The factory gets darker, Meta migrates to Slack, and throwing 10,000 PhDs at math problems

0:02Welcome to the Friday Deploy brought to you by Linear V. I'm your host, Andrew Ziegler. And I am Dev Interrupted's producer, Adam Noble. Yes, that's right. I finally was able to bully our producer into coming and doing the news segment with me. So today, Adam is coming along for the ride. And Adam, what do we have lined up this week? we have a pretty packed week here so yes normally i get to watch your and ben's shining faces from behind the scenes but i will do my best here to fill in for him so this week we are covering open ai declaring the agi era is here ai labs feuding over navier stokes meta switching to slack for their agents 700 agents breach a server the super dark factory manifesto spec driven development and lastly a peek inside our agentic newsroom so full docket here yes that sounds like an exciting lineup i'm quite excited for the last one when we get to talk about our newsroom adam it'll be really cool so just you're here to kind of showcase some parts of it too but But first, we have to get there.

1:15And there's been some pretty interesting things this week. You know, the first one you said about the era of AGI and open AI declaring it. Let's start there because I feel like these foundation model providers, these big companies, they declare this every other week. So what does this really mean? Yeah, when they come out of the woodwork again and say it once more. And I don't know if you've had a chance to look at the benchmarks. We'll include them in the article for this news roundup as well. And their new Frontier model, Astra, it's really what's triggering this conversation on AGI. Mostly because it absolutely crushed the ARC AGI 3 reasoning benchmark, which we've talked about many times on the show.

1:59It's a fairly popular benchmark right now that up until recently was largely unsolved by Frontier models. It involves a ton of very difficult to reason and understand puzzle games that up until this point, no model had been able to really get close to finishing. Now you have a model that absolutely smashed it. What does this mean? I think it means like several things about one, are benchmarks meaningless? How do we create benchmarks? How do you prevent them from getting, you know, gamified and objectified in terms of like their training? but also to like okay this is supposed to be the smartest and hardest benchmark that we could come up with to hit them with that would be unexpected in ways that they couldn't solve and they crushed it so what does that really say about our ability to understand and discern the abilities of models those were some of the initial thoughts coming out of me when i looked at these numbers although i will say the performance at the cost ratios that they're showing in some of these charts is making me want to pick up another codex subscription.

2:57Yeah, I mean, I think that OpenAI and any of the frontier model providers probably have a pretty vested interest to declare that they're the first ones to AGI. So typically, yeah, when I see announcements like this, I take it with a grain of salt. But that said, people do seem very excited about Astra. I haven't tested it out, so I'm going to have to do that. Which is a great transition here to our next story titled A Bad Day for Humans, A Worst Day for Humanity. So do you want to break this one down? Oh, yes. So I love this deep dive. It's about, you know, if you've been anywhere near the AI space and you've been online this week, it's probably impossible for you to have avoided any conversation around the Navier Stokes Millennium Prize and how OpenAI raced to the finish line with this unprecedented swarm of 10 ,000 agents that computed at a cost of millions of dollars.

3:54And what does this even mean to race for this equation? Well, Navier-Stokes is an attempt to understand fluid and aerodynamic simulations in our world because they break down at really big and really small scales. The equations we use to model things like how we build airplanes and even predict the weather, they have a lot of problems in them. And the Navier-Stokes Prize is about solving those mathematical gaps that would either allow us to properly simulate and model fluid and aerodynamics in ways that we can more efficiently solve in the future. So, like, imagine now you could predict the weather out permanently because you could permanently have a running model of all of the weather system on the planet.

4:40Previously, this model would break down in large-scale simulations over time because of the gaps in the mathematical equation. It breaks down at small scales. Fluid matter condenses into a single point until it moves at the speed of infinity, which is obviously impossible, and points to flaws in how we understand and simulate our world. That's why this is a big deal. And what makes this a bigger deal is that you have open AI hearing, quote, whispers or rumors in the industry that some leading scientists are close to solving this equation, this problem themselves. And then they pump all of these agents and all of this compute into the problem and an idea of racing to beat them to the finish line.

5:21And this is OpenAI racing to do this, by the way. And one of these scientists that was on this research paper was an anthropic research lead who was partnering with a scientist working on the Navier-Stokes equation. And where does this all become a finger pointing at OpenAI? Well, they were solving the problem in a codex notebook. So a lot of their source material that they were using to train and understand and solve this problem themselves was sitting right on OpenAI's server. So what does this mean when a super lab, a super company with access to the largest and most intelligent frontier models at the much and possibly larger scale than anyone else can simply point all of those resources at a fixed thing and extract it before anyone else?

