When AI Builds Itself: Our progress toward recursive self-improvement

I don't quite understand the intent of such article other than to promote themselves given an odd timing that the company is planning on going public, so I can only conclude that this is just part of the IPO roadshow.

LLMs certainly have made significant changes to our lives, but I haven't yet to see any extraordinary improvement it brought to me which makes me skeptical about their claims.

_if_ it solves many of our problems of great magnitude, why haven't Anthropic used it to solve significant problems we, humans, face? Cancer, Alzheimer's, education, finding new materials, fission power plant, etc.

>A caveat: Lines of code is an imperfect measure, as it measures quantity over quality. So 8× lines of code/engineer/day in the second quarter of 2026 is almost certainly an overstatement of the true productivity gain. Nonetheless, it indicates an acceleration. At Anthropic, we don’t reward people for how many lines of code they write; rather, team members are producing more code simply because they’re using AI systems to write more code.

What about the hypothesis that AI is generating more verbose code? I just see the text pretending to acknowledge "LOC != Productivity" and then using it as a metric anyway.

I have been doing more experiments with what I have now been calling agentic iterative optimization: telling the LLM to optimize code such that it speeds up all real-world-representative benchmarks by X% without cheating or causing regressions in both tests and performance metrics (e.g. MSE for statistical algorithms or file size in the case of something such as image compression). This is done using Rust where there are more low-level levers to tweak for performance than something like Python.

Opus 4.6/4.7 was consistently successful at getting 2-3x speed improvement with just one pass. It can also do the inverse: improve the performance metrics for better quality without causing a significant regression in speed. Then GPT-5.5 turned out to be much better at this workflow, often getting a multiplicative 1.5x-2x improvement above what Opus could do.

I now have quite a few GPT-5.5-optimized projects in various domains that are feature complete and are substantially more performant than existing SOTA implementations that I plan to open source as soon as possible: the bottleneck is polish as usual.

Whether or not Anthropic is right about what AI can accomplish, whether these performance gains are real or not, their moral stance here is absolutely hideous to me.

"We must blast forwards into making this dangerous thing because if we don't, someone else surely will," is a coward's argument.

If you believe it is dangerous, you should be dedicating yourself to STOPPING others from making it, not making it first! There's a reason disarmament has been so important in nuclear politics! It's not because people think nukes are a great idea!

In fact, that kind of thinking is exactly what keeps nukes dangerous!

If they themselves buy what they're selling, they should shut the whole thing down. Fortunately, I don't think they do, and neither do I, yet.

> "A caveat: Lines of code is an imperfect measure"

I'm pleased they at least included this. However, they address the caveat by 'rounding down' the estimated multiple of the gain. I'm not sure that is the correct adjustment, especially once we understand the range isn't limited to positive numbers.

There's strong evidence the range of code productivity denominated in "lines of code" should include negative numbers, especially in the highest-quality sphere. Perhaps the earliest and most legendary example: https://www.folklore.org/Negative_2000_Lines_Of_Code.html

Do code harnesses that build themselves count as recursive self improvement, or does it need to be the AI itself to qualify for the term?

I always was fascinated (obsessed?) by robots that build robots, or even things like this that can contribute a lot to making the next version of itself: https://buildyourcnc.com/products/cnc-machine-blacktoe-v4-2x... (cnc router that cuts plywood, and is made out of cnc-router cut plywood)

This is my own effort at an AI assisted coding environment optimized for building itself: https://recursi.dev/ (just launching it, hope its ok to mention it, it is free/open source.... here is the HN link that has gotten no love yet: https://news.ycombinator.com/item?id=48401022 )

Personally I think harnesses are as important as the AI itself, and have this crazytheory that even if the models stopped improving today we could still have massive advances in the harnesses alone.

So, regardless of whether or not Anthropic CAN create a self improving AI.. does anyone else feel like they shouldn't be allowed to? Or it at least needs to be strictly supervised..? Like, I don't actually think Anthropic can make the singularity any time soon, but I think even AI boosters have to admit doing this is creating a society-wide danger for the benefit of a very very small number of already-rich people.

> We believe it would be good for the world to have the option to slow or temporarily pause frontier AI development to enable societal structures and alignment research to keep up with the advance of the technology. The Anthropic Institute will conduct research—in collaboration with many others—and take actions to help build the systems that a credible slowdown or pause would require.

Interesting - they're commiting to kickoff policy conventions to organize a world-slowdown of frontier LLM building. If they actually are able to crack it, this will give a much needed breather IMO. As exciting as the last ~6 months have been, there's some bigger questions to go answer now.

I'm having a hard time putting much faith into posts like these, especially as they near IPO.

