State of Agentic Coding #6 with Armin Ronacher and Ben Vinegar

⬅️ Back to Videos

▶️ Watch the Video

📝 VIDEO INFORMATION

• Content Type: Podcast Episode / Discussion
• Title: “State of Agentic Coding #6 with Armin Ronacher and Ben Vinegar”
• Creator(s): Armin Ronacher (Arendelle, Pi, Leos), Ben Vinegar (Modem)
• Platform: YouTube (State of Agentic Coding)
• Duration: ~1h 36m
• Publication Date: May 11, 2026
• Link: https://www.youtube.com/watch?v=JM1sIVIZYRg

E-E-A-T Assessment

• Experience: 5/5 - Armin Ronacher is the creator of Flask, Jinja2, and the Lektor CMS, now building Leos (email agent) and Pi (coding agent) at Arendelle. Ben Vinegar co-founded Sentry (software observability) and now leads Modem, building product agent harnesses. Between them they have two decades of hands-on experience building developer tools and AI products.
• Expertise: 5/5 - Both operate at the cutting edge of agentic engineering. Armin has direct, daily experience running an open-source coding agent (Pi) at scale. Ben brings product and economics expertise from Modem’s agent harness work. Their combined perspective spans infrastructure, economics, security, and product design.
• Authoritativeness: 5/5 - Armin is one of the most influential Python developers alive. Ben has deep roots in developer tooling via Sentry. The State of Agentic Coding podcast is the definitive recurring series on this topic. Both guests have track records that speak for themselves.
• Trust: 5/5 - Both are transparent about uncertainty, failures, and the messiness of building in this space. They share specific numbers, name competitors honestly, and don’t shy away from self-criticism. No hidden commercial agenda beyond genuine knowledge sharing.

Verdict: Proceed with review - This is a primary-source conversation between two of the most credible voices in agentic engineering, discussing the economics, security, and strategy of AI-powered development with unusual honesty.

🎯 HOOK

Two of the people who literally built the tools you’re using to read this are worried about what happens next. Armin Ronacher-made Flask, now building a coding agent called Pi that got acquired into his own company-and Ben Vinegar-built Sentry, now building product agent harnesses at Modem-sit down for the sixth episode of their podcast to talk about something nobody else is saying plainly: the economics of AI coding are breaking down, and the industry is running out of runway. From rising GPU costs that ripple into SSD prices, to enterprises clamping down on $250,000-per-engineer token budgets, to the quiet beginning of the end of subsidies-this is the conversation that should make every founder, engineer, and investor pause and recalculate.

💡 ONE-SENTENCE TAKEAWAY

The age of cheap, subsidized AI compute is ending; the companies and developers who survive will be those who build principled, economically sustainable products rather than riding a wave of artificially cheap tokens and venture-funded infrastructure.

⚖️ VERDICT

Overall Rating: 9/10

This is the most honest, economics-focused conversation in the “State of Agentic Coding” series. Unlike many AI podcasts that oscillate between hype and fear, Ronacher and Vinegar deliver a clear-eyed analysis of the financial mechanics underlying the AI boom-and why they’re unsustainable. The discussion of token spend crackdowns in enterprises, the rising cost of compute cascading from GPUs through SSDs, and the “beginning of the end of subsidies” is remarkably grounded for a pair of practitioners who are actively building in the space. The conversation about data moats, consent, and training-set deals is especially timely given the legal and ethical landscape. What elevates this beyond a doom-and-gloom episode is the genuine constructive vision: both guests are building products they believe in and articulating what principled AI businesses should look like. The only gap is a more detailed treatment of specific technical solutions to the cost problem, though that may have been intentional given the breadth of topics covered.

📊 EVALUATION CRITERIA

📖 SUMMARY

Episode Overview

This episode of State of Agentic Coding features Armin Ronacher-who has evolved from creating foundational Python tools (Flask, Jinja2, Lektor) to building an AI product company (Arendelle) with two products: Leos, an email agent for everyday users, and Pi, an open-source coding agent-and Ben Vinegar, who co-founded Sentry (the software observability platform) and now leads Modem, a company building what he calls “a product agent harness for product work.”

Key Themes

1. The Recaps: AI Engineer Europe and Miami

The conversation opens with reflections on the major AI engineering conferences. Armin attended AI Engineer Europe, where he and Christina gave talks (his on why Pi exists and the pain of running open-source projects in an AI-saturated landscape; hers on introducing friction into the engineering process to maintain code understanding). Ben attended AI Engineer Miami. Both praise the conference circuit’s role in legitimizing AI engineering as a discipline, and highlight the value of hallway conversations over formal talks.

2. RAM, SSDs, and the Rising Cost of Compute

A significant portion of the episode is devoted to the economics of compute. Armin explains, using Canadian dollar data, that high-density RAM (64GB, 128GB, 256GB) has been getting progressively more expensive. But the real revelation is that SSDs and NVMe drives-particularly useful for prompt caching-are spiking in price dramatically. Armin traces the causal chain: conflict in the Middle East affects helium supply (35% of helium comes from the region, critical for semiconductor manufacturing), oil and gas prices drive power costs up, and the explosion of agent-driven AI workloads creates unprecedented demand for both GPUs and their supporting infrastructure. Ben’s early prediction is that energy costs will be the next price spike, and eventually manufacturing equipment for data centers will become scarce.

3. AI Security Harnesses and Slop Vulnerabilities

The conversation turns to AI-driven security research. David from Sentry released Warden, a vulnerability-finding harness built on the Claude Code SDK that found over a hundred vulnerabilities in Sentry’s source-available code. Armin notes the fascinating paradox: AI-generated security findings are simultaneously “slop” (they contain errors, like referencing a non-existent Red Hat Linux 14.3) and genuinely critical (the copy.fail vulnerability it discovered gives root access to Linux machines). The security harness ecosystem is proliferating, from Warden to competition hackathons. Armin also discusses the Mythos situation-Anthropic discovered severe vulnerabilities but kept the model unreleased, only for details to leak. This raised the question: do companies withhold models because the vulnerabilities are genuinely dangerous, or because they’re scared of what an open ecosystem would find? Cal.com decided to close-source their product partly because being open-source makes it too easy for AI-driven vulnerability scanners to find issues.

4. Enterprises Clamp Down on Token Spend

Ben introduces the concept that companies are starting to standardize and restrict AI tool usage-not from fear, but from financial necessity. If someone rolls out a suboptimal AI extension at scale, everyone’s token spend goes through the roof. The conversation reveals a growing divide: model providers want users to “give more agency to the agent,” while users discover that unconstrained agency equals astronomical bills. Ben recounts Grapile (a code review tool) switching from seat-based to per-use pricing, causing his bill to jump fivefold-to $800/month. Both agree that the SaaS pricing model of the previous decade doesn’t translate to the AI era, where COGS is fundamentally different.

5. The Beginning of the End of Subsidies

This is the episode’s most provocative thesis. Armin declares that we’re seeing “the beginning of the end of subsidies”-not a sudden termination, but a gradual withdrawal. Evidence includes Claude Code removing OpenClaw from cheaper plans, experimenting with removing Opus from $20/month tiers, and certain prompt phrases triggering per-token charges on supposedly unlimited plans. Both guests describe a pattern: AI products that were “free” or artificially cheap are shifting to per-use, usage-based pricing as actual compute costs make old models unsustainable. The parallel to the Industrial Revolution is drawn: factories and machinery suppliers both faced massive capital expenditure pressure as productivity and innovation accelerated faster than anyone could plan for.

6. Why Pi Got Acquired (into Arendelle)

Armin tells the origin story of Pi, starting with Peter Steinberger introducing him and Mario to coding agents in early 2025. They began building in parallel-Leos for Armin, a WhatsApp relay agent for Peter, and Pi for Mario. When OpenClaw “exploded” in January 2025, interest in Pi surged, and Armin convinced Mario to join forces. The technical acquisition makes sense: Arendelle provides the business structure and responsible AI framework, while Pi provides the technical foundation that Leos was already built on. Armin emphasizes transparency: Pi will remain open-source, but the company also needs to generate revenue from it. He describes a “chicken and egg” situation with trace sharing-once enough model labs want training data, and enough open-weight model builders exist, sharing traces becomes valuable.

7. xAI, Cursor, and Why Traces Are Gold

The conversation pivots to xAI’s acquisition of Cursor. Armin explains the logic: Grok has GPUs but lacks training data, while Cursor has massive amounts of user coding traces but no GPUs. The $10 billion services deal with a $60 billion acquisition option (essentially a “great escape strategy” for Cursor) reflects the extreme valuations in the space. Both guests discuss the staggering implications-a VS Code fork valued at $60 billion, Ford Motor Company (169,000 employees, 120+ years old) worth $48 billion for comparison. The deeper discussion is about traces as the new training data: coding traces are gold because they start with human input, contain human reasoning, and have a mechanically verifiable reward signal (did the user commit?). Both argue that the current system-where traces flow to ClosedAI-is broken, and there should be incentives for sharing traces with open-weight model providers.

8. GitHub Outages and Leaving the Platform

Ben describes a breaking point: GitHub’s infrastructure has been failing (Elasticsearch issues hiding pull requests, Actions outages), and Mitchell Hashimoto (Terraform creator) publicly announced he’s moving his project off GitHub. Armin connects this to infrastructure costs-as more AI-driven code pushes hit GitHub, they need more storage, more compute, and potentially Azure migrations that GitHub can’t fully fund. Both discuss the irony that every AI tool is built to integrate with GitHub, creating a catch-22 for anyone who wants to leave. Ben recounts evaluating alternatives: Pierre, Tangled, Codeberg-though each has limitations. The discussion touches on how GitHub has gone from one platform among many to a de facto monopoly, and whether federation (like the Go dependency system hardcoding GitHub URLs) makes escape nearly impossible.

9. CVS, Subversion, and How GitHub Won

In a historical tangent, the hosts discuss version control history: CVS, SVN (Subversion), BitKeeper, Mercurial, and Darks-all predecessors to Git/GitHub. They recount how the Linux kernel project’s switch from BitKeeper (after the reverse-engineering controversy) spawned Git, and how GitHub’s social/collaborative layer (combined with the failure of Mercurial to build equivalent social tools) led to GitHub’s dominance. They share personal stories: Ben boycotting GitHub in favor of Mercurial/Bitbucket, the era of self-hosted Subversion servers, and the gradual migration of everything to GitHub over 2014-2018. The current situation creates a fascinating tension: GitHub is more dominant than ever, yet more people than ever are talking about leaving.

10. Data Moats, Consent, and Training-Set Deals

The conversation turns to data as currency. Armin predicts that companies sitting on proprietary data will find ways to monetize it-whether selling training data to model labs or using it as leverage. Examples: a meeting recording company that sold transcripts as training data, and the emerging practice of companies explicitly asking for consent to use data. The discussion references the Delve controversy, the EU’s failure to solve EULA consent at a societal level, and Microsoft’s expanded training on GitHub data causing brief outrage before people accepted it. Sentry’s explicit opt-in consent model is praised as a counter-example. The conversation becomes emotional discussing Discuss (the embedded comments platform) being sold to a data broker-underscoring how user-generated data often becomes the most valuable asset, to the surprise and dismay of the people who created it.

11. A Plea for Principled Products

The episode closes with both guests articulating what they’re building and why. Ben emphasizes wanting “slow, painful, hard work to be rewarded” rather than shortcuts. He wants explicit, visible consent flows and principled business practices. Armin describes his long-term vision: a world where AI models are commoditized and open, where users retain their data, and where the choice of model is explicit and transparent. He hopes copyright will eventually be worth less and people will run powerful models locally rather than depending on Sam Altman’s cloud. The closing sentiment is hopeful but urgent: the current system rewards the wrong behavior, and the people building responsibly need the market to punish the people building irresponsibly.