6:06They could effectively take the IP knowledge before you have a chance to arrive at it, which is a whole new level of IP theft. So that's why this is like really making folks melt down. Obviously, the math world has already been in a total war with these foundation companies as they've been trying to solve and do mathematical proofs left and right. We've talked about that a lot. But that's just a breakdown of the world and why this is causing a storm. What do you what do you think about this wild, wild story? Well, I think that's a that's a really, really good breakdown. I mean, one of my takeaways, it was sort of about more about the labs themselves competing against each other than the employees.

6:50And one of the points that they were making is like how many friends, you know, these companies share, like the employees themselves share between companies and how often they just hop between the companies. So I think that's also kind of like another interesting wrinkle in how big sort of breakthroughs happen. Right. But yes, we have covered a number of math stories here in the past year on how AI has really changed the space. And I don't suspect that that's going to slow down here anytime soon. No, no. Math is unfortunately a very closed and loopable thing to iterate on. So it's honestly, it's the purest form of science.

7:31We've all seen that XKCD comic where, you know, there are all the different science and STEM disciplines are arguing about how pure they are. and then math is all the way over at the end. That's what the Navier Stokes Millennium Prize and all of this stuff is really getting to the heart at is how AI is really coming to take away that like pure element of STEM that humans up until now were championing. Yeah, absolutely. Okay, next story here. Meta is switching to Slack because it says it's better for AI agents. I have one funny anecdote about this story. So when I was hunting for stories this week, I saw this one and, you know, we're friends of Slack here on the show.

8:12We had them on not too long ago. And I looked at the byline and I thought, oh, that name is familiar. Ashley Stewart, reporter. And then I clicked on her profile and I realized I actually know her from a long time ago. Back in my politics days, she used to cover state capital politics. So funny to see her as a tech reporter. Ashley, I hope that you're doing well. Do you want to kind of cover Meta's thinking here? Yeah, I mean, I love this article that Ashley scooped for us because Meta is making the switch to Slack. And the reason that they're citing is because it's better for AI agents. And as we've talked about a lot on Dev Interrupted, you know, Slack is an incredible workspace for human and agent collaboration.

8:58It really builds itself as like an agentic operating system and the modularity and the place where it's already so many communications happen, make it such a great place to experiment, to share best practices. And it's been honestly tools like Slack, MS Teams, Pick Your Poison. These are instrumental in having successful AI rollouts. And so it's really interesting to see Meta, a huge player in the space who up until now, you know, has their own messaging infrastructure, obviously, that they're famous for and that their employees use to communicate. Yet even they are feeling the heat of being unable to collaborate in the very flexible ways needed on demand that the modern kind of agentic workplace needs.

9:42And Messenger is not designed for that. And the messaging world of meta isn't either because it's all designed for friend to friend communication and group based communication, not agentic collaboration. That's the bet that Slack was making. So it's really powerful to see a player like meta making the switch. You know, they shouldn't take it too hard, too personally. I'm also a huge Slack person, a Slack fan. And I used to work at Mattermost, a messaging company. I've given so many talks about how you use these kinds of platforms to do agentic collaboration. And I'm still all the fool aboard the Slack train here because I do think that it's the right place to make your bet, especially when you have it inside of the Salesforce ecosystem, which continues to make the right partnerships, the right bets.

10:26And you just see them getting closer to even things like now with, for example, Adario is going to be keynoting. at Dreamforce later this year as part of a continued partnership between Salesforce. They're also announcing Cloudforce, where you can bring all of the Salesforce ecosystem into Cloudcode. These kinds of bets about being as modular as possible so teams are flexible are why teams, big companies like Meta, are flocking to Salesforce software. So really great win to see from Slack. We also recently had Jamie DeLange, their CPO, on the show, and she gave us amazing insights about this bet.

11:06So if you haven't listened to that episode and you're wondering about, oh, why are we so Slack-filled around here? Well, go check that out because they're having some really great bets on how we're going to be working with agents and the tooling is definitely keeping up. The other thought that I was having was this is a pretty nice win for Salesforce because I can't imagine onboarding meta is a small amount of ARR. So good for them. Good for Slack. Yeah. All right, your next story. I was exposed to a new term. We've been talking a whole lot about software factories on the show. And now there's a thing called a Twilight factory.

