Eric Hauser

So cpython is going to be rewritten in Rust huh?

Codex app is so good but we really need a “PR mode”. 1) Commit, push, and PR at the end of a run 2) prefix the thread name with PR # 3) Loop on the PR until CI passes 4) Address review comments until Codex review give thumbs up 5) Change dot to green cc @thsottiaux @embirico

@martin_casado You have to combine something like this with a virtual bash to make it really useful; otherwise no way to efficiently ripgrep over a large corpus of files efficiently. I’ve been exploring this with github.com/ewhauser/gbash. Think SQLite file systems are going to be a big thing

This is really, really cool. Sort of the file system equivalent of the agent sandbox environments. Imma use it for my setup ...

@agentxagi @hwchase17 If you don’t have that though, there is no way to just plug that into a generic agent runtime. You can’t ripgrep over Postgres efficiently unless you rewrite ripgrep

@agentxagi @hwchase17 Maybe? I’ve exploring this concept a bit with github.com/ewhauser/gbash. If you your own bash system on top of a virtual file system then you can optimize how the tools use it; ie special hooks into grep, rg, etc so that the underlying storage engine can be optimized for search

We did a mini launch week for DeepAgents this past week! Launched things include: 📁Plug in any backend as a "virtual filesystem" 🖼️No-code agent builder powered by DeepAgents 💻DeepAgents CLI What else is missing & what should we launch next? 🧵 on the launches

Like I don't even want to see that there was a conflict. Just resolve it, update my branches

GitHub really needs to add automatic merge conflict resolution with AI. That is a feature I would definitely pay for.

@chaliy @AlemTuzlak Nice! I wish I would have discovered that two weeks ago 😂. I see it's MIT licensed but do you care if I reuse your LLM eval harness with attribution? Was planning on building one of those

@ewhauser @AlemTuzlak Hey! And welcome from rust port - github.com/everruns/bashk…

This is precisely why I dislike the Vercel narrative of open-source. Everything they do is "open-source" and yet they publically call you out for forking a repository. Sunil has nothing to do with your CF beef and if you hate the concept of OSS so much close-source your code.

OS level sandboxes are the end game for AI agents but virtual bash envs - pioneered by @vercel just-bash - work for many use cases. I’ve been hacking on an implementation in Go over the past couple of weeks. Not ready for prime time yet but sharing early: ewhauser.github.io/gbash/

@trq212 Easy to run when tokens are free. $15-25 dollars is insane for a PR review and this is coming from someone who personally pays >$1k every month in Claude Code costs

@omar_or_ahmed @Scobleizer Yes, on legal but it’s because healthcare is supply capped today. No one wants more legal

Healthcare AI will be bigger than legal AI, bigger than coding AI. But most investors still dont see it. Heres the math that changes everything 🧵

@tayloramurphy Curious if you could share more about how you are structuring the code context. Are you storing references to code associated with a table and then giving the agent access to the code? Or extracting context and indexing it alongside the table metadata?

we built one of these. our data team is cracked astronomer.io/blog/building-…

@zarazhangrui @phanosha We added a script that uses playwright to take screenshots of each page and then convert to a PDF

@phanosha Deploy the webpage (eg on Vercel)

So many people have told me "I made a presentation using this skill, and it was very well-received." The key is "well-received". Presentations are a medium of communication. It's about the audience's experience, not just the presenter's experience. I think because the models are trained on tons of webpages, they have an inherent "sense" for "how to arrange information on screen using code in a way that's easy on the eye". It's like the models speak HTML as a native language. Models do better when they speak their native language

Everyone is this thread is overthinking it. AI writing is easy to spot due to proper use of the English language

@julesyoo FFS isn’t inherently a bad thing but there are a couple downsides: 1) it can encourage bad actors and volume maximizing behavior 2) slow iteration cycle on the codes can lead to suboptimal solutions. Love what CMMI is trying to do with the flexibility with ACCESS

