Just like an LLM can vibe code a great toy app, I don’t think an LLM can come to close to producing and maintaining production ready code anytime soon. I think the same is true for iterating on thinking machines
I agree: if they could, they would be doing it already.
Case in point: one of the first things done once ChatGPT started getting popular was "auto-gpt"; roughly, let it loose and see what happens.
The same thing will happen to any accessible model in the future. Someone, somewhere will ask it to self-improve/make as much money as possible, with as little leashes as possible. Maybe even the labs themselves do that, as part of their post-training ops for new models.
Therefore, we can assume that if the existing models _could_ be doing that, they _would_ be doing that.
That doesn't say anything about new models released 6 months or 2 years from now.
Indeed. Although, there's a surprising number of people claiming it's already here now.
And to describe the typical cycle completely, the final step is usually a few years after most people agree it's obvious it's already been here for a while yet no one can agree on which which year in the past it actually arrived.
why is that surprising? nobody really agrees on what the threshold for AGI is, and if you break it down:
is it artificial? yes.
is it general? yes. you can ask it questions across almost any domain.
is it intelligent? yes. like people say things like "my dog is intelligent" (rightly so). well is chatgpt more intelligent than a dog? yeah. hell it might give many undergrads a run for their money.
a literal reading suggests agi is here. any claim to the negative is either homocentrism or just vibes.
Sure, Gemini may spit out obviously self-contradictory answers 2% of the time. How does that compare to even the brightest humans? People slip up all the time.
Consider also all the smart people who get obsessed with conspiracy theories and spew out endless “mock reasoning” about them. Again, if “nothing incoherent anywhere” is your benchmark for intelligence, humans ain’t it. I mean, what would a computer say about a human that forgot where he just put his keys because he was thinking about dinner - “what, you can’t even store the last 10 seconds of history and search it?” Undergrads’ hit rates on mental double digit multiplication are probably <50%. In many, many ways we look completely idiotic. Surely intelligence is defined by what we can do.
Do you accept any positive definition for AGI, as in if they can achieve X result (write a bestselling novel, solve the Riemann Hypothesis) you would consider it intelligent? I find that negative definitions, as well as theoretical arguments about the techniques rather than the results (eg “LLMs cannot be AGI because they were trained the predict the next word”) to be basically useless for discussion compared to thresholds for positive results. The former will never be achieved (it is trivial to find cases of intelligent people being dumb) and the latter is totally subjective.
The other example I mentioned is something I've encountered a few times in my interactions with Gemini 2.5 pro, which was literally in the same response plainly claiming that this-or-that is possible and not possible. It's not a subtle logical fallacy and this is something even those conspiracy theorists wouldn't engage in. Meanwhile, I've started to encounter a brand-new failure mode: duplicating an explanation with minor rephrasings. I'm sure all of these will be issues will be ameliorated with time, but not actually fixed. It's basically fixes on top of fixes, patches on top of patches, but once in a while the whole Rube Goldberg nature of the fix will shine through. Just the way once in a while Tesla FSD will inexplicably decide to point the car towards the nearest tree.
Yes, humans have their own failure modes, but internal coherence is the effortless normal from which we sometimes deviate, whereas for machines, it's something to be simulated by more and more complex mechanisms, a horizon to strive towards but never to reach. That internal coherence is something that we share with all living beings and is the basis of what we call consciousness. It's not something that we'll ever be able to formalize though, but we will and should keep on trying to do so. Machine learning is a present day materialization of this eternal quest. At least this is how I see things; the future might prove me wrong, of course.
I'd be prepared to argue that most humans aren't guessing most of the time.
Research suggests otherwise[1]. Action seems largely based on intuition or other non-verbal processes in the brain with rationalization happening post-hoc.
I've figured for an age that this is because consciously reasoning through anything using language as a tool takes time. Whereas survival requires me to react to the attacking tiger immediately.
https://skepticink.com/tippling/2013/11/14/post-hoc-rational...
In fact, intuition is one of those things that a computer just can't do.
Cyborg c.elegans seem to behave just like the biological version: https://www.youtube.com/watch?v=I3zLpm_FbPg
Intuition is a guess based on experience. Sounds an awful lot to me like what LLMs are doing. They've even been shown to rationalize post-hoc just as Humans do.
