51:54David Ondrej · Interview
$75M founder reveals his Agentic Engineering setup
10x's Alex Lieberman and Dan Zakon on why the markdown you write for your coding agents is starting to matter more than the code itself.
August 20thThe argument in one line.
Pi's coding agent won by betting minimal, extensible bash-based tooling beats feature-bloated harnesses, and durability, agent-native UI, and database access are the next fronts left to solve.
Read if. Skip if.
READ IF YOU ARE…
- You use or build coding agents (Claude Code, Codex, Pi, OpenCode) and want to understand why a minimal, bash-first harness is outperforming feature-heavy ones.
- You maintain or are considering open-sourcing a project and want an honest read on why people actually do it now versus why it lasts.
- You're building agent infrastructure and want a working list of genuinely unsolved problems: durability, agent-native UI, agent database access, portable session state.
- You're skeptical about AI coding ROI and want to hear that skepticism from someone running an AI tooling company, not just critics or hype accounts.
SKIP IF…
- You want a step-by-step tutorial on installing or configuring Pi Agent — this is a philosophy and systems conversation, not a how-to.
- You want quick AI news takes rather than a long, meandering two-person conversation that wanders into European politics and economics.
The full version, fast.
Armin Ronacher, creator of Flask and now steering Pi Agent at Arendelle, argues Pi's minimal bash-first harness won by betting early that models would get good at raw shell use rather than needing bespoke tool counts, a bet Claude Code and OpenCode 2 are now converging toward. He lays out what's still unsolved: non-portable server-side session compaction as vendor lock-in, agents that can't survive being paused, no agent experience that's genuinely web-native, and no safe way for agents to own their own data. He's skeptical enterprise AI-coding spend is showing up as revenue rather than cost, and argues durable software still needs a human someone can blame, which is why plain-text formats keep winning over opaque automation.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Who's talking.
00:00guestArmin Ronacher
Where the time goes.
00:00 – 02:23
01 · Why Pi beats Codex and Claude Code
Armin argues models are simply getting good at raw computer use; Pi's minimal bash/read/write/edit toolset was ahead of the curve, and even Codex, despite showing elaborate file-discovery UI, is really just shelling out to ripgrep, piping commands together for context efficiency.
02:23 – 04:09
02 · Where the low-level trend goes next
Armin says he called this trend a year ago by asking what's easy to train into models via reinforcement learning. Interleaved system messages, mid-conversation system-level nudges enabling things like deferred tool loading, are a new inference-layer capability changing what harnesses can do.
04:09 – 06:30
03 · Sponsor: PostHog
David reads the ad for PostHog, product analytics and session replay used by YC, Supabase, Fireworks AI, ElevenLabs, and Railway; Supabase reportedly used PostHog to spot a wave of AI-app-builder signups and pursue partnerships off that data. Whisper's transcript renders the brand as "PostHawk" throughout; it's PostHog per the on-screen graphics and the video description.
06:30 – 10:20
04 · Why Pi took off: extensibility over bloat
Pi got popular around Christmas for being a minimal agent you could make your own, right as competitors added massive tool counts. Armin notes Claude Code has since shed tools from its peak, and OpenCode 2 rebuilt itself entirely around plugins.
10:20 – 13:50
05 · The real story behind Arendelle acquiring Pi
Arendelle was courting Mario Zechner, Pi's creator, since summer 2025; acquiring Pi was really about landing Mario, not the software. Arendelle looks like the Pi company for now, but Armin insists that's not the long-term identity.
13:50 – 18:00
06 · The mission: make AI work for everybody, not just programmers
Arendelle's goal is bridging what a programmer can extract from an LLM versus a non-technical user. Armin isn't confident it's even possible yet, but says a good agent experience has to eventually deliver value to non-programmers too, pointing to things like Claude Cowork as early attempts.
18:00 – 22:00
07 · Pushback: the technical keep pulling ahead
David pushes back that technical users are compounding their advantage with agents. Armin reframes with a DOS-vs-Unix analogy: today's coding-agent interfaces are a crude, power-user-only phase, and he doesn't expect people to still be driving raw coding agents in four years.
22:00 – 27:10
08 · What's actually still unsolved
Armin lists open problems: model-layer competition undercut by lock-in features like non-portable server-side session compaction; durability (suspending and resuming an agent without a human babysitting it); nobody has made a truly web-native agent experience; and giving agents real database access.
27:10 – 30:50
09 · Agents trapped in a chat transcript
Armin's example: an agent managing his smart home shouldn't just take text commands, it should render a durable custom UI of the house. He argues this isn't a model-creativity limitation, it's a systems-engineering problem (state management, component libraries, safe agent-writable databases).
30:50 – 35:50
10 · Why banks will never vibe code
Armin's accountability argument: once something goes wrong, someone has to be blamed, and humans won't accept the machine did it. That demand for blame forces introspectability, which is why plain-text formats like Markdown, JSON, and Unix pipes are winning.
35:50 – 40:20
11 · Armin's actual daily workflow
Despite running an agent-tooling company, Armin says he isn't running a software factory, Arendelle doesn't even have an internal auto-fixing bot. Most of his day is coding agents on his own machine, doing investigative work rather than pure code generation.
40:20 – 45:50
12 · Local over cloud, and the Postgres that never ran
Armin stays mostly local because his Mac is faster than his cloud options and moving data to the cloud has too much friction; he uses Tailscale and SSH to a desktop Linux box when needed. He tells a story of debugging with someone whose cloud dev environment never actually had Postgres running.
45:50 – 49:10
13 · The unsolved problem of ephemeral dev environments
Armin recalls onboarding Sentry engineers used to take two to three days to get a working dev environment, and a team of three engineers once worked full-time fixing that. Running several agents in parallel multiplies the problem: seven agents can mean seven databases and Kafka clusters.
49:10 – 53:20
14 · Is any of this actually paying off?
Armin is skeptical enterprise AI-coding rollouts show up in revenue versus a clear new cost line, like extra security-review spend. He thinks solo side-project builders are getting more real value than companies doing mass rollouts, evidenced mostly by commit counts rather than outcomes.
53:20 – 56:00
15 · No visible societal shift yet
Unlike the smartphone shift, Armin says daily life around him hasn't visibly changed from AI coding, but he expects it will, the way it's taking far longer to show up in outcomes than in cost and infrastructure spend.
56:00 – 59:20
16 · Why side projects and jokes become big software
Using Flask, Ghostty, Django, curl, and Sentry as examples, Armin argues successful open source isn't about being cutting-edge at launch, it's genuine curiosity plus years of sustained, unglamorous commits until reliability compounds into trust.
59:20 – 1:02:40
17 · Open source's new incentive: free CI, not ideology
Armin says a wave of AI-era projects go open source mainly for free GitHub Actions infrastructure and marketing, citing Daytona and Cal.com as projects that later walked back open-source status. He flags staying in model training data as a strategic incentive, and argues a project's quality can only be judged in hindsight.
1:02:40 – 1:04:20
18 · Inference eating the software budget
Armin's own AI spend keeps climbing rather than falling, and he says he'd stop using AI heavily if he had to pay raw API prices instead of a subscription; the future split between human labor and inference spend depends on where token costs land next.
1:04:20 – 1:05:29
19 · Can Europe catch up?
Armin, who is Austrian, says Europe is genuinely falling behind: 27 separate legal systems, labor rules, and VAT regimes fighting each other rather than one market, with a culture oriented around preserving the past rather than enabling the future, and capital increasingly flowing to US companies.
Lines worth screenshotting.
- Codex's file-discovery UI is largely theater: under the hood it mostly just shells out to ripgrep, the same bash-first approach Pi was built around from the start.
- Pi's early bet was that it's easier to train reinforcement learning data for a model driving raw bash than for a harness with hundreds of bespoke tools.
