if "just press the magic button" is what the industry wants, then i'm out. not because my programmer feefees are hurt. i've written enough code in my life to be very tired of it. but i couldn't imagine a more boring waste of my time than just pressing the magic button.

super refreshing take! i think the "software engineering is solved" perspective forgets that one of the most core parts of the job is turning smaller unstructured messiness/goals into a candidate (but likely imperfect and iteratively improved) artifact that helps more complex goals get solved. writing code used to be part of this recipe, but the supervision of the artifact is definitely still the same with agents code generation throughout the years has always been about putting bigger and bigger "holes" in your spec and and having an automated solver try and fill it in (it's a feature not a bug!), and while we should definitely strive for Good Code™, what makes code "good" has always been contextual: not just whether an implementation works, but whether its tradeoffs make sense for the messy, evolving goals of the system it's meant to serve. benchmarking this will be a super fun thing to do :) more interestingly, the frontier of software engineering research might increasingly be about understanding how much ambiguity we can safely delegate. making the holes in specs bigger is only useful as long as we can have reasonible feedback loops that tell us whether the solver filled them correctly, and the hardest part is often that we don't even know what correctness means until we've built something and seen how it behaves.

Give your app real-time decision-making with Decisions API, powered by GPT-6 Luna. Define questions and possible answers to classify content, route requests, or choose an agent’s next action. Available in limited preview.

Okay. I’ll take the bait. I went to work for Peter in 2013 with opposite politics. He sat me down right at the beginning. PETER: “I don’t want the (Thiel) Foundation to do good that the market can do. That is the job of the market. Likewise, I don’t want to do the good work that is seen as good. That work can be done by people richer, weaker and more needy of approval than we are. I want us to do only the work that is *deeply* good but somehow confused for being evil, because that is the work of the courageous and the strong. We strive to be both and that is where a small foundation can make the greatest positive difference. Go look for such good and let’s do that. But remember. Good has TWO different antonyms. What is “bad work”? It could be seen as shoddy work. Or it could be seen as evil work. And we can’t do shoddy work. Why? Because evil that isn’t competent never gets to be evil. It’s just pathetic. So we have to be exceptionally competent at the good work which we do to make sure it is always confused for evil.” I thought it totally logical. So that’s the argument. And that’s classic Peter. This is an old idea of his. And now you know.

I used Opus 5.5 to formally verify the Claude Agent SDK using Lean. A couple short prompts = 16 PRs fixing various bugs and race conditions. Video attached. TLA+ also works well. I sometimes combine Lean and TLA+ to look for issues around data flow, concurrency, and state mgmt. I don't know either language well, but Claude is excellent at both. This approach is super useful for formally modeling your code and finding bugs that a human probably wouldn't have spotted. Is formal verification the future of coding (or at least, bug finding)?

MCP vs CLI for LLMs is the wrong discussion. People are debating at the wrong layer. Both work incredibly well as long as they are run through a repl like environment that can persist state. The funny thing about CLI is that CLI tools are usually accessed through BASH which happens to be a repl with persistent state (the file system), therefore cli works much better than mcp. But this isn't even close to being an intrinsic property of MCP. Just means that we need better harnesses. Right now the best repl for this are: - bash + fs - jupyter kernels - codemode type repls, usually quickjs LLMs understand the concept of forward evolving a system to solve a need very well. This comes from the agentic RL. They understand how to change the state of a codebase or acquire data from databases and then operate on it similarly to humans. But without an execution environment, they actually cannot do this properly. My bet? Sometime soon someone will (or has already?) create a embeddable, sqlite style mini execution environment that parses bash, typescript, or tool calls into a common IL execution plan that's easy to security check before executing. And design this specifically for durable execution environments. Then we will just connect cli, mcp, webmcp, whatever to that and it accepts any of the input modalities as they can all be represented as each other.

Qwen3.8-27B is now live at Cerebras speed. The dense, open-weight model from @Alibaba_Qwen scores 34 on the Artificial Analysis Intelligence Index—making it comparable to models such as GPT-5.6 Luna, DeepseekV4 Pro, and Claude Sonnet 4.6.