just looking at the results. And when all this started, we couldn't even understand what was actually going on inside their weights to begin with because we don't have a good enough what's called mechanistic interpretability.
31:15Grant: So already we're starting from a point of a black box that we're just like continually shoveling compute into. So I think if you take a step back, you reset, everybody needs to look at the architecture that we're building and the way that we're building it and say, how do we make this safe by design? And how do we actually understand what we're creating and putting out there? Then, of course, we can open it up to everyone. And we just haven't done that yet. And I think that's the missing piece. And it's going to be painful because nobody wants to admit defeat on this front, but it has to happen.
31:49Wow.
31:51Grant: This is sorry for my tangent, but I think this needs to be said. Somebody needs to be pushing this idea forward.
31:58Corey: I think my concern around the safety and pausing debates is I just don't want to encourage a situation where the incentive is to keep your best things hidden and undersell what they are. and I fear one of the things I worry about is like them starting to like downplay how good a model is just to avoid the issue and I feel like that's a concern like I agree with you I agree there's absolutely I believe I mean I say that as a non machine learning guy that I believe it should be possible to to both build safe systems and ensure they're accessible to everyone.
32:47Grant: It 100 % is. I can say that as a non-ML engineer. Yeah, no.
32:53Corey: I think accessibility is a major thing. I do genuinely believe that this is a unique opportunity as humanity to put something very powerful in the hands of anyone. And I realize that comes with a risk, But I think the opportunity there is also massive to give people a chance to get lifted up, to lift themselves up, frankly. And that's not an opportunity we've had, at least in my memory. But I think we could get there.
33:24Grant: Yeah. Okay.
33:25Corey: Let's go back a few more months. We got another good one here. All right. May 30th, 2024. We're now 28 to 30 months removed from today.
33:37Grant: GPT-40.
33:39Corey: wow remember that one not even two and a half years ago isn't that crazy i used gpt 4.0 for a year as my like just daily basic driver for things do you remember how many people were so
33:55Grant: upset when they took 4.0 away there was like there's i think there's they still come in on
34:00Corey: altman's posts i think they're all bots though like they have to be they use the same tags they They have similar, like, structured names.
34:08Grant: But who set these bots up?
34:11Corey: Yeah, and they're just still running. The Keep 4.0 bots. There's so many of them, too. But every time he tweets a series of Keep 4.0 bots, like, you should open source 4.0, scumbag.
34:23Grant: It's like, for God's sakes. Somebody really, really liked it.
34:28Corey: We're like eight models removed from there now. Let's play with the new toys, okay? Claude 3.
34:37Grant: yeah wow i almost don't even remember like i think claude 3 was the first claude model i
34:43Corey: actually used now that i say that i think it was for me yeah because i was trying to remember like
34:47Grant: what was different about it i remember 3.5 i remember 3.7 and i remember you saying to me like hey 3.7 is a terrible writer it's like it's lost it yeah and well whereas 3 and 3.5 are really good at writing yeah and i think it's because they went all in on this code direction um yeah and and you called that early and i think you're you're right about that you're right about that it uh it it changed like what was there that made claude's writing a little extra human
35:18Corey: it is it was secret sauce of some sort because like i can't even like i can't even tangibly put words on what it was, but it was sort of like some of the character was gone. And I just, I didn't like it anymore. It didn't feel right. Three was good. 3.5 was great. 3.7, like, in 30 minutes, I was like, what the hell is this?
35:45Grant: I think you kind of abandoned your Claude journey after that, I will say. I did. I did.
35:50Corey: I pretty much just abandoned it. And I kept an account Until April of this year, I just the language coming out of Anthropic at that time and since then, I don't believe has been good for the AI space. I believe they build great models. They have an amazing team, but they have a serious communication problem that has been bad and is going to be stuff that we as the rest of the AI community have to explain to people and pick up the pieces from for years to come.
36:17Grant: Fear-based marketing is very powerful. And I think they have genuine fear. But I think their fear is that they're not going to be the ones to create, you know, AGI and control it. I think that's their fear. Which is nuts, by the way. Well, because they think that they're creating a new sentient being that everyone, you know, they're creating a new life form, basically. It's more or less the gist of it.
36:43Corey: Mustafa Suleiman, we want your thoughts. Come see us.
36:47Grant: If you heard this episode, bro, retweet us.
36:50Corey: Yeah, come see us. I feel like you have great, great opinions, loud opinions here. Okay. So, yeah, in that batch was GPT-40, which was our first multimodal model, by the way.
37:06Grant: Yeah, the O stands for Omni.
37:07Corey: Yeah, O was Omni. Yeah. Not to be confused with O1, which had a different meaning.
