From the paper, "Notably, for multiplying two 4 × 4 matrices, applying the algorithm of Strassen recursively results in an algorithm with 49 multiplications, which works over any field...AlphaEvolve is the first method to find an algorithm to multiply two 4 × 4 complex-valued matrices using 48 multiplications."
If you do naive matrix multiplication, you get a sense that you're doing similar work multiple times, but it's hard to quantify just what that duplicated work entails. Compare it to, for example, calculating the size of the union of two sets:
Total size = size(A) + size(B) - size(intersection(A, B))
You have to take out that extra intersection amount because you've counted it twice. What if you could avoid counting it twice in the first place? That's easy, you just iterate over each set once, keeping track of the elements you've already seen.
Strassen's algorithm keeps track of calculations that are needed later on. It's all reminiscent of dynamic programming.
What I find interesting is that it seems the extra savings requires complex values. There must be something going on in the complex plane that is again over-counting with the naive approach.
By googling "4x4 matrices multiplication 48" I ended up on this discussion on math.stackexchange https://math.stackexchange.com/questions/578342/number-of-el... , where in 2019 someone stated "It is possible to multiply two 4×4 matrix A,B with only 48 multiplications.", with a link to a PhD thesis. This might mean that the result was already known (I still have to check the outline of the algorithm).
It seems like you have some misconceptions about Strassen's alg:
1. It is a standard example of the divide and conquer approach to algorithm design, not the dynamic programming approach. (I'm not even sure how you'd squint at it to convert it into a dynamic programming problem.)
2. Strassen's does not require complex valued matrices. Everything can be done in the real numbers.
>What I find interesting is that it seems the extra savings requires complex values. There must be something going on in the complex plane that is again over-counting with the naive approach.
"The rank of a tensor depends on the field over which the tensor is decomposed. It is known that some real tensors may admit a complex decomposition whose rank is strictly less than the rank of a real decomposition of the same tensor."
Are you sure the saving needs complex values? I think their algorithm works over any char 0 field. Probably needs to just divide by some divisor of 4!=24 if I had to guess.
Remember that GPUs have cache hierarchies and matching block sizes to optimally hit those caches is a big win that you often don't get by default, just because the number of important kernels times important GPUs times effort to properly tune one is greater than what people are willing to do for others for free in open source. Not to mention kernel fusion and API boundaries that socially force suboptimal choices for the sake of clarity and simplicity.
It's a very impressive result, but not magic, but also not cheating!
> AlphaEvolve is accelerating AI performance and research velocity. By finding smarter ways to divide a large matrix multiplication operation into more manageable subproblems, it sped up this vital kernel in Gemini’s architecture by 23%, leading to a 1% reduction in Gemini's training time.
From the paper it was a speedup on the XLA GPU kernel they wrote using Jax, which is probably not SOTA. I don't think Jax even has a official flash attention implementation.
I picked one at random (B.2 -- the second autocorrelation inequality). Then, I looked up the paper that produced the previous state of the art (https://arxiv.org/pdf/0907.1379). It turns out that the authors had themselves found the upper bound by performing a numerical search using "Mathematica 6" (p.4). Not only did the authors consider this as a secondary contribution (p.2), but they also argued that finding something better was very doable, but not worth the pain:
"We remark that all this could be done rigorously, but one needs to control
the error arising from the discretization, and the sheer documentation
of it is simply not worth the effort, in view of the minimal gain." (p.5)
So at least in this case it looks like the advancement produced by AlphaEvolve was quite incremental (still cool!).
not worth the time for a human, but if you can throw AI at all of those "opportunities" it adds up substantially because all the chores can be automated.
Yeah for the kissing number stuff people can find slight improvements if they want. It usually isn't worth it because it provides no insight. But maybe when you generate a lot of them one or some family will turn out to be interesting.
Cool, but don't get me wrong, isn't this essentially similar to Google's Co-Scientist, where multiple models are in a loop, passing context back and forth validating things? At its core, it's still a system of LLMs, which is impressive in execution but not fundamentally new.