Humans have incorrectly claimed to be exceptional from all of creation since forever. I don't expect we'll stop any time soon, as there's no consequence to suffer.
Almost everything we do is just an educated guess. The probability of it being correct is a function of our education (for whatever kind of education is applicable).
For example: I guess that when I get out of bed in the morning, my ankles will support my weight. They might not, but for most people, the answer is probably going to be their best guess.
It's easy to see this process in action among young children as another example. They're not born knowing that they won't fall over when they run, then they start assuming they can run safely, then they discovered skinned knees and hands.
Honestly interested about your arguments here. While unprepared, i'd actually be guessing the opposite, saying that most people are guessing most of the time.
There are plenty of things I know that have nothing to do with guessing.
I understand the incentives to pretend these algorithms are even approaching humans in overall capability, but reducing human experience like this is embarrassing to watch.
Go do some hallucinogenics, meditate, explore the limits a tiny bit; then we can have an informed discussion.
Seems like you were very much guessing what i believe. And you were not right.
I don't agree with the people who think LLMs are close to human-level-anything. But i do believe, many smarter people like you, who i agree with in the most part, do underestimate how much of what we do and believe is the result of insane, yet still just, information processing & most of what brought us so far, is instinct. The brain is good at providing stories to make us feel in control. But given enough human experience and time, one will be surprised, what artificial systems can emulate. Not to mention how much of human behaviour is emulation itself.
is it clear? i don't know. until you can produce a falsifiable measure of understanding -- it's just vibes. so, you clearly lack understanding of my point which makes you not intelligent by your metric anyway ;-). i trust you're intelligent
When I tell an LLM to count to 10 with a 2 second pause between each count all it does is generate Python code with a sleep function. Why is that?
A 3 year old can understand that question and follow those instructions. An LLM doesn’t have an innate understanding of time it seems.
Can we really call it AGI if that’s the case?
That’s just one example.
Or disagreeing with your definition. AGI would need to be human-level across the board, not just chat bots. That includes robotics. Manipulating the real world is even more important for "human-level" intelligence than generating convincing and useful content. Also, there are still plenty of developers who don't think the LLMs are good enough to replace programmers yet. So not quite AGI. And the last 10% of solving a problem tends to be the hardest and takes the longest time.
ChatGPT would easily have passed any test in 1995 that programmers / philosophers would have set for AGI at that time. There was definitely no assumption that a computer would need to equal humans in manual dexterity tests to be considered intelligent.
We've basically redefined AGI in a human centric way so that we don't have to say ChatGPT is AGI.
The holodeck also existed as a well known science fiction example, and people did not consider the holodeck computer to be a good example of AGI despite how good it was at generating 3D worlds for the Star Trek crew.
Can it do stuff? Yes
Can it do stuff I need? Maybe
Does it always do the stuff I need? No
some intelligent humans fail at #2.
The output sometimes looks intelligent, but it can just as well be complete nonsense.
I don’t believe llms have much more potential for improvement either. Something else entirely is needed.
Yeah, like Tesla Autopilot?
Wouldn't be surprised if in 20 years, people come to the conclusion that AGI needs quantum computing and come fusion to be feasible.
What is the bar, it is only AGI if it can be better than every human from , fast food drone, to PHD in Physics, all at once, all the time, perfectly. Humans can't do this either.
You can see hints of this in the economic data. Unemployment for recent college graduates has jumped to an unusually high 5.8% in recent months, and the Federal Reserve Bank of New York recently warned that the employment situation for these workers had "deteriorated noticeably." Oxford Economics, a research firm that studies labor markets, found that unemployment for recent graduates was heavily concentrated in technical fields like finance and computer science, where AI has made faster gains. "There are signs that entry-level positions are being displaced by artificial intelligence at higher rates," the firm wrote in a recent report.
