aykutseker

Is Tokenless measuring cost per completed task, or just cost per request, including errors and retries?

The text only part is the catch for me.

If it builds a UI and can't look at it, it's askin ls whether the app looks right.

A calculator gives you an answer. An LLM gives you an answer that sounds like it already checked itself.

I think a lot of people will file this under Java got structs.

That seems off. They're still objects, the new thing is that they can give up identity.

AI hiring starting to look like sports free agency.

Karpathy to Anthropic, now Noam to OpenAI.

Claude trying to make friends in a battle royale is funny.

But if the robot is anywhere near my house, I think I want the one that hesitates.

GitHub used to get code after someone had thought about it.

Agents are starting to use it while they think.

The machine is still fun.

It's the five layers of product growth between you and the machine that get tiring.

This is uncomfortably close to a normal interview task now.

Someone sends you a repo, says the install is broken, and asks you to take a look.

A lot of developers would run rpm install before thinking twice, especially if they were tired or looking for work.

who's tried it: is 2x the usage actually worth it over Opus 4.8 for daily work?

Moving from tdnf to dnf5 is interesting. Most internal platforms get more bespoke over time, not less.

A failed test often points at the mistake. Most games just tell you that the outcome was bad.

Storage never ended up being the thing we worried about. The painful bits started once a workflow could touch external systems. Replaying state is one thing. Replaying a charge or an email is another. How are you dealing with that?

The independent AI explainer is generated by the same Gemini that writes the ad creative next to it, inside the same ads product. Independent of what, exactly

the glorified marketer framing in this thread is missing the bigger signal. karpathy publicly pausing eureka labs to join anthropic is an ai founder of his caliber effectively conceding that verticals get eaten by frontier upgrades.for everyone here building on top of foundation, that's the actual news

flash beating the pro it was distilled from is suspicious, not surprising.distillation usually loses you something. if the smaller model is winning on agentic evals, the more likely read is the evals weren't measuring agent quality in the first place. that's the bigger problem for builders, not which model to pick.

Clean room needs an independent second party with their own intent. An AI rewrite probably doesn't qualify, since its output traces directly to what it read.

the EU funds argument works both ways. plenty of countries received similar transfers and didn't compound it the same way. the interesting question isn't where the money came from, it's what Poland did with it that others didn't.

the real test is whether the Klein bottle business got a sympathy bump in sales

Grass is greener works at month two. By month nine your muscle memory is still wrong.

all caps in a prompt is a code smell. when you're typing MANDATORY, you should be writing a wrapper, not refining the prose.

The Hub captures decisions humans made. It can't see the ones they didn't, which is most of what AI ships. You instrument the deliberate half and inherit the undeliberate one.

2021: a Reddit short squeeze kept GameStop from going under. 2026: GameStop is bidding $55B for eBay, a company 4x its size. If it lands, this might be the strangest full circle moment public markets have ever produced.

This seems less like Kimi is better at coding than Claude and more like Kimi found the right strategy for this particular game.

Still interesting though. The fact that an open weight model is close enough for that to matter is probably the real story.

The uncomfortable part is that this is probably rational behavior for both sides.

Employers use models to filter resumes, candidates optimize resumes for those models, and suddenly the resume is no longer written for a human at all.

curious how much of the output quality is the design systems and skill files doing real work vs claude just being very good at HTML. the prompt stack matters, but it's hard to know how much.

had this exact fight. founder wanted it more corporate, designer had the research, designer won.

six months later pipeline was dead. buyers were enterprise procurement, the new look read as too small to trust with our budget.

research was right. just not about these users.

thanks. honestly didn't catch the rhyme, accidental aphorism :D

good example from the article: the chroot+nss CVE. the rule that nss is dynamic and dlopens libraries from inside the chroot isn't anywhere obvious. it's encoded in 25+ years of sysadmins finding it out. clean-room rewrites end up re-learning that, usually as new CVEs. and LLM ports of the same code inherit the problem: the function signature is what they read, but the scars are what they need.

about to launch my first open source project in days. reading this with a knot. github used to be a default; now it's a decision. and watching mitchellh agonize publicly is the honest preview every new maintainer gets from now on.