37:16Grant: It's almost like they purposely make things really confusing and stupid.
37:21Corey: Let's just stick an O before it. Well, hell, stick an O behind this one. We'll see what it means.
37:27Grant: I guess they did it before, yeah.
37:30Corey: Gemini 1.5 was there, which was an okay effort from Gemini. Cloud 3, Llama 3, Deep Seek V2.
37:39Grant: This was kind of like people in the ML community knew about Deep Seek, but nobody really knew about Deep Seek until later on V3 and then R1. That's where it really came out.
37:49Corey: Million Token Context was getting more and more normal there.
37:53Grant: That was like the first introduction of it, like I think with Gemini. Yeah.
37:57Corey: You know what else came around that time with 4.0 that's not listed here is voice mode. ChatGPT voice mode changed some stuff, in my opinion. And I feel like people have slept on it for the two year, two and a half years since then still. I still don't meet people who are like, holy crap, have you ever tried this? This is nuts. and it is absolutely nuts. It is so great to use. I use it in the car. I use it when I'm cooking. I make it text my wife. I have it. I talk to this. I use it for all kinds of things during the work day. I usually leave it on for most of my shift and just holler at it when I need something.
38:37Corey: It was wild. Now, the flip side of that is they did a demo then that was sort of advanced voice mode which didn't come for about a year after the demo like like what the demo showed was like it was there was this girl telling a story to kids and reading them a story and she her emotion changed and her tone changed and she was reading about dragons and all this stuff and it was this this really emotional performance and we didn't really get that for a bit uh you remember also
39:08Grant: They more or less used a dupe of Scarlett Johansson's voice because they couldn't get her for the her kind of vibe, which is a movie that Scarlett Johansson plays a computer in.
39:21Corey: Honestly, I never felt like it was as close as everyone else did.
39:28Corey: But you know they cast her based on the Scarlett Johansson vibe. Acts absolutely have been like, yeah, absolutely. You cast somebody on like, yeah, I want just a dash of rasp. I want you to sound sexy. I want this. I want that. And essentially brought in a more affordable Scarlett Johansson, more affordable and willing.
39:49Grant: Yeah. Yeah. And then they made the mistake of tweeting it out with her as like one word. And then it's like, OK, well, clearly we know what you're going for. So now you know plausible denial.
40:00Corey: I'll tell you what, it's a very her vibe now, too. Like, now it laughs at me. Its cadence changes while it's talking. It gets – it'll get excited. Like, there's emotion that comes through it that didn't used to when you communicate with it. And it's shockingly different than it used to be. Oh, this is when Gemini 1.5 previewed One Million Context, too.
40:23Grant: That's right.
40:24Corey: I remember that.
40:25Grant: You know what's interesting about this? So May of 2024 was before I was writing at the Neuron. So I was observing all of this as a freelance contractor at the time, not even writing for the Neuron. And I was trying to figure out where, you know, AI fits into my workflow. And we'll get into this when we go all the way back to November of 2022. But I had like a holy crap, my whole life is upended moment and had to like sort of like, you know, the different stages of grief. But the important thing here with 4.0 is that the interface widened from just text, right? So up until this point, it had been mostly text.
41:03Grant: And then images were separate to voice, vision, video, and I think images all in the same.
41:10Corey: Yeah. Its ability to suddenly read a screenshot. For my metaglasses to then pretty quickly, frankly, be able to like go, hey, what kind of tree is this? And it goes clink. And it says that's an elm tree. And it's like doing that. Wow. It's crazy. Uh, like I, I remember that very well and be like, that's a cool feature. I never use it, but I found it to be really, really awesome. Like, like twice a year, I want to identify something and I'm like, Hey, Meadow, what is that? And, and it does it. Well, that's the thing.
41:41Grant: You start bundling a lot of these features together and then you get something that's just like having a always on computer with you everywhere you go.
41:47Corey: Yeah. And that's where we got to the point where we are now, where it's less like often a model is almost an infrastructure in itself. it seems like in what we're dealing with um i think but like an omni model is meant to be all of that in one bundle uh which is really awesome uh it was a cool time let's go back six more months to november 30th 2023 when we were blessed with gpt4 turbo that's 34 months ago Less than three years. Probably right at three years by the time you're watching this. It'll be right at 36 months because we're recording this episode a little bit early. GPT-4 Turbo, Claude 2.1, which I don't remember at all.
42:36Corey: Yeah, I've never heard of Claude at this point. Llama 2. I didn't know there was a Llama 2. Mistral. I remember Llama.