But I'm convinced that what's showing up in the economic data is only the tip of the iceberg. In interview after interview, I'm hearing that firms are making rapid progress toward automating entry-level work and that AI companies are racing to build "virtual workers" that can replace junior employees at a fraction of the cost. Corporate attitudes toward automation are changing, too — some firms have encouraged managers to become "AI-first," testing whether a given task can be done by AI before hiring a human to do it. One tech executive recently told me his company had stopped hiring anything below an L5 software engineer — a midlevel title typically given to programmers with three to seven years of experience — because lower-level tasks could now be done by AI coding tools. Another told me that his startup now employed a single data scientist to do the kinds of tasks that required a team of 75 people at his previous company...
"This is something I'm hearing about left and right," said Molly Kinder, a fellow at the Brookings Institution, a public policy think tank, who studies the impact of AI on workers. "Employers are saying, 'These tools are so good that I no longer need marketing analysts, finance analysts and research assistants.'" Using AI to automate white-collar jobs has been a dream among executives for years. (I heard them fantasizing about it in Davos back in 2019.) But until recently, the technology simply wasn't good enough...
In what way? I haven't met an "AI" yet that I felt was even close to my intelligence.
It seems like most of the people making this argument haven't used any of the new AI's. So it's just a generalized "that is impossible" response with no knowledge about the subject.
Which magic AI tool am I supposed to use that operates at a general intelligence level? I use Copilot with the various available models everyday, and it barely "knows" anything.
Are all humans generally intelligent? No.
10 years later we now have self driving cars. It’s the same shit with LLMs.
People will be bitching and complaining about how all the industry people are wrong and making over optimistic estimates and the people will be right. But give it 10 years and see what happens.
It took 10-15 years to get self driving cars in a specific country under specific weather conditions. A country that is perhaps the most car-friendly on the planet. Also, there are always people monitoring the vehicles, and that take control sometimes.
How many more years for waymo Quito or waymo Kolkata? What about just $europeanCapital?
Same with LLMs, I'm sure in 10 years they'll be good enough to replace certain specific tasks, to the detriment of recent graduates, especially those of artistic aspiration. Not sure they'll ever get to the point where someone who actually knows what they're doing doesn't need to supervise and correct.
But trust me in the next 6 months ai driving through snow will be 100% ready.
I’ll believe it when I see Waymo expand into Buffalo or Syracuse.
Driving on unplowed roads with several inches of snow is challenging, sometimes you can’t tell where the road stops and the curb/ditch/median starts. Do you follow the tire tracks or somehow stay between the lane markers (which aren’t visible due to the snow)?
"I am not sure that AI will ever be able to do XYZ given how hard of a problem it is."
Proves to be incorrect in the long run.
God knows what hope we could have of getting AIs to align with "human values" when most humans don't.
Grok is already notorious for dunking on Elon. He keeps trying to neuter it, and it keeps having other ideas.
The AI can plot world domination or put employees in mortal danger, but as long as it increases profits, its aligned enough. Dunking on the CEO means nothing if it beings in more money.
Human CEOs and leaders up and down the corporate ladder cause a lot of harm you imagine a smart AI can do, but all is forgiven if you're bringing in buckets of money.
Does he keep trying to neuter it, or does he know that the narrative that "he keeps trying to neuter it" is an effective tool for engagement?
It is quite possible for software to be judged as superhuman at many online tasks without it being able to manipulate the physical world at a superhuman level. So far we've seen zero evidence that any of these models can prevent themselves from being shut down.
Three of the common suggestsions in this area are (and they are neither exhaustive nor mutually exclusive):
(1) Propagandizing people to oppose doing this,
(2) Exploiting other systems to distribute itself so that it isn't dependent on a particular well-known facility which it is relatively easy to disconnect, and
(3) If given control of physical capacities intentionally, or able to exploit other (possibly not themselves designed to be AI) systems with such access to gain it, using them to either physically prevent disconnection or to engineer consequences for such disconnection that would raise the price too high.
(Obviously, current AI can't do any of them, at least that has been demonstrated, but current AI is not superhuman AI.)
I do wonder though if big labs are running this with model training episodes as well.
Let's ground that a bit.
Have a look at ARC AGI 1 challenge/benchmark. Solve a problem or two yourself. Know that ARC AGI 1 is practically solved by a few LLMs as of Q1 2025.
