The binary is running. Maxwell is watching it: presumably there is some cost to the watching, but the LLM is not running in the binary process. Maxwell also has dynamic information about the running environment. Maxwell can then try and superoptimize the running binary using the dynamic information it has, but to do so, it needs to create semantics-preserving patches to the running binary.
What should be clear here: dynamic information is strictly more information than what a static AOT compiler has access to.
So that's the problem: we're going to patch the binary while it is running, and we need the patches to preserve the semantics as expressed by the high-level program, possibly assuming that the compiler was semantics preserving, etc.
I'm assuming they have custom compilers or something for all of this.
Getting this to work probably involves creating some sort of source map, perhaps a custom linker, etc (just guessing) from high-level program to binary -- but assuming you handle all of that, then the job of Maxwell is just the job of a normal superoptimizer: propose semantics-preserving patches, see how they do, etc.
In theory, you could propose terrible patches frequently - superoptimization normally discards these, but you with bad proposals you might spend a long time searching for good patches. But LLMs are good proposers now, so you'd probably mostly get good patches (to a limit).
I fully understand the merits of AOT vs JIT, I've worked on compiler tech, built JIT engines, and done lots of high performance realtime stuff for decades.
But if:
- we're talking about compiled code, where the compiler has already squeezed out all the low hanging fruit in terms of performance
- it's running some kind of external process somehow watching the data flow around in memory, and is figuring out what code is doing what. This is at best incredibly expensive, even if happening in parallel
..then the chances of actually finding useful micro-optimisations that are also safe to patch while running without changing the memory layout and messing up.. is surely pretty close to zero.
I can only assume that it's not actually native code, and they've built some kind of python JIT engine which is finding efficient replacements for chunks of common python, where it starts off so slow that even the massive amount of effort required manages to pay off in a long enough run.
- superoptimization specialized to a particular piece of hardware
- patching binaries while they are running
all great
That being said, I'm getting some weird vibes from the prose on this site. Little phrases and claims that seem intended for the purposes of marketing, but conflate things with clear technical meaning in confusing ways.
For instance: I watched the video ("a code execution model has to be right every time") which implicitly communicates something like "the model cannot hallucinate" ... of course, the agentic LM is still probabilistic, but it is wrapped with verification tools (to reason about semantic equivalence), and is likely post-trained to be good at using those tools.
This strikes me as a conflation between the model and the harness which is intended to strike a marketing point (like: "it doesn't hallucinate") but which is clearly a ridiculous claim to make (falling down the slope: any harness which does error correction for improper tool calls might also be a "model which makes no mistakes")
Also: "self-improving code"
that's not really it, is it? Self-improving code would be code that runs itself and improves itself. This is code I run, which Maxwell watches and improves ... no?
Related: I dislike the usage of the word "model" in the above discussions on the site. Do we want to call everything a "model" now? Can anyone invent a new frontier "model" by post training an agentic LM with tools?
Also, let's say you are optimizing kernels on GPUs. This can be subtly fraught: for instance, it was well-known for awhile that one should only use certain pieces of software (vLLM, for instance) to reproduce papers (which used vLLM) there are non-deterministic sources of randomness (including, of course, floating point) that would prevent one from recovering the performance claimed in a paper ...
I agree with you that Maxwell is like a JIT (with superoptimization inside of it: a very smart one which can propose smart things using the underlying agentic model, and reason about semantic equivalences to figure out what to commit or not) and that that is a natural analogy.
But people who write JITs don't call the overall process (original code, JIT) "self-improving code", do they?
What is the "self" in (original code, JIT)? Is it the entire process? But the JIT doesn't make itself better as it runs? Is all code that runs in a managed runtime with dynamic analysis self-improving?
A much better term might be "dynamically-improving code"
But related and to the point of my original message: the phrase "self-improving" has been used in vogue by AI research / labs ... and this is another example of something which is not a clear term, and seems intended for marketing
I'm confused.. are you suggesting that Claude Code / Codex don't work? Because if you're still saying that in October 2026, it's a you problem. You're doing something wrong.
- Prime Intellect (and all their agent experiments)
- Geoff Huntley (see Jiti, for instance)
(many more)
There's a ton of interesting software being developed with these models by people outside the in group, but I find most of the software from these big guys to be ... bland. Buggy copies of copies.
I don't see the Cloudflare stuff as groundbreaking, let alone people building companies around it. "Workers" are the wrong primitive, IMO. What if my code isn't in Javascript? They sandboxed the wrong part of the machine.
I have a version of "Cloudflare OS" running in production, using Temporal as the orchestration and MicroVMs for the sandboxing.
Sorry, I get your point but personally I don't get the hype around them.
I don't use either myself but I'm in a position where I get to see what it produces for other people. Here's a recent (2-3 months old) example: A zsh script that immediately invoked a Python interpreter that immediately invoked a shell command and iterated over each line of output looking for one of a few strings to match... So, ` | grep "(stringA|stringB|stringC)"`, but a whole lot dumber.