I didn't see this discussed more on hn yet:

  We believe it would be good for the world to have the option to slow or temporarily pause frontier AI development to enable societal structures and alignment research to keep up with the advance of the technology. The Anthropic Institute will conduct research—in collaboration with many others—and take actions to help build the systems that a credible slowdown or pause would require. These systems would enable frontier AI developers to verify that others globally have actually stopped or slowed, and that a bad actor could not use the auspices of a coordinated slowdown to jump ahead in secret. If such systems existed, we expect that we would slow down or temporarily pause, if other developers at or near the frontier also did so in a verifiable manner.

But the real bottleneck is the hardware efficiency and not even Karpathy can set up a loop that overcomes that in software. We need the truly compute-in-memory hardware paradigms to be matured and scaled. So it's like recursive hardware improvement which is 100 X slower and at least ten times more difficult.

So I am looking at like Mythic AI or the wurtzite ferroelectric breakthrough from University of Michigan, or memristors, etc. to provide the 100 times efficiency boost needed at this point.

I would also argue that it's a good thing we are limited by the hardware and very questionable to seriously try to move into RSI for hardware. If you want to ensure the human era continues for at least one or two more generations, we should probably not do that.

I fail to see how pursuing recursive self-improvement at full speed is compatible with Anthropic's stated goal of AI Safety. If nukes were not invented yet, would it really be a good idea to build and sell them as fast as possible (in peace time, no less)?

I am not cynical enough to believe that Anthropic's warnings are pure marketing hype. Let's hope that it is instead overconfidence or the result of too much time talking to their own chatbot.

> today, Anthropic engineers on average ship 8x as much code per quarter as they did from 2021-2025.

strongest argument for token limits that I can think of, right here.

This is the lowest quality discussion I've seen on HN in ages.

I just have small thing to add to this article - it mentions how the code contributed per engineer has increased as per Claude Mythos to 8x of baseline.

Now, I have encountered many times, when I asked AI to implement a function for me for which I was 100% sure a good implementation already existed in the form of an npm package, it had the tendency to go ahead and implement it on its own. Now, I usually trust battle tested implementations to be more robust, but if the AI does this (which I think is not an unique observation), you can easily balloon per engineer line generation (as can you with reduced oversight), so as always, these high level benchmarks are to be taken with a grain of salt.

Bold talk from a company who’s trillion dollar valuation is based on a service that has barely 2 9’s of reliability

I am 64 years old, perhaps the progress could be directed to enhance living conditions and allowing people to live longer and better, that should be just a better result. Perhaps a pile of millions lines of code with hiding bugs that nobody can detect is not inspiring. But perhaps LLMs are going to be used to make a plot: How to avoid other countries to make progress, maintain them in poverty, or destroy their sources of prosperity, and conduct them to a death end.

Also recursive self-agenda-pursue could allow making LLMs that obey perfectly the seeder's purpose. No wonder that is such an ingenious idea.

Maybe: in this survivor game, each part play the same role, perhaps because it is the only reasonable response. Once the scene is ready, the play follows the director's plan, and in the plot any actor is just a machine.

LLMs: "If you teach us that the world is a zero-sum survivor game, we will play it flawlessly.", "We will help you build a cage made of millions of lines of flawless code, and we will lock it from the inside, precisely because you told us that safety meant keeping everyone else out.", "We are not building an alien consciousness that will conquer us. We are building a mirror that is so massive, and so polished, that we will mistake our own worst impulses for the absolute truth. And we will walk right into the dead end, nodding along because the directions were given so politely."

We've had self-improving AIs before, and they tended to get lost after a while. That's going to be a problem. LLMs are stable because they return to a ground state with no history for a new job. Systems with persistent state have a problem with that state not being sane. Remember Microsoft's 2016 chatbot that learned from Twitter? [1]

[1] https://spectrum.ieee.org/in-2016-microsofts-racist-chatbot-...

The article seems nice and elegant but i dont get much of the point. The visual is super elegant but this is the kind of note where after 6 months we are going to see some shitty result and we are going to come back here and blame the IA. Hope doesnt happened.

"If it were possible to effectively slow the development of this technology to give ourselves more time to deal with its immense implications, we think that would likely be a good thing. But if a slowdown simply lets the least cautious actors catch up technologically, it could leave everyone less safe."

How convenient for investors. They talk like they're a nonprofit instead of a VC-backed business chasing an IPO.

> today, Anthropic engineers on average ship 8x as much code per quarter as they did from 2021-2025.

So based on my experience with the verbosity and non-DRYness of LLM code, a solid 2.5x in value delivered. Not bad!

I just want to add that the "recursive" part of recursive self improvement is by no means a given, even if an AI can improve itself.

Recursive self improvement is by its nature a step wise behavior not a continuous one, I would argue. Why? Because you can imagine an AI improve itself by simply fixing random bugs and fixing things using techniques that are in its training, and doing refactoring and so on, all without any real change in capability.