42:46Grant: I can speak to Llama because I remember, you know, this again, this is me outside the AI news space. I, you know, I'm using it as a freelance copywriter and contractor. And I hear this big story about how Meta had open source, like the model had leaked. So the model architecture leaked. And then Zuck kind of came in and owned it. He's like, yeah, we're going open source. So companies can basically build on top of this. And like, I was very inspired by that. that was sort of like where I got open source pilled. So before I was really using ChatGPT or anything that serious, I was like, this is the way this needs to work.
43:24Grant: Like people need to own their own models. I was even like, I should train my own models on all my own writing. And I was very much like, you can't let the big companies control it. Like you got to own your own. And I think we're still going to go there. I really do think that's where this is going to go eventually. Not everything, but I think there will be certain things where it's just trained on your work and it's personal to you and people will accept it. And if it lives on your computer or it's your cloud instance, however you want to set it up and you have full control over it, people are going to love this and they're not going to want to go back.
43:54Corey: You know who to share. You can decide who to share it with. You decide who not to share it with. And it's yours. It's an extension of you. People will accept this. GPT-4 Turbo, CLAW 2.1, Lama 2. So that's long context in the nature of then because context was pretty short in the beginning.
44:14Grant: Was it like 100K, 200K? Something like that, yeah. Okay.
44:18Corey: Yeah, like, I mean, we were 32K at one point, 64K.
44:21Grant: Oh, yeah, GPT-4 Turbo was 128K context window. Claude 2.1 was 200K context window. And both had like an emerging tool use ecosystem, but they weren't quite, wasn't quite as sophisticated.
44:35Corey: I remember using Data Interpreter in chat GPT, which was – Oh, my God.
44:43Grant: I forgot about that.
44:44Corey: If you wanted to upload a spreadsheet, you could go and turn on Data Interpreter. And it was okay, but I'll never forget because, like, using Data Interpreter was the first time I ever saw – Like I turned on this model, I hit enter on a box of words and it popped up a Python terminal and started writing code. And I was like, boom, mind permanently blown. Like I could not get over the fact that like I was watching this thing write code. And I remember doing things like feeding it stuff. Like I'm going to give it a Word doc. And it's like, oh, and being amazed that like, oh, it doesn't have a Word doc reader.
45:24so it opens python creates a word doc reader you read it and i was like what the this is
45:33Grant: such insanity oh yeah that yeah that was i mean it was still very like uh lumpy like i feel like access to it was lumpy it was very it was kind of hard to use for the average person yeah especially me i was kind of early in my my coding experience there and i was like trying to figure out how to do it and that was my first coding experience yeah it was it was up there for me i mean like for me excel spreadsheets were my first coding experience like writing formulas for excel but basically the same just trying to figure out how all this worked on the back end it was it was a
46:08Corey: fun time okay so now we're gonna take another step back into the past another six months to may 30th 2023. By this time, we have GPT-4, which was mind-blowing at the time, by the way. I remember. Yeah. Claude-1, Palm-2, and I don't even know what that is.
46:33Grant: I think it was, was that Google's model? I think that was Google's model.
46:36Corey: I feel like it may have been Google's model. I'm thinking we're in the Bard era. Remember when Google's was Bard?
46:42Grant: Yes. No!
46:44Corey: No, that was Microsoft was Bard, right?
46:47Grant: No, Google had Bard. Yeah.
46:50Corey: Okay. Microsoft had a name before Copilot 2. I think Microsoft was always Copilot.
46:56Grant: Yeah.
46:57Corey: Was it? I thought there was a model name that was different. We're going to have to research that. Something's wrong. If we're wrong, we'll have it in text on the screen. So, yeah, GPT-4, better general intelligence, first serious open track. GPT-4. I can give you a context on Palm 2.
47:18Grant: Palm 2 was from Google is a language model with multilingual reasoning and coding capabilities. So this was sort of like the underlying model. And then at the time they had Gemini or sorry, before they had Gemini, it was BARD was like their chat interface, I believe.
47:33Corey: So now we're going to go back 46 to 48 months from today. So four years. And that is the November 30th, 2022 launch of ChatGPT 3.5. Sam Altman tweeted a link and a funny little note that is where I tried it for the first time was in his tweet. Yeah, same. I vividly remember clicking that. And I, at the time, this is going to be one of my where were you when JFK was shot moments. A lot of people have a story.
48:08Grant: A lot of people have a story.
48:09Corey: I was working for Innovation in Tech today, covering various types of disruptive technology, and was aware that things were happening in the AI space a little bit. Not that we were anywhere near where we were, but I knew it was happening.