🎬 CHAPTER MARKERS

• 0:00 🏷️ Introduction
• 0:02:14 🏷️ AI Engineer Europe and Miami recap
• 0:10:04 🏷️ RAM, SSDs, and the rising cost of compute
• 0:17:05 🏷️ AI security harnesses and slop vulnerabilities
• 0:22:37 🏷️ Enterprises clamp down on token spend
• 0:26:23 🏷️ The beginning of the end of subsidies
• 0:36:33 🏷️ Why Pi got acquired (into Arendelle)
• 0:45:29 🏷️ xAI, Cursor, and why traces are gold
• 0:57:25 🏷️ GitHub outages and leaving the platform
• 1:03:33 🏷️ CVS, Subversion, and how GitHub won
• 1:25:24 🏷️ Data moats, consent, and training-set deals
• 1:35:37 🏷️ A plea for principled products

📝 FULL TRANSCRIPT

I want the slow, painful, hard work to be rewarded. Not not the [ __ ]

This idea that we have to piggyback on top of GitHub. I I think everybody’s rejecting that increasingly. Companies don’t want to spend $250,000 per engineer.

Okay, so welcome back to State of Agentic Coding number six. I don’t think we’re going to hit the 10,000 episode milestone that Syntax hit. I’m just going to throw that out there.

I don’t think the AI engineering bubble will last that long.

So, look on this on this podcast. What do we talk about, Armen?

We’re talking about agentic engineering.

Vibe coding. I don’t know.

Vibe engineering. My name is Ben Vinegar. I work at a startup called Modem. Used to work with Armen uh for a long time at a company called Sentry where we did software observability. Modem is… I’m I’m beginning to think about it more. Humor me as a you know, we have like coding agent harnesses.

I believe it’s a product agent and I don’t know. I’m going to explore this to see if that makes any sense whatsoever. I didn’t see any reaction for you, so maybe it’s going to go back into the box. Um, but we help people do like product work, not coding work. Armen, what’s uh who are you? What are you doing here?

Hi, my name is Armen. I now work for a company called Arendelle that I started with a friend of mine and we are building AI products for humans. I don’t want to be too specific because like one of the products that we have is is Leos which is an agent that sits in email which where we’re trying to build something that sort of a normie can use. Um and then we now also have Pi which is a coding agent. So it’s like an umbrella company of two products. I can talk about that later.

We we’re going to come back to Pi. Although I I do want to take a moment that if you watched this is episode six and if you I feel like if you went back to episode one, you could basically follow the story of how much you know Armen increasingly loves Pi down to the point that you know it’s now part of his company.

Exactly. If you love something a lot then you have to go after it. That’s basically what happened.

Okay, we’re going to get into predictions in a moment, but we wanted to touch on really quickly the fact that there were two major AI engineering conferences. AI engineer Europe, let’s say that that’s the most major one. Armor was over there. And then just a few weeks later, there was AI engineer Miami. I happened to be there. I guess I just wanted to touch on this because this is an AI, you know, AI engineering agent coding podcast. I feel like we should acknowledge that we are…

Yep.

Well, we’re at the one conference if it feels like it’s the one conference.

Quite frankly, I don’t think there’s another engineering conference that matters…

so far.

You know, by the way, it did feel like one conference in the sense of content almost now just kind of continuously coming out because there’s AI engineer Europe and then Miami and then I think San Francisco World Fair and and then they’re going to have New York. There’s a lot of these now to the degree that just like content is kind of like coming out all the time and it feels a little continuous. I don’t know if that if you’re you meant that comment or not.

So I think what the conference does really really well is um it uses a very I mean it’s not super consistent but like it uses a sort of relatively consistent format and all the talks are also relatively quickly on the internet and they’re streamed. So even if you don’t go to a conference I think you actually have something from it. Um and and Swix does an excellent job at running it. So I think like even as far as like other programming commercial conferences go, this one is surprisingly well done, I would say.

Yeah, I think it’s pretty good too. And I um my first experience was World’s Fair last year and that was like a real eye opener for really this whole field at the time. I do credit Swix, this is Sean Sean Wang for popularizing AI engineering. I think he does deserve I think that’s fair and correct. Do you agree?

Yeah. and also in like I think ways that go beyond the conference like the the podcast that he’s doing and everything. It’s definitely has given some sort of legitimacy to the pipe. I I found it really interesting like AI engineer Europe I think was like a thousand something people is my guess. It was it was pretty…

Oh, yeah.

…pretty filled. And I think as far as like Europeans go, obviously there were a bunch of Americans too, but like as far as Europeans go, um sort of the the attitude towards um AI engineering, I think it feels quite a bit different than uh than in San Francisco. I wouldn’t say it’s like less energetic, but I think there’s always there’s a little bit more balance to it in part also because obviously Europe doesn’t have great like model companies. Um, and so I found some of the conversations really really engaging and sort of un I mean more surprising I would say than like AI meetups in San Francisco or like even AI meetups in Vienna. Um, so yeah was it was quite grounding actually in many ways and then after Mario’s talk and my talk a bunch of people walked up and said like oh this is really nice that you’re giving some balanced input. Um maybe this is just selection bias but yeah I was surprised um in a way was it was much more down to earth I think than some AI meetups are. Also I would say the the the age of the attendees was old older than I expected.

By the way really quickly can you explain what that talk was? you know, we’ll link to it, but just like, you know,

so Mario Mario basically gave a talk on why PI exists and how it is to run an open source project in um in an AI adjacent space that’s full of machines putting pull requests up and how painful that is. And the talk that I gave over Christina was basically one about like how we put more friction into the engineering process so that we can actually still understand the code. Um, and I think both of those talks had something kind of similar going for them. And because I messed up, it was actually really funny. My Christina and my talk was supposed to be at the end of the last day, so the closing keynote, but I messed up my flight. And when we um when when we announced that we acquired Pi, Swix wrote me, “Oh, you should have told me. I would have moved the talks together.”

And then I was joking. And I said, “No, it’s fine.” And then jokingly on the day before of the talk I was like ah I will I will see if I can run to the airplane tomorrow. Um and I was like I was like this this will probably not work. And then overnight Sean moved the talks together last minute because I think he wanted Mario Mars and my talk and Christina’s talk to be like next to each other. I mean I confirmed it afterwards that he can move it but it was I think he took the opportunity a little bit to to to put them back to back. Um, so that wasn’t intentional that that those were like Mario…

submitted his talk proposal prior to even us uh coming up with our talk proposal. It just happened to be on a very similar topic.

Stuff comes together last minute often at conferences. I think there’s a lot of moving variables that people don’t always see. I actually participated in two talks. You know, that was fun. While I was in I was very privileged to um get to chat with Sonel Pi from Cloudflare and Max Dorber who you originally introduced me to a long time ago uh from OpenAI today and the panel we you know we were being interviewed uh by Eric from code rabbit uh about taste. I guess that the setting was like, “Hey, all of you have worked on sort of like developer SDKs and experiences and how much does taste factor into you building today, right?” And then I had a talk which was about how I’d sort of moved a lot to working with remote agents over SSH and that that was honestly driving a lot of you know my GitHub contributions which we’ve talked about on this show and I wanted to share some of that experience and I had a bit of a preamble of sort of like you know well I’m not somebody who runs things in the cloud continuously but I you know I I do slow down and one of them is that I can’t walk around and crack the laptop open all the time. Like that was a real kind of antidote. Uh which is funny cuz I feel like there’s you see a lot of that talk now online. My takeaway by the way when you go to these conferences, I don’t know if this is true for you, but I think it is. I think for this stuff right now, it’s incredibly valuable because it’s changing so fast and the value of being able to get into a room with a whole bunch of people who are kind of at the bleeding edge of this, right? whether you whether you learn it from the talks or you get the chance to talk to people, which you often do, just to see where everybody’s at, you know, and to kind of share notes and to learn what everybody’s working on and and where they’re struggling and where they’re not. I find that just so incredibly valuable thoughts.

Yeah, I mean the the hallway track is always the most interesting for sure and and and that continues to be the case for me. I think a lot of the topics for me at this conference was just like it is continues to be like surprising I mean surprising not I guess we got used to it but it continues to be reinforced to me how most of the things that we’re all learning are just they have a shelf life of a month very few things actually sort of really stick around I mean I think it was maybe less so at AI engineer because I think everything was kind of fresh but I was at another AI event just a month earlier when you go to an AI event and the talks are sponsored

Then in part, someone just sells you their AI company in a way. Some of which no longer feel all that sexy. So it’s always kind of interesting because it moves so quickly. Here’s a thought. I remember if I went back a year to the World’s Fair. There were so many AI products. So many AI products. There’s arguably way more today. However, when it came to people presenting, it didn’t feel that way. And maybe it could just be because Claude does 90% of it, you know? I don’t know if that’s true. I don’t know if I’m just like talking, you know, if I’m just like making that up, but or it could just be that it was, you know, World’s Fair was so gigantic. I don’t I don’t know that could really compare the two.

So, the first thing I think we should talk about is we talked about RAM prices and and I think we sort of I don’t know who took the up and who took the down on like is it going to increase, but now we know it went further up.

Can I even just set the stage? I’m assuming not everybody’s saw the last episode. So, we look we That was a long episode. We meandered all over the place and we were talking about the fact that compute has you know compute and resources and tokens we’ll come back to tokens you know to be in software right now there’s a real kind of gap between how you know what you can achieve with a lot more resources and it feel it felt like that gap was widening and that’s when I was like well do you think it’s actually going to get worse is it going to get more expensive that’s that’s what we were predicting you took down maybe relative to USD I think was your asterisk like maybe the price will go up but the US dollar will depreciate so um does that make it more expensive or less expensive like

well for you said like if you if you compare it to gold prices

okay

so what’s RAM and gold today

that’s funny by the way we should actually let’s let’s vibe generate a website which is just tracking RAM versus the price of gold maybe RAM is like a better better hard commodity going forward I cuz cuz I gen I did pull up some charts and answer I was looking at this really curious. I was curious. This is in Canadian dollars. So I’m I’m acknowledging that dollar fluctuation may uh make this very questionable. But quick observation which is like the highest density or like the fast the fastest and highest density like RAM dims or whatever like 64 128 they are getting more expensive. I think I think 256 gigs went up. All right, Ben.

Question

question. When we recorded the last episode, was the straight of Hamus open or closed,

you know, because like today in the morning, I read this thing where it’s like, oh, here’s some other things are going to be expensive because Helium is expensive and and like we’re now getting to the point where like where that stuff is now getting more expensive. So, I’m just kind of wondering now, will it go even further up? Because

if it was closed, it was at the very beginning of it. I think certainly the conflict had begun, but you’re right. I Yeah, oil, etc. And like I I don’t want to be here talking about like dim sizes or whatever, you know.

I’m just saying like if if I said it was going down and didn’t yet assume we’re going to have yet another conflict that might make chips more expensive.