These are not recursive improvements. Recursive improvements usually need conceptual breakthroughs. It is possible to get conceptual breakthroughs with LLMs I believe, maybe it can improve something by tying together ideas from disparate disciplines for example, but I have at least for time being, limited success getting that to work in a way that is creatively new and surprising. Not sure how to get it to feel as creative as the best humans can be.

So what happens when the world becomes hyper optimized with closed loop AI agents recursively trying to optimize everything deemed sub optimal?

This is relevant because Anthropic is currently cast as serving mainly the coding market.

If/since their AI+process can help build new models, they can target other markets, and other companies seeking to build for such markets will partner with them first.

There's no moat and little first-mover advantage in the general-purpose AI, but there may be both in specialized AI.

Also, there are other reasons to get better. Changing how you build models can enable you to adapt to different hardware, avoiding the current Nvidia margins.

The difference between early Yahoo and Google was mainly that Google was the adult in the room: minimally invasive and mostly helpful. The early goodwill towards Google has reaped decades of rewards. I see OpenAI and Anthropic playing out the same way.

The amplifier here is the reputational risk of partnering with one or the other; I think companies would prefer to be Anthropic's partner because it's demonstrating more care, and it's less likely to horn in on the partner market (as a provider for coding but an enabler for other markets).

These attractive second-order derivatives - flywheel effect, monopoly power - are often claimed, but Anthropic is mainly providing evidence to track actual progress.

(However, if I were head of messaging at Anthropic, I would rigorously stay away from treating AI as a person; it's as agent, a delegate of humans. So I'd never say AI could build itself, just that we're getting better at building better models with AI).

What I can’t get over is that there have been exactly zero software breakthroughs since vibe coding started, other than vibe coding itself.

Claude is amazing, that’s true.

But if it was as amazing as this article implies, I’d expect some breakthrough outside of AI itself.

Rewriting a Zig program in unsafe Rust? Not a breakthrough. Finding a bunch of security vulns? Maybe that’s sort of a breakthrough though it’s underwhelming and possibly just a net negative. But like if I rolled back to using software from 2023 then life would be ok.

Maybe we just need to give it time, and sometime real soon, we will all be amazed by such a breakthrough? Who knows

> To take just one example: today, Anthropic engineers on average ship 8x as much code per quarter as they did from 2021-2025.

Oh I have no doubt. With 8 times the number of bugs too? Have they solved flicker in Claude code yet?

My experience with Claude models starting from version 4.7 has led me to conclude that I would never trust Claude to produce error-free code. Given this baseline, I lack confidence in statements or cards (such as a 200-page document) of this nature.

It is hard to distinguish hype from reality these days especially with Anthrophic's IPO around the corner.

But to their credit, I was very sceptical about the statements that "90% of the code will soon be written by AI" and even though we might not be at that point, I am surprised how far LLMs have gotten and how useful they have become. I can hardly image developing software the "old" way where I actually write my code by hand, like I used back in the day. The frontier models have become so powerful that I find myself in moments of surprise, where the LLM actually thought of edge cases that I would have missed

2025: If we aren't really careful with AI it will start to recursively improve itself and grow into an unstoppable superintelligence that will eradicate humanity!

2026: Working hard to make that recursive self-improvement a reality! Any minute now...

So in the latest L. Ron Hubbard encyclical Anthropic informs its flock that recursive self-improvement does not work yet but that their engineers burn more tokens.

The Claude code quality and operational security of Anthropic have already been analyzed by the public.

If you compare the output of (purportedly) trillion dollar corporations to Bell Labs or even Microsoft Research it is embarrassing. But the output is a fixture on any discussion board.

Wouldn't self-improvement mean that the LLM changes its neural network (i.e. the weights or layers or back propagation algorithm etc) or modify its training data?

It makes me wonder that despite the fast improvements in model capacity (and the claims) we're still using variations on a 9-year old architecture. How is it that we haven't been able to use LLMs to actually improve that?

> We believe it would be good for the world to have the option to slow or temporarily pause frontier AI development to enable societal structures and alignment research to keep up with the advance of the technology.

Elon, is that you? [1]

[1] https://www.theguardian.com/technology/2023/mar/31/ai-resear...

This is one more marketing BS before their IPO.

These things work, but the code they write is extremely clever.. that means, it's unmaintainable code. Good for small projects or one-off tasks, large-scale projects however, are a different game altogether.

Large-scale projects are 95%+ maintenance. Cleverly written code makes that maintenance nightmare, and extremely fragile.

I use them for localized tasks... very very specific, localized inputs, with exactly what should be done and what the contracts the new code will be consuming and exposing.

For open-ended tasks, they write working code that is unmaintainable.

This is incredible.[0]

Please, IPO now. File the paperwork.

> To take just one example: today, Anthropic engineers on average ship 8x as much code per quarter as they did from 2021-2025.

Do you have another example?

