Haha this is such a misunderstanding of law. At least the loophole part. I do like the idea of formalising it though.

Was this written in 2025? 😄

Nothing. Search is AI based now. I expect even the traditional list results actually use AI under the hood (probably not LLMs though).

That’s the job now. If you need a job you’re going to have to get with the program. If not just quit!

In terms of maintaining skills, I have ~4 hours on a train each week that has shitty WiFi so I can’t use AI there - I spend some of that time doing manual coding on a side project.

Funny I’ve seen the opposite. I want it to rewrite half the codebase to do things properly. A human shouldn’t do that because of the effort, but it’s no issue for AI.

However it very often says “that’s too big of a change so I’ll do this workaround instead” which is not what I want at all.

But I agree they sometimes implement things that don’t actually make sense to do. Mostly you can solve that by just asking how to program / code works at a high level without needing to read every line.

Yeah so Opus 5.5 (which is what I use basically 100% of the time currently) gets 100% anyway. As do most of the other frontier models.

Would you care to show some evidence, or is it a vibe?

It’s an very conservative estimate based on my experience. Almost all open source software these days keeps their docs in the same Git repo as their source. I haven’t checked these, but go and look at e.g. LLVM, Git, libpng, SDL, Qt, etc. if you don’t believe me.

Show us the evidence.

Nah, too much work and you wouldn’t believe me anyway. I’ve seen it with my eyes though. Hell, frontier models are happy to reverse engineer binaries without prompting in order to get things done. Reading a bit of source code is nothing to them. (Does it violate a “no reverse engineering” license agreement if AI does it of its own volition lol…)

Would you like to find this out at runtime

No I’m not dumb enough to use dynamic typing.

I can’t say I agree with this.

AI writes shallow tests.

It definitely did in the past. But not any more. It’s far better at writing tests than humans and I haven’t seen a single nonsense test out of Opus 5.5.

AI writes overly complex ccode

This one I will kind of give him. AI does seem to like being super paranoid - writing tons of checks for things that we don’t really need to check, and adding config flags for absolutely everything. You can just tell it not to do that though. Annoying that you have to, but it does work.

Reliance on AI decreases human understanding

Yes but does it matter? Does your manager understand the codebase? Does his manager?

AI moves too quickly to maintain a high quality bar

It’s too quick for thorough human code review definitely, but I don’t think that means you can’t have a high quality bar. There are other ways to ensure a high quality product. Traditional QA testing for example.

The moat in CoDoc is that the documentation ships inside the package, version-matched, offline, deterministic. Can you tell me what proportion of software libraries today ship complete AI-optimized documentation within the library’s artifact? I’ll wait.

Wow what a moat. I would say maybe 50% of software already does that. For example 100% of Rust crates ship the docs with the code. In any case I don’t think this really matters because AI will happily skip the docs and just figure things out from the source code, which is usually better anyway.

Vercel’s evals

Using which model? They don’t say? Huh.

A 27B on your laptop with the right context can beat a frontier model that’s hallucinating your v5 API signatures.

Frontier models very rarely hallucinate anymore, and they’re actually really good at writing tests now so I really doubt this.

there is a difference between using it a tool and claiming that it will replace human programming

Yeah the difference is about 6 months of model releases.

Seriously, you would have been totally correct 6 months ago (or if you are still using models from 6 months ago). Now… not so much.

Remember when we all laughed when Anthropic tried to vibe code a C compiler and it was laughably broken? Yeah now they can vibe code an entire OS.

(They should really run that C compiler test again just to show how much progress they’ve made.)

Pure drivel.

docs/README.md as entry point

Wow how innovative.

The big labs’ and hyperscalers’ moat was never the model

Uhm, yes it was. Open weight models are nowhere near the capabilities of frontier models. I’m guessing you didn’t use a frontier model to generate this slop.

Do you think Valve isn’t using AI? lol

this is no doubt all vibe coded which I hate

No doubt, but why do you hate it? Vibe coding actually works now, and it means we can ditch Electron!

Ha yes I always thought this is how assembly should look. Although why not go further - instead of move a, b which is ambiguous about the move direction, why not just copy 99% of other programming languages: b = a.

Also it’s a real disappointment that even new ISAs like RISC-V have the same “assembly must be unreadable” brain disease. auipc? Come on. You can’t even claim that they want a strict 1:1 assembly:instruction language because they added a ton of pseudoinstructions. Even add can turn into add or addi.

Anyway although AI means you can easily make a tool like this, I think it also means there is not really any point to it anymore. Oh well.

This is pretty much the “I fancy her but I don’t want to reveal it unless it’s reciprocal” problem right?

Impossible because anyone can just lie. Say they fancy everyone, see who they match with and then say “haha joke!”.

Also £5 says your application is dating/sex/kinks lol

I work in silicon verification where we do extremely thorough testing, and that is nonsense. It’s perfectly fine to test internals of a system because often it is difficult or impossible to get good test coverage of internal parts of a system just using the public interface.

LLMs love advertising / speechwriting techniques. Wikipedia has a good list here: https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing

If you use AI regularly these things are very very obvious.

Yeah maybe, but usually if I’m using Make it’s because my use case is “I want users to be able to build stuff without having to install any more build tools”. The curse of network effects. That’s the only reason I’d ever use Bash too. Otherwise I’d switch to Nushell in a second.

Can that scale down to individual .cpp files though? My understanding was that it works more at the package level.

I don’t think that’s unique to LLMs though. Traditional learning material is very often like that too, and I actually find it worse because you can’t ask it questions.

You need a project that you want to achieve. Maybe automating something, or some kind of Arduino thing? The latter might be especially good for learning because you can’t trivially delegate it to AI, but you can still ask AI for help.