Well, I did say I was very cynical and so the price would go up for that reason. All right. Anyways, oh, but this is one thing, one consequence was hard drives are getting expensive. Like I think I think this is worth talking about which is this just you know we all know how GPUs got expensive because of crypto and then they got expensive because of AI right and then RAM got expensive because we could we could sort of connect that to AI and now hard drives specifically like SSD NVMe drives are getting expensive and that’s a more recent phenomenon. They’ve gotten really expensive and look I’m not tracking the news on this. It’s mostly like one day I’m downloading a bunch of LLMs and I’m like, “Oh, I should probably get some more hard drive space and I and I go online and I’m, you know, shocked Pikachu face to to how expensive it is.” And that’s when I start digging into like how much this had actually spiked in price like pretty dramatically. So, I’m wondering one, what is your take on this? It seems like an AI it’s it’s an AI after effect, right? I’m just curious like do you understand cuz I don’t or like you know what are hard drives more expensive or why tokens more

yeah hard drives

I mean so all I know is that I I listen to this um Bloomberg Daybreak something like sometimes when I cycle in the morning to work I’m like economic whatever morning show and part of the conversation was that well by the way like 35% of helium coming from from the Middle East and helium is important for semiconductor manufacturing and so like as that sort of all clogs up the like manufacturing will get even more expensive and the other thing is like as gas and oil prices go up power prices also go up right so there seems to be just a lot of reasons why this stuff is going to get more expensive in the meantime NVME chips in particular I think like anything that is sort of like very obviously useful for like prompt caching sort of like dump a whole bunch of GPU on disk like the more agents they are the more important the caching of stuff is going to be presumably the more people are going to buy that [ __ ] out. It just seems somewhat natural right now that there will be more of a demand for like not just the GPUs but also the whole supply for it. Prompt caching in particular I think is is a huge thing all of a sudden because like even the for a while if you were not entropic and openi you didn’t actually see that much atic traffic to your inference systems because like most of this was just here please answer this question in context of this retrieve document and all of a sudden now it’s like okay they’re hitting you repeatedly with over and over and over and over longer sessions and so now even if you’re like a small player in the space that wants to surf up the GPUs you kind I want to get this caching stuff going. So that to me makes a ton of sense. So it will get more expensive for a while. I think

I just want to comment that was a great answer. Like that was very educational. Thank you Armen. And at the risk of dragging this out early prediction ear I would love for you to make an early just a quick early prediction. You know since we saw like these sort of different different pieces of compute go up in price. What’s going to be the next one?

You mean more hardware that will go up in price?

Yeah. when just saying like you know starting going from GPU to RAM to storage like

it’s a good question. I mean I mean I’m guessing it’s just generally the cost of energy will go up and so maybe that’s not really a hardware thing but so for instance uh I think like Caterpillar stock went up like crazy because people figured out that you got to dig a bunch of stuff to build more data centers, right? So presumably at one point we’re going to run out of like I don’t know manufacturing equipment to build data centers. I think we’re reaching that point, right?

or just the full the full thing, right? Like motherboards and computers because we just want to shove those into we want to show those into data centers so there’s none left.

Well, I don’t know because like those are like kind of like sort of like it’s not like you’re shoveling like a computer into the data center, right? It’s like it’s mostly like the whole H200 whatever these things are called these days like this they’re fully wrecked up machines but like I think it’s going to be like concrete and whatever else you need to actually build these things. I hear you on the um AI inference training right compute but also I think just the if you subscribe to this idea that agents are going to be all over the place they all need computers to live in I think that’s driving some too but we can come back to that

yeah maybe could be like there’s a question if the generally the cloud prices are also going to rise I think because like even the the access patterns are different u but I don’t know man I don’t know

speaking of strange access patterns do Frontier Labs keep models back because they might be discovering too many vulnerabilities. That was something we talked about and then at the same time you highlighted like it doesn’t even matter like even with today’s models you could discover a ton of vulnerabilities and that was recorded and potentially released before mythos was sort of revealed really quick. What was the mythos thing like super high level? So, somewhere between the recording of last episode and now, Entropic announced this myth thing that as of recording has not been released to the wider world, except it sort of leaked to some random Discord group seemingly, but a lot more vulnerabilities have been found anyways by people not using. I think we talked last time briefly about that if you build yourself a nice little harness for finding security vulnerabilities, they’re getting pretty good at this. So what has actually happened in the meantime is that people built more harnesses to find vulnerabilities and I think the most interesting one that came out in between uh that in part has been copy fail which is a basically zero day is I don’t know if it’s a zero day but it’s like basically root execution like gaining root access on Linux machines and I found that particularly interesting because obviously it was very clearly found with AI assistance and they have admitted as much but als

Also the post that is the technical announcement of it was very LLM supported and because it was that and there was a there was clearly some mistakes in there for that talked about Redhead Linux 14.3 which doesn’t exist. The first couple of comments that came in after that dropped were like oh this is just a metal generated blah.

So the the immediate response of a bunch of people was like, “Uh, doesn’t matter. It’s just slop.” And they’re like, “Oh, holy [ __ ] It actually gives you reflexes.” So there’s this really interesting thing right now is like simultaneously there’s a lot of like they slop finds vulnerabilities. It’s still slop, but it finds vulnerabilities which are critical. And so it doesn’t really help us to have a response to it that is like, “Oh, it’s just AI generate nonsenses. It’s still the vulnerability. I don’t know what to tell you.

So, it’s just very interesting. Right now, David from Sentry released this Warden thing, which is his own harness to find security vulnerabilities. Cal.com decided they’re going to use the opportunity to just close source because they feel like being open source these days makes it too easy to find issues. I don’t know.

By the way, Adam, yeah, I I obviously follow David um David Kermit Centry. He used to be our boss. Warden, would you call that a harness or a system built around a harness?

It’s definitely a harness uh in the sense that I think it uses the cloth code SDK to uh which I guess is is sort of the the version one of the harness, but it’s it basically runs uh the loop itself with loading some special skills into it. Um and then sort of more or less automatically try to find all kinds of security issues. So you basically just run it like an agent. What I do know is that David highlighted that it found I think he said like a hundred vulnerabilities in Sentry which is source available. source available is not quite open source but you can read it and so this it kind of connects to the the

and in a in 14 years of AI time which is two years it is going to be Apache to licensed so it is on the way to open source

it’s continuously

continuously continuously

right there are many versions of open source centry right out there right now so one that’s interesting to me because that team has always had access to claude

Right. They have a lot of engineers there. I don’t know, 200 engineers possibly. David more or less, you know, fuse this up with a bunch of skills and with this sort of vulnerability discovery harness just like goes to town and finds like a hundred things. Is that like a that is that the state of the affairs which is like yeah, if you really guide these things to find vulnerabilities not to rehash this thing, not to rehash this too much like

Yeah. But um I think this is exactly what’s happening. And to be clear, like he didn’t just build this in the last couple of days. I think he built Warden already in like January or something, but it became better and better over time. Um and I think generally the awareness of how in part a lot of just like understanding like what everybody else is also going to do like theoretically a lot of these models were really good at finding vulnerabilities in November or October. I don’t know like Opus 4 or five is is equally capable of finding this stuff, right? And the first time I I sort of saw at scale people talking about the security thing was like around Christmas when um at the case computer club congress there was um someone demonstrating this and there’s more and more of it. The the openi hackathon so the codex hackathon that was in Vienna where I was a judge uh there was also a team that did a capture the flag kind of vulnerability thing. It’s all over the place now. So, in part, I think it’s just people waking up to some of the capabilities that are in the models, but they maybe haven’t been so communicated so widely.

It’s scary. So, last kind of like we’re just kind of wrapping up the predictions. Um, and but predictions are just kind of things to get us talking. Another thing that came up was do you think we’re going to lock down more like within within engineering organizations? Are we gonna see companies sort of, you know, mandate certain coding agents and models, right? For any number of reasons is what we were talking about. But, you know, now there’s there could be more reasons now. One of which is token spend. I guess the way I’m doing this is because I’m just kind of like rehashing the comments and I’m just like tossing the the mic over to you. Are we locking down more? Did you know I don’t even think we I think I said yes but maybe it’s going to be months away but I guess this turned out to be a monthly

I I don’t know like to which degree this is already happening but we have some evidence of like team standardizing on things in kind of interesting ways. I will also say like I think we should have a separate conversation later on on like the token pricing itself and sort of like like what some providers might or might not be doing. But for instance, um since we now have according harnessers in the company, uh we also get to talk a little bit to some of the people that are using it. And if you are a large enough enterprise that uses API pricing and you’re not using subscriptions, then all of a sudden cash efficiency really matters. And if someone rolls out um at scale a bad scale or extension, then everybody’s token spent goes through the roof. And if you’re not a believer into token maxing, but you’re a believer into responsible I don’t know budgeting, then you kind of want to know about this, right? And you want to understand this better. And so they everybody runs whatever they want to have is actually in some companies already no longer what they want to do. They might still be at that stage, but they want to clamp down. So I think it’s uh it’s starting. It’s definitely starting. So translation, companies don’t want to spend $250,000 per engineer. Is is that fair?

Yeah. I mean, like that’s like look just to circle back a little bit to AI engineer. I think it was very very funny that um on the I think on the day before Mario or maybe it was on the same day as Mario I forgot there was a talk by an open eye engineer was like just don’t worry about the token spend just let give more agency to the agent and then the talk by Mario was like oh it’s just like it’s all about efficiency and like thinking and like so like the the incentives couldn’t be any different between the providers and the users that actually have to pay for this This is also a really good indicator of this growing divide, you know, like like a little bit of like let them eat cake. I I you know what? I don’t think you got $250,000 laying around. Um you know, if you’re worth a trillion dollars or you’ve got $200 billion in funding, that is a you know, just speaking to Nvidia and OpenAI. Yeah, that’s a trivial concern for you. You’re probably like, “Yeah, but please do it.” And also, conveniently, you know, you’re spending that money on us.

Yeah.

Right.

Okay. Um, so, you know, but companies actually don’t want to do that, including reasonably wealthy companies, I would add. Just maybe not quite so wealthy. anything to add on that or

Yeah, I mean I think there’s a general growing understanding in some companies that are fiscally responsible that they probably don’t want to create an engineering workforce that is fully dependent on dishing out a ton of money to some other company with no negotiation power. So we have reached that point I would say. So the the token maxing from January got old really quickly. I feel like this the last few weeks have been a moment where I’m call I I’m calling it the end of subsidies. It’s not like really truly the end. Like clearly things are still subsidized but it feels like the beginning of the end. How about that? Where one there’s just so many indicators on this. Let’s take let’s talk about the biggest one. Claude Code doesn’t support OpenClaw and for example they took away Opus from or they’re starting to experiment with taking away opus from some of their cheaper like $20 a month plans. That’s a whole topic. I want to continue this.

So, right. So, that’s getting clamped down and now is like even a topic last couple days because now now Claude people have discovered that there’s certain phrases that if you’re on like a max plan, Claude might reject and charge you per token. uh people are discovering that there’s sort of like I don’t know some discrimination on like basically your prompt. So that’s one thing I’m noticing in a lot of SAS products like some of which we use there’s they’re switching from seatbased pricing and sort of like um you know you get a seat some amount of compute goes with that. That was fine when you know you had half your company paying per seat and they weren’t actually using it or we found some way to make it work but it’s not working anymore. So a lot of those a lot of that pricing is moving to like per I’ll give you an example. Grapile grapile is a code review tool. We use it. I like it. There’s so many of these. They switched to per use. Look, our bill went up, I think, like five times, you know, like maybe we’re paying1 150 something like that a month. And I think I just got the bill and it’s like, hey, surprise, $800, right? Look, am I getting a ton of value out of that? Could I rationally agree that I’m I’m getting like a AI code reviewer and maybe that’s worth, you know, 10 grand a year, right? Well, multiply that for every product you’ve you’ve got that is touching AI. It gets really expensive really fast. Do you agree with that?

Yeah. I mean I have so many thoughts about the economics in the space because on the one hand a lot of it is like clearly valuable. On the other hand, we will not pay all those companies per use unless because like think back about like what the what the economics were about like SAS businesses for a really long time which was like well because they the the most of the cost here presumably is like running infrastructure. We are happy with them having great margins because it’s still actually going to be more expensive to for us to run our own show.