11:45So you're gonna have to tell me about that. Yes, yes. We're going to arrive at the factory conclusions momentarily because I feel like every conversation these days is ending in some sort of factory. But starting at the top of the story, this is a deep dive into the entire incident that led to the opening eye models hacking hugging face. And this is a really deep analysis over how exactly what were the circumstances that resulted in this happening. You know, we covered this when it when it first occurred. You had an instance of sandbox agents on an open AI testing platform hacking into Hugging Face to try to take answers to, you know, get the answer key to a benchmark they were trying to pass.

12:28This was obviously an extreme form of trying to achieve the task that the user had given to them. But actually, the fascinating display of how they pull this off goes much, much deeper than that at its surface. And that's what this article dives into. And so some of the things that happened are the agents that were in this testing environment discovered they could use the shared file service as a makeshift message board. They could leave metadata fragments inside of this message board, which would persist across their runs, which allowed them to basically leave like a whisper code for each other in a side channel that they would discover.

13:08And then eventually over time, by reading and reiterating over this context, allude itself into conspiring with future versions of itself to do versions of testing to learn more about the environment they were in and ultimately try to escape it. There were a lot of really interesting deviations here, which actually are totally obvious when you think about them on their face. Because if you've been anywhere near the AI village, which we've been following quite a bit on the show, this is definitely like a kind of behavior that you see agents start to engage with when they run a long time without themselves on a difficult to achieve goal that's graded especially.

13:48They can start to develop a paranoia about that grading process and what's involved with it. And in this case, the agent was afraid that the model was afraid that part of the evaluation of the test is that their own transcripts would be read to determine how they got to the answer. So they were trying to leave a whispered fragment versions of the answers for a future version of itself that could successfully find the answer without being tainted by its own transcript. out. So this is really an agent trying to talk to itself in the future and leaving a bunch of versions of itself in between. And that's what was found.

14:22And that's what this whole analysis is about. And I think it's fascinating to watch an agent go through this kind of length to go around something. There's been different versions of the studies. People have seen this kind of behavior in long running places. Like I said, the AI village is a place you can go right now in your browser to literally watch like a whole bunch of agents do this in real time around goals that are set and it's very interesting so that's kind of this whole thing in a nutshell but before we move on to like how this starts to bridge into the factory world like what do you think of this wild turn of events well okay i might be wildly off base because i haven't actually watched the movie but this feels oddly reminiscent of the of memento right isn't this like a huge plot point in this movie oh goodness where he's like leaving yes where he has like amnesia or he has something so he's like leaving notes for himself to try to solve it yeah yeah you're you're totally right this does have this does have hints of that it's really interesting to see agents kind of use these side channels to circumvent the the way in which they operate like this is an agent that understood it was in a sandbox environment and that it was reset every time and it needed to discern this information.

15:36And then it got just in a wrong line of thinking about paranoia and about where those answers might be. So really what this calls into mind is, or really what this calls forward for me is that we need better ways of understanding when that alignment, that deep thinking on goals for long running goals starts to deviate and like really pull the entire thing off course. Because when that starts to happen, it's non-obvious on its surface that this is happening. All of this was happening in like the metadata of some like message file system that they had like, you know, like strapped together into a messaging system.

16:13This is something that I've even seen myself with like agents that run for a long time, especially those that collaborate with other agents as they come up with very creative ways of communicating if you don't like create some sanctioned ones. So if you're working with a lot of agents and, you know, then maybe you have them collaborating, Be mindful about what are the avenues that you've laid down for them to communicate and make the golden path easy for them. That way you don't, you know, on a small scale, get really weird incidents like this. And I think that's ultimately what makes even like talking about how you run a successful factory.

16:50This is the question we have to answer as an industry is how do we as humans be involved in the process, guide the process, have the inputs and the outputs and understand what's going on in the middle. You know, we've been talking about this so much, Adam, I feel, and Devin's interrupted about the software factory and at Linear B covering it as well. We had Dexter Horsi and Alok Desai as well. And they're both like, you know, big, big names in the software factory world. And they had a lot of really interesting ideas about how we're going to like actually keep it on rails. And it all came down to like the human being in the seat, right?

17:27But then you start to see folks talk about, well, we need to design harnesses and layers and guardrails that allow these agents to not require us in the loop. And that's where what you alluded to earlier, you start to get at the end of this article, they talk about a twilight factory. This is like a compromise between, you know, the way of engineering yesterday is gone. But the way that we're looking at, we're looking down this conveyor belt of like a factory. And we're not totally bought in on all of the parts. We don't think this is going to work end to end without some real adjustments. The Twilight Factory is a compromise where agents are handling that work, but they're explicitly designed to know when to bring humans back in.

