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

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.

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.

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.

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

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.

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.

"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.

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 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.

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.

  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.

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

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

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

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

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

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.

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."

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.

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

I disagree with this. Good code is easy to change, which is much harder to accomplish than code that can be added to.

"If technical trends in advancing capabilities continue, and AI systems are able to develop the capabilities inherent to transformative human ingenuity, then it is plausible that AI systems could design and refine themselves."

I find the first premise weak and implausible, and the second one is obviously false. To me it comes across as an insult to the reader.

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).

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

Or is this?

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

Elon, is that you? [1]

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

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.

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

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..

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.

You will forgive me when, between muted snickers, I express considerable doubt that Anthropic will be able to bring its AI to a point of "self-improving" any time soon.

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.

If was used in writing the article, why not list it? If it wasn't used, that seems to go against Anthropic's whole message.

Obviously readers value human-written content more, but isn't it their interest to attempt to destigmatize llm output as much as possible?

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

Aye.

Living organisms evolve towards some notion of "better", and "better" is an incredibly multifaceted notion (many facets of which we simply cannot even capture in language).

The metric being tracked, code commits, is hilariously one sided. Philosophically, if you had one part of your work now practically free, you'd like to utilize that freedom to maximally cover for the other parts, for instance:

Instead of thinking about edge cases with brain and whiteboard, you can have the LLMs to simply generate most possibility including tests for it, because that is cheaper. There's probably 50x more commits of which 40 will be revert pairs but we are only twice as fast. And in reality nothing did change because the outcome remain the same. I can't see how it is necessarily different in the LLM space.

https://safebots.ai/declarative.html

I for one, believe that we should pause all work on AI for the forseeable future. This is almost impossible to orchestrate - but we should still try nevertheless. Maybe we are not able to pause, but we are able to slow down. That might give us more room, to maybe able to pause in the future. But going ahead is too dangerous.

And its not just Anthropic which is saying this. Even Geoffry Hinton has said the same thing. If there is a non-zero chance that AI can kill all of humanity, and both Geoffry and Anthropic have the same position, then it makes sense for us to be hundred percent sure before we move ahead. Dario/Anthropic have already made their money from AI, maybe they are just being honest about what they think lies ahead.

One of my focuses now is my own model-agnostic, harness and workflow orchestration (I know everyone is building these) , baselining on opus, and aiming to transition to Chinese models like deepseek in the short term and hopefully open, self hosted models in the future (which I plan to open source).

The nonstop marketing fluff from anthropic while their service quality and availability noticeably degrades... just continues to destroy my trust in the company.

One of the examples they provide, of giving Claude the task of training a small AI model, then asking it to improve certain benchmarks, is essentially Karpathy's AutoResearch. This is already known to work. While calling it "self-improvement" is perhaps a stretch, it is describing a capability current gen AI has, that anyone can test and I have been using to great effect.

I disagree with their conclusion, I think this kind of self-improvement will hit an asymptote, where every subsequent model can only make smaller and smaller improvements.

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

Month 1 - 6 months to AGI

Month 2 - We will Replace all jobs

Month 3 - Okay maybe only the SWEs, programming is solved

Month 4 - Announce model that is too dangerous to release

Month 5 - Releases dangerous model

Month 6 - This is it! We will replace AIs with more AIs (*secretly files for IPO)

AI is here to stay, like it or not but it is not the solution to everything. If it is, what is Anthropic's moat? A better model? I don't see any ecosystem being built by them, as MCP is almost obsolete except for some very niche use case. And they're doing stuff that a non-profit version of OpenAI would do. Can we trust a for-profit company to stand against their investors during a conflict of interest? Because running a company for maximum profit versus being ethical is two different end of the spectrum.

If they wanted to they could have convened an international forum with commercial and political stakeholders years ago. Less talk, more do.

I simultaneously think the AI revolution is making real revolutionary gains and am mystified by the lying.

An accurate Translation seems to be “we made this shit up, but it feels right”

If we ever get to a point where the centaur period is over (when human + AI is not better than just AI) then what competitive advantage ANY human can have other than

- the money they already have

- luck?

- a good idea and good taste but if we assume AI can do better than any human, that also goes out the window

So, this whole singularity goes into a place where no one is really needed, the only thing that will "save us" (other than "The Expanse" like world / UBI) is if there will be no demand to the supply of AI work. Even if it's better. (example is - there is demand to seeing Magnus Carlsen play, there is no demand to the Stockfish on my phone getting into a stalemate with another Stockfish on another phone. Also people like to watch humans compete with humans, there is no demand to see a race between Usain Bolt and a rocket). So if people will not buy AI generated stuff (we'll get to a point where everyone will assume something AI generated because AI might get to a point where it is not as easy to identify it. E.g. it will stop looking like slop... but I believe services that give you a "human generated" 3rd party evidence can happen, again all based on supply and demand...)