- Interleaved system messages, which let a harness inject instructions mid-conversation instead of only at the start, are a new model capability enabling patterns like deferred tool loading.
- Pi got popular around Christmas for being an agent minimal enough that users could extend and customize it themselves, right as competitors were adding massive tool counts.
- Claude Code has started shedding tools from its peak usage, and OpenCode 2 rebuilt itself entirely around a plugin system, both converging on the extensibility bet Pi made early.
- Arendelle's acquisition of Pi was really about hiring Mario Zechner, who had been courted since summer 2025 — the software was secondary to landing the person.
- A cloud coding-agent environment silently never had Postgres running, so every database test in that project had been mocked out without anyone noticing.
- Running multiple coding agents in parallel multiplies infrastructure needs directly: seven simultaneous agents can require seven separate databases and seven Kafka clusters.
- Agents are trained disproportionately well on Linux because so much of the public internet documents it, making a Linux box more reliably agent-controllable than Windows, which barely appears in the training distribution.
- Non-portable server-side session compaction on some closed-weight models is a lock-in mechanism, since an agent's compacted history can't be exported to a competing tool.
- Because humans need someone to blame when software fails, plain-text formats like Markdown, JSON, and Unix pipes are winning as the shared interface between humans and agents.
- Enterprise AI-coding rollouts show up clearly as new cost, like added security-review spend, well before they show up as measurable revenue gains.
- A new incentive for open-sourcing AI-era projects has nothing to do with philosophy: open source repos get free CI/CD infrastructure that closed-source repos have to pay for.
- Whether an open source project is good can only be judged in retrospect, by whether it's still maintained a decade or more later, not by how it looked in its first release.
Minimalism, accountability, and unsolved agent infrastructure.
WHAT TO LEARN
Pi's bet that raw bash beats bespoke tools is now industry consensus, but durability, agent-native UI, and honest cost accounting remain wide open problems.
01Why Pi beats Codex and Claude Code
- Codex's elaborate file-discovery UI is mostly theater: it largely shells out to ripgrep, the same bash-first approach Pi was built around from day one.
- Betting on what's easy to train into models via reinforcement learning, like raw shell use, was a better long-term bet than building hundreds of bespoke tools.
02Where the low-level trend goes next
- Interleaved system messages, a newer inference-layer capability, let a harness inject instructions mid-conversation instead of only at the start, enabling deferred tool loading.
- This capability doesn't change how you architect an agent much, but it quietly unlocks patterns that weren't previously possible.
04Why Pi took off: extensibility over bloat
- Pi's popularity came from being minimal enough to make your own, not from feature count, right as competitors were adding massive tool counts.
- Claude Code has since shed tools from its peak usage and OpenCode 2 rebuilt itself entirely around plugins, both converging on the extensibility bet Pi made early.
05The real story behind Arendelle acquiring Pi
- A high-profile acquisition can be less about the product and more about the person: Arendelle courted Pi's creator for months before the deal closed.
- A company can look like it's defined by one product while explicitly not wanting that identity to be permanent.
06The mission: make AI work for everybody, not just programmers
- The gap between what a technical power user and a non-technical user can extract from an LLM is real, and it isn't obviously closing on its own.
07Pushback: the technical keep pulling ahead
- Today's terminal-and-chat agent interfaces are compared to the early DOS era: functional for power users, but not the form the technology settles into long-term.
08What's actually still unsolved
- Real competition on model pricing is being undercut by proprietary lock-in features, like session compaction that can't be exported to a competing tool.
- Durability, the ability to pause and resume an agent without a human babysitting it, is still an unsolved problem even in mature coding harnesses.
- Nobody has yet built an agent experience that's genuinely native to the web rather than just a terminal wrapped in a browser tab.
- Giving agents real database access, so they can store and share their own data safely, is treated as a bigger open problem than model capability itself.
09Agents trapped in a chat transcript
- An agent that can only respond in a chat transcript is architecturally limited for tasks like home automation, where a durable visual interface would serve the user better.
- The obstacles to agents producing rich custom UI are ordinary systems problems, state management, component libraries, safe data access, not gaps in model creativity.
10Why banks will never vibe code
- Because someone has to be blamable when software fails, humans will keep demanding introspectable systems rather than trusting opaque automation.
- Plain-text formats like Markdown, JSON, and Unix pipelines are winning precisely because both humans and agents can read them without translation.
11Armin's actual daily workflow
- Even a company built around coding agents doesn't necessarily run an internal 'software factory' of unattended agents fixing issues on autopilot.
- A lot of real agent work is investigative, turning a bug report into a reproduction, rather than pure code generation from a blank slate.
12Local over cloud, and the Postgres that never ran
- A faster local machine can beat cloud convenience often enough that friction, not capability, is what's actually blocking cloud-first workflows.
- A cloud coding environment can silently fail to run a required service, like Postgres, meaning every database test in that project was mocked out without anyone noticing.
- That kind of silent failure is a strong argument for periodically verifying what your cloud dev environment is actually running, not just trusting green checkmarks.
13The unsolved problem of ephemeral dev environments
- Making a dev environment good enough to onboard new engineers in hours instead of days is still a multi-engineer, multi-week problem even at mature companies.
- Running several coding agents in parallel multiplies infrastructure needs directly: more agents can mean proportionally more databases and message queues to provision.
- It's worth asking, before scaling agent concurrency, whether your supporting infrastructure can actually scale with it.
14Is any of this actually paying off?
- Enterprise AI-coding spend shows up clearly as a new cost line, like added security-review spend, well before it shows up as measurable revenue.
- Solo builders doing side projects may be extracting more real value from AI coding tools right now than companies doing large-scale rollouts.
- A surge in commit volume is not by itself evidence that AI coding is creating proportional business value.
15No visible societal shift yet
- Big technology shifts can take far longer to show up in daily life and outcomes than they do in cost and infrastructure buildout.
16Why side projects and jokes become big software
- Successful open source projects are rarely cutting-edge at launch; they're built by people scratching a genuine itch and sustaining unglamorous work for years.
- A project doesn't need to be unique to succeed, it needs someone who keeps showing up, making commits, and fixing issues long after the initial excitement fades.
17Open source's new incentive: free CI, not ideology
- Newer projects choosing open source may be doing it mainly for free CI/CD infrastructure and marketing rather than open source philosophy.
- There's a real strategic incentive to keep code public specifically so it ends up in model training data, shaping which tools and stacks agents recommend by default.
- An open source project's quality can only be judged in retrospect, based on whether it's still maintained and trusted a decade or more later.
18Inference eating the software budget
- Personal AI inference spend can keep climbing even while per-token prices fall, because the tasks people ask for keep getting bigger.
- Even a founder running an AI tooling company says raw API pricing, versus a flat subscription, would change how much he personally uses AI.
19Can Europe catch up?
- A fragmented market, many separate legal systems, labor rules, and tax regimes rather than one unified one, is a structural drag on a region's ability to compete on new technology.
Terms worth knowing.
- Pi Agent
- A minimal, bash-first coding agent built around a small tool set (read, write, edit, bash), now developed by Arendelle after acquiring it from creator Mario Zechner.
- Harness
- The wrapper software around an LLM that gives it tools like file access and shell commands and decides how it calls them. Claude Code, Codex, and Pi are all harnesses around different models.
- Interleaved system messages
- A model capability that lets a harness inject system-level instructions mid-conversation instead of only once at the start, enabling patterns like deferred tool loading.
- Deferred tool loading
- Only exposing a tool's definition to the model when it's actually needed, instead of loading every available tool into context up front.
- Server-side session compaction
- Some closed-weight model providers compress an agent's conversation history on their own servers, producing a session state that can't be exported or resumed in a different tool.
- Durability (in agents)
- The ability to pause a running agent and resume it later exactly where it left off, without a human needing to babysit it the whole time.
- Software factory
- Running large numbers of coding agents in parallel, largely unattended, to work through a backlog of engineering issues automatically.