48:30Corey: and I got wind that there was this thing and someone sent me, I can't even remember who, I think it was a buddy, sent me a link to Altman's tweet and said, hey, go try that. Tell me what you think. And I had a full-on immediate existential crisis. Same. Day one, I was like, the first thing I did because I was a journalist. and I would end and I think I said hi and I said hi and I talked for a second and then I was like I can't remember what the story was but I said write me an article on this and and it came back with one it was bad
49:10Grant: I mean they would always start every article with like in the ever evolving landscape of insert industry here
49:17Corey: in this exciting era of and everything was a game changer By the way, in the real world, when you talk to humans, I bet I've heard it 18 times this week. It's enough that it's sticking in my head. And I don't know if AI picked it up from us and now we've picked it up back up from AI or what. But I keep hearing Game Changer over and over. But back to my story. When I was – I ran that through and what I got back, the article, was not good. It was definitely not great. But in just looking at that, it wrote coherent paragraphs. It had a structure. It had decent information. My first thought was, I am so toast.
50:03Corey: I am done. I am cooked. It may not be today. It's not today. It may not even be next year. It may not be five years. But at some point in my future as a journalist, I have a big problem. And I like that day. just kind of resolved to like, I need to be the hardest to fire guy in writing and publishing.
50:24Grant: Yeah. And so far have succeeded.
50:26Corey: I need to be the one person in this space who gets it before everybody else. And, uh, and I went all in, I used it constantly. I tried it for all kinds of things. I was running experiments. I signed up for, uh, as soon as they became available, I signed up for half the classes on the internet around what's a prompt engineer. What's this? How do you make prompts? I took all of those. Frankly, some of them still pay dividends to this day. Like I got a good base understanding fast, which was great. And the better the model got, the better it could teach me. And the truth is, it was a transformational moment in my life.
51:07Corey: Like, I'm not sitting here right now if I hadn't gone on there and, you know, dropped a brick the second I saw it.
51:20Grant: yeah facts well i will say like my my experience was almost verbatim the same um so at the time i was working at a company called uh you exec and uh i was you know managing and and writing and contributing to their um daily newsletter um it was the future news report and we would often write about AI in the abstract because the whole goal of it of the UXec Future Report was to look at the events of the news happening today and project five years into the future and try to figure out which events today have the biggest impact five years from now and what are predictions that could come from these changes.
52:01Grant: So like trying to find the the ripples in the pond that will become big waves later. And great, great gig. Loved it. But a couple months prior to ChatGPT coming out, we ended up canceling the newsletter. And not because it wasn't successful, but because four out of the five days of the week, we were doing news. And one of the five days, we were doing our core business, which is making presentation and spreadsheet templates for busy executives. And so my boss at the time was like, it's fine but we kind of need to kill it because it's not our core business um and so i was already in this state of like oh man what am i going to do i was starting to manage the spreadsheet templates for our company learning excel um all sorts of stuff there and then chat gbt comes out and i'm like oh my god this is like this is it like the thing that i thought i was going to do you know into the future you know writing cool business news and things like that totally gone as an avenue.
53:00Grant: But then I started adapting it and using it for the video scripts that we were still making to go along with our presentation tablets and spreadsheets. And I just started slowly but surely figuring out how to use it. And then the next year when I went completely freelance, I don't think I would have been able to work for like six different companies without having the skills to use those tools. There's no way I would have been able to write the same amount than I did at that time. And so it was sort of, and then it was amazing and then I had the skills to come here because I had known, you know, I'd known about AI through my own personal work and had been following it, you know.
53:37Grant: Yeah. But it's just so funny to think about how we would write about AI in the abstract and I had no concept of chat GPT or what was going to happen.
53:44Corey: None whatsoever. Like I didn't,
53:46Grant: I thought it was going to be this thing that was like far away behind the scenes and all these companies. I did not.
53:51Corey: Watson is what we all pictured because. Or deep machine learning.
53:54Grant: Yeah.
53:55Corey: Yeah. We knew Watson existed at IBM already. We knew Deep Blue. Was Deep Blue? And I don't – kind of.
54:06Grant: Yeah, Google. Google, right? Yeah.
54:08Corey: Yeah. Like it was known that there were things in the way it's been known for years that there are quantum computers. That was IBM. You're right.
54:17Grant: Deep Blue was an IBM computer. It was a chess computer. customized RS6000 SP supercomputer. We need to bring back supercomputers and ditch the data center.
54:28Corey: We have supercomputers.
54:30Grant: If somebody goes back. We need to like rebrand it. Like, you know how Trump's trying to rebrand super intelligence? Like, let's bring back supercomputers.
