Dan Kubb

What would a project, framework or language need to provide to maximize effectiveness for AI? I’m thinking: a large corpus of working examples, the ability to establish fast feedback loops via strong types, good testing culture, good linters/formatters, well written user docs and plentiful examples. My gut says something like tigerstyle.dev would probably help too. IME the more guardrails I establish the more effective the LLMs are because there are fewer degrees of freedom for them to get lost in. The faster they can identify when they’ve left the golden path e better the results.

@jdegoes I think with AI all roads eventually lead to formal verification. I’d not even be surprised if LLMs not only excelled but thrived at working with formal verification, to the point where not doing it is less efficient.

Ironically, AI coding is going to do far more for advancing and operationalizing formal verification than Haskell ever could.

When you do a direct port you want a stable reference so you can cross check things at every step. I’d run differential tests with a fuzzer and assert behaviour is identical in every way I could measure. You need a stable oracle and if you make the mistake of porting *and* refactoring at the same time you no longer can directly check one side against the other. In the age of LLMs I don’t see this process changing, except I could automate more, and generate a ton of characterization tests to lock in observable behaviour. Only once I had behaviour locked down I’d then begin refactoring sweeps and/or larger rewrite of critical components.

@penberg i dont think many people have experience of porting large codebases. even if it was no ai, likely i'd do same do same: mechanical also, i can't keep the branch open for long where i have to keep porting new stuff idiomatically (or worse have the master branch blocked)

It’s both funny and sad that many people are complaining about how ”terrible” the Rust code in the Bun port is. It’s like that on purpose because Jarred took the safe and correct path of mechanical translation from Zig. Idiomatic Rust should be incremental!

@thsottiaux @romainhuet Closing the app with a long running session takes forever even on a very fast computer. Also opening a previously closed long session is also very slow.

For those of you living inside the codex app, what should we prioritize among features, reliability or performance?

@glbrief Sleep deprivation probably contributed to this.

This man went viral after losing sight of his child for just a few seconds 😭😂

@DellAnnaLuca @pmarca I would be embarrassed to say those things in public. Can you imagine passing off work as your own without checking to make sure you understand it? Hmm, wait a sec… 🤔

The reactions of many researchers on finally being held responsible for having read the very paper they submitted are... something.

@Bennieeexyz Save those notes!

My dad discovered voice notes last year and has fully replaced phone calls with them, He just sends long, unedited, stream of consciousness voice notes sent at random hours. I receive them like dispatches from another world. Voice note received 6:47am Dad: (sound of kettle in background) Good morning. I was thinking about you. I saw something on the television last night about young people and stress and I wanted to check on you. Are you eating. I hope you're eating. Your mother made stew yesterday it was very good. I told her to send you some but she said you live far. I said that's what couriers are for. She said I don't know how to use the courier app. I told her I would do it. I haven't done it yet but I will. Anyway. Call me when you wake up. Or send a voice note. Your mother says you never call. I told her you're busy. Are you busy? Don't be too busy. Okay. Have a good day. Eat something. Bye. (paus It's dad by the way. Me: (listens to entire two minute voice note) Me: (it is 6:47am) Me: (sends voice note back) Me: Dad I know it's you. You're the only one who sends voice notes this long. I'm eating. Tell mum I'll call this weekend. Also you don't need to say bye and then keep talking and then say bye again. Voice note received 6:51am Dad: (laughing) Okay okay. Bye. Me: (smiling at phone at 6:51am) Me: (this is actually the best start to a morning)

@haider1 Isn't GPT 5.5 Pro comparable already? They'd had pro for months now. It seems to be 6-12 months ahead of the current SOTA model.

openai wouldn't serve a mythos-sized model publicly either for now but considering that gpt-5.5 is already close to mythos on some benchmarks, i doubt they're too worried the compute bottleneck won't disappear but it still helps companies train larger models over the next few years

I'm making my agent to write lints before being allowed to fix style issues. Every time I find a style problem I ask my agent to identify all the locations, then write a lint to catch them, and only then it's allowed to fix them. This does result in lots of special purpose lints, but it's much better running a program to find style issues than delegating that to the LLM.