- Vibe coding
- Building software by iterating conversationally with an AI coding agent rather than writing and reviewing every line by hand.
Things they pointed at.
11:40linkArendelle
00:00toolFlask
57:00toolGhostty
58:00toolDjango
58:20toolcurl
57:20productSentry
1:00:00productDaytona
1:00:10productCal.com
10:00toolOpenCode 2
15:50productClaude Cowork
Lines you could clip.
03:30
“Codex, directly, if it says it's looking for files, it basically just calls RG to find files... it's getting quite creative in being relatively context efficient.”
11:45
“The question was less of Pi as it was like what we have to do to get Mario on board.”
19:40
“My mother would probably be more successful on using a coding agent for personal Linux customization on Linux than on Mac... Windows is just completely outside of the distribution.”
29:40
“None of those really are AI problems. They're just classic system architecture problems.”
31:40
“I don't think banks will ever vibe code. The moment something goes wrong, their customers will not be happy... you're not going to be able to blame a machine.”
44:10
“How do you run your database tests? He doesn't actually know... none of the database tests ever worked in his cloud environment.”
56:40
“If you force yourself to build something you don't believe in, it's not going to succeed. If you have an itch to scratch, even if it's just for fun... that's the real indicator.”
1:00:10
“Is that project going to be there in 10 or 15 years and still open source? If yes, that's a good open source project.”
1:03:20
“I would feel weird burning that much of my investor capital on AI... I would personally feel weird having such an expensive expenditure to inference within my company.”
1:04:50
“We have 27 armies, we have 27 legal systems... starting a new company you have to VAT in every fucking country. It's terrible.”
Where the conversation goes.
00:00 – 04:09densePi's minimalist bash-first harness design and where the field is headed
04:09 – 06:30sparseSponsor: PostHog
06:30 – 22:00densePi's rise, Arendelle's founding story, and its bigger mission
22:00 – 30:50denseOpen problems in agent infrastructure: durability, UI, memory, lock-in
30:50 – 35:50steadyTrust, accountability, and why plain-text formats win
35:50 – 49:10denseArmin's real-world local-first agent workflow and dev environments
49:10 – 56:00steadyDoes AI coding actually pay off? Token economics and ROI skepticism
56:00 – 1:02:40denseWhat makes open source software last
1:02:40 – 1:05:29steadyInference costs and Europe's competitiveness
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
metaphoranalogystory
This is Armin, the creator of Flask and the founder of Arendelle, the company currently in charge of PyAgent. In this podcast, we talk about why Py is beating CloudCode and Codex, his own custom agentic engineering setup, and how to build great software in the era of AI coding. This is the David Andre podcast.
Enjoy. All right, Armin, I want to start by talking about PyAgent and how it's winning a lot of the different benchmarks and comparisons between harnesses, despite being... one of the most minimal harnesses out there.
Why do you think that is? The models themselves are getting very good at using computers. And Pi just gives you basically bash.
For the most part, harnesses are getting really good at just doing the basics. And in a way, over time, it became more and more obvious. So at one point, Pi was sort of...
maybe at the forefront of doing this, but I think at this point this has become pretty standard. A good example of this is, I think, that if you are using Codex, even though it shows you at this point that it discovers a bunch of files and it found a bunch of files, it doesn't actually have many tools left anymore. Codex as a harness basically also relies on Bash for a lot of stuff.
So it directly, if it says it's looking for files, it basically just calls RG. to find files. It has been a very good approach in general to just rely on Bash.
And a good reason for this is that rather than pulling stuff into the context, it can pipeline stuff together. So it can do, find me some files, but in the same command also add some separation markers so that it knows this is program one to run, then it executes program two to run, executes. If you look into how it actually executes the Bash commands, you can see.
It's getting quite creative in being relatively context efficient, I would say. So where is that going to leave us? Like in the future, maybe say like six, 12 months, you think it'll just be like models going more and more low level?
Look, I think I felt reasonably confident last year to sort of say like, okay, this is probably where it's going to go. Because what I felt like... probably tells us a little bit in the future is like, where is the training data?
What is easy to train into the models in a reinforcement learning setup? So I felt some level of confidence that Pi's approach probably is a pretty good approach. Now that it has become like the center of where everybody's focusing on, like coding agents turned from, oh, this is one approach for AI to basically this is the one approach of AI.
And so now I think at least temporarily it has become a little bit less clear where it's going because the model labs themselves probably are going to compete on also getting other training data in too. This is sort of the part I think where it becomes maybe a little bit less clear. At the same time, I think all of these labs are now also competing on just being really good programming agents and they're not diverging that much from each other in terms of what the basic ingenuity is in tool calling.
But we're definitely seeing some changes in the capabilities of the model on the inference layer that are maybe indicative of some other approaches. I think the biggest change that I noticed is that, as an example, you can now have interleaved system messages in many of the state -of -the -art models, which allow you to do deferred tool loading, as an example.
It's not a huge change to how you would write an agent harness, but it is a meaningful enough change that certain patterns now become possible that previously were not possible. But I don't know if it's a huge difference. in what we would all be doing.
The most common mistake I see new founders make is not enough feedback from real users. Sure, with AI, you can build software much faster, but the faster you build, the more important real customer data becomes. After all, you need to know whether you're building the right thing.
That's why I want to tell you about today's sponsor, PostHawk. PostHawk shows you how people are actually using your product, like where they get stuck, which features they're using and whether they are coming back.
And with their session replay, you can even see what the users actually did. Y Combinator, Supabase, Fireworks AI, Eleven Labs, Railway, all of these companies use PostHawk. In fact, Supabase used PostHawk to spot a wave of new users coming from AI app builders.
And based on that data, pursued partnerships with those companies. As one marketing expert at Supabase put it, PostHawk has literally helped us get 10x more weekly new users than we did a year ago.
When you're shipping new features every day, your feedback loop needs to keep up. And that's exactly the problem PostHawk solves. So visit go .posthawk .com and learn how PostHawk can help you understand your users and build better products.
That's go .posthawk .com. The first link below the video. What do you think is like the reason?
why pi took off i mean obviously people like it for its customizability minimalism but it feels like the whole world is trying to build a harness right like we see literally every week multiple new harnesses being released and it's speeding up so like why do you guys think like you know relatively small team especially outside of silicon valley like why are you able to stay on the cutting edge is it just like thinking about it deeper is it you know taste how would you how to describe it Look, I don't think we have done a ton of going into ourselves trying to figure out why is Pi popular.
I'm not particularly preoccupied with keeping people to be Pi users of what we have today. I'm more interested in where should we be going and what it will look like then. I think if you ask me why did Pi become popular around Christmas, I think the first thing that Pi demonstrated people hooked onto was the extensibility part.
It was a very minimal agent, but it was one that you could make your own. And that was also taking place at the time when some of the other agents were turning the other way, where it basically turned into massive amounts of tools. Every single update from Cloud Code, I remember being a pretty meaningful impact to one's workflow.
And you can see now that Cloud is actually slowly moving the other way because Cloud has fewer tools now at peak of Cloud tool usage. And if you now look into OpenCode 2, for instance, it's all based on plugins. So I think this idea of self -extensible software, maybe Pi was rather early at this.
But yeah, I don't know. Honestly, at the end of the day, I don't necessarily know why things are popular. I think Pi struck a pretty good balance and that's as far as the software goes.
I think beyond that... And one thing that probably set Pi apart, and I think maybe to an extension sets Arendelle apart, is like this idea is like, okay, we have a pretty powerful machine here in the form of an LLM. How do we actually make it work for you?
Which is a positioning question, right? He's like, okay, we're sort of sitting in the middle of like, we think AI is really, really great. Simultaneously, we have a lot of skepticism about it.