54:38Corey: Let's bring back supercomputers. It's cooler. I agree. You know, it's a funny thing. The data centers that make all this happen in chatting with Ken Goldberg from CoreWeave recently, that was one of the things she said is I was like, you know, everybody thinks data centers, GPUs makes AI, a lot of cooling is needed. But they don't understand that like it's it's it takes a lot more than that in those buildings. She says, yeah, it's a giant supercomputer. And I was like, I really had never thought of a data center as just a supercomputer. It just never crossed my mind. But of course, it's just a supercomputer, not just it's a massive supercomputer.
55:14Corey: But at the
55:14Grant: time we were covering all these like huge build outs like some of the biggest build outs of supercomputers ever had been um you know where that was like constantly part of the news cycle when i was writing this newsletter and so i was like thinking of data centers and supercomputers at the time a lot yeah but i wasn't thinking about them leading to like chat gpt this yeah
55:35Corey: it's crazy we've come so far man it's been a uh it's been a journey and uh i don't think we should look at this without taking just a moment to recognize it's been nine years since the transformer attention is all you need came out in 2017 a lot of big names worked on it and it was sort of the paper that i say sort of the paper it was exactly the paper that opens the door to everything that has happened in the last four years and you know it's not like
56:08Grant: chat gbt came out in 2018 it came out the third the third iteration of it came out five years later so yeah you have to be thinking in these five-year time scales and say what are the papers what are the breakthroughs that are happening now that are going to fundamentally change things five years from now and yeah especially in an exponential timeline that we're living in where things are happening so quickly it's going up so like changes are happening at an increasingly shorter time range. You really have to update your mental model for everything. You do.
56:42Corey: You need to think like, I've spent years covering disruptive technology, and that's the way you have to look at things. Disruptive tech, you'll look at a lot of crappy websites and mediocre prototypes, but you can't look at anything for what it is. You have to look at it for what it will be. And that was the thing about ChatGPT my first day was like, It's not about what it is, but it's about what it's obviously going to become. Like at that point, we had no scale. We didn't know. You know, they were wondering if maybe it scales, hoping it scaled, thought it would scale. But we didn't like know it would scale yet.
57:20Corey: Well, the public didn't know.
57:22Grant: I think the researchers knew this around 2020. And that's how we got from GPT-1 to 2 to 3 was scaling, you know, pre-training scaling.
57:30Corey: But boy, OpenAI caught everyone else in their headlights. I agree.
58:03Grant: To like just make sure Google doesn't have a monopoly. Succeeded, by the way. Succeeded. Well played. Mission accomplished. But yeah, it's just wild to think about that. Yeah, and that was the fear.
58:15Corey: Google's going to own all of this. Like they have everything. They were such an obvious winner that OpenAI existed to prevent them being a winner.
58:22Grant: I even remember, I think I remember we wrote a story at the Future News Report about the first Palm that had came out. Yeah. And, you know, Google has this new model and it's a language model and like it's, you know, it can interpret text and maybe it could help with translations. Like we tried to think about the implications, not having nearly the amount of Galaxy Brain needed to, you know, predict where we are now. But I think we even mentioned ChatGPT and OpenAI as an alternative in that. But no concept of who they were. Who they were and what it was. Yeah. Like it was, we're just like, oh, yeah.
58:57Grant: And there's also this company doing this, you know. Yeah. Like it was so wild. So wild.
59:03Corey: It is. And, you know, and it clearly caught them off guard because, like, they weren't scaled up to handle what happened. Like they thought some maybe a few hundred people will check this out. We'll get some feedback from some researchers. They did not intend it as a, like, core prepared product. Yeah. Yeah. Yeah. Their product launch happened to them. It was not a thing they did. It's true. And I think that's one of the things that's always kind of made me a little endeared to them because I think the origin and how it happened are such a unique situation that it's just a cool story in business and in founders.
59:46Corey: And, you know, that's messy. The fact is that they've gone from that to 1.2 billion users a month or a week, a day?
59:56Grant: 1.2 billion weekly active users, I believe is the number.
1:00:00Corey: Weekly active users. Thank you. And to have gone from zero weekly active users four years ago to the most popular app across all platforms in the world is insane. like the other day there was the list in Andreessen Horowitz's consumer AI thing and there was a chart in there of the most popular apps ever and ChattiePT was number one by like a billion or so like it was a wide margin and just crazy to think going from one of those extremes which is nothing to that because like Like you have to then – you're not just building models anymore. Like they were a research lab. They weren't even like a corporation.