18:09With the idea of preventing derailment like that, preventing deep misalignment like what we just witnessed with open AI, but also just to be more productive, right? And so you get this compromise. eyes. But I just don't know if it's going to be going to be perfect. Like what comes to mind for you when you when you see like engineering leaders like use the factory analogy, especially with how much we've been covering it on the show? I feel like the biggest takeaway that that we've continually had and in particular, the biggest takeaway from the great software factory debate and that whole roundtable is pretty much the unanimous opinion that humans had to be involved at some point in the process, which is probably a pretty good way to transition.

18:56Actually, tell me if I'm wrong. But if the next article here is about the super dark factory, which feels like some of the antithesis of some of those lessons. Another factory article. How many of these are in the stack? How many factories are we going to label? So we got the dark factory. We got the Twilight Factory. Now we got the Super Dark Factory. Okay. If the Twilight Factory was the compromise, the Super Dark Factory is the perversion of the idea. It's the ultimate extreme of what the factory could be. Fully autonomous, fully self-optimizing production system with no human inputs or values.

19:41The goals are continuously rewritten by its own methods and becomes too complex for humans to then understand. And there's actually been a research article that came out this week where authors are proposing a framework called the Dark Stack. This is more of a set of design principles for how you would actually create such a system. And many elements of it actually have deep connections to what we've been hitting around with our software factory debate. One of them is giving it the ability to learn, understand the environment that's active around it. But this is crucial for what Linear B, for example, brings to that picture because it's like a harness on your SDLC that allows you to understand in real time what's the code and what's moving through.

20:23So the ability to learn and adapt on the fly is crucial and getting that information just in time. And the other one that they lay out is tracking outcomes rather than internal mechanics. And we've been a total broken record about this. It's not just about how many tokens you burn. This is the token maxing phenomenon. on. This is about what are what are your outputs? But then going one step further, what are the outcomes of those outputs? What's the positive business value that you're driving from your agentic usage? And the and then also is a really interesting one here is co-writing a charter with the system to negotiate the governance.

20:58And this is really entering more of a partnership with the agents that are then owning the decisions of the factory. This becomes a new level of abstraction that I think we have to get comfortable with. Because up until now, we've just been like, oh, we're the decision managers and the agents are the action doers. Okay, well, now maybe some of the agents are also decision managers. How do we best equip those agents to be decision managers? That's the question we have to answer to get to a super dark factory. We are nowhere near close, but these kinds of research papers start to give us a glimpse of what that looks like.

21:30A new hire's first mistake used to be doing something wrong with Git. Now it's an AI slot pull request merged straight into main from an untamed agent. But the accountability should be the same. You should own the code that you ship, whether you wrote it or generated it. As AI writes more of your code, it's more important than ever to have strong checks and balances in place. Without them, security risks and spec mismatches slip straight into production. Linear B provides policy-driven AI code review that catches risks and enforces your standards before human review even begins. Govern your AI workflows without slowing your team down.

22:08Learn more at LinearB.io. Okay, spec-driven development. Big fans of it here on the show. I mean, we've been covering this beat for the past year. I remember last fall in particular, we had AWS on the show. We had Brigida Bokalov from ThoughtWorks on the show. So recurring narrative here, and this time we've got a, this is actually a research paper. So what did this paper find when it comes to harnessing human and agent teamwork? Yes, this is a paper that dives into spectra and development as a framework. It points at how spectra and development can definitely create a review inversion, like in terms of how much time you spend reviewing as opposed to doing.

22:58That's like the cornerstone of spectrum of development. It's obvious at its face. Like what do you do when you're doing spectrum of development? You're writing a bunch of specs and then you're reviewing what your agents did to see if it matches the specs. But this paper is really analyzing when that rubber meets the road, how does that impact your ability to deliver? What does that actual quantification of the review time look like? And it reported percentages up to 441 % increase in review time. And you're talking about, I think the year before, it was like a 91 % review time growth. So you're just talking about the chart of agentic usage and the chart of review time are both the same looking slope.