And we're trying to navigate the balance between. this um and i think that there's probably a more european approach in some ways yeah so you started you switched i think in december right like fully to buy and then uh you like four months later you acquired it so like what was the like deciding factor that like hey i want to have it under aaron drill was it like mario was it the project taking off like how did you make that decision We were talking with Mario even like in the summer of 2025 if he wants to join the team.
So like in many ways I think like the challenge with hiring people in an early in a startup is often like how do you how do you meet like how do you how do you cross the paths. So in many ways I think that the question was less of pie as it was like what we have to do to get Mario on board. So that to me at least is like my version of the story I think Mario wrote.
quite a bunch about on his blog, how he joined Arendelle. Arendelle right now looks like it's the Pi company, but that's really not our goal. It might continue to be the Pi company for a little while longer, just because we have a bucket list of stuff we need to do.
But our future is unlikely going to be, we're going to be the Harness company. I think we just want to be a company that makes AI work for everybody. And we just happen to start at the Harness.
So what do you think that will look like going forward? How do you think about it? Is it you're talking to people on the lab?
Is it like you're deep testing yourself? How do you try to see where the world is headed?
That's a pretty loaded question. Yeah, it is.
Look, I think the very simple version of this is like someone puts money into a data center, data center peers, Some other company puts money into models, keeps training models and disappears. And then both of those are really great ingredients to hopefully drive a whole bunch of really interesting interactions that an LLM can do together with a human driver.
In many ways, we're incredibly early in this.
I don't think we have a particularly strong vision of this is exactly how it should unfold, but I don't think we would be satisfied if the experience that you get from an agent doesn't also deliver value to a non -programmer. The fact that this seems to be... We're moving this way with Cloud Cowork and things like that, but in many ways what you can do with an LLM is...
A programmer can do a lot more. than a regular user can. And we're trying to figure out how do you bridge that?
Is that even possible? I need to push back on that because people who are more technical, they can use more agents, they can figure out how to manage dozens, hundreds, thousands of agents, what models to use, what harnesses to use, which ones to run locally, which ones to run on the cloud. So it would be ideal to make everybody a power user of AI, but like...
what we're seeing is that the people who know how to use it are getting more and more out of it. You know, like the saying, the smarter, the smarter. Yeah, I mean, I think like at the, I'm not saying like everybody's going to be a programmer.
I don't think that that's either the way we're looking at this, nor is that really the goal. I think that, look, I happened to grow up with DOS. I don't know if you've ever used it.
It's basically the 3D version. I know what it is. It was basically like, like nowadays you would basically like, Funny enough, you look at a coding agent and they're very good at writing bash.
And Unix was really a great operating system from the start for a power user. DOS was just a really crappy text -based computer, personal computer, very optimized for resource utilization and so forth.
Nobody would use DOS if they have a choice to use something else. And it took a bunch of iterations until... A desktop computer turned into something that became really something everybody could use and wanted to use.
And I think we're a little bit there with the agents right now. The agents are basically not the user interface that really anyone should be using. I have a hard time imagining that we are all going to use coding agents in the current form four years from now.
That to me doesn't seem... particularly logical because on the one hand all the people that are token maxing are going to deploy this thing into like software factories where they're not going to spend a lot of time working with any agent individually and the people that are going to want to use this for a bunch of like domain specific work they're probably not going to hang out in like an agenting interface of today either right and there are a bunch of technical reasons why like What we have today is not going to be a sufficient solution for really good software of the future.
But someone is going to build it, and that to me seems very obvious, and we just want to take a shot at that as well. And at the very least, build a really good infrastructure that someone else can do that too. We would like to end up in a world where, from Aaron's perspective, we would like to end up in a world where people can use the stuff we build to build really good agents themselves.
Simultaneously, we want to... push forward the foundations of um all the bits and pieces you need for really good agents so that we can also build our own solutions on top of it that we can make available to non -engineers i see so apart from the harness what do you see as some of the biggest pieces that still need to be figured out i mean i think that a lot of them like one of one of which is obviously that i am It's not so much that a piece needs to be figured out, but the entire economics in that space have to move towards real competition on the models.
The good thing is I think that seems to be somewhat happening at the moment, so you have a choice of a whole bunch of models. But one thing that is maybe suboptimal right now is that a lot of the new features on some of the closed -weight models in the US, they're kind of locking into the ecosystem more and more.
Classic examples are the server -side compaction right now creates a non -portable session. That's suboptimal. One of the things that we found need -solving is just durability, that you can take an agent, suspend it, and continue where it left off.
We're doing rather well in many ways with coding harnesses today, but it's just not quite at the level where you can build durable systems that don't require human in the loop. And then...
I think nobody has quite managed so far to take an agent from a terminal user interface into the web and make it the most powerful experience that it can be. Most agents in a web look a little bit like just a web user interface to a terminal user agent. But there's a lot to be explored in that space.
And then obviously... the whole topic of like, we're not really strong believers in memory, but I think we are believers in like giving agents ways to manipulate data. And so the question is like, how do you make, how do you give agents really good database access so that they can store their own data and also make that available to a human?
So there's just a lot in that space and there will be more that we will find. But it's like, there's an almost like never ending list of technological challenge that everybody's working on right now.
It's just to make these agents actually deliver value to everybody and not just be a coding agent, I guess. Yeah, I think the interface problem is an interesting one because right now, basically, it's like terminal or some GUI. But I imagine...
It could be even something more like a video game where you have teams of agents, you have your resources, maybe your limits. I don't know. I don't even care about this as much.
I think on a very basic level, I think it is ridiculous that... I mean, it's great that OpenClaw and Hermes and all the systems are sort of demonstrating the value of agents, but many of the problems that I have are not to be solved with text only. and these agents are not really able to bring up custom UI in a durable way.
It's like if I want to use my agent, I don't know, to manage my home assistant solution, then surely my agent should not just be able to follow my commands but also visualize me in my household and all my smart home devices. And so the fact is these agents are, yeah, they can sort of give you a little bit of UI but not a lot.
That in itself I think is a problem that needs solving.
So it's the fact that an agent is kind of constrained within a chat transcript is one of the things that is not optimal. So is that an issue on the architectural level then? Because I feel like some things like the lack of taste, lack of creativity, it's within the transformer, right?
So is this even solvable or do you think we need new AI breakthroughs? No, I don't think you need new AI break. I think at the end of the day, a lot of the things right now that restrict what an agent can do is just like we moved from humans are writing code to agents are writing code.
And some of the things that are cheap for humans are expensive for agents and the other way around. And so you can notice, for instance, now that Linux has a resurgence, not just because of Omachi, but also because like... agents are really good at Linux because they found a lot of stuff on the internet about how Linux works.
So they're very, very good at remote controlling a Linux environment. And you can basically take an off -the -shelf Ubuntu or Arch Linux or NixOS installation and just fully remote control it and customize it with just an encoding agent. So that is a huge change.
I would make the argument that... Even nowadays, my mother would probably be more successful on using a coding agent for personal Linux customization on Linux than on Mac customization on Mac, just because the Mac is not trained as much, let alone Windows. Windows is just completely outside of the distribution.
And so there are certain things that previously very few people were very good at and used, even though they're technologically maybe really a good choice. I'm pretty sure we're going to see more people write Rust. I'm pretty sure we're going to see more people use NixOS.
I'm pretty sure we're going to see more people, I don't know, try really complex databases or work with Kafka or event streaming or stuff like this, just because that was always a better solution, but it was just very complex for humans to work on it. But for an agent, not so much. And so we're in this phase now where...
It's not really a question of do we need an AI breakthrough. I think we just need a breakthrough of figuring out which systems work really well with what we have. And so the question of how does an agent bring up UI is not a question of creativity or anything like this.
It's mostly just state management, component libraries, how do you build. databases that an agent can work with without destroying data like this none of those really are like ai problems they're like just classic system architecture problems i see so we almost like took the in the you know last 20 30 years like a detour optimizing everything for humans but now that's kind of crashing down because it doesn't matter and the agents are becoming the user Well, I mean, I think both of them are becoming users.