Look, if someone is reviewing your code please do not just cut and paste what they are saying into Claude/GPT and tell it to fix the problem. Engage with the reviewer to get real understanding before asking your agent to make the change. You're not adding any value being the middleman. If you're just blindly cut/pasting review comments, then you can be easily replaced by the reviewer running the agent themselves.

I think you could rig that up by telling it something like: - investigate the problem deeply and document your process and findings into a doc in /tmp - spin up a subagent to read the doc and go one level deeper and append their findings to the doc - the subagent should stop if it finds and proves the root cause, or cannot proceed any deeper. And then set up /goal to run this until the root cause is found.

I need a /goal-but-why where you just keep asking the LLM but why until it actually gets to the root cause.

The mental models you develop when learning Regexps deeply are invaluable. It’s often people’s first exposure to state machines. I think Regexps are one of the most beautiful concepts in computer science and you’re missing out not learning them.

Nah, the AI is excellent at reading Regexs but only above average at writing them. They can do better if you instruct them how, but they always leave something suboptimal if your instructions are lacking. I think it’s because the training set and evals are not at good as they could be.

Imagine spending years mastering Regex right before LLMs arrive

@adamemedia1 @DanielleFong They should only be allowed to buy and sell index funds on a predefined schedule. This ties their income with the overall prosperity of the market rather than individual companies.

“WE NEED TO LET THEM DO INSIDER TRADING TO FEED THEIR FAMILIES” That was the actual argument just made in defense of politicians trading stocks. Members of Congress make $174,000 a year. The median American income is roughly $63,000. (3x less) Meanwhile the federal minimum wage is $7.25 an hour… also unchanged since 2009. And you’re seriously being told politicians need access to insider trading to survive. The system is designed to lead earth to neo-feudalism. And it’s becoming more brazen by the year. Asset ownership consolidates upward. Living standards decline downward. The middle class gets squeezed from both ends and disappears while the elite class accumulates more wealth, more influence, and more protection. Leaving behind a permanent underclass. And a permanent political-financial aristocracy at the top…

@iridescence_dev @croloris But seriously, I'd be curious if GPT Could audit the Zig to Rust port and then plot a better path from Zig to Rust than Claude.

@iridescence_dev @croloris

I wonder if GPT5.5 would have done a better job at translating Bun. Now Bun not only has a hard dependency on token expenditure to be maintained, but it's also locked into one model brand, while the competition can choose whichever model happens to be the best at the time.

Yesterday I ran through a whole week of GPT Pro 20x credits in a single day. Did something change? Maybe something break with caching?

@VictorTaelin Would it be possible to use your Mac mini cluster to run a lightweight model, and then train it to convert Rust to Bend? The AI labs are providing the inputs to the process.

Bold bet: Bend2 rewrite in 2027 Bun will run natively in your GPUs

@jardel1307 @simpsoka Why wouldn’t it look through the Codex logs to find how you’re using it? I was doing this kind of self-introspection in spring 2025 before any harnesses had memory.

@simpsoka But it is sad that it works only if you can use Memory feature on Codex. And sadly, the memory feature is not available for the EU.

Just a fun thing to do every once in a while, ask codex, "study how i've been using codex and make recommendations for how i can improve"

I almost always start a new project and just focus on declaring all the constraints. Then I have the agent read and interrogate me to surface more invariants I didn’t think of. I keep it technology agnostic and write it in IDEA .md. Then I write an ARCHITECTURE .md with technology choices and design. Then I continuously cross-check both against each other. Every point in both docs must connect, and not contradict; otherwise I have to refine or remove it. As I work on the system I’m always reorienting against those two docs. Every refinement or fix gets added to the docs as a snapshot of current understanding. Sometimes I audit the code against the docs too and either fix the code or the docs when they diverge. So far it’s been working quite well. The agents do a better job than if I had been driving only or on-demand planning.