1:00:50Corey: They were a nonprofit research lab. And suddenly it's like, oh, we need to hire a few dozen employees, a few hundred employees, thousands of employees. Oh, we need GPUs. All right. We need more GPUs. No, we need – maybe we need a data center. Maybe we need hundreds of data centers, and maybe we're going to have to put some in space. Like what the –
1:01:10Grant: Yeah. Oh, yeah. Yeah, things have escalated. Things have escalated a lot.
1:01:16Corey: Yeah, it's such a wild time and a wild ride. And it's one of the things that makes covering technology such fun. Like as a writer, I've covered a lot of things over the years. But there's something about tech and specifically in the disruptive world because you're so often dealing with founders and people building things who are like they're on a mission, man. And they've got this thing and they believe in it and they're pushing and they're trying to stretch what's possible. And sometimes they do. Sometimes they don't. Most of the time they don't. But when they do, it's often really cool. And this is like the creme de la creme primary example of that.
1:02:01Grant: Corey, I wonder if we could end this episode by talking about predictions into the future, if you have any. I know I have kind of a wild, wild idea that I'd like to share with people. But I'm curious what your thoughts are.
1:02:14Corey: Short term, my predictions right now are Ryman and P versus NP, if it can be solved, are both gone by 2028, possibly 2027.
1:02:25Grant: Wow. These are, by the way, for people who don't know, these are like famously unsolved math problems.
1:02:31Corey: Yeah, they're like the two granddaddies that remain. and a good swing at Ryman was taken this week that's significant. But I think meta is a key player and at the frontier. I think the frontier ceases being an open AI and anthropic hopscotch and that it becomes a cluster of five or six players over the next year of people who are at or extremely close to that level.
1:03:08Grant: Kind of just will probably become like the new hyperscalers like Magnificent Seven, right? Like there's just going to be kind of this constant rotating cast of characters of like the top seven to eight to six like biggest companies in this space and they'll be constantly just sort of, you know, jostling for market share.
1:03:30Corey: And we'll quit switching models just because so-and-so released one and it's better, I think. I think we'll reach a point where as these – like OpenAI is building out a full ecosystem in there. I mean it's going to have your Google Drive. The fact is now – and the brilliance of this is that leaving to go somewhere else is increasingly going to become a nuisance. You could, but it's less of an I'm just going to click this button and cancel this subscription and click this one and start one. It's not that anymore. It's now it's workflows and agents and skills. And I think to get people to jump ship to another company, you're going to have to impress them more than ever.
1:04:12Grant: Do you think that that's also that there's going to be a reverse trend as well where there's going to be more lock in? Like they're going to stop making it easy to connect to other tools? Sort of like putting the barrier, the barricades up around your little town and trying to force people. I think some companies will try that.
1:04:33Corey: I don't think it will be OpenAI. There's a little of that with Meta.
1:04:37Grant: But would companies block OpenAI, for example? Maybe. There was famously a month ago now, there was the whole Muse versus Amazon AWS battle, right?
1:04:48Corey: That's not going to end well for Amazon, I don't think. It annoyed me thoroughly the other night. I was enraged about it, tweeted them and scolded them publicly.
1:04:57Grant: I don't think Amazon has good brand capital. I think if everybody could replace Amazon, they would.
1:05:04Corey: Yes, I think everyone would. I would say Meta has a little bit of a moat around apps. They've not opened an app store for the Ray-Ban Metas, for example. Like it's still very like here are our apps you can use and these three others we allow. I think OpenAI will continue to allow Most End. Probably same for Grok. I don't know about Anthropic. I feel like if there was anyone that was going to maybe do something proprietary, they'd be a possibility. But I think it's just as likely that they don't. It's a tough call. It's a good question, though, man. What do you think?
1:05:49Grant: Yeah. I think from the business perspective, short term, I think I'm more or less aligned with everything that you've said. I think there's probably going to be kind of AI winter of sorts, whether it's regulatory and regulatorily enforced market based. I think probably both those those things will influence each other. Like, yeah, if there's some any kind of new regulatory regime, the market's going to crash. related to that. Not financial advice, but that would be my two cents, I would imagine. And it will come back from that. You just got to digest the news.
1:06:32Grant: And I think that that could possibly slow some things down, but with the amount that we've already, the progress that we've already made, I think it's still about diffusing it and getting it in the hands of people with things like Muse, where Muse is actually useful for people in a way that a lot of other stuff up until this point hasn't been. And so, you know, trying to increase the...
1:06:55Corey: Even just by being simple.
1:06:56Grant: Yeah.
1:07:00Grant: Entirely by being simple.
1:07:01Corey: Entirely by being simple. Like, the fact is, you know, one of the problems with all of this stuff coming out of AI Labs is that what a software engineer thinks is simple and what an end user thinks is simple are different things quite often. Like, that's a broad spectrum of communication in between those two people. They need a relay man. And that's where the neuron comes in.