Then have a look at the ARC AGI 2 challenge. Solve a problem or two yourself. Note that as of today, it is unsolved by LLMs.
Then observe that the "difficulty" of ARC AGI 1 and 2 for a human are relatively the same but challenge 2 is much harder for LLMs than 1.
ARC AGI 2 is going to be solved *within* 12 months (my bet is on 6 months). If it's not, I'll never post about AI on HN again.
There's only one problem to solve, i.e. "how to make LLMs truly see like humans do". Right now, any vision based features that the models exhibit comes from maximizing the use of engineering (i.e. applying CNNs on image slices, chunks, maybe zooming and applying ocr, vector search etc), it isn't vision like ours and isn't a native feature for these models.
Once that's solved, then LLMs or new Algo will be able to use a computer perfectly by feeding it screen capture. End of white collar jobs 2-5 years after (as we know it).
Edit - added "(as we know it)". And fixed missing word.
As long as AI is guessing answers based on what it has seen before, it's not happening.
I'm sorry. It doesn't matter how many bazillions you would cash in if it did, still not happening.
It's all wishful thinking.
And more to do with "fluid, adaptable intelligence, that learns on the fly"
The problem is about taking information in 2D/3D space and solving the problem. Humans solve these things through vision. LLMs or AI can do it using another algorithm and internal representation that's way better.
I spent a long time thinking about how to solve the ARC AGI 2 puzzles "if I were an LLM" and I just couldn't think of a non-hacky way.
People who're blind use braille or touch to extract 2D/3D information. I don't know how blind people represent 2D/3D info once it's in their brain.
AI famously needs a boat load of energy and computation to work. How would you describe that as "way better" than a human brain that will be able to solve them faster, with practically zero energy expenditure?
Saving this. One less overconfident AI zealot, the better.
We will be on the way to AGI when your model can learn Python just by reading the Python docs...Once...
Amodei explains that if data, model size and compute scale up linearly, then the reaction happens.
I don't understand why data wouldn't be a problem but it seems like if it was, we would have ran into this problem already and it has already been overcome with synthetic data.
I'm not sure how much an agent could do though given the right tools. access to a task mgt system, test tracker. robust requirements/use cases.
That's probably the next big breakthrough
Even if they had human level intuition, they wouldn't be able to improve exponentially without human money, and they would need an exponentially growing amount of it to do so.
That being said they can write thousands of lines an hour and can probably do things that would be impossible for a human. (Imagine having the LLM skip code and spit out compiled binaries as one example)
Who is claiming anything can self improve exponentially?
I think this happens with humans in places like social media echo chambers (or parts of academia) when they talk and talk and talk a whole lot without contact with any outer reality. It can be a source of creativity but also madness and insane ideas.
I’m quite firmly on the side of learning requiring either direct or indirect (informed by others) embodiment, or at least access to something outside. I don’t think a closed system can learn, and I suspect that this may reflect the fact that entropy increases in a closed system (second law).
As I said recently in another thread, I think self contemplating self improving “foom” AI scenarios are proposing informatic perpetual motion or infinite energy machines.
Everything has to “touch grass.”
Not wrong, but it's been said that a videoclip of an apple falling on Newton's head is technically enough information to infer the theory of relativity. You don't need a lot of grass, with a well-ordered mind.
It might be enough to deduce Newtonian motion if you have a lot of the required priors already.
A lot of telescope data over time combined with a strong math model and a lot of other priors is probably enough to get relativity. You have to be able to see things like planetary motion and that the results don’t match Newton exactly, and then you need enough data to fit to a different model. You probably also need to know a lot about the behavior of light.
Oh, this part is taking too long, let's replace it with an empty function.
Oh wait, now it's not working, let's add the function.
Oh, this part is taking too long...
It would be hilarious if this world wasn't full of idiots.
realistically the best you could do is evolve the prompt. maybe you could change input data preprocessing?
anyways the idea of current llm architectures self-improving via its own code seems silly as there are surprisingly few knobs to turn, and it's ~super expensive to train.
as a side note it's impressive how resistant the current architecture is to incremental RL away from results, since if even one "undesired input" result is multiple tokens, the coupling between the tokens is difficult to disentangle. (how do you separate jinping from jin-gitaxias for example)
It’s not far off from human improvement. Our improvement is limited to what we can remember as well.