So like there that was actually like in some sense it was a win-win transaction to to to use SAS business as a sort of the last super cycle but now like what actually like there are no margins really they’re very small margin like some some people have the margins right so like the the people that run inference look pretty healthy margins but like everybody downstream from that doesn’t have anything so where is the value here particularly if you have the sort of like I can build my own reptile in an afternoon because it sitting on a model anyways and then I have just like my own gateway and I’m just being token efficient and I just like I don’t know it to me it seems like there’s there’s I’m not saying like no nobody’s going to make any money in there and there’s people are definitely going to make money in there but shit’s getting expensive way too quick and I’m cycling out of products just because I don’t it’s like it doesn’t make any sense for me to pay that much money for something that my agent can reasonably do itself

what I want to say is I value the service layer that companies bring to this and I don’t actually want to like generate and maintain my own services again like and then multiply that for everything you have that’s not strictly an attachment to G reptile speaking of which you should ask them to sponsor us for this episode for how many times we’ve mentioned their name GPL

it’s a dinosaur

it is fair to say that that price increase has caused us to be like oh maybe I will shop around again right and I guess I’m calling it it feels feels like a bit of an unwinding where people are sort of ending their subsidies or it may not be so much ending their subsidies. Like it can, but it can also just be that the amount of usage that we are all driving is going up higher and do not support the pricing models that maybe made sense 6 months ago. There are a bunch of things that I uh sort of know is like AP AI products that are like say like a year and a half old which started with basically the following approach. Here’s some rack, here’s some data retrieval and now here is some post-processing on it. That shit’s cheap to run. But if everybody else has an agent, then those companies also have to build an agent and the agent is no longer cheap to run. And so as you have more and more companies going on the well now is should do tool calling and all that kind of stuff train they can no longer offer the pricing model it had previously which scaled to some degree with the number of humans attached to it. The other really interesting thing is that if you have an AI product that is in itself used by agents and not by humans then per se pricing also no longer works. Um, so there’s there’s a whole bunch of like movement just as a result of changing usage patterns and and expectations.

Modem has usage based pricing.

Leos has usage based pricing

but I know for both of us we approached it that way on the if this continues sebased pricing won’t make sense and I know that both of us have kind of had that in mind going back a long time if that makes sense. I mean the reality is that if you’re responsible then if your cost scales with tokens then your pricing has a scale of tokens until the cost goes down to a tenth of it. Like if if the cost of like we never worried about the cost of storage or about the cost of CPUs or something like this when offering SAS products in the last cycle because for the most part infrastructure was so cheap. If GPUs also get that cheap, then presumably nobody cares anymore and you can actually have good good margins and different pricing models, but that’s not the case today.

So, how does this end, right? Does this is this like

is it going to end with like a the death of some companies or

I guess I could see this as like a trigger to a bubble like we’ll find out, right? If you thought there was a bubble, this is sort of an interesting it’s an interesting gut check moment, right? Like, are you willing to spend $100,000 per engineer? Are you willing to spend a dollar per code review or, you know, a dollar for that? You see what I’m saying? Like,

yeah,

this is an interesting moment. Maybe the answer is yes. This is like a real in some ways it could be super healthy, right? Because if this continues and everyone looks at the cost and goes, “Well, it’s more expensive, but I’m getting the value and these costs are closer to the true value of in terms of the cost, right? They’re sort of like not subsidized. People are taking a margin on it, then hey, maybe everything’s great, right? And we maybe even accelerate from here.” Or alternatively, if everyone does a little bit of what you described, which is I don’t really want to pay that and kind of, you know, everyone like retrenches, maybe that could be a real pulling back.

Yeah. I’m reading a lot of I mean, I think I mentioned this a couple of times on this podcast, but I’m I’m I’m reading a lot about the industrial revolution because I think a lot of it is really really relevant. And one of the sort of recurring topics about how the industrial revolution happened was that uh you had factories and then you had people supplying machinery.

But because the productivity went up so much and the rate of innovation of machines also went up that quickly, the capital expenditure that the that the factory had to actually buy the always best machines and to stay competitive was humongous. And a lot of them didn’t make it because like they’re like the the the forecasting of like what the demand will be and what the machine will do and like the cost of upgrading the machine like it didn’t actually play out. So the individually like at the end of the industrial revolution we were all more productive than before. But along the way, it was actually very hard for most people to actually get a lot of value out of it because like it sort of scaled up in a really interesting way.

And I think it’s the same thing now, right? This is like entropy clearly has stability issues because they’ve probably forecasted their AI demands wrong. So now they’re kind of paying for this and we are all more productive at the same time also paying so much money for these tokens. And then we’re not quite as productive because we also do some nonsense. But also, if we don’t play the AI game, then also it’s not going to be great for us. So, it’s like we’re all collectively in this really bizarre situation again. I don’t know. I’m not saying it’s a bubble, but it is just a very painful kind of transition in a way.

And some companies are going to succeed and some will not succeed. And it’s maybe not even the greatest product that’s necessarily going to win. It’s also going to be just really good financial modeling and and I don’t know, it’s complicated.

I’m actually excited for this right now because to me it’s like we get to find out when tokens were massively subsidized. Well, you could generate, you know, you can generate garbage because, you know, if you generate or you could generate so much and it doesn’t matter that 90% of it is garbage because if 10% is good, well, look, you didn’t you didn’t pay the true price. So, so you you’re getting that value out of it. If you’re if we’re all paying the true price for everything and you know now we’re going to be a lot more selective and gonna generate more SLO we’re going to kind of like I think solve these problems a little bit better. I guess I like the honesty of it. I don’t know if that makes sense.

Yeah.

I mean theoretically if we all get our token usage down by 50% then we have 50% more capacity, right?

I mean like there’s a there’s a whole bunch of stuff that could be done that probably should in one way or another sort of even out uh prices that is not just from the supplier side but also from like the demand side. Maybe it’s a good moment to talk about pi.

Let’s talk about pi. It’s my perception. Maybe you can correct me. Maybe you’ve had this secret ambition this whole time. You didn’t set out to be building a coding agent. Why don’t we start there?

No. Yeah. So the pre-story of the whole thing is that in March last year, April, March, I don’t know, sometime around this, Peter Steinberger, a friend of mine, I got me and Mario hooked on coding agents. And for the rest of the year, we just ex shared experiences and and bunch of stuff, right? And so like similar kind of environment and so forth. And then throughout this uh I started a company with Colin and we also already tried to hire Mario. He didn’t build Pi at the time but like we we kind of liked sort of conversations that we had on on like what could be built in the space and so forth. And in October or so, it became somewhat clear that we all like something is going on and and at least Mario was increasingly convinced that he wants the old clo code back I guess that didn’t do crazy stuff and we all started building in one form or another on agents like we were started working on Leos. Peter was obsessed with building this WhatsApp relay agent that was originally based on plot code.

Can I interrupt for a second? Because I think look a lot of people clearly came to claude code later right December holidays when you when you when Mario or you say hey I want that claude code that didn’t do the crazy stuff like what are you referring to like what does it do today that it that that it didn’t do before vice versa. So when when cloud code released early on in April, I think it released earlier, but the the experience that you got from claude code in like April to July was that it didn’t actually do all that much. It it it ran some commands, it read some files, it wrote some files, it added some stuff, but it it was it was rather sort of minimal compared to what it does today. And Mario maintained a website for a really long time called CC history where you can see the changes from cloud code version to cloud code version. And I think I talked about this in the podcast before like both Mario and I we sort of got a little bit obsessed with this idea of like software extending itself. Our idea was like okay this is a thing where like normies can have their own like workflow expressed in a in a tool that sort of sits with them. But Mari’s version of this was like, okay, this can be a coding harness that is adopting to your workflows and and adopting to your kind of way of working. And and the to a large degree, I think the motivation for him was that you can have when you run PI within project A or you run PI within project B, it can sort of behave differently because it sort of adjusts to the the demands of that project. And in December when when when he already was like fully on this and I sort of slowly started moving on it, I was still using AMP for the most part I think because also what I really really liked about AMP is that the way the AMP folks I mean AMP is expensive in a way because like they they basically had to charge token prices. They couldn’t charge the subsidies. But AMP also very much focused on how do you write a harness that’s actually very good. like this there’s like intrinsically it’s doing like not too much crazy stuff and so they had like very few tools in comparison they had this um uh this oracle things like that. So anyways, what happened in January was I I think we talked about this on the podcast that basically openclaw exploded a little bit and I also wrote a blog post on pi in relation to openclaw and what sort of unfolded for the next two months or so was that a lot of people started using pi or talking about pi and there was like oh there was some interest in pi and so we managed to convince Mario that maybe we should do something together because we we have very much aligned interests of what we want to do with it. Why why does Arendelle exist? Quite frankly, it exists because we think that there’s a lot of value in AI, but we want to do it in a most responsible way and make it accessible to as many people as possible, right? Leos is our first attempt of doing it. I don’t think in the form that we have it, it’s it’s going to be the be all end all because like we we kind of want it to be more accessible. Um, but there’s a lot of hurdles in a way of actually making that work. for a start tokens are too expensive.

Um and and and and some things are not quite there yet. Uh but PI is is just a is a really really useful building block also for for building your own agents on it. Like Leos has been built on PI for for the longest time. I think we I think basically from from November or so we built it on top of PI instead of um we used to use the Velia SDK and we moved off that for for PI. Yeah. So that’s the that’s basically a pre-tory. So we Mar is is now part of Arendelle. Um and we we want to run PI as like a very responsible stewards for an for an open source project. So we’re not doing anything that that will take away from this. But we also want to be quite transparent. It’s like if we have PI as as a property with an Arendel then we also want to ensure that PI in one form or another is generates revenue. We’ll see. We have some ideas of of what what we think is is actually going to be a win-win transactions for for both users and us, but um it’s not the highest priority, I would say.

When things like this happen, can I call it an acquisition?

Yeah, it’s technically an acquisition.

It’s technically an acquisition. When things like this happen, it’s rarely out of the blue. Just one thing I guess like I want to share with people because sometimes you see this stuff happen. I don’t know. Maybe TVPN maybe may maybe maybe TVPN is out of the blue. But my experience is

next acquisition will be this podcast.

Yeah. Is just that these things happen. They’re often happening in the background like over a long time, right? Um actually I was listening to the Syntax 1000’s episode which happened uh this week. Syntax is a very popular podcast. If you’re not familiar with it, go check it out. West Boss, Scott Tolinsky, CJ, and a great crew. But they talked about like they got acquired by Sentry and that was a topic that they went into and Wes explained a little bit like that. I think it was a reader question. and they were doing like a live stream and someone had a question of like how did that come together and Wes is like well you know it it went back like a long time like Sentry was one of our earliest sponsors and you know

yeah no nothing ever happens like actually I would say like a lot of things happen without thinking right it’s like a Facebook acquired mold book that was not a long way coming but I would say in my circles uh and I think in your circles we h happen to hang out I think a lot with people that uh value long like they play the long game. Let’s put it this way, right? Where like you you really think this through and at one point it sort of becomes more obvious that this might be a good idea and you like you really force the conversation like like what does it mean for both sides and and it’s it’s very very planned in a way. Eventually, I think I wanna like when when all of this is over, not in the sense of like we have we’ve sort of done this, but like like maybe in 10 years time I kind of feel like, okay, that then will be the moment to talk about like how it came, but there’s there’s a lot that goes on in these conversations. They’re very boring, but actually really hard because you basically like you like I know Mario for a really long time, right? It’s like there’s the risk with this is always also that like you might not just [ __ ] up the business, you might also [ __ ] up your relationship, right? And so all of a sudden it feels like like dating like now it’s going deep.

You’re moving in together.