This idea that now there's only agents doing stuff, to me, seems a little bit questionable. But if you take it to the extreme, like if the models keep improving, harnesses keep improving, input -output is easier, isn't it going to be obvious that agents are going to be 99 .9 % of the users of any software?
I don't think so. Why is that?
Well, I mean, it probably depends a little bit on how you measure it.
On a very basic level, the number of humans in the world, if anything, is stagnating. Agents have not created more humans. So the addressable market of people that give a shit is the same.
Software doesn't matter if only an agent is using. At the end, there has to be a human. The GDP is driven by people buying stuff and not by...
like agent doing things for fun like that like an age like a computer system on its own doing something is just like burning energy for for nothing so so someone has to benefit and that has to be a human maybe you can extend it to animals or something but like so at the end there is going to be a human and like this idea that now we're all just and we're talking live devices to an agent every once in a while that doesn't seem very likely because like this is always going to be inferior than the person that is doing the same thing and puts a human in front of it like this because like if machines are driving everything then like the person that uses the machine that everybody else has plus a human is going to have an advantage.
Yeah, I agree with that. So there will always be software that has a human somewhere there.
Obviously you can make the argument that there will be a lot of software to support agents. That's probably true, but there has always been a lot of software to support more software to run. Every observability company ever, most of the code is not there for a human to look at, it's for processing all the data points.
And now, okay, maybe previously 70 % of queries that were sent to an observability platform were driven by humans and 30 % by automation. Maybe now it's going to be 30 % automation, 60 % AI, 10 % human, but that an observability platform would never give a human access to data to me doesn't make any sense at all because anyone who then still provides that for a human is going to be...
There's also this idea that if agents get really good at just writing assembly or a version of it, I don't think why that would ever be true. Because if you lose the ability as a human to understand what's going on, then you also don't have a ton of trust into the system. And the moment something goes wrong, someone's going to complain.
I made this argument recently. because the moment something goes wrong, their customers will not be happy. The accountability part, like you need to blame somebody.
Exactly. And you're not going to be able to blame a machine. That's not how society works, and that's not how society will ever work.
But isn't that just like you can blame the human that used the machine? Yeah, but once the idea sets in, like, okay, you have to blame the human that used the machine, then the human that used the machine is going to demand better introspectability. And if the human is left with like, okay, I'm just looking at binary shit right now, and it's just...
I don't know anything, that human will no longer use the machine. But I think that's mutually exclusive. If we have super intelligent agents with 200, 300 IQ models, surely they can cook up a custom user interface.
There might be a future very far out there, but we are so far away from this that I just don't think it's particularly important to think about this right now. In the foreseeable future, humans will want to understand what's going on, and so they're going to have the same interface, give or take, as an agent has, just because...
They want to look at what the agent did and understand it without having to ask another agent that might lie about it.
That's why Markdown is winning. I mean, it's ironic that all these really simple things like Markdown, JSON files, Unix pipelines are becoming the underpinning of everything that's going on. But in a way, it's not surprising because it's also what humans understand.
Walk us through your typical usage of agents. When you sit down in the morning, how do you work? How does their setup change?
What does it look like right now? I'm not particularly interesting, I think, as a use case because I don't run a software factory. It's not that I don't try to run one, but I haven't succeeded.
Arundel actually might be the only harness company right now that doesn't have a RoboPi or something like this that just automatically works on issues. I think we might want to go there, but so far a lot of our engineering is basically still the same, which is we still mostly use the coding agents on our machines. Very little actually sits in a cloud.
I have a desktop Linux that I run via SSH some things on if I want to keep them running. But most of it is still where the code is. We have a little bit of automation in GitHub Actions so that it can run Py there.
A lot of what I'm doing actually is not so much code generation. It's actually trying to figure out how stuff should work. So a lot of it at the moment in particular with the new harnesses, like investigating with the machine together.
Here's a report, how to get repo out of it. stuff like this. It's very traditional engineering for us still, I think.
Compared to, I think, what a bunch of people are doing, we're very conservative still. Definitely, I am very conservative still. I see.
So you mentioned that you still are heavily local, but we see a huge push, especially by the big companies like OpenAI, Cursor. towards cloud agents, do you think it's like valid argument to say, you know, soon enough, your machine won't be enough. You need to move everything to cloud.
Or is that like more of a, you know, ecosystem locking tactic where the moment people move all their environments and everything to the cloud is like very hard to migrate away from that. I don't think it matters, honestly, because like there will be a sufficient competition that you can do whatever locally or in the cloud.
I think the question is, what do I get from having it in the cloud versus what do I get from having it local? And so, for instance, one of the reasons why I don't really care that much about the cloud right now is because I have a really kick -ass Mac. And it's just way faster doing it locally.
I do have a pretty decent B -Link Linux machine, and it's fine. But it runs the tests at half the speed of what my Mac is doing.
There's this inconvenience still that certain things I want to have on my machine and then bringing that data into the cloud is just a little bit more annoying than it should be. In some ways I just have a little bit too much friction to do more of that and it's just a little bit too slow. One thing that I noticed, for instance, I had this conversation, this was already a month or two months ago, where I was helping someone debug a problem and they were using, I think it was Codex, the Cloud Codex version.
Or maybe it was Cloud Code, I forgot. But it was one of the desktop apps and he was programming with this. And I was like, how do you run your database tests?
He doesn't actually know how you run the database tests. And then we looked into it and actually none of the database tests ever worked in his cloud environment. Missing some environment variables or what?
No, because there was actually no Postgres running. The cloud environment didn't actually manage to set up Postgres ever. So it just mocked out everything.
And even nowadays, that continues to be a pretty significant issue. I found pretty good success in having a machine set up with a dev environment that's just SSH and 2. There are better solutions than SSH, but to SSH and 2 and just have my stuff running than trying to get any of those cloud solutions work for day -to -day engineering.
They will get there for sure, but I just have too much. that needs running. And I found it easier to just, here's a machine that I control.
Everything is on there.
We're obviously going to move to this.
Historically, the biggest reason why this didn't work was just because people did not spend a lot of time building ephemeral dev environments. I remember when we onboarded Sentry engineers in the first week, it took them two days, three days to have the dev environment going. We had an entire team at one point of three engineers that did nothing for a little while to get the dev setup working better.
But even then, you ended up in a situation where you had one dev environment. So if you want to run seven agents on your machine, then you need seven databases, seven Kafka clusters. There was complex enough software that it was just very, very hard to actually get all of that set up.
I don't know how the century engineers are doing this nowadays. I'm pretty sure it still would take you like seven minutes to bootstrap a dev environment for nothing, even with caches, which is pretty expensive for an agent to do still. That means like every single change now has this extra lead up time or maybe you have a pool of them, but it's also getting really complex.
I'm pretty sure people working this, I'm pretty sure that some people have this perfectly figured out for themselves. It just happens to not be yet there for me. Or at least I found something simpler.
that continues to work. But I do think that moving stuff into the cloud makes a lot of sense for a lot of reasons. And in particular, even just me having access to my Mac, I just use Tailscale for this.
It's pretty straightforward to do, but that still requires me to set this up. And if you look at, for instance, what AMP does with orbs or the codecs in cloud stuff. that is much easier than figuring out how to do tail scale and SSH and all of this.
So like from a UX perspective, my setup obviously sucks. But yeah, it will get better. Yeah, I mean, right now I'm the same.
Like most of my stuff is local. I'm just, you know, maybe I'm asking because I'm getting this FOMO of seeing, you know, all this talk about cloud agents and then like thinking from first principles, like if we have a hundred times more agents, you know, a thousand times more agents. like there's going to be a point like you know i have also like a really expensive mac and it's sometimes overheating you know if there's like multiple agents running multiple work trees and like they're all decided to run tests at the same time like you can see the fans start to spin up which is unusual on like a top spec macbook so i'm wondering like where is the limit you know like is the next generation of models or the next after that is going to be that good where like i actually physically cannot run all the agents on my computer.