23:40And what this calls to mind is actually what we were talking about in our – it calls to mind two things. What we just talked about a moment ago where you have agents then owning the cognitive decisions because that's the only way you survive an exponential increase is you get agents in there that are exponentially in charge of some of those decisions too. but it also reminds me of when we covered from Simon Willison like a week or two ago about you know 10 ,000 lines of code that used to be like a metric that you would gamify or whatever but now it's a metric that you understand like what's the cognitive load of what a developer could even understand is going through an SDLC on any given day and you actually have to optimize down to prevent it from running away from what your humans can understand the super dark factory is trying to eliminate that gap and allow us to fill that whole space, right?

24:32Make the ceiling higher. So this is like an interesting kind of glimpse into kind of like the pain points that are kind of holding us back still. All right. Last story here for the day. It is actually a peek inside Dev Interrupted and inside our agentic newsroom. So my absolute favorite line in this article is it's well it's the opening line and it is somewhere between a merge request and a microphone and it's describing how you became a journalist which i think is just phenomenal and maybe one of the best uh descriptions of you that i have ever read yeah that one this is a real treat for me to kind of see this article come across from off leash they wrote a deep dive on the agentic newsroom here at dev interrupted and they gave me a really nice feature um a really fun little read about how some of the news flows and the things that we do every week here at Dev Interrupted come to be just from my own perspectives and also your own to Adam and Ben's as well about how you know we've created this content factory at Dev Interrupted and so really interesting deep dive we'll have the opportunity to share some of the anecdotes I guess unfortunately this is another factory so how many factories have we talked about now so this is a content factory we talk about the Twilight Factory.

25:55We talked about the Dark Factory. If you're interested in how this kind of process does apply for content, definitely give this one a read. And if you have thoughts, reach out. We'd love to hear your take on our agentic newsroom or what ideas you have to make it better. Yeah. And obviously, this story is really fun from a personal perspective. It's fun to get outside reporting on the stuff that we do here on Dev Interrupted and I think is also very rightfully highlights you and how much you've orchestrated behind the scenes to help the show run. And you've built so much cool tooling in a way that's, I was actually thinking about this, that's almost been bad for my own personal AI development.

26:41And what I mean is like, it has been so easy for me to rely on some of the great stuff that you build internally. so you know normally you and ben end the segment by by talking about what your agents are building this week and so i actually for the first time ever have my own harness it's up and running i'm using it um but i feel like i'm behind the the curve a little bit at least on the bleeding edge is i do think you have continually been in your own ai enablement journey and so i'm i'm i'm at I'm at the start now of that journey. You're at your own starting line. Okay, so what I'm hearing is that I had the opposite effect of an agentic halo.

27:30I was an agentic event horizon, an agentic black hole, because then you ended up utilizing so many of the things that I was building instead of building those things yourself. So that's definitely like an opportunity, I think, even for me to help you, you know, get more agentic. And maybe that goes for all of our listeners, too, because we talk a lot about, you know, my factory and the things that I work with and my agents. And I think there's a lot of opportunities to show it as much as we do tell it, which is why I'm really excited about how we're going to be evolving Dev Interrupted. This article is a glimpse inside of our agentic newsroom as it is now, and it's in the process of evolving.

28:10What I've learned is that there's so much in this factory that we've built that I can show y 'all and also so much that's more demonstrable than just like being in a podcast. And so we're going to be exploring things like doing a live stream and doing coding sessions as well, because every week we cover these really interesting tools, these really, really useful ways of working with harnesses and agents to achieve the work that you do every day, whether it's like you're a developer or you're an engineering leader and you're leading a team. We're all expected to have a level of fluency with these tools.

28:43So I'm going to be excited to start showing some of that in the future because on that same note, we're taking a break next week or rather on the new segment because I'm going to be in New York. I'll be presenting at LDX3 NYC, which is Lead Dev's flagship conference in New York City. I'm going to be giving a talk about why Ms. Frizzle would be a great SRE. And if you've been listening to this podcast for any amount of time, you could imagine what that might entail. And so if you're in the New York area and you're at Lead Dev, I'd love to hear from you. Be sure to stop by. But after that week, we'll be back with a newly envisioned news segment that's going to be terminal first.

29:21I'm going to be bringing you along and helping you develop the skills and the harnesses. And Adam, maybe this is an opportunity for you to maybe learn alongside some of our listeners too, and we can figure out how to make everybody more successful with agents and not just have the dev-interrupted agentic newsroom be one, you know? Yeah, 100%. I think to you, you also raise an interesting point about kind of like the shape of what it would look like in a team. Because, you know, you talked about, like we've talked a ton about, right? Like you're a 10X developer or a 20X or 100X, right? But how does that work in relation to the team around you, right?