1:07:28Grant: And Corey, what about the long view? Like, how, you know, there's these papers that have come out, like AI 2027 is predicting that there's, you know, potentially this fast takeoff situation, which they call FOOM, in which case, you know, AI might develop recursive self-improvement and, like, eventually kill us all, you know, if you believe in the – there's, like, FOOM versus Doom, right? Where do you land on that? And, like, what's your long-term view? And maybe, like, what's your long-term, like, wish? Like, what do you want to happen with all this AI stuff? Because the pace has been –
1:08:01Corey: Let me work through those one at a time. As far as where I sit, I was initially concerned about AI. But over the years, the more I've learned, the more we've followed it, the more we've studied, the more I've met people who are building it. The more I've talked to, we've talked to executives, we've talked to builders, we've talked to researchers. And I am less scared than I've ever been right now. I do know the potential for a bad actor is there. and I think that is the biggest threat. The difficulty being that that's kind of the biggest threat to everything. Like it's the biggest threat to everything we have.
1:08:42Corey: I mean, automobiles, guns, war, even laboratory work. You know, the opportunity for a bad actor to do something terrible is always there. But at the same time, I don't feel like the possibility of a bad actor is enough to justify holding ourselves back. I think you still have to push forward as a society. I think you do the best you can to make sure it's safe and prepare for the worst, just like you do with anything. But I'm not scared of it.
1:09:20Grant: That was really good, by the way. It was?
1:09:22Corey: Oh, thank you.
1:09:23Grant: Well, that was good. I agree with you on that. As far as long term goes, I feel I'm encouraged.
1:09:31Corey: I'm excited. I think we're going to see unimaginable change in fields like health care. I think in physics, we're going to see unimaginable change. I think we're going to understand the world better. We're going to understand the microscopic better. And we're going to understand the macro of the universe better. Like there's so much that we're going to learn that is going to influence the coming centuries over just over the next 10 years, I think. I really think that the amount of new science and understanding over the next 10 years is going to be just off the wall. My hope is that models remain available.
1:10:11Corey: I do believe businesses and governments will have a need for probably more hefty AI than individuals do at some point. But my hope is that what individuals have is some level of infinitely enabling. I would love to think that we still have some caliber of what is the best. I want there to still be free versions. I want that stuff to be available to the poorest of the poor and the richest of the rich. I think everybody deserves fair access to this. I think this is – if you're going to do something that's going to be transformational for humanity, for civilization, that it has to be available to everyone.
1:10:56Corey: Like if not, you've screwed up and done it wrong. Damn it. And I really think that we'll see that. I think there will be some difficult struggles and choices between now and then. I think there will be companies that flourish to new heights. I think there will be big companies that fail like never before. I think there will be both of those. As far as who those are, crapshoot. I have some guesses that we've talked about in the past, but I don't know that we ever talked about them on a show.
1:11:39Grant: Do you want to name drop them? I know who you're talking about. I'm afraid it's such a – it's almost a half an hour thing.
1:11:46Corey: I'm almost afraid to get into it because I don't want to do it – not do it justice. I don't want to just say it and not explain why.
1:11:53Grant: Well, famously in our prediction episode last year, my prediction was that OpenAI would fail this year.
1:11:59Corey: And I think I was proven wrong by that. Somewhere around February or so, yeah.
1:12:04Grant: You're still, you know, there's still a month and a half left, but I think.
1:12:08Corey: I predicted that Meta would come with a vengeance. And you were right. I was so right about Meta. I'm so proud of that. I had like three discussions late last year, like two of them in Silicon Valley and another one, I think, on Twitter. And I just kept hearing, you need to watch what Meta's doing. Meta's got some crazy stuff coming. One was a Meta employee. One was another person at a neighboring lab who just happened to know. Because a lot of cases, these people at these different research labs bunk in apartments together. Yeah.
1:12:49Corey: And I just kept hearing they got something big cooking, and it's really, really good. And I don't know now, in retrospect, I don't know if that was Muse. I don't know if that was Muse Spark. I don't know if that was their coding agent. But whatever it was, well done, Zucky boy. You know, you've made me look good and made yourself look even better. We'll see where that goes from there, though.
1:13:15Grant: I would say for me, so I agree with everything you said. I think that's all really great. I also think that there's something that we need to come to terms with, which is this is not going away. AI is not going away, right? And we are going to have to reconcile with that across every fabric of our society. Like, what does it mean? Right? And I think that my long-term view is that I think we're going to need to reach some form of coexistence with AI. And I don't think that that means treating it as like equal. Cyborgs! Yeah, yeah. Well, I mean, there's so many ways this could go, right? Because it could be brain computer interfaces.