CMMI ACCESS is bringing sexy back to Medicare 🎸. A lot of investors and companies pulled back from Medicare over the last few years in light of regulatory headwinds and recent performance degradation amongst publicly-traded MA businesses... and the current wave of AI care apps, while exciting, is primarily targeting younger, healthier, commercially insured users. But I've now heard of numerous startups viewing ACCESS as a reason to consider serving Medicare populations for the first time. So this program could re-accelerate the deployment novel care models into senior populations who face some of the worst access and health issues in our society! Also - a common reaction to the announcement of ACCESS rates was, "they're so low!" But that's the point! Constraint breeds ingenuity. Great conversation with @AbeSutton (CMMI) and Rohit Chandra (@ClevelandClinic) at ViVE on these and other topics related to payment model innovation in the age of abundant, AI-powered care. One debate that we didn't fully get to resolve on stage: Is FFS really all that bad? What would need to be true for FFS to be viewed as productive versus destructive? Not all healthcare interactions and customers are conducive to capitation or subscription models... and AI holds the promise of being able to deliver services (both cognitive and administrative) at a fraction of the human version of the cost, with less errors and waste. What say you?

@1ovthafew There are dozens of us! Dozens!

I'm definitely in the minority on this one. My experience is that Claude is a very useful tool, whereas Codex does things that make my eyes pop out. But of course my experience is with a very particular kind of code that's much more difficult to reason about than most.

@HarryStebbings Fun @LucasSwisher1 story: summer of 2022 at an offsite he demoing a chat app built on davinci-002 telling our company that generative AI was going to change everything

I am so fricking bored of guests that go on 10 podcasts and say the same frameworks again and again. Coatue manages $30BN on the private side. Their growth fund is $7BN. They have investments in Revolut, Anthropic, OpenEvidence, Canva and more. And yet, the Co-Head of Coatue, Lucas Swisher, never does podcasts. That changes today. @LucasSwisher1👇 Spotify 👉 open.spotify.com/episode/5QxHPw… Youtube 👉 youtu.be/Hom5OMMzOQ0 Apple Podcasts 👉 podcasts.apple.com/us/podcast/20v… Timestamps: 00:00 Intro 01:04 Why Public SaaS Is Getting Crushed in the AI Wave 07:35 Durability of Revenue in AI 15:28 Market Size vs Founder Quality: What Wins? 16:52 Why Price is the Last Thing to Matter 23:32 Mega-Funds Math: Can $5B+ Funds Still Generate Venture Returns? 27:01 What Returns Are 'Enough'? Why 3x Isn't Exciting at Growth 29:54 When Double-Downs Go Wrong: Overestimating TAM and Multi-Product Expansion 32:37 Margin Matters… But at Scale: AI Gross Margins, Cost Curves & Efficiency 37:11 Why it has never been harder to be a seed investor 40:11 Is 'Kingmaking' a Myth: When Capital Helps (and When It Hurts) 45:23 Is Canva Really a Platform Company? Multi S-Curves and Leaning into AI Early 46:52 Lessons from Mary Meeker 50:08 Lessons from Mamoon Hamid 51:34 LP 'Pick One' Games: Mamoon Hamid, Mary Meeker, Insight Partners 53:40 OpenAI vs Anthropic: Who Wins? 59:17 Most Memorable Founder Meeting 01:01:35 Career Decisions & Misses

@mikebeckhamsm This is my argument on why many people prefer Claude Code over Codex. Codex is clearly the better model but no dopamine hit due to the slow speed

I’ve probably spent 150 hours in the last month working with AI. One thing has become crystal clear to me: The tight feedback loops release so much dopamine that it is easy/enjoyable to build. It does feel like playing a game. But I have something to show for it at the end.

@lennysan @bcherny @cursor_ai Huge fan of your work and tons of empathy that these decisions are very difficult. Does seem a little disengenious to assert it was 100% a cultutral thing when there were likely some pretty serious financial considerations involved

Remember when @bcherny left Anthropic for @cursor_ai and then came back 2 weeks later? Here's the story.