Oskar Wickström

Wake now, my merry friends! Forget the nightly noises! Ring a ding dillo del! I recently joined Antithesis, where I'm building a new browser testing framework called Bombadil. We're doing this openly, no stealth mode shenanigans this time: github.com/antithesishq/b…

@kingprotty @tritlo Any examples? :)

@owickstrom @tritlo Im getting a similar experience with cryptographic computing. The things ppl have done to sculpt the math to the computer is very fun to learn about

studying math was probably the most fun i had. every week they show you theorems and proofs that have been refined for centuries to be pure works art

@suis_siva @AntithesisHQ Fantastic! Let me know how it works out and what we can improve!

just added @antithesishq' github.com/antithesishq/b… to harmont.dev. loving it, ty for the good work @owickstrom <3

@_laurynas My dreams are unclear, but yeah, probably refs tagged somehow.

@owickstrom How do you imagine storing the comments on jj? Dedicated refs or some other metadata that travels with revision information?

In my dreams, I can use jj and nvim locally to review and comment on changes in a first-class (all stored in jj) and decentralized manner. How far is this from reality?

@tritlo True.

@owickstrom you get a bit of it with binary search and quicksort etc, but it’s still just such a young field comparatively

@tritlo Also great

@owickstrom Conal and Conor papers honestly

@tritlo In some sense anything that comes from Dan Friedman and Co.

@tritlo What would be the programming equivalent, if any? Functional pearls? Various lisp lore?

@SeanTAllen Yes, more!

@owickstrom We should chat

Maybe I'm dating and deprecating myself in a single stroke, but I genuinely struggle seeing a productivity increase from agentic coding. I try, I retry, like every 2-3 weeks. I always end up tired and slow, either from hand-holding in real time or spending hours undoing the slop.

@ZPostFacto Honestly quite tricky code in some cases. It's like 80% there a lot of the time, but the bad 20% take way too much time to fix.

@owickstrom What type of code are you trying to get it to write? I have had great success in several categories of tasks: - bug hunting - web code - eriting mock framework and tests - triage my issues and PRs But it cannot write C++ very well right now

@kamatsu8 Yeah, being in the training data is not always a good thing it seems. Actual expertise stands safe for now.

@owickstrom the one time I tried the bots it was to figure out how to program to AWS API's. Dismal failure. I thought it would be well represented in the training data, but it turns out that the training data includes a lot of broken, insecure or heavily out of date AWS code.

@Sirupsen Go Simon!

dangerously close to the last few features needed to replace even the trickiest queries from gen-1 search engines...

@nicolaygerold That probably contributes, yeah!

@owickstrom Just from our conversations, the stuff you're working on is so out of distribution in multiple ways that the models will have a hard time with it.

IT'S IN! I might be able to rest.

@jussisaur I swear I didn't read this tweet, but posted pretty much the same sentiment 7 hours later.

i am quite close to going back to an autocomplete-only AI coding style. dead serious. i'm not sure the ostensible speed of agent-first coding is worth the brainrot, the laziness and the loss of code and architecture comprehension

This is not a moral stance. It's not me saying these things aren't sometimes useful, or that they aren't very impressive anyway. It's an observation of the current practical applicability in my overall work. YMMV.

And this is like a broken record, but yes, some things it does help with (the silly scripts, the CI YAML hell, etc). I just can't seem to get it to hit my quality bar without a lot of effort, which obviously isn't the purpose.

@pachacutec_vibe I'm literally building tools that satisfy my quality bar. :)

@owickstrom I think that perhaps you could focus on how you could make better tools that satisfy your quality bar

@ChrisSeltzer Claude Code enterprise and opus 4.7, skills, claude.md, Rust, property based testing and other e2e testing, etc. Pretty much what people are raving about. It fails on maintainable and coherent coding regardless of how much I scream at in in claude.md.

@owickstrom Every time I see that I wonder what your set up is like. What tools are you using?

@AndrasGerlits That's been my experience as well.

@owickstrom I don't think agentic coding is salvageable if you need the code to meet any quality standard. I had good outcomes with doing bits on the side like 'give me a JWT token federation module' and then doing the things it missed. It's useless for any core work.