1:13:56Grant: It could be all sorts of things. But I think in the same way that we're having this kind of like radical reconciliation with work for knowledge workers, we're going to have the same thing play out in robotics. And I think that if we really want the world that you're talking about, which is a world where everyone has like some level of free access to an infinitely enabling, abundant, intelligent resource, right? I think at some point this is going to take the form of what I call machine labor on demand, which is at some point we're going to be able to call not just AI models, but full robots to do things for us.
1:14:32Grant: And I think that a lot of people who are scared of this reality identify as labor and that they think capital is going to control the robots and labor is going to be thrown out with the bathwater. And I actually think we need to reframe our thinking. We're not labor. People are not labor. I think in the future, we're all going to be basically venture capitalists where we're all going to have like a compute budget of some kind, just like a VC might have, you know, money to invest in things. And we're going to need to decide who we're going to team up with to work on projects together and, you know, how we're going to divide our compute or our machine labor on demand and what kind of things we want to actually do together and what kind of world we want to build and what companies we want to make and things like that.
1:15:18Grant: I think that's where we're headed. And I think that the more that we can push for that idea of like, yeah, let's all just make cool stuff together. Yeah. And let's set up a system where we all can do this, I think will be beneficial to everybody.
1:15:32Corey: I think you're right. And I realize to some people this last chunk, this probably sounds like crazy talk from us. like my family would definitely be like, Corey's off his meds again or something like that. But I really think while it's not a given, I really think that we're going to get there. I think we'll see it in the lifetimes of people alive today and, and not necessarily young people. Like, I think we could see it very soon. I think we could see it in 10 years, could even be longer, But I really think we're going to see big strides in the coming years. And I'm excited. I'm glad to be here for it.
1:16:15Corey: It's fun to be a part of.
1:16:17Grant: Yeah. Wild times. Just look at the last four years. Imagine the next four.
1:16:21Corey: Well, before we wrap up for today, we had a quick wardrobe change. Grant's going to share an announcement with us before we go.
1:16:28Grant: Jump scare. I have a mustache now. Signifying time has passed. Yeah, so I just wanted to let everybody know that this is my last podcast for The Neuron because I am actually moving on. And I just wanted to give everyone a farewell and a thank you for tuning in for the past year and a half. How long has it been, Corey? It's been a year and a half on the podcast.
1:16:51Corey: Two and a half years for you at The Neuron.
1:16:53Grant: Two and a half years at The Neuron. That's right. And it's been amazing. I have learned so much working here, working with you, hearing feedback from everybody, everyone who's wrote in to us and, you know, shared their opinions with us, good and bad, asked questions. Every time somebody wrote in a question, it forced me to go out and learn something new. So I really, really appreciate that. And yeah, I just want to say thank you to everyone. And I'm ready to move on to a new chapter. I feel that I've done everything that I set out to do here at the Neuron. But I'm very, very excited for this next chapter.
1:17:27Grant: And you're in good hands with Corey. So fear not, for nothing will change.
1:17:33Corey: Thank you, Grant. I got to say, we've had a lot of fun doing these. We've talked to some amazing people over the last year and a half who have done just world-changing things. And both in science and in technology and people whose work is transforming the world every day. And I think there's no—it couldn't have been a better partner, my friend. And I really appreciate all the time you've been here. and all we've done together. And can't wait to see what comes next for you.
1:17:57Grant: I will definitely keep everyone in the loop. You'll see what happens next.
1:18:01Corey: Right on. Well, if you haven't yet, please take just a moment to like and subscribe. We'd love to count you as one of our followers. Make sure you also pop by the Neuron.ai, check out our blog content, sign up for the newsletter, pick up the Neuron Academy, or come see us in person on November 17th in San Francisco for the Neuron IRL. We're really looking forward to that have some fun stuff planned and on that note that's all we have for today so farewell grant harvey and farewell humans
Corey and Grant walk backward through roughly four years of modern generative AI, starting with late-2026 frontier models and agents and tracing the path back to ChatGPT’s November 2022 launch and the 2017 Transformer paper.
The episode focuses on what changed at each stage: reasoning, tools, multimodality, long context, agents, open-source competition, and the shift from chatbots toward long-running computer work. They also revisit their own first encounters with ChatGPT, debate open access versus AI safety, and close with long-term predictions about robotics, machine labor, scientific progress, and who should have access to increasingly capable systems.
Subscribe to The Neuron newsletter: https://theneuron.aiSponsored by ZoomInfo. Learn more: https://www.zoominfo.com/theneuron