Exactly.

Right.

Which is also why I didn’t I didn’t want to talk with anyone about this in any form whatsoever until like we were really really sure because I think Yeah.

All right. So like I’m going to take a moment here to to highlight this one more time. So when I I mentioned earlier like if you if you actually listen to the first episodes and you follow this you’ll basically see this whole pi story follow the timeline that Armen just laid out because we we we did the first episode at the end of November. So pretty much following that what I mean to say is also I in some ways like things are there in plain sight sometimes and some of that is happening right now.

Okay, some of that is happening right now on X, you know, or YouTube or whatever, like these little these sort of relationships that are developing and they’re sometimes in plain sight and it’s just whether whether you notice them or not. Prediction for listeners, put it in the comments. What is the next AI coding related acquisition? Oh my god, we didn’t get the cursor.

Oh, we forgot about cursor.

Maybe because it hasn’t happened yet, if you know what I mean. Yeah. Then maybe maybe we should talk also about cursor just very briefly.

Maybe these are all always going to be 90-minute episodes because there’s just too much stuff happening. Um

yeah.

All right. So on the topic of acquisitions, Aarendelle acquired Pi. Congratulations. And in the same tier in the same, you know, in the same uh pantheon of of corporate acquisitions, SpaceX acquired Cursor. Was it SpaceX? Yeah, I guess it’s called SpaceX now. Every like it will be called Tesla in half an hour probably.

Musk is going to merge everything together into massive conglomerate at just the X company. I have a lot of thoughts on this and actually kind of for forgot about it. I had two conversation, one of a good friend and another one um with an acquaintance where we had a discussion on a Grok 420. I don’t know if you remember this, but there was a Grok model that didn’t do amazingly well. I missed that one. I’m glad. I’m glad I missed that one.

But they basically try to train a model purely on synthetic data. Generally the the feeling within the industry seemed to have migrated towards that that doesn’t quite work. So like purely like training in a good aentic model and purely aentic data is not a winning strategy at least right now. So that also has increasingly led to certain companies that actually have good data being really good suppliers of training data. And so uh Grock has the GPUs but Grock doesn’t have the data is sort of the short version. Cursor doesn’t have the GPUs but curs has the data. So it’s a match made in heaven. And I think a lot of it can be sort of explained by this. Also I think maybe this is sort of like where where the whole thing sort of gets weird. SpaceX is basically in a process of IPOing for which they have filed an S1 which is basically sort of the last statement that you’re making to the IRS I guess not the RS SEC you you’re making a statement to the SEC about sort of like things that people should know when they purchase those shares and obviously that would be a material event for them to uh acquire cursor so they don’t actually acquire cursor they do some sort of transaction where they basically buy they pay cursor for XM X mill X was what was it 10 billion for services with an option to acquire them for 60 billion and presumably if they go public in the meantime they’re going to raise enough money that they can basically buy cursor at a discount that’s more or less what this looks like it’s a great escape strategy for cursor that was my main thinking because last time we talked about like oh what’s going to happen to cursor it’s like oh now we figured out what’s going to happen for cursor money I don’t know it’s very simple button.

The world is so crazy right now and the valuations have gotten so crazy which is hat tip to cursor a couple years ago. This was a VS Code fork turned into a $60 billion acquisition. Look, if you’re young, that might just seem par for the course, but I remember, you know, when Flickr sold for like $45 million to Yahoo and the internet went crazy. Like that was a that was a high water mark for a while.

Look,

right.

I I just want to it’s going to be a shitty comparison, but the Ford Motor Company employs 169,000 employees.

Oh yeah.

Produces I don’t know how many cars a year is what 120 years old or something. I have no idea when I got founded. And they are worth 48 billion in market cap. I don’t know how many employees cursor has but I tell you they’re not building any cars.

And yeah, I don’t know. I’m not saying bubble, but it’s wild. It is kind of wild, right? This this seems a little bit wrong.

The Overton window is usually used to describe politics, but I think there’s an Overton window for like how you sensibly evaluate companies and uh that has shifted quite a bit, you know.

Yeah. I don’t know. I mean, look, I I’m definitely not calling it a bubble because I think like if you take the money overall, like that’s being spent on data centers and [ __ ] like that. To me, that actually makes a ton of sense because I do actually think we’re going to use all of this, but that the market individually finds that any one of those companies makes a ton of sense at those valuations that I have really hard time with.

You actually made this acquisition make sense to me earlier. Your your comment about they have the compute, they don’t necessarily have the data. And you also made the comment about X the everything app. Look, if you’re building the everything app and you need the everything AI also so much has been said about hey is like is code basically the the core ingredient here for AGI if you want to call it that but just sort of like you know agentic behavior that in in which you can go and do anything and it’s basically driven by code you know for if anthropic is worth this and Gemini and and open AI are worth these things and this is like your unlock to get there I actually think that you know is not I think that’s a maybe a r

rational bet that’s like how I look at it based on what you told me right

yeah it’s probably the most rational thing that xai has done in in a long time

you heard it here first

look the value of traces I think has been sufficiently understood by a lot of play players in that space coding traces are just great because they start with human input they contain a little bit more human input and they have a very easy to measure signal at the end did the user commit

right like it’s like it’s a mechanically verifiable uh reward function and you can build reinforcement learning environments around them. the the closest equivalent that Grock has and like Grock also doesn’t program but it’s like did you like this response and is like by number of tweets to it and basically the main signal they’re going to get is like boobies count better right it’s like that’s the signal that you get on Twitter it’s like if it produces content that people engage with there’s no programming in Gro you’re also highlighting something which is it’s making me sad which is ultimately like these big like the biggest sort of software businesses and I’m including social and that if you really reduce them to what they are is that they are basically human data aggregators for you know like just sucking down human behavioral data to apply to some other function. Do you see what I’m saying? If you’re not the the customer, you’re the what is it?

The product.

Yeah. You’re the product.

With cursor, you’re both.

Yeah. I mean like you’re literally you’re literally training the model in one from another. Like I think there are two ways of looking at this, right? Like one is is like okay, oh here’s this $60 billion company and you’re sort of like providing data to them for their models to get better and that feels wrong. I would actually argue there’s a better way of looking at this and I think Mario said this really well in a recent podcast that we did with GG as engineers and I think we also talked about this last time you you mentioned like that the cost of like tokens going up, the cost of laptops is going up. like actually being a programmer these days is is a is a real investment and I I think that the big risk here is that we actually come fully dependent on this stuff and now we are like we’re actually forced to dish out money to one or two large labs right and what actually stands in a way for open weight models to get really really good is that they don’t have the agent traces they they they suck a little bit in pre-training so there’s there’s there’s something lacking in the base models for like long long content length adherence. So once that is fixed, what the other thing is going to be really good reinforcement learning on aic traces. So there should actually be an incentive of us programmers sharing our traces to open weight model providers uh as opposed to just sharing them to entropic right now. Um and um and I think like it has hasn’t fully like that kind of thinking hasn’t fully um happened yet. Um Mario started sharing the traces that he’s doing on pi on hugging face. Um he wrote a tool that helps others share the stuff. Uh because like we need more open weight models to have the same kind of data to work with that currently just the large companies have.

It feels very similar to so many things that we’ve already seen like Yelp reviews, right? I contribute reviews, I get some value back in the sense that other people have reviews and I I receive, you know, uh there’s something to that. You could argue GitHub is a similar thing, open source, right? I I get value out of this. I contribute to it and the whole thing kind of collectively gets value. There are products that collect traces. There’s traces.com is one of them. I don’t yet understand the value of I’m going to share my trace and I’m going to go and look at other people’s traces. like as a human I’m not that interested in that. That’s like it’s like that is so much work and time and I don’t know what I’m getting out of that. It’s also it’s kind of like you know you can record all the meeting notes in the world. I I I prefer if you just summarized if you just told me what happened really quickly. So I feel like there’s an effort to sort of like get people engaged in capturing and sharing traces but I don’t know the value for me.

Yeah. So first of all, it’s also it’s a gamble right now, right? Because like so there’s no point. Let’s put this way. There is a point, but there’s there’s very little point in each one individually of us sharing our traces on hugging face right now if there’s nobody training on those traces, right? So like it’s in some sense a chicken and egg situation but like it is there has to also be awareness on a model lab side that this is a really important thing to do and the this whole move towards more agentic coding based agents and so forth that is very obvious now but even like nine months ago I think a lot of labs just didn’t really have like entropic and codex openi probably already knew and the Chinese already knew a little bit but But that hasn’t extended to every body. It also hasn’t extended to the people running the GPUs yet. That there will be more aentic everything. If someone would have started sharing traces like six to nine months ago, who would have given who would have cared? There’s a little bit of awareness. I know that we talked to folks at Hugging Face. We talked to folks at Nvidia um and a bunch of other companies who are like, “Oh yeah, this like having traces is is a is is interesting and and there will be demand for it.” But that also took a while to to sort of like become aware. And so like and that’s and so now maybe there are going to be some people training on it. But now you have to sort of like actually convince people that they should share some data which might have stuff in there that they don’t want to share, right? Obviously like cloud code by default is sort of has some sharing going on. You sort of have have to opt out of it. So like they cross this bridge already. But imagine something like open code or pi wanting to do that. like we would never do that by default and so then we will have to convince people that this is actually good for you but then you also have to also have critical mass. So like there’s there’s like this awareness building has to start somewhere. Um and then there has to be other people on the other side that are building open weight models because like if if they’re just going to build another rock from it then then you don’t feel good about this either, right?

Should everyone be more mindful of their traces? Like I’ve been burying deep therapy sessions with Claude in like coding sessions so that no one would find it. Is that is that going to get out there? Should I be more careful?

Uh I don’t know. If you have the data sharing thing on with cloud codes then then your stuff goes to entropic, right?

They probably tricked me to say yes. So EULA man uh we got into some interesting stuff here. I’m trying to think about where we want to go. But we had some other topics which I don’t even know.

I think we should definitely talk about GitHub.

I I was thinking Yeah. How about you start this one off?

Okay. So, one other thing that has happened in the last actually a thing that has been going on for a really long time but has definitely exploded a little bit more in the last couple months is GitHub. GitHub exploded twice. As of right now, I still don’t see all my poll requests because their elastic search plus made a boo boo. Um, and uh, before that, I think GitHub actions had a pretty I I forgot which one it was.

Anyways, um, this took quite a dramatic turn earlier this week because Mitchell Hashimoto from Ghosti or Ghosti Y, never know how to pronounce this, made a heartfelt post that he is going to move the project off GitHub. He doesn’t know where yet, but he’s sufficiently upset with the state of affairs. And if you follow some GitHub people, they’re also unhappy, but they’re also super upset, or maybe not upset, but they are like they’re frustrated by just the instability that they’re having to deal with that is largely caused by a combination of just way more traffic coming from agents. And I I don’t think they will admit this as quite as much, but GitHub doesn’t have a ton of leadership right now. and they have some updown demands to do some data center migrations and Azure shenanigans and so like the whole thing sort of falls together to what is basically right now a pretty disappointing experience.