Thanks, but no thanks, I don't need that kind of code in anything I'm working with.
The killer software is one-off custom things that are written for an audience of one user, or maybe a small handful of them. The software is on their computers, where you don't see it.
None of it does anything particularly amazing, but that's kind of the point -- the programs that people are generating are for the use cases that are too personal, trivial or specific for anyone to have bothered making a software package for.
There's several things I disagree with in this post (a matter of perspective, I think):
> The discourse around software engineering has gotten dumber. It’s like everyone in the industry lost 30 IQ points. People used to talk about compilers, type systems, logic. Now they talk about “prompts”, “harnesses”, “loops”. The discourse is narrower, shallower, and more repetitive. There’s only so many times I can hear about “agentic harnesses” before I lose my mind.
I agree with the frustration here :)
> You design a revolutionary new programming language: who cares? Maybe Claude cares, for what that’s worth. But humans don’t write or even read the code anymore. The programming language is an implementation detail the humans no longer have to care about.
Perhaps the people who weren't inclined to think about compilers, type systems, and logic before coding agents will never think about those things. If anything, I think about those systems significantly more now, because I view them as ways to describe the search space for the agent ... and perhaps, in the future, even tools to structure agentic latent representations.
So I propose rotating your viewport: yes, perhaps entertain the idea that it doesn't make sense to think about these things for human consumption ... but instead, what if we think of programming language and compiler design as being a tool to control the search space, and suddenly there's lots of new ways to think about these things.
Is this a valuable way to think? I'm not sure, but we should try and find out.
I agree with you and disagree with the OP in similar ways.
I work on my own "agentically-engineered" projects and the pursuit of certain ideas for what you're calling intrinsic motivation, and I don't particularly care if others find it interesting or even find it at all or not, because the ideas themselves are cool to me.
It's kind of like I can suddenly tinker at 2,000% efficiency and share what I'm doing, _and if someone else finds it interesting, that's cool too_
Isn’t the point you’re making what OP is saying - that there are of course some humans who have deep intrinsic motivation, and they’ll continue to do what they do, but most humans aren’t wired this way, and for that, society will collapse?
There are obviously exceptions (e.g., the specification for the Java programming language is well over 100 pages [1] and is extremely low on filler), but most ideas don't work like this. In my area, 10 pages really is enough most of the time.
I love these commercials where someone has chosen to show how the AI product will essentially be used to slop out some garbage piece of corpo communication ... and the user basically says "looks good" with barely any thought, then mixed with some sort of real life thing (wedding planning here).
I mean, I feel like I'm going crazy -- but I was struck by this jarring blending of experiences ... shitting out some growth plots followed by autopilot on your wedding. Nice OpenAI. The only thing missing is a moment of self-reflection where I contemplate where exactly I lost what makes me ... me.
Is this what SV wants the world to look like? Mixing fucking cake batter while a bot shows me a regression to the mean website? Pretending like I have any sort of intentionality in my life, while a nameless entity (given quirky form) sort of walks me through my life?
I'm not sure why it gave me this impression, but strikes me as vaguely reminiscent of soma (from Brave New World).
Kind of sad, because the tech is actually incredible: who are they hiring to storyboard these commercials?
Meat proxy at work, meat proxy in your personal time. The utopian visions of AI futures strike me as alternative forms of hell. Humanity accomplishes everything and it means nothing, actions have no real consequence, the bumpy friction and individuality of life smoothed to a plane of perfect optimization.
Ultimately they are selling a lifestyle. It's success without the need to have any of the skills or knowledge of a successful person.
The famous John Deer lifestyle ad from years back was selling running a successful farm without actually farming; it's some other schmuck out in the fields on the tractor.
With dots you can be a shot caller, a taste maker. Forget learning things; dot will learn for you. Forget making things; dot will make for you.
Notwithstanding the duplication of these posts across social media, OpenAI, and Anthropic ("it's all the model, they just tell it to keep going" if you beat anything with RL enough ... it's going to do the thing)
Here's my issue with this post:
> True to the spirit of the challenge, they didn’t use millions of dollars in computer power. They used Fable 5.1, working within Claude Science, a platform scientists can pay to use.
Okay, billions of dollars have been poured into these agentic LMs, right? Each training run to get the next increment is costing millions of dollars?
This feels like an obvious jab at Navier-Stokes, but where we get to shift the numbers around to hide where the compute actually is being spent ... compute is being spent. It's either being spent in amortization to make the search smarter ahead of time, during training, or its being spent after.
Also love: scientists get to pay Anthropic to work within their special science harness to do science. That's exactly what I dreamed of doing when I pursued physics in undergrad, one or two companies holding the keys to "progress" for a monthly subscription price.
> Also love: scientists get to pay Anthropic to work within their special science harness to do science. That's exactly what I dreamed of doing when I pursued physics in undergrad, one or two companies holding the keys to "progress" for a monthly subscription price.
