Ask HN: How much of OpenAI code is written by AI?

Not OpenAI, but Anthropic CPO Mike Krieger said in response to a question of how much of Claude Code is written by Claude Code: "At this point, I would be shocked if it wasn't 95% plus. I'd have to ask Boris and the other tech leads on there."

[0] https://www.lennysnewsletter.com/p/anthropics-cpo-heres-what...

TFA says "How Anthropic uses AI to write 90-95% of code for some products and the surprising new bottlenecks this creates".

for some products.

If it were 95% of anything useful, Anthropic would not still have >1000 employees, and the rest of the economy would be collapsing, and governments would be taking some kind of action.

Yet none of that appears to be happening. Why?

They are likely lying:

https://www.anthropic.com/candidate-ai-guidance

> During take-home assessments Complete these without Claude unless we indicate otherwise. We’d like to assess your unique skills and strengths. We'll be clear when AI is allowed (example: "You may use Claude for this coding challenge").

> During live interviews This is all you–no AI assistance unless we indicate otherwise. We’re curious to see how you think through problems in real time. If you require any accommodations for your interviews, please let your recruiter know early in the process.

He'd have to ask yet did not ask? A CPO of an AI company?

Sure but what did the CTO say ? Also was he shocked ? There's no definitive answer, this is an evasive one.

I don't doubt it, especially when you have an organization that is focused on building the most effective tooling possible. I'd imagine that they use AI even when it isn't the most optimal, because they are trying to build experiences that will allow everyone else to do the same.

So let's take it on face value and say 95% is written by AI. When you free one bottleneck you expose the next. You still need developers to review it to make sure it's doing the right thing. You still need developers to be able to translate the business context into instructions that make the right product. You have to engage with the product. You need to architect the system - the context windows mean that the tasks can't just be handed off to AI.

So, The role of the programmer changes - you still need technical competence, but to serve the judgement calls of "what is right for the product?" Perhaps there's a world where developers and product management merges, but I think we will still need the people.

Standard CxO mentality. “I think the facts about our product might be true but I won’t say it because the shareholders and SEC will hang me when they find out it’s bullshit.” Then defer to next monkey in circus. By which time the tech press, which seems to have a serious problem with literacy and honesty (gotta get those clicks) extrapolates it for them. Then analysts summarise those things as projections. Urgh.

The other tactic is saying two unrelated things in a sentence and hoping you think it’s causal, not a fuck up and some marketing at the same time.

Been using claude code almost daily for over a month. It is the smartest junior developer I've ever seen; it can spew high-quality advanced code and with the same confidence, spew utter garbage or over-engineered crap; it can confidently tell you a task is done and passing tests, with glaring bugs in it; it can happily introduce security bugs if it's a shurtcut to finish something. And sometimes, will just tell you "not gonna do it, it takes too much time, so here's a todo comment". In short, it requires constant supervision and careful code review - you still need experienced developers for this.

> If it were 95% of anything useful, Anthropic would not still have >1000 employees

I think firing people does not come as a logical conclusion of 95% of code being written by Claude Code. There is a big difference between AI autonomously writing code and developers just finding it easier to prompt changes rather than typing them manually.

In one case, you have an automated software engineer, and may be able to reduce your headcount. In the other, developers may just be slightly more productive or even just enjoy writing code using AI more, but the coding is still very much driven by the developers themselves. I think right now Claude Code shows signs of (1) for simple cases, but mostly falls into the (2) bucket.

That's not evidence of anything to do with their hired developers. Interview practices have never reflected on-the-job practices

Weasel words. No different than Nadella claiming 50%.

When you drill in you find out the real claims distill into something like "95% of the code, in some of the projects, was written by humans who sometimes use AI in their coding tasks."

If they don't produce data, show the study or other compelling examples, don't believe the claims; it's just marketing and marketing can never be trusted because marketing is inherently manipulative.

It’s worth pointing out that the statement is about how much of Claude Code is written with it and not how much of the codebase of the whole company. In the more critical parts of the codebase where bugs can cause bigger problems, I expect a lot less code to be fully AI generated.

95% of Claude Code was actually written on a whiteboard!

It could be true, the primary issue here is that it's the wrong metric. I mean you could write 100% of your code with AI if you were basically telling it exactly what to write...

If we assume it isn't a lie, then given current AI capabilities we should assume that AI isn't being used in a maximally efficient way.

However, developer efficiency isn't the only metric a company like Anthropic would care about, after all they're trying to build the best coding assistant with Claude Code. So for them understanding the failure cases, and the prompting need to recover from those failures is likely more important than just lines of code their developers are producing per hour.

So my guess (assuming the claim is true) is that Anthropic are forcing their employees to use Claude Code to write as much code as possible to collect data on how to improve it.

Anthropic should have a check for CC, that runs “rg ‘todo’” after every turn. Then tells Claude to fix it if not.

95% of Claude Code is actually just inverting a binary tree

> I would be shocked if it wasn't 95% plus.

This is classic marketing speak. Plant the idea of 95+% while in actuality this guy doesn't make any hard claims about the percentage. It can just as well be 0 or 5%.

"You lied" "No I didn't. Look my face is shocked can't you tell?"

Did he ask Boris?

This guy is so full of shit. Anthropic’s leadership are all talk and hype at this point. And they’re not the only ones guilty of this in this hype cycle by far.