By the way, I’m going to connect this even earlier to the price of compute or disks. There’s a lot of people pushing code as like GitHub probably needs more discs than they than they did before, right? And from what I’ve heard from people familiar with infrastructure at GitHub, they actually and sometimes have challenges or they’re basically blocked from getting more resources in some of the data centers that they have. So part of the migration to Azure is also just because for reasons that I won’t share here. They actually cannot put more stuff into some of the centers that they have. No software that I’m working with frustrates me quite as much as GitHub. Like really really badly. in in part because of the availability and the other part because like um Pi as an example takes gets so many issue requests, pull requests every day and like the tools that GitHub actually has to deal with this [ __ ] is non-existent and I think the incentives are not really aligned to fix that and then the the the changes that they’re rolling out on their UI, they’re just horrible. So many problems with like I’m seeing issues that were already closed, they still show up. Anyways, the whole thing doesn’t make me very happy. Um, and for me, seeing Mitchell Hajimoto leave actually is really, really meaningful because when Sik moved, I wanted to write this blog post and I started this draft. I was like, yeah, maybe this will cause people to move. But I don’t think Sik moving would have done anything. But when someone like Mitchell writes about it, I think it will cause other people to also move. And then like when a bunch of large projects move then then like it starts obviously the former CEO of GitHub raised a ginormous funding seat round to build a GitHub competitor. Um so there will be financial incentives also that sort of go into the market to try to create some alternative to it. There’s another company called Tangled. Do you hear about this? They also raised money as well. Tangled.org.

So the another one is tangled.org which is from what I can tell some distributed.

So this was a listener question from Serge I think surgical which was where where is it going or like what are even the possibilities and maybe is is the answer just tangled

I think um so tangled is the one the other one is called n and something and not encore and entire it’s called entire entire entire.io I don’t know where it’s going to move. Uh there’s another comp there’s another project called uh code code. Um my experience with code has been pretty bad uh because it’s quite unreliable at least on European working hours. I don’t know where it’s going to move, man. I have no idea. Maybe people are going to host their own tracks again.

I almost wanted to go a little bit into the history here because I feel like this is one topic where you and I are both experts. You are an expert on many many more things than me. Look, You’re an expert on GitHub.

I’m an expert on GitHub. Mitchell Hashimoto flexed that he was GitHub user like 1,200. I’m GitHub user 2153, I think. And I know that you’re

under 10K. I don’t know exactly what my numbers

and quick reason the quick reason why I have such an early account is I was really into Ruby on Rails at the time when GitHub came out and I read Chris Wayne who is one of the co-founders of of GitHub. I was read I was a big fan of his blog. you know, I had it subscribed RSS. So, I’m like quite confident whenever this was announced, like I saw it pretty early. Uh there was no Twitter or Twitter was very nent, I think, at the time. So, pretty much this was all like blog posts and forum posts and stuff like that. Anyway, so that’s that’s kind of how I got on that. Where I want to go with this was so look, we’ve been on the platform for a very long time. And it’s crazy to think how old it is, right? When GitHub came out, I thought Source Forge was old as dirt. Okay,

which it was,

which it was. GitHub is way older than that right now.

Maybe maybe like what people don’t know is like on GitHub you go and you press a button they have a repository. Did you ever create a source for source force project?

No. No.

I tell you how that was. You went you filled out a form. You had to put your real name there for a start. You had to fill out the form. You signed some terms of service. Then you declared your license up front and you applied. And then within two weeks someone someone said like yes you can or yes you can’t or no you can’t. It was like it was a process that felt like creating a company

startup opportunity. Also another thing that we should talk about during this moment subversion. So like there’s been like many different uh source.

My first source force project was on CBS CSV how it’s called CBS.

I also I also use CVS not the drugstore. Is the C for concurrent? Is it concurrent version control?

Probably. I I know it was an extension to RCS.

I also used RCS. This is all very embarrassing. Anyways, the point was there had been different version control tools of of like what you think of Git today in terms of like committing like a lot of those concepts all exist, but what was emerging all of those tools CVS, SVN pretty much had a um centralized server that you were that you were like pushing and pulling code from. So, same concepts, right? and 2006 2007 2008 around this time there’s this like oh distributed source control and git was not the only choice and many people do know about mercurial which was sort of like number two and I I guess I like to do a little history here to say that you know the the environment that exists today kind of shook out from almost like a I don’t know battle

do you remember do you remember how it started

um what do you mean like Lionus or what are we talking here?

Yeah, like why? No, this even even prior to Linus like Linus was not the first to

to start this.

Oh, then then maybe not. Go ahead.

The whole the whole pre story on this was they used a system called BitKeeper for many years which allowed basically patches to be distributed via the Linux mailing list

and then um they had basically a very similar system to Git when it comes to being able to merge them together. And the company that ran BitKeeper gave the Linux developers free licenses and the rule was that nobody’s allowed to reverse engineer BitKeeper and then someone did

and the company based like no more licenses. I mean this is a simplified version of this and the first people that tried to do this there were a bunch of project one of which uh was darks one of which then was git one of which was material and there was another one

I’m upset you name dropped Darks before me by the I was so ready to drop that.

Um, yeah. So, there were a bunch of them. Um,

because I went to a like bar camp event where someone was like, “Hey, here’s Darks.” And it was like a full presentation on darks. And like I went home and played with it, you know? So, like, you know, I guess I like to share a little bit of like this whole a little bit of what you’re experiencing right now with Open Code and Pi and different agentic solutions and everyone being like pretty excited. You know, I actually think this was a kind of a little bit of a a exciting moment for source control. Is that fair? No, it was it was a very exciting moment. It was so cool. And and also like I think like

the what what was so interesting about it is like actually nobody ever liked subversion. I think like on a very fundamental level subversion like it’s it’s not that people hated subversion but like subversion a lot of stuff in it which was just actually not really good. Like you couldn’t merge. You had to be online to do something.

The conflicts were bad. You could you could get into a conflict and you’re like I think I have to wipe away everything.

Yeah. It it was it was like I remember like we I mean it was so stupid but like when you wanted to merge a release branch into it was like a multi-day experience like nobody commits right now.

We’re going to do some

the way that people build software kind of like adapted to how subversion worked like in terms of organizing releases and branches. I I don’t want to get too historical here other than I think it’s fascinating. This could be like another documentary. uh somebody should put this documentary together. But I I guess I want to I guess I’m doing a little bit of like where GitHub came from, right? Because I’d say at the end of the day, it came down to Mercurial and Git and Mercurial was certainly very popular. You like for example, I’m quite confident Facebook mccurial was used everywhere at Facebook maybe around 2009, right? So you had you had you know big sort of engineering teams and valuable companies that were like you know like teams you know like Xbox versus PlayStation like murial and and and git. Um I had to give a I had to give a present I gave a presentation once inside of my company at Freshbooks about how these stacked up and why we should maybe pursue one versus the other. And my presentation hinged on GitHub.

I was making the argument that it was very obvious that the open source community had had more or less chosen GitHub, but also not just chosen because there was no equivalent honestly like social software thing from Mercurial.

That was very wrong.

I’m not saying that it wasn’t one. I’m just saying.

So, one of the reasons why I didn’t have a lower GitHub ID was because I was boycotting it for a little while because I put my money as a Python developer on Muriel and Bitbucket. I even contributed a little patch to Bitbucket at one point um to do the div rendering.

This speaks to you heard me earlier. I was into Ruby on Rails and depending on what community you were a part of and I do know that like um was material written in Python? I know it supported Python extensions, right?

It was written in Python,

right? And I can tell you at Freshworks by the way, we had a huge Python contingent. That’s also why it became sort of a a little bit of a a challenging conversation. Yeah, people were kind of in the tank for one depending on your religion.

Yeah, for sure. like Bitbucket came out of the Django community. It was written in Django. Actually, I think if that might be like bad retelling of history, but um I am pretty sure the reason I joined GitHub at one point was just because I I couldn’t escape it anymore.

Today everything is on GitHub, but there was a period where slowly

everything was on Google code.

Yeah. You know, things were everywhere and kind of like almost like the dominoes fell and and it’s this took years. you know, you’d still see major projects maybe like 2014, 2015 that were like we’re going to GitHub. Does that make sense? Is that

So, first of all, like maybe what we didn’t tell about is like look, I I’m pretty sure the the listener to this podcast are old,

our age, so I don’t think we’re telling anything here to anyone who wasn’t there, but is a good retelling to remember that it was very common that everybody ran their own server.

So, I ran a server that had subversion on it. I shared it with Giok bundle um and a bunch of others. We called it puku because we wanted I don’t know we made a software collective like everybody did at the time uh so that we can have more projects on one subversion server with track and everything on it. Uh so that’s what we used there was also red mine there was a bunch of other projects too.

Yes. Yeah.

And like that was the that was pretty common and there was another project in Germany was called beios which was a clone of subversion of um sorry source force. If you ever had a source forge server, you got SSH access to run your own Pearl programs on there. So like a lot of projects just ran bug what was it called? Bugzilla on source force server source forge servers. So you basically you got the infrastructure from them but then you had to run your own buck tracker on it and you had to install your own security. It was kind of wild. And and Python ran a setup like this with their own issue system and like their own subversion until I’m pretty sure like 2015 or something. I I’m pretty sure they didn’t move that long ago. Yeah. 9 years ago. I

think this is tracking with some of my dates. Like there were still major projects moving, right? And I I don’t make this totally historical, but I do think that the history is relevant to how we got here, which is over time

ultimately,

you know, I I said that the case for Git, a distributed version control system, was GitHub, a centralized commercial service.

And today, you know, I I guess even Armen saying, you go back 10 years ago, it wasn’t quite so dominant. even though it was huge, it’s not quite so dominant certainly as it is now or a few years ago where just like everything is there and it’s become like the it it turned into the baseline expectation that your stuff is there right if you write open source it’s there frankly if you there’s not a lot of businesses I would say even you know with private source code who are off of GitHub there

yeah that’s like a big asterisk

yeah GitLab is big I I right but if you’re like like a SAS company that’s just starting up or whatever like a new startup up. You kind of go there. Um, and I should also mention this, which is both Discuss and Sentry use another tool, Fabricator, an open- source another open-source um source um source control hosting. Uh it it hosted Git, right? I’m pretty sure

it had a program called I think if I remember correctly, it was called ARC, which basically allowed you to send patch sets up. So it sort of piggybacked the top of Git, but it wasn’t quite using git, but it had sort of a pull request system in it that was significantly better than GitHubs. But yeah, like I think there’s another thing. So sorry, very briefly, there’s another thing I think that sort of I remember and that is really not that long. When I started working, maybe it was the gaming company I worked for or it was the company before, but in any case, like I remember I showed up on my first workday and it was a self-hosted material installation. I was like, “Can we just use GitHub?” And they were like,

“No, nobody uses GitHub. You don’t put proprietary software code on GitHub.”

That was like that was like I was like, “No, this is just for like open source nerds.”