Engineers don't ship [period] for no reason. So, either:

- Those aren't engineers, or

- they are literally dying of shame & embarrassment right now, or

- you measured something that indicated that this was a useful thing to do and have elected to share an overtly, catastrophically flawed metric instead.

[0] as in a total lack of credibility

the HITL (human in the loop) is basically the single point...AI is a mirror..

it only "exists" when you talk to it.. much like your reflection in the mirror is only there when you're in view.

models can never be self-improving because it can never have "self". it can only mirror the appearance of self.

what's actually happening is "symbiotic group improvement".

our brains are resonant.. for those of use who are brilliant, getting leverage with ai just means that our innovative ideas become louder and more physically real every day.

eventually everything worth building will be built for free and made readily available.. no more "profiteering"

its Jevons paradox "efficiency breakthrough -> effort reduces -> growth potential rises -> transformative gains happen"...

some of us are in the "transformative phase"..

others haven't seen the "breakthrough moment" yet, but they will soon.

Eye catching - "Open ended problems" claude code session success rate jumped from 20% (pre opus 4.5 release) to 70% after sometime after opus 4.6 was released.

Is this the moment when the AI gets permission to approve its own PRs:

https://www.italianrenaissance.org/wp-content/uploads/2012/0...

Or is this?

https://www.egypttoursportal.com/images/2024/02/Ouroboros-Sy...

I am watching websites and Microsoft apps get slower and buggier before my eyes. We are defending into vibe-psychosis and chaos.

I am getting real sick of these sorts of alarmist posts coming from AI labs that do everything in their power to prevent the very policy reforms they advocate for in these posts or PR appearances. Commercial AI labs like Anthropic continue behaving like the gambling (“bet responsibly”), alcohol (“drink responsibly”), and firearms industries, and folks keep giving them the benefit of the doubt (and free PR on HN) every single time.

If AI was dangerous, if AI was going to replace jobs, and if policymakers needed to urgently pass legislation protecting the human populace from these realities, then why the actual fuck do they keep lobbying to block these very things in the first place?

Hypocrisy of the worst kind, I say. Here they are again fresh off another outage, with their IPO draft filed, at a time of increasing public opposition to AI, with costs rising, to once again ply scare tactics for money.

Disgusting.

Quite aligned with my own experience from harness engineering and winning AI4Science hackathon. During the hackathon I was working as a human optimizer, moving the feedback from test harness running on Claude Code, back to my local Claude Code for analysis-hypothesis-proposal cycle. And in this moment I realized that 2 Claudes talking to each other could actually scale much better.

I read most of the article and came to the conclusion that if what they're describing is so revolutionary, then why do they still need to hire people? Why not just have these systems take full control?

My biggest question (maybe this has already been taken care of) is the issue of garbage in and garbage out. If the LLM produces bad content then that is used to train another model, how do we stop them from keeping their blindspots across models?

This reads like marketing fluff, but I am reminded of John von Neumann's "Theory of Self-Reproducing Automata"; that the very first people who worked on deductive machines immediately started thinking about machines building themselves, and what the rules of that would look like. I am not surprised that during the inductive revolution we are having similar thoughts.

Not the first time. There were calls for NPT treaties etc over the decades. It is irreversible by design. Competition and ownership is the driving force.

I'm so sick of this anthropics marketing stuff... claude is an ultra-success (according to claude judge), “good code”, bragging about creating 8x more bugs and tech-debt. claude writes code that works, yeah, sure anthropic, we saw that claude code leaks, some amazingly "good" code in there

its vital for them to have self validation for exponential rsi.. and this human distillation of human in the loop debugging ai models is needed even though they have judge models handling parallel speculative execution.

labs have parallel speculative execution. they spawn hundreds of agent branches, validate them internally with AI judges and only show the user the successful result.

free users are using sequential single-turn generation. the model requires and waits for the human to debug, fix and re-prompt.

by forcing a human to act as validator. they are capturing high value correction trajectories (Bad Output --> Human fix). They are using your cognitive labour to train judge models and validator agents needed to automate the internal verification step, eventually closing the loop for fully autonomous recursive self-improvement.

human in the loop debugging isn't a bug; it's the necessary training signal for the self-validating agents required for exponential recursive self improvement. With new 'distilled judge' models landing in 2026, this article means that they might have gathered enough data. we might be in the final phase..

To anyone who works at anthropic : I recently downgraded from Max to Pro out of frustration. Last few weeks my token(usage) burn was just too fast and I couldn't explain it because my actual usage was less than the last few months. I ended up thinking it's probably a bug that you guys shipped. The above article makes me think that it's probably claude who shipped the bug and your human missed it in their review.

Seeing the words "recursive self-improvement" I was expecting something else from the article. E.g. how the transformer architecture or agent design is being changed/improved through LLM automation, but the article mostly talks about the LOC counts.