I mean, I guess the question is like, why are we doing all of this? I think it's a pretty important question. I think that...
I think the simplest way to look is like, for how much money enterprise is now spending on AI coding? Did their... Yeah.
But particularly, did their revenue actually go up? I'm not questioning that there are a lot of AI startups and they're making a ton of money. I'm not questioning this.
But if you take a traditional enterprise and they're rolling out coding agent en masse, does it get reflected in the numbers or did it just get more expensive? Because we now have a line item that we all pay, which is basically... these models are getting so good at security research that we also have to pay money now to permanently check our code bases against security issues because someone else is doing it against us otherwise, right?
So we have primarily made it more expensive right now. But on the other hand, I think a lot of engineers are finding on their free time a lot of motivation to build side project of a side project of a side project. And that also is in many ways positive for GDP, even if...
it's not necessarily being used that much. But like all of these sort of side experiments are not useless. And so it's like, who is benefiting of this right now?
Actually, I think like it is people with the side projects are having way more success than companies actually rolling this out. And I think this, the fact that it happens, like this is sort of my biased view of this. Like some people might actually say like, hey, we rolled this on the company and actually here's some numbers that really demonstrate that this works.
But for how much more expensive the workforce has become, I don't yet have these great examples of... Our token spend went up 500%, but maybe the number that we have moved in the company is like... Maybe there is some number that we can demonstrate.
X amount of support tickets closed or something, but it's very, very hard for a lot of people to see anything other than a number of commits. Do we even need that much concurrency and all of it? We are producing a lot more code.
That's true. It's very noticeable. GitHub is struggling with staying up because so many commits are going up.
In the morning, I'm taking a train, I'm taking a bus, I'm cycling somewhere. All the stores around me behave more or less the same as they did three years ago. I haven't yet seen this huge change in society.
On the other hand, the last shift before that was mobile phones. Whenever I go to a concert now, 100 % of people standing in front of the stage hold up their mobile phone. That has permeated society aggressively.
I think ChatGPT has also, you see ChatGPT everywhere. But yeah, I don't know. I haven't yet seen that impact.
It will come for sure, but I think so far we're always trying to figure out.
It is exciting times. It is absolutely exciting. It's fun to explore all of this, but at the same time, it's not very hard to be at least a little bit skeptical about the impact of it because so far, it seemingly takes a lot of time to show up in results, but it doesn't take nearly as much time of showing up in terms of cost, infrastructure build -out, all of it.
Yeah, I mean, this is the classic response, you know, when somebody was on Twitter flexing that they are burning more tokens than... I think somebody from OpenCode is like, why aren't you more successful then? You know, it's like there's so many people running their 10 software factories and like, you know, they're not producing anything.
And actually, your point was a great lead up to my next question, which is like, why is so many of these, you know, successful projects start as like either a joke, either a side project, either something that... that people don't think could be successful, and they just kind of put it out there, and then it turns out to be, like, capturing interest.
Why is this a pattern? I mean, your story with Flask is a great example, but, like, I want to understand what's the magic there, where, like, the creator thinks, like, ah, other people won't like this, you know, whatever, and then it turns out to be the thing that grows. Yeah, I mean, I think to some degree it is maybe not so much that it is the joke as it is genuine curiosity.
If you force yourself to build something you don't believe in, it's not going to succeed. If you have an itch to scratch, even if it's just for fun. Most systems basically are just because people keep doing stuff.
You show up in the morning, you do something, you keep doing that. That's the real indicator. Flask was...
maybe a joke initially, but after that it turned into just work. You just had to do a lot of commit, had to work with the issues that came up and eventually make it work and keep doing that for a long enough period of time until eventually... It's not an overnight success.
You have to keep doing that. Even, I don't know, everything goes so much faster, but if you look at how many commits Mario put into Pi early on, that was not a tremendous surprise that Pi It wasn't just that it was unique.
It worked really well for proportionate to everything else out there on the market at the time. I don't think that's a big surprise. You put time into something and obviously you can't put time into something that nobody cares about.
I think that is true. A lot of my open source projects are like... this is my thing because nobody else cares and then you can sort of try to do something with marketing to see like is there enough people that care about it but like there have to be enough people that stick around and give you all the problems that you can see like what the real issues are and then you go from there so I guess it's being like very specific very passionate about a single thing like specific thing being on the cutting edge of that and that giving you like the unique insights like this could be better, that could be different.
I don't know if you have to be at the cutting edge. I think a classic example of something recently is Mitchell built Ghostity. I never know how to pronounce it.
Really ghosty.
Ghosty now might be cutting edge in many ways because it also does so much. But for a really long time, Ghosty was... Look, I don't want to say anything wrong here because it was always fast, but it had something to catch up to.
I remember when I used it very early on, there was a lot of stuff missing. But it wasn't good enough compared to some other things that was really interesting and it was very noticeable. It puts a lot of energy into it.
So it felt like if I keep using that, eventually it's going to be great. But it was not cutting edge in the sense that there was so much stuff missing. that the cutting edge was kiddy or the cutting edge was a bunch of other stuff, right?
So I don't think the cutting edge part matters. It's just like there has to be something. One unique insight.
Yeah, one unique insight, one unique take, something new.
It doesn't necessarily even have to be new. I'm trying to like... It just has to be good.
I'm trying to explain it for people because a lot of people, they want to build something and they start off by building something that nobody wants. I think there has to be something that gets people excited about it. The idea of why am I building this?
Ghosty was built for... I want it to be fast. I remember I watched him very early on when he started doing the dev blogs.
I remember, okay, he wants to make it native to the platform and fast. It's like, okay, there's two things that... There has been native or fast, but not both.
It's like, okay, cool. If he's going to achieve that, there's going to be enough of a reason for me to use it. But in that sense, it was a good story.
There was a lot of energy behind it.
But for instance, Django is a web framework. I still adore it to this day because primarily what Django really did was, yeah, it was a little bit new in that it has the admin and everything, but primarily it was... People kept building it and maintaining it and like putting energy into it.
And it felt like you can actually fully rely on this thing because it's going to be there a decade later. And it is still there like two decades later. And so in some way, or curl is a perfect example.
Like everybody uses curl. Curl is not unique or like... cutting edge, it is just very reliable.
And it makes it a very good choice. And maybe that was not obvious in year one or year two of that project, but it's very obvious nowadays. So some things are just literally because people keep putting energy into it.
Century is the same thing. Was Century the best choice when we started? I would argue it was.
But there was a lot of competition on the market for this. But many of them just didn't have the longevity. continue putting energy into it so like eventually only century remained um so like some projects that's just literally this is like keep doing it um but then it's like obviously not an overnight success and it's it's going to take you years so if you're like after what can i build that i can sell to like another bag holder after six months then you you work differently but that has never been my my motivation yeah that's very helpful Let's talk about, I guess, open source because I think there's a trend of more and more of the projects that are being released and that are taking off are open source.
Do you think it is because people trust it more if it's open and they can build on top of it and they can look inside and they can fork it easily? Do you see the percentage of software in the world being more open or do you see it constant? How do you think about this?
Are you noticing any interesting trends? in this direction? I don't know, but here's a bunch of arguments that I think why...
So first of all, I think the general move is not necessarily there is more open source. A very classic move, very annoying move, has been you start at open source and then you close it. I mean, this has been the move forever.
In the AI space, incredibly common. Like Daytona started out open source and then found an argument why they shouldn't be. Even Cal, which was a company that...
or cal .com or whatever. It was a company that said, oh, we are great in open source. Even them are no longer open source.