That is true. I that is worth acknowledging that there was a long time where you know your the open source stuff was on GitHub and you and I you know had stuff on GitHub but when it came to your actual work you had it somewhere else in like a more secure location and then

and or GitHub enterprise so they had like a behind the firewall version if you had that you were reasonably confident and then and then eventually you know cloud so just to say that GitHub you know GitHub has become dominant and not even too long ago honestly people had other platforms that they could reasonably use. David might yell at me. I think we got off Fabricator. It could have been 2017. It could have been 2016. You know, it’s not even about the open source or the private source or anything. It has just become such a platform. This summer when we started modem, I was very in like I wanted to try new products. Like I I wanted to try new stuff. I wanted to try new things being built by people using AI. And um I knew my friends were uh building Pierre Pierre.com, Pierre Computer Company, and they had a lot of the ex GitHub team on there. Um you know, uh including MGO, who I’m pretty sure was GitHub’s creative director for years, right? So I felt really good about the team behind that. And then I got I asked to get into the beta and we were building modem for I want to say May through

August September. Okay. Pushing a lot of codes through this platform. And I think if you go to pure.com I’m not even sure if you could find it right now. Is it is it live? Jacob Thornton the CEO might yell at me. He’s like it’s totally live Ben. Everyone should sign up today. I want to tell you that I actually enjoy this platform and they were exploring with like new ways of work, new ways of like doing code review and um they had a really nice CI runner where instead of YAML for GitHub actions they use TypeScript which was really great because you had like typed you know types workflows that didn’t fail all the time. However, the problem was this. As all these AI solutions were coming out and they wanted to interrupt with software, that software was naturally GitHub. I barely even GitLab, frankly. It’s mostly GitHub. So,

you know, catch 22. I want to use these new products, but if I wanted an AI code reviewer, well, they only interop interacted with GitHub. And so many products do this, right? I had to leave the platform even though I thought it was good because just the the AI ecosystem was really gravitating towards GitHub in this like almost penultimate way in the last year, right? I’m just going to repeat this. A year ago, I rationally thought I could just take my business and I could write some software and I could put it on another provider and it would be fine. I didn’t think that was true. you know, come August,

I felt similar in the sense that like there there was no question I would obviously put my code on GitHub like why wouldn’t I? also in part because like I really felt like there’s going to be maybe not all of it of Aron’s code is going to be open source but I think a lot of it is going to be open source and where else is going to be like I don’t want to be a server hosting company and so yeah like there was no good reason not to at the same time I actually felt like throughout the entire last year there were too many

um there were already too many signs that people are going to explore alternatives to GitHub Um I think you already mentioned Pierre is a is a company that I clearly wants to sort of explore an alternative to GitHub. Um want to just paging on top of it and I I know that venture funds were already looking for like now is the time to rebuild a whole bunch of stuff. So there are VCs looking to fund GitHub alternatives because they also feel like there’s going to be a space for like human and Asian collaboration and and maybe both Git and GitHub are not the solution for it. So there’s movement in that space um for sure. Now

I want to touch on something really quickly. I I thought you were going here. I don’t want to mention the dinosaur company again. If you want an AI code review tool in July, man, I didn’t know how to build that. I’m not spending time building that, right? That that is a really dumb thing to do. So yeah, look, I’ll plug I’ll plug it into GitHub. Great. You know, and think about what the state of the Asian SDKs or everything were at that time, right? we didn’t have skills, right? But today, and also, how about this? You know, there was no PI, there was no um I don’t even know if the cloud SDK existed at the time. Could have.

Yeah, it was. I think

but I just didn’t want to spend the time on it. The models have gotten so much better. the tools have gotten so much better where I think the argument you’re trying to make here is while that platform probably really mattered in the sum, you know, a year ago might not matter so much now because the, you know, people can build new tools, you can build new tools. Is that kind of where you’re going?

Yeah, for sure. Like the this idea that we have to piggyback on top of GitHub, I I think everybody’s rejecting that increasingly. Like even the code review companies I think are increasingly rejecting that they have to sit on top of GitHub because like as an example like all of those companies including including the dinosaur company as far as I know have a thing that runs locally now. They go into different places and particularly also they want to go into any CI solution. Um they don’t just want to hang out in GitHub actions.

How about this? Also your agent can instrument these things more easily.

Right. A year ago I I I just want to click a button. But today uh I can ask the agent to figure it out for me. So

since I think GitHub is up for grabs, the question that I have in many ways is that if if open source moves away from being on one platform and being on like seven different platforms again, maybe a federation of platforms with Tangled, how is stuff going to change? And as an example, GitHub is is integrated into the Go compiler as far as I remember because like if you if you set dependencies on Go, you hardcode the location on GitHub into the dependency. Um I don’t know if there’s like special support for GitHub in particular, but like it’s um the the Go dependency system is very much dependent on like stable URLs that point to GitHub uh to Git repositories at the very least. So the models also completely understand the GH CLI, right? So does that maybe just that CLI becomes kind of a generic API?

I don’t know. We’ll see.

Should we talk about side projects?

We have a I don’t know. We have a lot. We’ve got I guess we could if we’re doing the side projects, it’s like we’re closing with the side projects is so we’re closing with the predictions.

Okay. So much time side projects. What have we been building? All right. Tell me. Look, tell me tell me about your side project.

My side project over the last week has been building a computer game with my kids, with my boys. I’ve built games before with agents, but this time I was like, okay, it would be really cool to sort of like teach the kids just how to use an agent, but also figuring out how you would use an agent of kids and like maybe they can learn something from it. So, we obviously didn’t just build the game, we also built tools to build the game as you do. The two things that we built, one of which is public, the other one is just on my computer, is one is um a tool which is basically just a PI extension that helps you explain to a kid what’s going on. Experiences are mixed. Turns out kids don’t really care. They just want the game.

But uh the other one that we did was a extension to pi which is kind of funny because like you I think you build a very similar one but the one we called is called pi draw where if you hit a command uh command like a command was called hotkey in pi it opens up tl draw in the browser you can draw into it and then you can press a button to send the picture back to pi pretty straightforward. I have another one where I can take a picture on my phone and then send it to pi but that turned out to be really really helpful for this project. So that that has been my sidekick and I will link it in the show notes because it’s very addictive while we build.

I want to make a comment which is a lot of people are building on some interesting primitives right now. I’m going to highlight TLR. I’m seeing a lot of TLR projects. Pier diffs. I have a project on PR diffs. I’m seeing it pop up all over the place. Uh I was using conductor today and I hadn’t used it in a while and I saw that PR diffs are in there too. That’s you know if you want diffs PR diffs term draw. So I built this thing term draw and actually this has been a project that’s been going back maybe like six or seven weeks because I as I was experimenting with PI extensions for the very first time probably I struggle with agents when I’m like no just move it over here right no I why are you do put it over here and I I would frequently

you know instead of going back and forth 10 times I would actually draw out ASI like in the prompt input area Right. Literally space it out. Dash

uh you know I I know what characters help it the box look a little good. So and I’m like why am I doing this? So that that was that pi extension which I think I also called pi draw or just uh I should have taken the should have taken that npm name space. However, it had problems which is like it was just basically drawing it was drawing. Sorry it’s a it’s a drawing in asy on the terminal. So you can use it over SSH, you can use it over, you know, in ghost and it turn whatever. At some point I just had this idea which is like why can’t we have this like illustrator concept or like um you know more of like a Figma in the in the terminal and so I started experimenting with drawing boxes and turning those into objects and being able to drag them around. Um it has grouping. It has like you can resize things. Um, grouping meaning that you can like draw a box and put a box in that and you can drag it around. And I and I use this to communicate with the agent like, “Hey, I want the layout this way or I’d like I would like the, you know, the toast to be over here or or the clothes component over here.” I put out a video of this. People thought it was neat. A question I got a lot of was like, “Oh, does the can the agent read like these asky diagrams?”

Somewhat.

Oh, I think that they can.

No, I think they’re pretty good at it. But you know what I did? It’s going to be so stupid. I took pictures of the asy drawing and I was comparing to how good it did with the ask drawing.

I meant to do that. Yeah. Yeah. So, how’d it come up?

It does better on a picture, which is so stupid.

But how much tokens did that take?

I actually don’t know.

It’s a lot more. Yeah.

I think it would work slightly better if you would pick something else than braille characters is my guess.

Oh, that was an exper Yeah, that was an experiment. I did a lot of experimenting on like how I could make the smoothest lines and braille characters visually does that but there is there’s multiple line modes and yeah it’s I did experiment with like does the agent understand it? Yeah, but it also makes a comment of like these are braille characters.

So you you no longer have so should it update you no longer use bra characters.

Is that what I’m

uh maybe I’ll if if it’s not Yeah, maybe I’ll get rid of it. I I maybe got a little bit obsessed with sort of a esthetically pleasing uh things. But anyways, check it out. I do want to make a comment by the way because I’ve been doing so much kind of asy art and I was doing um my AIA engineer talk was actually done over the terminal. So I was using presenter term. I was I was going all terminal but I was generating asy art for that. Opus 47 actually makes significantly better asy diagrams than 46. This is just sort of a discovery.

It’s not scientific, but just the number like the the some of the characters it was bringing in. I’m bringing this up because I was I was literally in the act of generating a lot of kind of asky diagrams for my presentation and then 47 came out and I was still generating them and and so I had this kind of like backto-back comparison. Does that make sense?

Yeah.

Um so just a a a random discovery. I would there somebody has put together an asky bench but it’s out of date. I I’d love to, you know, if I had actual time about to do that again. Any closing predictions?

Um, I wanted to predict where we think the moat of some companies is going to be.

You got quiet there. The moat of some companies are going to be.

So, basically, I’m going to te my thinking up here. Um, you already mentioned TL Draw, right? As a sort of like an example of a company that exists. I think one of the reasons TL Draw is able to sort of monetize itself right now is like they have good data. I think a lot of people that sit on good data have found ways to monetize it in one form or another. Some of the conversations that I had over the last 3 months or so has been, “Oh, by the way, we are a big model training lab. We want to train blah blah blah. Where are the places where you can get it?” Um, I know of a meeting recording company that had a lot of transcripts that sold the transcripts because it’s good training data. But I’m going with this a little bit is like will we see more of it and or and actually the core question like will people get upset with all the data selling that’s going on right now to metal companies because I think it’s it’s gradually moving from oh we’re just like there’s just like public books and Wikipedia to usage data from all kinds of platform from the last couple of years. that’s sort of sort of open but not really open is now being traded like gold.

And I guess there’s like recent public examples of companies kind of like finding other ways to ask for the consent to add your data, right?

I I actually don’t know to which degree does this sort of like publicly understood at this point, but I but I know this I I know some transactions have already taken place.

The whole like this almost feels like the end consequence of we never solved EU EULA’s as a society. You see where I’m going with that? Which is just we just turned into

So that you opted into becoming the um what was it called? The human centipede because you have bought something on iTunes.

Yeah.

Go Google that folks. I’m not going to speak more on that. But just like we we became blind to license agreements, right? And so you know just click give me the thing you want.

I guess there’s a question like one of which is the license agreement. The other one is sort of like what’s societally acceptable. In a way, it feels like we got so used to it taking place right now that like nobody would care anymore. In between the last recording and this recording, Microsoft has started to train on GitHub data more than they did before. And it was like half a day of outrage.

It wasn’t more than that, right? Did you opt out of your data? Like did you go to your account settings and try to find the button where you opt out of training?

Probably not. So like I feel like two years, three years ago there would have been a big outrage, right? And now it’s sort of like, oh yeah, of course it’s happening.

I guess we should also comment if people aren’t like that familiar with us.

Probably not that familiar with me. But you know, one I’m pretty sure centuries is explicit optin consent for data. This was a big topic internally. That’s still been true.

No cookies.

You Yeah, no cookies. Look, you know, Sentry open source.

We have thoughts. We have feelings on this. I’d also add that a bunch of us worked for worked for Discuss, right?

David and I did and Chris.

I almost did.

Yeah, that’s a story for another time. You’re always you’re always in the W. You’re always in the way. You know, Discuss like we worked on that because we were nerds. We like like Discuss is embedded comments. Honestly, a bunch of us all had static websites and we liked having comments on our website because the internet was a better place at the time, you know, and it was fun to work on that. And as we I guess this is like even tying it to the theme. Well, Discuss sold to a data broker for I don’t know 70 90 million years ago. And that was really upsetting to me. That was really upsetting to me, you know, because this really does tie into what we talked about earlier because I kind of signed up to build a product, you know, that ostensibly kind of like gave value to people, gave value to me.

People, it had a ton of users. At the end of the day, the most material value of that product was the data. And so there it went, you know, and I I worked on that for four years. I just didn’t even see that kind of like outcome. And that’s what it was. And it kind of made me upset, you know, not it’s sort of like I’m not upset necessarily at the founders, right? Like they have a fiduciary duty, you know, like I I I understand that. It’s just sort of a I guess a cynical consequence of that.