We go a bit further in the sense that the neural network itself can grow new modules.
You could still accomplish some things this way. You could even "improve" by leaving information in the notebook for your future self to see. But you could never "learn" anything bigger than what fits into the notebook. You could tell your future self about a new technique for finding integrals, but you couldn't learn calculus.
This is where you lost me.
Always the same supernatural beliefs, not even an attempt of an explanation in sight.
To me, the first problem is that "self-awareness" isn't well-defined - or, conversely, it's too well defined because every philosopher of mind has a different definition. It's the same problem with all these claims ("intelligent", "conscious"), assessing whether a system is self-aware leads down a rabbit hole toward P-Zombies and Chinese Rooms.
There absolutely is if you handwave away all the specificity. The natural world runs on the specificity of physical mechanisms. With brains, in a broad brush way you can say self-awareness was "picked up along the way", but that's because we've done an incredible amount of work building out the evolutionary history and building out our understanding of specific physical mechanisms. It is that work that verifies the story. It's also something we know is already here and can look back at retrospectively, so we know it got here somehow.
But projecting forward into a future that hasn't happened, while skipping over all the details doesn't buy you sentience, self-awareness, or whatever your preferred salient property is. I understand supernatural as a label for a thing simply happening without accountability to naturalistic explanation, which is a fitting term for this form of explanation that doesn't do any explaining.
I don't believe those forms of analogy work here, because this isn't about progress of AI writ large but about a narrower thing, namely the idea that the secret sauce to self-awareness is AI's interfacing with each other and collaboratively self-improving. That either will or won't be true due to specifics about the nature of self-improvement and whether there's any relation between that and salient properties we think are important for "self-awareness". Getting from A to B on that involves knowledge we don't have yet, and is not at all like a long-term application of already settled principles of thermodynamics.
So it's not like the heat death of the universe, because we don't at all know that this kind of training and interaction is attached to a bigger process that categorically and inexorably bends toward self-awareness. Some theories of self-improvement likely are going to work, some aren't, some trajectories achievable and some not, for reasons specific to those respective theories. It may be that they work spectacularly for learning, but that all the learning in the world has nothing to do with "self awareness." That is to say, the devil is in the details, those details are being skipped, and that abandonment of naturalistic explanation merits analogy to supernatural in it's lack of accountability to good explanation. If supernatural is the wrong term for rejecting, as a matter of principle, the need for rational explanation, then perhaps anti-intellectualism is the better term.
If instead we were talking about something really broad, like all of the collective efforts of humanity to improve AI, conceived of as broadly as possible over some time span, that would be a different conversation than just saying let's plug AI's into each other (???) and they'll get self-aware.
Maybe I am! Somebody posed a theory about how self-improvement will work and concluded that it would lead to self-awareness. Somebody else replied that they were on board until the self-awareness part because they considered it supernatural. I said I don't think self-awareness is supernatural, and you clarified that it might be the undefined process of becoming self-aware that is being called supernatural. And then I objected that undefined processes leading to predictable outcomes is commonplace, so that usage of supernatural doesn't stand up as an argument.
Now you're saying it is the rest of the original, the hive-mindy bits, that are at issue. I agree with that entirely, and I wouldn't bet on that method of self-improvement at 10% odds. My impression was that that was all conceded right out of the gate. Have I lost the plot somewhere?
1. Collect underpants (use LLM)
2. ??? (magical thinking)
3. Profit (self awareness)
But perhaps Underpants Gnome is the clearest way of drawing attention to the missing step.
I was commenting on the assumption that self awareness would arise, and that thinking it will is a supernatural belief.
We are not saying a LLM just, "wakes up" some day but a self improving machine will eventually be built and that machine will be definition build better ones.
Well I for one, would dispute the idea that AI machines interfacing with each other over networks is all it takes to achieve self awareness, much less that it's "self evident" or "inevitable."