So I think the careful thing is who builds an open source project and continues being an open source project is a very different thing from who started an open source project. One reason I think you're going to see a lot of AI code being open source is actually a completely unrelated reason to if they wanted open source or not.
It's because if you're an open source project, you're going to get free infrastructure. So the cost difference of having GitHub Actions on open source versus closed source is a lot of dollars.
And presumably the cost of someone cloning someone else's code is very little. So if you want to have free infrastructure, you build it open source. That is the status quo right now.
How long Microsoft wants to continue doing that is a different question. But obviously there's a huge monetary advantage right now to start at open source. It's marketing and you get free infrastructure.
I'm just afraid that longer term it destroys open source a little bit, or at least it's going to make some, it will have some negative impact on how open source works because a lot of the people that are now open sourcing software have no idea about licenses, what it means to be an open source project, how to actually make finances work for an open source project.
For them, it's mostly just a marketing channel. And as someone who has been working open source for a long time, that kind of upsets me a little bit. But maybe that's just a temporary thing that will play itself out in one form or another.
I was asking from the end point. Again, I'm not saying it's going to be a year or two. I'm saying, like you said, if cost of replicating software is cheaper and cheaper, doesn't that track as...
you know agents get better models get better people will be like able to replicate more and i'm not making the argument that like everybody will you know replicate their own slack or their own crm like i'm not trying to make that argument i'm trying to make the argument of like as agents write more of the code and it's easier to build basically anything do you think more of that will be released open source even if it's for the marketing reasons like it doesn't matter what the reason is i'm just thinking like is the trend i really i don't know I wouldn't even know how to know it because I don't know what the percentage of closed service open source code was in the past, let alone how it's now.
One thing is pretty clear. There's an inherent advantage if you build something that you want the agents to build when it's open source because it's maybe, unless they're changing something, going to end up in the training data.
That is actually a pretty significant issue. I had a conversation even with people within Microsoft where they were evaluating the quality of what the models were doing for outside code versus what it means to write an Xbox game. Because very little console code has ever been publicly shared.
In fact, it was pretty much disallowed. And there was also pushback within Microsoft as far as I know to make the Xbox training, the Xbox code available to models because they were afraid of it. So I don't know where it landed, but this is sort of implicitly there is a benefit of being, if you want others to build on your stuff with agents, it's great if your stuff is at least.
in parts open source because then it will end up in the training data.
But again, this space is moving so quickly. Who knows what's going to happen in two to three years? And also, it doesn't really matter.
Nobody's choice is going to tremendously change. They're going to do whatever they want to do anyway. So I think people are doing what comes naturally.
I hope that people are going to continue maintaining really great open source software. That's what I hope. and I don't think that a lot of the projects that are starting out now have the right reasons to doing stuff open source.
That probably was always the case. It's just very noticeable now, I would say.
So is it the raw technical capability? Is it the thought? How would you put a mediocre open source project versus a great one?
What are the things that actually separate them? Is that project going to be there in 10, 15 years? It's like the only question really that matters.
Is that project going to be there in 10 or 15 years and still open source? If yes, that's a good open source project. If not, then it's not a good open source project.
A bad open source project is one that goes away, has no future, turns closed source, gets abandoned. That's a bad open source project. Or it has no users.
But if it turns out to be useful to society, continues to be maintained, people build stuff on top, then it's a great open source project. So it's purely just something we can judge in retrospective. It's not something you can judge now.
As an example, I don't know if you've ever downloaded PHP 1 .0 or PHP 2 .0. You can go to museum .php .net. You can download the old tarballs.
It's horrible. It's objectively horrifying what the old PHP looked like. And I remember when I started doing PHP early and there was PHP three times, I think, and then there was already the good version of PHP.
There were so many people who were saying, oh, PHP is trash. The guy doesn't know how to make a programming language. But look at PHP nowadays.
It is an actually really good programming language. And all the people that worked on it learned a ton. And you can build very productive...
code with it. Like Laravel is a great ecosystem. And so like if you were to sort of judge PHP as an open source project on like year one, based on what Lars did when he started out, you would have a horrible opinion of that language.
But nowadays you can look at this and say like, this is in fact a great open source project because for a start, it's still around. It's still evolving. It still has people supporting it, right?
So we can really only judge it from after the fact.
Yeah, that's a really great point. Like basically half the success is just sticking with it and keep improving over time. And that's how you gain the trust of people to build on top.
Pi will be a great open source project if it has meaning a couple of years from now. If it turns out to be not needed anymore, then it was an interesting blip in a moment of time, but it was not a great open source project. Let's talk deeper about being in a training data because that's a very interesting point.
We're starting to see it with like, you know, when you start a new project, what the models are pushing you to choose as the tech stack, as the deployment platforms. Do you think like a lot of it has been already written where like, you know, version control, it just goes for Git, right? Or like you mentioned, the operating system, they're great at Linux because there is a lot more of that in the training there compared to Windows or macOS.
So like, how do you think this impacts the future?
present day model architectures being in the training data is a great benefit and at least that is the general I think agreed upon interpretation and I don't have a counter argument to this and that goes as far as I have observed that people that are in fact in the training data are banking on this. I know that there are teams in companies that are underrepresented in the training data trying to figure out how do they get into this.
Because if you're the top choice of an agent to do something, I don't know which company it was, but I think one of the companies for their space figured out that they're really good in all the models. And I believe, if I remember correctly, they had a billboard. it wasn't a billboard, it was an ad on a bus in San Francisco where they said, hey, do you need, I think it was like observability, ask your AI what it would use.
So there was a power move that you could do because you had a default choice.
So there's obviously a huge benefit to this, particularly because people are increasingly using these models as a replacement for Google search.
Obviously, to some degree, if you have web search also in it, then it's also an SEO question, right? Because the agent might do some investigation rather than just going through the weights. But there's obviously benefits.
That said, that alone doesn't really help you fully.
One, because these agents will always have to solve problems which are in your code base, which is hopefully never in the weights unless you work on open source code. So enough work and training will have to go into in -context learning and then making that really good. So maybe it's not so necessary because it can point people at documentation and so forth.
But yeah, I mean, in general, I think people will do everything they can to get into the weights in one form or another.
I don't think it's going to change. I want to talk about more future things. Recently I had the realization that inference is going to become more and more of the spend because I mean first of all we you know I can look at myself and like I don't really spend more money on random vibe coded AI SaaS but I spend more money every month on AI inference right and I don't see this stopping I mean sure everybody wants to like make it more efficient and you know spend optimize their tokens get subscriptions not like pay API prices but like do you see this again assuming model progress will continue harnesses, everything will keep getting better.
Do you see this becoming like eating more and more of the software market? Whereas, you know, people in the past were paying some consultant or some researcher. Now they can just do two prompts as a deep research and get their favorite agent to like answer that question.
So that spend goes from that consultant to the inference. Like, do you think it will just eat the software market? I think it will depend on what the cost of tokens will be.
Presumably if the cost of tokens will go down. So cost of tokens is going down today. But simultaneously, the actual cost of solving tasks still seems to be going up.
At least from my personal observation, my sessions are not getting any cheaper. My general spend on what I'm doing is not having a downwards trend. Let's put it this way.
And if I were to have to pay API prices rather than subscription prices, I don't think I would actually be using AI much. I would feel weird burning that much of my investor capital on AI. Maybe I shouldn't, because that's what they want.
But I would personally feel weird having such an expensive expenditure to inference within my company. So I think it would... purely be a question of how expensive is it?
Because if the human in comparison can be both trustworthy and be more reliable and they themselves then use AI to augment their capabilities, it might actually be a better deal than me trying to get rid of all the humans in a company and then just purely using AI for it, especially at the prices where we already are. But I mean, I can't foresee the future.
It's also going to greatly depend on what society wants. It's very clear that outside of programming, AI is a mixed bag. People are using it for sure, but they're using it in the same way that they're using smartphones and TikTok and YouTube, where plenty of those are saying, yeah, I'm using it, but...