So as we near singularity... All it takes is one open weights model, and one open harness that is capable of self improvement, and Anthropic's entire moat is gone. That open weight model might even be built with Claude Code + Mythos (once it's released).

But don't worry, all moats will be gone and we'll all just do yoga, read books and connect to each other because AI will produce everything for free using renewable energy, right? Or we'll all become batteries in a simulation, probably something in between.

This whole set of imaginary scenarios is based on a single company writing code that isn't even that complicated and represents a single product line for a single company in a single industry. You might wanna see this replicated in at least one other scenario first before you call it on the AI gods enslaving humanity. These imaginary scenarios also depend on a logistical, financial, & geopolitical system that is unsustainable & will be curtailed in the near-future one way or another.

They keep referring to this as intelligence - it isn't. It can't actually learn. It can just code in a loop. That isn't learning. It can't do real RL with meaningful persistent semantic memory in a realistic timeframe or cost, and it can't reason accurately outside of predetermined scenarios (hell, most of the models still can't tell time). It still can't do what a 4 year old can do. So let's cool it on the dreams of benevolent god-machines or whatever.

The tech industry has been a farce for years. We sit here in this bizarre artificial echo chamber and imagine that the whole world revolves around us, when in reality the whole world is limited by us. If a recursive self-improvement loop replaces us all, it will be a boon to the world, as the world won't be limited by this industry's stupidity anymore. But considering that the world is not actually run by tech bozos, harms and uncertainties brought by AI will be pushed back on and reigned in by normal people, as always happens with new technologies. An AI can't engineer its way around politics. The self-improvement loop is just as likely to be outlawed as it is actually working outside of Anthropic's walled garden.

Shifting their focus from Training new models to instead serving inference, they would greatly reduce their spend. In fact this is something being reported on that they are already doing, which is the reason for their first ever profitable quarter.

Its awfully convenient that the company which has greatly reduced its spend on training is now asking for a slow down in this area.

Come on guys...

That is making me less impressed not more impressed!

To me, unattended agentic coding is not RSI, in the same way a self-reloading "Unattended 3D printer" is not at all a "3D printer that recursively prints complete 3D printers in which each generation is significantly faster and more advanced than the last." The "unattended" part is obviously necessary but hardly sufficient. The article tacitly assumes LLM progress to be something like 1: Unattended agentic coding, 2: AGI, 3: RSI. I suspect that third step should be labeled "not to scale."

I'm increasingly convinced that actual Full Foom RSI (FF-RSI) is on a radically different scale than the first two. Just leaving it unaddressed is like assuming: Step 1: Manned space station, Step 2: Manned Mars base, Step 3: Manned Alpha Centauri base, are "just logical next steps." FF-RSI requires sustaining superlinear, recursively amplifying cognitive returns along a specific directed path - and we currently have no empirical evidence that such returns can exist for artificial OR biological intelligences. Large collectives of the smartest humans alive (Bell Labs, IAS, etc) haven't just failed to get anywhere close to reliably sustaining that, we can't even reliably predict non-recursive, single occurrences or even imagine any way all 8B humans could fully mobilize to predictably achieve non-recursive, single occurrences.

The only prior we have for open‑ended intelligence improvement is biological evolution which shows extremely slow and unreliable sublinear returns at best. And even if unbounded, recursive self‑improvement is physically possible, it may be practically unachievable due to asymptotic economic, resource and other barriers in the same way approaching light speed requires exponentially more energy. I think it's plausible, and maybe probable, that AIs achieve true super-human intelligence in a decade and yet still won't achieve FF-RSI for centuries, if ever. To me, absent compelling evidence to the contrary, that's the reasonable Null Hypothesis. Even if you feel that's too pessimistic, it seems reasonable to expect any serious discussion of "Progress Toward RSI" to first discuss why it might even be plausible that 1: Miles, 2: AU (Astronomical Units), and 3: Light Years belong on the same scale, instead of just assuming it like the meme's empty "Step 3. .... " before moving on to "Step 4. Profit!" (or "IPO!" but very, very responsibly).

I really can't stand these guys anymore...

Even Anthropic wants to Pause AI now. There must really be not much time left for "edging". Please write to your lawmakers, no matter whether you are in the US, Europe, China, or elsewhere. Only an international agreement between governments can enforce an AI-Pause and eliminate the necessity to dangerously push the frontier.

https://pauseai.info/

https://intelligence.org/agi-ruin/