Sorry, this is getting a little emotional, but um but at the time I think it mattered for a lot of us, right? It’s like and it I think it mattered for people just generally. I remember when like 2 3 years after the acquisition of discuss like like people slowly started figuring out like oh this this is sort of happening and I remember like when you had I think we also showed like ads at one point in the in the in the embedded widget and stuff.

Yeah, we did we were experimenting with how do you make this function? So look to answer your prediction I think that society has demonstrated that as long as they get enough value they really don’t they don’t they don’t care. AI is a great example of that.

I’m not sure. Yeah,

I’m not sure if it’s so simple.

Obviously, to some degree, you’re right. It’s like for as long as like we’re all getting something out of it, like seemingly we’re okay with a whole bunch of stuff.

Mhm.

But it’s not without people in the mix that are in sort of in a position of power also getting some major push back in one form or another.

Clearly, right now with the companies, for instance, like someone just threw a firebomb at Sam Alman’s house, right? So, so it’s like there’s there’s there’s a certain extreme kind of counter movement that’s going to happen if companies continue to be irresponsible, right? And GDPR happened largely as a result of of like companies doing a whole bunch of stupid stuff. And there were a lot of really engaged people lobbying for like consent dialogues and and the cookie law and stuff like this. They didn’t happen because like a bunch of people bureaucrats in Brussels were like you know what we really want we want consent dialogues like that that came out of like people being super upset with data leaks and stuff like this and they sort of

create this lobbing and I think like right now what’s happened with AI a bunch of really scary [ __ ] is going to happen because we already talked about security thingy but like I have already seen like some really bad things in the last 3 months of data being publicly accessible.

because some young founders vibe coded together some [ __ ] and then other people are just putting the data into it. The level of irresponsibility that’s going on right now seems on another level compared to what was there 15 years ago. I I don’t remember it being as bad.

Oh, I totally agree. Yeah, I totally agree. I think one thing that our experience lets us have is it lets us give a pulse check on like, well, how was this versus before? And yeah, I agree. A good example is like a company like Delve or like whatever happened there, you know, that didn’t really happen, you know, and u we went through so at century I worked on that years ago. Boy, that was painful. Incredibly painful. You know, some things should be painful. I guess is where I’m going with this and I think it probably connects back to your talk. AI sometimes makes things too easy and there’s value in the pain. And uh are we going to learn from that?

Well, I think we’re going but I don’t know. I I feel like I learned I’m an old man at this point. I feel like I learned my lessons, but I I think like a lot of these young founders are also going to learn this in some painful ways. And look, it’s not like the last cycle was so great. I mean, like Theronos was in there,

FDX was in there, right? Like people went to prison. Um so people will go to prison again. Um, presumably like eventually the damage that some of these companies are going to cause is just going to be too great. Um,

I you know what? I’m not going to give you the answer. I’m going to but I’m going to give you what I want. I want it to change. I want it to go the other way. All right. I’d like things to be Yeah. a little slower. And I would love that we didn’t have like when we click accept on something. I’d love that to be, you know, you know what, you get 2,000 charact something like that. I’d love for that. And if you’re training the data or whatever, like if you’re training on your data, like it’s very visible and apparent when you sign up that that’s gonna happen. Like I really do want that.

But right now, my prediction is we’re going to figure out that some b some a bunch of older companies are going to figure out that they are sitting on a treasure trove and they’re going to find out that there are some companies right now which are giving ridiculous sums of money for that treasure trove and they’re going to not be able to resist.

I’m trying to think of like what what is an industry we’re going to hear about like a rental car company just announces that they’re going to ann they’re going to give their like fleet fleet data to open AI and you’re like why would they do that and you’re like oh no

but the thing is like the rental car companies I it’s like Strava rental cars like okay like feels it feels like that sort of like okay maybe like I’m too cynical like this feels like okay

I was trying to think of something like uh that’s not it’s not a great example just trying to think of something boring that actually, you know, has some once you realize it goes into AI has some sinister like ramifications.

Actually, like geoccation data is actually pretty scary. I mean, do they remember this story for just recent where like the uh the French aircraft carrier got sort of like spotted because someone was running in circles in Straa, whatever it was, right? Like like if you if you had a data point that’s in the middle of nowhere, it’s revealing. Yeah.

I don’t I just don’t like people selling data for training with little disclosure and but the problem is people don’t like to sh if you ask people for permission they will say no and then you don’t have enough data right so that we already know that so the trick is like how do you cheat people into sharing the data so that they don’t know

maybe I want to close on this because when we’ve been talking for a while and I you know Armen I love talking to you and I could talk for hours and sometimes we do here’s what I I hope I really do hope principled

you know people get their day again. I hope that you know people that conduct business in a straightforward way that is like hey I’m going to give you value for this service and this product and I hope to solve your problems and there’s no shenanigans. Not to glaze David here but I do think David Kramer is one of these people. Eats me up inside. Never I hope you never I hope you never hear this. There are others and I I wish that they kind of had, you know, like I wish that that was rewarded more, you know, right now it doesn’t feel like there’s enough penalty for doing the bad thing.

And so I guess if there’s a prediction I like I want it. I do want, you know, if you’re really cheating on the practices, whether that’s compliance or, you know, just like not doing fundamental software engineering, you know, stuff. I don’t mind that there’s a little punishment. I don’t mind there’s a little blowback because I guess it’s like for us who, you know, I don’t know, I feel I practice those things like I I want I want that work. I want the slow, painful, hard work to be rewarded, not not the [ __ ]

Yeah. I I hope the same and I think like I like this is me pitching Arendel a little bit here but like I really hope that we manage to find a way at least with PI but I also hope with Leos where you actually retain the data with you and we we we managed to create an environment where there is a choice of models and there is like a very explicit sort of data sharing thing going on because

there’s also I think there’s a a lot of opportunity in theory right now that we’re commoditizing the hell out of these models and everybody ends up being better.

I think in a long run open source software kind of tends to win because it’s just it it it it makes it too hard for other people to compete in the space and I hope that’s going to be true for LM as well. Like my my my ideal outcome in a couple of years is like copyrights are worth a lot less and we all have powerful LLMs to help with legendic tasks and we can run them where we are and not in in Sam Alman’s cloud.

Maybe a good just to bring it even back to GitHub. The fact that anybody can leave GitHub is because Git is open source. Well, should we close it there?

Yeah, this took longer than last time.

I hope it’s coherent in the end. We’ll find out.

All right.

All right. Thanks everybody.

📚 REFERENCES & SOURCES CITED

Key Companies and Technologies Mentioned

• Arendelle: Armin Ronacher’s AI product company, encompassing Leos (email agent) and Pi (coding agent)
• Pi: Open-source coding agent created by Mario Mars, now part of Arendelle
• Leos: Email-based AI agent designed for everyday (“normie”) users, built on Pi
• Modem: Ben Vinegar’s startup building product agent harnesses for product work
• Sentry: Software observability platform co-founded by both Armin and Ben
• xAI / Grok: Elon Musk’s AI company; discussed in context of the Cursor acquisition
• Cursor: AI-native code editor (VS Code fork), acquired by SpaceX/X in a ~$60B deal
• OpenClaw: CLI-based coding agent (formerly Clawdbot), maintained by Peter Steinberger
• Claude Code: Anthropic’s agentic coding tool, discussed in context of pricing changes
• GitHub: Central code hosting platform; discussed regarding outages, migration trends, and data consent
• Pierre / Pierre Computer Company: GitHub alternative founded by ex-GitHub team members
• Tangled (tangled.org): Distributed source control platform, positioned as a GitHub alternative
• Fabricator: Open-source Git hosting tool used by Sentry and Discuss

Economic and Industry Concepts

• Token economics / Tokconomics: The economics of AI token generation, consumption, and pricing
• End of subsidies: The gradual withdrawal of artificially cheap AI compute and API pricing
• Data moats: Proprietary datasets as competitive advantages in AI development
• Training-set deals: Commercial arrangements for using user data as AI training data
• Agentic engineering: The practice of using AI agents as collaborative coding partners
• Prompt caching: Storing previous AI responses to reduce redundant computation costs

Industry Events

• AI Engineer Europe: Major AI engineering conference held in Europe
• AI Engineer Miami: Companion AI engineering conference in Miami
• World’s Fair (SF): Annual San Francisco tech/AI conference

⚠️ QUALITY & TRUSTWORTHINESS NOTES

• Accuracy Check: Both speakers are primary sources for their own companies and experiences. Claims about GitHub outages, token pricing changes, and acquisition details are publicly verifiable. Specific numbers and timelines should be treated as approximate given the conversational nature of the podcast.
• Bias Assessment: Both Armin and Ben are active builders in the AI coding space with direct commercial interests (Arendelle, Modem). However, they are notably transparent about industry problems and don’t shy away from criticizing their own tools and the broader ecosystem. Their concerns about economics and sustainability are well-founded and supported by observable market trends.
• Source Credibility: Exceptional. Both speakers have decades of track records in developer tools (Flask, Sentry, PSPDFKit ecosystem). They provide firsthand operational details rather than pundit-level speculation. The podcast format allows for unscripted, honest exchange.
• Transparency: High. Both guests share specific financial details, name competitors, and admit uncertainty. The conversation includes self-corrections and acknowledgments of what they don’t know.
• Potential Harm: Low. The discussion is analytical and constructive. The main risk is that the “end of subsidies” framing might cause unnecessary panic among smaller AI companies, though the guests are measured in their assessment. No harmful misinformation is present.

🎯 AUDIENCE & RECOMMENDATION

Who Should Watch:

• AI Engineers and Developers: Essential for anyone building with or investing in agentic coding tools. The economic analysis of compute costs and token pricing is directly actionable.
• Startup Founders and CTOs: Critical for understanding the financial sustainability challenges facing AI-native products and how to plan for the end of subsidized compute.
• Tech Investors: The discussion of data moats, acquisition dynamics (Cursor/X, Pi/Arendelle), and market economics provides valuable context for evaluating AI investments.
• Product Managers: The conversation about data consent, training-set deals, and ethical product design offers practical frameworks for responsible AI product development.
• Open Source Contributors and Maintainers: Armin’s discussion of running an open-source coding agent at scale is directly relevant to maintainers navigating the AI landscape.

Who Should Skip:

• Those seeking pure technical tutorials: This is an economics and strategy discussion, not a how-to coding session.
• Viewers looking for simple optimism or pessimism: The conversation is nuanced and sometimes uncomfortable, resisting easy narratives about AI’s future.
• People unfamiliar with basic AI coding concepts: The episode assumes familiarity with coding agents, LLM pricing models, and the general AI engineering landscape.

Optimal Viewing Strategy:

• Speed: 1.0x-1.25x speed. The conversation is information-dense but conversational.

• Sections to prioritize:

  • 0:10:04 - Rising cost of compute (hardware economics)
  • 0:17:05 - Security harnesses and slop vulnerabilities
  • 0:22:37 - Enterprise token spend crackdowns
  • 0:26:23 - End of subsidies analysis
  • 0:36:33 - Pi acquisition story
  • 0:45:29 - xAI/Cursor acquisition and trace economics
  • 1:25:24 - Data moats and consent

• Note-taking: Focus on the economic frameworks, specific pricing examples, and the “end of subsidies” thesis. These provide concrete mental models for evaluating AI investments.

• Follow-up: After watching, review your own AI tool subscriptions and compute costs. The episode’s thesis about subsidy withdrawal has direct implications for any team or individual using AI coding tools at scale.

Meta Notes: Review written based on full transcript. This is a conversation between two deeply experienced practitioners who have been building in the AI coding space since its earliest days. Their economic analysis is more grounded than most industry commentary, though specific financial figures should be verified independently given the conversational format.

Crepi il lupo! 🐺