I still hate the companies that make it, or I hate the fact that it uses so much energy, or I hate the fact that my electricity prices are going up. So they can both find it somewhat useful, but not really like the technology. And I think depending on how much they like the technology, that will really greatly impact what companies will do, what society will do.
And that I don't think I can predict in any form. I think this can go either way.
Okay, I have two thoughts. I think first, you could see your token spend going down if you were optimizing for it. If you really had a small budget and you were stacking open code subscriptions, you could get more and more every month with better open source models for tasks completed that cost less.
So I think part of that is just maybe you don't care that much to optimize. for it to go down and, you know, you're comfortable spending what you're spending. Like if you didn't have the money, you would probably find a way to run some AI, maybe not the same quality of models, but like, you know, cheaper subscriptions.
And I guess the second thing I want to say is like that, yes, you still want to talk to a human in certain cases. And if that human is using AI, then maybe like, if we say like an accountant or whatever, tax advisor, that accountant that's AI enabled can serve more clients, right? And as a function of that, he can charge less price per client, but like except, 10x more clients and undercut his competitors who are not using AI.
So that is still like moving the cost to inference because he will make more money, but his margins will be lower because he's using AI to fulfill his work. Yeah. I mean, I think a general thing I'm worried about is like percentage -wise a greater part of the, like...
in proportion to European GDP, more money will be funneled to American companies. I think that's just like, that's a general worry that I have because like Europe doesn't really have. Yeah.
So like, I think that is most likely going to happen in one form or another. Where it lands, I don't know. As an example, for instance, I feel like, I don't know like what the trend here is, but at one point, you would find no designer using AI, but you could go to chat GPT and make yourself a picture.
Then there was a percentage of designers that would use AI with you. So you say like, hey, I want this and this, but you're also like 30 % for slop, right? Because like all of a sudden, like there was AI generated stuff in your image.
And at one point that became, at least in certain circles, to be such a negative sign that you now almost pay a premium for that person not to use AI in the first place. I know that is a topic, for instance, with ads. Ads are representing your product.
And there are people paying premium now for their ad not to show any AI in it, or at least obvious AI in it, because their customer is having a negative response to that. So I think that will greatly depend on what it is that you're doing. And I think there's a chance that if more and more lawyers are going to use AI and then maybe some of their court cases are actually losing their court cases because maybe they were badly researched or they made up some stuff, then there will be a new balance to it.
And on the upper end of it, then maybe they will still use AI, but you're not really paying... Because they're not doing more court cases, but in fact, because they're winning them better because they found a good way of using it. But I don't think it's so simple to say like, okay, if we give every doctor in the world now access to AI, they're going to handle, I don't know, 10 times as many patients.
And as such, the quality of healthcare will go up. That could be... Well, I didn't mention quality.
I just mentioned the span of inference, right? The quality as a whole is another topic. I meant the percentage of software spending just going to plain inference.
But presumably there's a counter move to that because if... Maybe it doesn't follow. Maybe it is in fact...
Most likely what's going to happen is that how much we're spending right now on tokens is a fraction of what we're going to spend a year, two years, three years, four years from now. Like as a society at the very least.
I think that is probably going to be true. However, are the tokens still going to be that expensive if they don't deliver quite as much value as we are assuming they will? It's a different question, right?
I think that it's not completely impossible that we're building too many data centers proportionate to how much it costs us to do that and how much the companies can actually charge. There's fierce competition.
So maybe the prices will just have to go down in the same way as the prices for electricity go down or should go down if there's sufficient supply. At the very least, solar is getting cheaper. So maybe tokens will eventually also get cheaper.
I don't know. I really don't know. What I'm hoping where this lands is that society rebalances itself around this somehow where it doesn't end up like social media did.
My suspicion is that at least in Europe... the majority of people hates it now and still uses it. Social media or AI?
Social media. Social media, short -form video content, everybody's using it. But if you make an opinion poll on that, people are incredibly frustrated with it, but they feel like it's almost like an addiction, like smoking is.
And so I would hate that we built... AI in a similar way where everybody's using it, but simultaneously everybody's hating it. That would be an unfortunate outcome.
Let's talk about Europe. You mentioned it a few times. You're from Austria, I'm from Czech Republic.
Are we just falling behind? How do we save this continent?
I mean, we are falling behind. I think that is probably true. How do we save the continent?
That's a good question. Well, I think like, hey, I know a bunch of ways in which we're not saving it. I think like right now what's definitely happening is that all the motivated people, or not all the motivated people, but a sufficient number of motivated people are just going elsewhere.
I think that probably is not great, right? So like the starting situation is not overly amazing.
And I think a fundamental issue that Europe has is that Europe, in the stage in which it is, is all about preservation of the past and not so much about the enablement of the future. We have a lot of examples of this.
But Europe also has some things figured out that maybe some other countries have not figured out. And if they get into that position, they will have to figure this out. A growing middle class in China will have different demands to their lifestyle than China did up to this point.
Europe is a society where people want a lot of rights, where things are balanced out more, and that makes it also harder to start new things because you have to run up against all the regulations in place. And it's not a question of politicians, it's a question of what does society want, and society wants it that way. Society doesn't want data centers to the degree that maybe some other countries want.
Society doesn't have a great... towards large tech companies and if anything is happy that they're not here. So given how the European societies largely work, I think it's sort of a tricky situation.
Europe's biggest issue is that Europe is not one country. The US is a massive single market. China is a massive single market.
India is a massive single market. Russia to some degree is a pretty big single market. At least Russia plus the other CIS countries together are.
europe is 27 countries that are constantly fighting against each other and then the uk fully left right so like it doesn't be a question of language it's just literally a question of like Does every European country give the full worker rights to every other European citizen? And the answer still is like there are a lot of, like all the countries are trying to put as much friction into it as possible to benefit its local population over everybody else.
Like we have 27 armies, we have 27 legal systems, we have 27...
worker regulation pieces um like starting a new company you have to like vat in every fucking country it's terrible yeah it's like this there's just a lot of it right and and it's it is because we haven't quite managed to come over this idea yet that um uh we're just europeans and and i think like right now the the mindset is not quite yet to enable that And so maybe that has to be the first step or at least it has to be a pretty significant step.
I think if compared to where I grew up, the times where I grew up to now, I feel like it did go in the right direction because I remember like Austrians were deeply critical of every neighbor country that was not German speaking. I think that has changed. If anything now, it's all about the countries that are not in the EU yet.
It's about Montenegrins, it's about Albanians, it's about... maybe to some degree Ukrainians. So if there is a suspicion towards other Europeans, it is sort of at the boundaries of the European Union and maybe like plus minus Norway, UK and Switzerland.
But there's less suspicion, I would say. But there's still a lot of suspicion. And I think that holds back Europe just intrinsically.
And I don't think Europeans... politicians can fix that i think this is just a question of like how how does society get to that and i think then society will have to start pushing back against some of the shit that happens on the european union level and that happens there not because european bureaucrats are idiots it's because like the countries themselves are like these compromises are coming because 27 different countries have to agree and they and they're doing what what what sort of comes out of it um that's that's the unfortunate reality All right, so I think we can end it here.
I want to be respectful of your time. Appreciate you coming on, Armin. Where should we send the people?
Where should we send the people? Send them to pi .dev, aarondill .com, and yeah, probably there. All right, I'm going to link all of that below, as well as your socials.
Again, appreciate your time, and have a great day. All right, thanks.
The Hook
The bait, then the rug-pull.
David Ondrej sits down with Armin Ronacher, the creator of Flask and the person now running Pi Agent at Arendelle, to unpack why a stripped-down, bash-first coding harness is beating tool-heavy competitors like Claude Code and Codex.
How they asked for the click.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
AFFILIATECommission earned if you click.
OTHER LINKSAlso linked in the description.