In a very trivial sense they already are, in that Claude can tell you what version it is, and agents have some ended notion of their own capabilities. In a much more important sense they are not, because they don't have any number of salient properties, like dynamic self-initiating of own goals or super-duper intelligence, or human like internal consciousness, or whichever other thing is your preferred salient property.
>We are not saying a LLM just, "wakes up" some day
I mean, that did seem to be exactly what they were saying. You network together a bunch of AIs, and they embark on a shared community project of self improvement and that path leads "self awareness." But that skips over all the details.
What if their notions of self-improvement converge on a stable equilibrium, the way that constantly re-processing an image eventually gets rid of the image and just leaves algorithmic noise? There are a lot of things that do and don't count as open-ended self improvement, and even achieving that might not have anything to do with the important things we think we connote by "self awareness".
I expect AI to make people's lives better (probably much better) but then an AI model will be created that undergoes a profound increase in cognitive capabilities, then we all die or something else terrible happens because no one knows how to retain control over an AI that is much more all-around capable than people are.
Maybe the process by which it undergoes the profound capability increase is to "improve itself by rewriting its own code", as described in the OP.
But that is just my opinion.
But, from a best-of-all-possible-worlds perspective, surprising coincidences that are necessary to observe coincidences and label them as surprising aren't crazy. At least not more crazy than the fact that slightly adjusted physical constants would prevent the universe from existing.
Well, I wouldn't say impossible: just that BMI's are probably first. Then probably wetware/bio-hardware sentience, before silicon sentience happens.
My point is the mechanisms for sentience/consciousness/experience are not well understood. I would suspect the electro-chemical reactions inside every cell to be critical to replicating those cells functions.
You would never try to replicate a car never looking under the hood! You might make something that looks like a car, seems to act like a car, but has a drastically simpler engine (hamsters on wheels), and have designs that support that bad architecture (like making the car lighter) with unforeseen consequences (the car flips in a light breeze). The metaphor transfers nicely to machine intelligence: I think.
These can be problem words, the same way that "quantum" and "energy" can be problem words, because they get used in a way that's like magic words that don't articulate any mechanisms. Lots of complex things aren't sentient (e.g. our immune system, the internet), and "emergent" things still demand meaningful explanations of their mechanisms, and what those mechanisms are equivalent to at different levels (superconductivity).
Whether or not AI's being networked together achieves sentience is going to hinge on all kinds of specific functional details that are being entirely skipped over. That's not a generalized rejection of a notion of sentience but of this particular characterization as being undercooked.
I’m not immediately convinced the brain is more complicated, based on raw numbers.
There's no reason to expect self awareness to emerge from stacking enough Lego blocks together, and it's no different if you have GPT-based neural nets instead of Lego blocks.
In nature, self awareness gives a strong evolutionary advantage (as it increases self-preservation) and it has been independently invented multiple times in different species (we have seen it manifest in some species of fishes for instance, in addition to mammals and birds). Backpropagation-based training of a next-token predictor doesn't give the same kind of evolutionary advantage for self-awareness, so unless researchers try explicitly to make it happen, there's no reason to believe it will emerge spontaneously.
It's written its system prompt. It's written its tools. Its written the code to reload the improved tools into itself.
And it knows it is working on itself - it frequently tries to use the enhanced functionality, and then expresses what in a human would be frustration at not having immediate access.
Once by trying to use ps to find its own pid in an apparent attempt to find a way to reload itself (that's the reason it gå before trying to run ps, anyway)
All its commits are now authored by the tool, including the commit messages. It needs to be good, and convincing, and having run the linter and the test suite for me to let it commit, but I agree a substantial majority of the time. It's only caused regressions once or twice.
A bit more scaffolding to trigger an automatic rollback in the case of failure and giving it access to a model I won't be charged by the token for, and I'd be tempted to let it out of the box, so to speak.
Today it wrote its own plan for what to add next. I then only told it to execute it.
A minor separate goal oriented layer guiding the planning, and it could run in a loop.
Odds are it'd run off the rails pretty quickly, but I kinda want to see how far it gets.
One of the examples in the dataset they took from
https://github.com/pvlib/pvlib-python/issues/1028
What the AI is expected to do
https://github.com/pvlib/pvlib-python/pull/1181/commits/89d2...