I get this anxiety, and am largely an AI skeptic, but at one point there were only a handful of computers in the world too (same for batteries, or engines, or crucibles, or stills -- it goes way back), and the organizations that had them had a stranglehold on progress in the field, as did the small number of companies who knew how to make them. It got better as they got cheaper and more plentiful.
I guess my point is that there are more important anxieties to feed when it comes to LLMs and the current state of the world.
> This feels like an obvious jab at Navier-Stokes, but where we get to shift the numbers around to hide where the compute actually is being spent ... compute is being spent. It's either being spent in amortization to make the search smarter ahead of time, during training, or its being spent after.
I think that argument is recursive? These posts aren't very complicated for either of us, but they're written on devices that are fabricated with billions of dollars of semiconductor equipment. At what point do we just acknowledge that we stand on the shoulders of giants?
To me, the distinguishing factor is that the expense not special-purpose but upfront. The model here is trained without foreknowledge of what problems it will solve. Solutions like nine loops are genuine expressions of a pre-existing model capability, even if that capability has not pre-existed for very long.
Oh I agree with you! I just don't see "we're standing on the shoulders of giants" in most of these marketing blog posts?
If I'm wrong here, I'd love reference links. I think of these companies as trying to inspire the idea that Claude (or GPT) are these special alien entities, in a sense?
> Also love: scientists get to pay Anthropic to work within their special science harness to do science. That's exactly what I dreamed of doing when I pursued physics in undergrad, one or two companies holding the keys to "progress" for a monthly subscription price.
The moat is very limited. Harnesses aren't crazy hard to engineer. Open models are quite capable.
Divide the compute of your argument (pre-training, training, post-training) through everything this model now can do and it will not look that bad at all.
One of the most interesting aspects of MiMo 2.6 is that they shared their RL costs, sub $5M (yes, million)
This stuff is going to get a lot cheaper, just like Stable Diffusion
I for one find the incessant side questing the moment something goes awry to be very annoying. Please stop and ask the human for clarification or guidence
The training costs get rolled into the usage prices. It's not hiding anything to talk about the cost of usage without the cost of training, any more than I'd be hiding something by talking about the price of a $1,000 CPU without mentioning the tens of billions of dollars in R&D and infrastructure needed to create it.
What a weird thing to be butthurt about. You don't have to pay them. Do it the old fashioned way, with elbow grease and pots of coffee. Or invest in local AI.
Or embrace the future and realize that you couldn't imagine everything that would unfold, when you pursued your undergrad.
There was a time when a generation of hackers got (rightfully) worked up about the Microsoft tax for every PC sold. It's not hard to see that somebody gets mad if they believe that an Anthropic/OpenAI tax is about to become the standard when doing any serious Physics/Math/... research or writing code.
I'm always surprised by statements about "things we don't need to learn anymore" with AI.
It's sort of a litmus test for how little someone understands how learning in humans works or AI itself even. "Agentic" AI acts as a mirror: the more insight have yourself, the more powerfully you can pull the distribution ... the more you get out of it. (there's a bloodborne-inspired parable in here somewhere ...)
I think it's why the software from AI companies has "garbage" qualities. They are assuming that AI is going to fill in skills ... when it is a reflective amplifier.
But back to the discussion of "things we don't need to learn anymore": do we go to the gym, and say ... oh I don't need to do pull ups anymore, because there's an assisted pull up machine and I can just set it to my weight ...
> "Hey Claude, can you be smarter than me and tell me what I should tell you to get the thing I don't know I want yet?"
I'm so glad you asked! Many smart people are now eating Snickers Bars(tm)! For you, I think, to get ahead of this trend, it would be the perfect smart person choice right now. It's nutty and has this amazing new invention called Nougat and chocolate!
Here's how it works. First you have to use your fingers to tear open the wrapper. Don't worry, the wrapper isn't going to hurt you. It might be a little effort, but you'll get it. Then you take a bite. Try to use your back teeth to chew it a little. Yeah, I know it's work, but it's good for you. And then just swallow it down!
The binary is running. Maxwell is watching it: presumably there is some cost to the watching, but the LLM is not running in the binary process. Maxwell also has dynamic information about the running environment. Maxwell can then try and superoptimize the running binary using the dynamic information it has, but to do so, it needs to create semantics-preserving patches to the running binary.
What should be clear here: dynamic information is strictly more information than what a static AOT compiler has access to.
So that's the problem: we're going to patch the binary while it is running, and we need the patches to preserve the semantics as expressed by the high-level program, possibly assuming that the compiler was semantics preserving, etc. I'm assuming they have custom compilers or something for all of this.
Getting this to work probably involves creating some sort of source map, perhaps a custom linker, etc (just guessing) from high-level program to binary -- but assuming you handle all of that, then the job of Maxwell is just the job of a normal superoptimizer: propose semantics-preserving patches, see how they do, etc.
In theory, you could propose terrible patches frequently - superoptimization normally discards these, but you with bad proposals you might spend a long time searching for good patches. But LLMs are good proposers now, so you'd probably mostly get good patches (to a limit).
Authors, correct me if I'm wrong.
reply