Win Wang

I strongly believe there are entire companies right now under heavy AI psychosis and its impossible to have rational conversations about it with them. I can't name any specific people because they include personal friends I deeply respect, but I worry about how this plays out. I lived through the great MTBF vs MTTR (mean-time-between-failure vs. mean-time-to-recovery) reckoning of infrastructure during the transition to cloud and cloud automation. All those arguments are rearing their ugly heads again but now its... the whole software development industry (maybe the whole world, really). It's frightening, because the psychosis folks operate under an almost absolute "MTTR is all you need" mentality: "its fine to ship bugs because the agents will fix them so quickly and at a scale humans can't do!" We learned in infrastructure that MTTR is great but you can't yeet resilient systems entirely. The main issue is I don't even know how to bring this up to people I know personally, because bringing this topic up leads to immediately dismissals like "no no, it has full test coverage" or "bug reports are going down" or something, which just don't paint the whole picture. We already learned this lesson once in infrastructure: you can automate yourself into a very resilient catastrophe machine. Systems can appear healthy by local metrics while globally becoming incomprehensible. Bug reports can go down while latent risk explodes. Test coverage can rise while semantic understanding falls. Changes happens so fast that nobody notices the underlying architecture decaying. I worry.

We asked a dozen DevTool founders from companies like @RevenueCat, @greptile, @firecrawl, @infisical, @ollama, @resend, @mintlify, @UnslothAI, @porterdotrun, and @recallai, about the state of AI agents and the future of software engineering. In this episode of Founder FAQ, we covered everything from agents as customers and the end of coding, to advice for founders starting out and what they're most excited about going forward. Their answers might surprise you. 00:00 – Meet the Founders 03:00 – Building for Agents First 04:22 – Biggest Early Mistakes 07:15 – Do Founders Still Write Code? 09:22 – Most Unexpected AI Discoveries 12:09 – What's Underrated Right Now 14:38 – Predictions & What's Next

@garrytan Not saying your test suite is bad at all, but I find that AI almost always writes very questionable tests seemingly aiming to increase test case count. And the actual test code itself could also usually be ~30% lower even without culling useless/redundant cases.

Also I’m a psycho about testing

What makes gbrain genuinely different? It's not a memory layer OR a code tool OR a search engine. It's all three unified under one graph with one query interface. Nobody else has that. And I use it all day with my 100k markdown file OpenClaw+Hermes Agent setup.

We're open-sourcing Cua Driver - our new macOS driver that lets any agent (Claude Code, Codex, your own loop) drive any app in the background, with true multi-player and multi-cursor built-in. 1/8

@GergelyOrosz I have a hook auto-blocking rm -rf with a reason + allowable safe (soft) rm, so the agent can continue work but actually get to think a second time.

AI agents are far more cable when they have full system access; but when they do, they can mess a lot of stuff up (not unique to any one model). AI harnesses have guardrails: but those can fail. I wonder if we’ll need OS-level “sandbox primitives” to deal with this better?

@garrytan I'm not sure that's exactly true. I feel like it made the weights of certain skills more extreme, which I suppose is still a change, but not super different.

The awkward truth is that what counts as a good engineer just became a different thing in the last 4 months

Shit LLMs say... 8 parallel reads, best-case latency: 8x sequential read. Hilarious! ...until you realize it wasted the next 30k tokens with that assumption in sentence #0 😅

@ChShersh Have we realized that AI could be a sufficient smart compiler which compiles FP to optimized C? I'm only half joking here :P

It was a good post

@ChShersh Indirection solves everything!

@Winium And in this talk, I'll be like "So here's what I was going to tell there", and just proceed with my original talk. Clever.

My talk proposal was rejected.

Holy fuck, it's so much easier to develop CUDA with AI help. ...Yes, I'm talking about CMake.

We doubled Claude usage on weekends, and outside 5–11am PT on weekdays for the next 2 weeks.

We just added /btw to Claude Code! Use it to have side chain conversations while Claude is working.

@lydiahallie Amazing. I did this all the time as a custom skill-like solution, glad to see it formally in CC!

You can now ask Claude a quick question about your current session without interrupting the main task! It's read-only (no tool access) and doesn't add to the conversation history. The answer just disappears when you dismiss it

If you're not seeing "Thinking about issues with this request" on Claude, are you even prompting?

@blelbach I was just thinking about making a mini-transpiler in Scala, lol. Hyper-typesafe cuda. (though Rust-like lifetimes/borrow don't seem completely possible)

I predict in the near future, the majority of GPU cycles in the world will execute code written in Python DSLs that lower to machine code without going through C++. Triton, JAX/Pallas, torch.compile, cuTile, CUTE DSL, etc The era of C++ and C as the foundational layer is over.

Love Scala docs, lol. I guess resolution decided to diverge at that line.

For you Claude Coders/Codexers, I suggest to use a different model for code review than the implementation agent. Also consider a different model to write tests to spec (do this before/in parallel with impl).

We’re excited to release TorchLean which is the first fully verified neural network framework in Lean. The Lean community has largely focused on pure mathematics. TorchLean expands this frontier toward verified neural network software and scientific computing. With the recent release of CSlib, we see this as another step toward a fully verified ML stack. We support features: 1. Executable IEEE-754 floating-point semantics (and extensible alternative FP models) verified tensor abstractions with precise shape/indexing semantics 2. Formally verified autograd system for differentiation of NN programs Proof-checked certification / verification algorithms like CROWN (robustness, bounds, etc.) 3. PyTorch-inspired modeling API with eager-style development + export/lowering to a shared IR for execution and verification Project page: leandojo.org/torchlean.html Paper: [2602.22631] TorchLean: Formalizing Neural Networks in Lean Work done @Robertljg, Jennifer Cruden, Xiangru Zhong, @huan_zhang12 and @AnimaAnandkumar. #MachineLearning #ScientificComputing #Lean

@fwbrasil I've actually been working with different CLAUDE_CONFIG_DIR with multiple agents in the same repo dir. I use a commit mutex with a prompt "macro" for single-time auto-approved commits. Also working on a fully dockerized solution with tmux'd persistence.

@Winium yeah, the new built-in worktree feature isn't great. Sub agents don't respect it and file searches end up returning content from the main repo or other worktrees. I'm managing worktrees manually in separate folders. It works ok that way

to be fair: I complained about Claude Code recently but the beast is back! working on multiple parallel sessions is truly a next-level experience 😍