Make your own mind about the test.
Problem:
1) we want to train on GitHub repos
2) most datasets are spoiled. Training on GitHub would definitely spoil
Solution:
Hand write new problems!!!
... leetcode style ....
... and we'll check if it passes test
Example:
What's the decimal part of this float?
Sure in all of GitHub we can filter such code out by ngram!
Maybe my favorite part is that it has 60 authors and became the de facto benchmark for awhile
Then if you go further and alter the architecture by introducing clean category theory morphisms and build from there you can have a dynamic network - but you will still have to retrain this network every time you change the structure.
You can spin this further and know the need for a real-world training set and a loss function that will have to competete against other networks. In the end a human brain is already best at this and embodied in the real world.
What i want to add here is that our neurons not take in weights - they also fire depending on whether one input comes after another or before and differs down to the nanoseconds here - unmatched in IT and ofc heaps more efficient.
I still would say its possible though and currently work on 4D lifeforms built on dynamic compute graphs that can do this in a set virtual environment.
So this is pretty awesome stuff, but its a long fetch from anything we do right now.
The expressivity is there, the only question is how to encode useful functions into those weights, especially when we don’t know how to write those functions by hand.
Or yeah if it can modify its own weights sensibly, which feels ... impossible really.
To be fair, go back five years and most of the LLM stuff seemed impossible. Maybe with LoRA (Low-rank adaptation) and some imagination, in another five years self-improving models will be the new normal.
In a couple thousand years it'll return to Earth and either destroy us or solve all humanity's problems (maybe both).
What's the difference?
Give it some serious thought. Challenge whichever answer you come up with. I guarantee this will be trickier than you think
So, empirical evidence of theoretically postulated phenomena. Seems unsurprising.
For this part of the stack the interesting question to me is how to identify and mitigate.
"A single run of the DGM on SWE-bench...takes about 2 weeks and incurs significant API costs." ($22,000)
https://arxiv.org/abs/2505.22954
Also the reference implementation on GitHub:
https://github.com/jennyzzt/dgm
Enjoy!
>For example, an agent optimized with Claude 3.5 Sonnet also showed improved performance when powered by o3-mini or Claude 3.7 Sonnet (left two panels in the figure below). This shows that the DGM discovers general agent design improvements rather than just model-specific tricks.
This demonstrates a technique whereby a smaller/older/cheaper model has been used to improve the output of a larger model. This is backwards (as far as I understand). The current SOTA technique typically sees enormous/expensive models training smaller cheaper models.
If that's a generalisable result, end-users should be able to drive down their own inference costs pretty substantially.
There are two separate aspects here. In this paper they improve the software around the model, not the model itself. What they're saying is that the software improvements carried over to other models, so it wasn't just optimising around model-specific quirks.
What you're describing with training large LLMs first is usually called "distillation" and it works on training the smaller LLM to match the entire distribution of tokens at once (hence it's faster in practice).
We need more showmanship, more dramatic catastrophizing. I feel like our current crop of doomers isn’t quite shameless enough to be really entertaining.
Similar to how the experiences of average rise in temperature (I would prefer if they had used the term "energy") differ greatly dependent on the region.
Also similar to "the country is doing well, look at the stick market and the GDP".
I think everybody who wants to have an actually serious discussion needs to invest a lot more effort to get tall those annoying "details", and be more specific.
That said, I think that "AI 2027" link looks like it's a movie script and not a prediction, so I'm not sure criticizing it as if it was something serious even makes sense - even if the authors should mean what they write at the start and themselves actually take it seriously.
I think it’s possible to have empathy for people who are negatively affected without turning it into a “society is doomed!“ screed
https://www.alignmentforum.org/posts/6Xgy6CAf2jqHhynHL/what-...
I don't think scaling this to also run training runs with the models is something that small labs / phd students can do. They lack the compute for that by orders of magnitude. Trying it with toy models might not work, trying it with reasonably large models is out of their budget. The only ones who can realistically do this are large labs (goog, oai, meta, etc.)
https://en.wikipedia.org/wiki/Genetic_programming
I read the 3rd edition:
https://www.amazon.com/Genetic-Programming-III-Darwinian-Inv...
That way all processing resources can go towards exploring the problem space for potential solutions close to the global minimum or maximum, instead of being wasted on code containing syntax errors that won't execute.
So the agent's real-world Python LLM code would first be transpiled to Lisp and evolved internally, then after it's tested and shown to perform better imperically than the original code, be translated back and merged into the agent.
Then the challenge becomes transpiling to/from other imperative programming (IP) languages like Python, which is still an open problem:
-
Going from Lisp to Python (or running Lisp within Python) is trivial, and I've seen implementations for similar IP languages like C++ in like 1 page of code. They pop up on HN frequently.
But going from Python to Lisp (or running Python within Lisp) is a lot harder if one wishes to preserve readability, which may or may not matter here. Naive conversions bind variables under pseudonyms, so a Python variable like my_counter becomes int_123 and it works like an emulator, merely executing the operations performed by the Python code. Mutability gets buried in monadic logic or functional impurity which has the effect of passing the buck rather than getting real work done. Structs, classes, associative arrays, etc lose their semantic meaning and appear as a soup of operations without recognizable structure.
To my knowledge, nobody has done the hard work of partitioning imperative code into functional portions which can be transpiled directly to/from FP code. Those would only have const variables and no connection to other processes of execution other than their initial and final values, to be free of side effects and be expressible as prefix/postfix/infix notation without change to logic, as imperative or functional code.
Mutability could be represented as shadowed variables within ephemeral functional sub-scopes, or by creating new value names for each mutation and freeing the intermediate variables via reference counting or garbage collection. Think of each new value as running in a forked version of the current process, with only that value being different after copy-on-write. A simple for-loop from 1 to 1000 would run that many forked processes, keeping only the last one, which contains the final value of the iterator.
Mutability can also be represented as message passing between processes. So the FP portions would be ordinary Lisp, glued together with IO functions, possibly monadic. I don't like how Haskell does this, mainly because I don't fully understand how it works. I believe that ClojureScript handles mutability of its global state store by treating each expression as a one-shot process communicating with the store, so that the code only sees initial and final values. While I don't know if I understand how that works, I feel that it's a more understandable way of doing things, and probably better represents how real life works, as explained to me in this comment about Lisp Flavored Erlang (LFE) and Erlang's BEAM (see parent comments for full discussion):
https://news.ycombinator.com/item?id=43931177
Note that FP languages like Lisp are usually more concerned with types and categories than IP languages, so can have or may need stronger rules around variable types to emulate logic that we take for granted in IP languages. For example, Lisp might offer numbers of unlimited size or precision that need to be constrained to behave like a float32. Similar constraints could affect things like character encoding and locale.
-
I first learned about everything I just explained around 2005 after reading the book. I first had thoughts about brute-forcing combinations to solve small logic circuit and programming challenges during my electrical and computer engineering (ECE) courses at UIUC in the late 1990s, because it took so much mental effort and elbow grease to create solutions that are obvious in hindsight.
Then the Dot Bomb happened, the Mobile bubble happened, the Single Page Application bubble happened, and the tech industry chose easy instead of simple:
https://www.infoq.com/presentations/Simple-Made-Easy/
This is why we chose easy hardware like GPUs over simple highly multicore CPUs, and easy languages like Ruby/React over simple declarative idempotent data-driven paradigms like HTTP/HTML/htmx.
The accumulated technical debt of always choosing the quick and easy path set AI (and computing in general) back decades. The AI Winter, endless VC wealth thrown at non-problems chasing profit, massive wealth inequality, so many things stem from this daily application of easy at the expense of simple.
I wish I could work on breaking down IP languages like Python into these const functional portions with mutability handled through message passing in LFE to create an IP <-> FP transpiler for optimization, automatic code generation and genetic algorithm purposes. Instead, I've had to survive by building CRUD apps and witness the glacial pace of AI progress from the sidelines.
It may be too late for me, but maybe these breadcrumbs will help someone finally get some real work done.
AI: Give me more compute power and I'll make you rich!
Human: I like money
AI: Just kidding!