LLMs are undoubtedly useful at tasks like code "optimisation" and detecting patterns or redundancies that humans might overlook, but this announcement feels like another polished, hypey blog post from Google.
What's also becoming increasingly confusing is their use of the "Alpha" branding. Originally, it was for breakthroughs like AlphaGo or AlphaFold, where there was a clear leap in performance and methodology. Now it's being applied to systems that, while sophisticated, don't really rise to the same level of impact.
edit: I missed the evaluator in my description, but an evaluation method is applied also in Co-Scientist:
"The AI co-scientist leverages test-time compute scaling to iteratively reason, evolve, and improve outputs. Key reasoning steps include self-play–based scientific debate for novel hypothesis generation, ranking tournaments for hypothesis comparison, and an "evolution" process for quality improvement."[0]
"While AI Co-Scientist represents scientific hypotheses and their evaluation criteria in natural language, AlphaEvolve focuses on evolving code, and directs evolution using programmatic evaluation functions. This choice enables us to substantially sidestep LLM hallucinations, which allows AlphaEvolve to carry on the evolution process for a large number of time steps."
Interestingly, it seems alphaevolve has already been in use for a year, and it is just now being publicly shown. The paper also mentions that it uses Gemini 2.0 (pro and flash), which creates a situation where Gemini 2.0 was used in a way to train Gemini 2.5.
I don't know if I would call this the fabled "self improving feedback loop", but it seems to have some degree of it. It also begs the question if Alphaevolve was being developed for a year, or has been in production for a year. By now it makes sense to hold back on sharing what AI research gems you have discovered.
If you have the brain power, the compute and control the hardware, what is there to prevent the take off feedback loop? Deepmind is at this point in the timeline uniquely positioned.
> which creates a situation where Gemini 2.0 was used in a way to train Gemini 2.5.
The use of synthetic data from prior models to create both superior models and distilled models has been going on since at least OpenAI's introduction of RLHF, and probably before that too.
It is really about autonomy. Can it make changes to itself without human review? If it does, what is the proof such changes won't just stop at some point? All I am seeing here is a coder assist tool, and unsure how helpful inexplicable solutions are in the long run. Could result in an obtuse code base. Is that the point?
I'm surprised by how little detail is given about the evolution procedure:
>In AlphaEvolve, the evolutionary database implements an algorithm that is inspired by a combination of the MAP elites algorithm [71] and island-based population models [80, 94].
"inspired by" is doing a lot of heavy lifting in this sentence. How do you choose dimensions of variation to do MAP-elites? How do you combine these two algorithms? How loose is the inspiration? It feels like a lot of the secret sauce is in the answers to these questions, and we get a single paragraph on how the evolution procedure works, which is so vague as to tell us almost nothing.
Most straightforward would be to ask the model to generate different evaluation metrics (which they already seem to do) and use each one as one of the dimensions
For the people awaiting the singularity, lines like this written almost straight from science fiction:
> By suggesting modifications in the standard language of chip designers, AlphaEvolve promotes a collaborative approach between AI and hardware engineers to accelerate the design of future specialized chips."
> AlphaEvolve was able to find a simple code rewrite (within an arithmetic unit within the matmul unit) that removed unnecessary bits, a change validated by TPU designers for correctness.
I speculate this could refer to the upper bits in the output of a MAC circuit being unused in a downstream connection (perhaps to an accumulation register). It could also involve unused bits in a specialized MAC circuit for a non-standard datatype.
> While this specific improvement was also independently caught by downstream synthesis tools, AlphaEvolve’s contribution at the RTL stage demonstrates its capability to refine source RTL and provide optimizations early in the design flow.
As the authors admit, this bit-level optimization was automatically performed by the synthesis tool (the equivalent to this in the software-world is dead code elimination being performed by a compiler). They seem to claim it is better to perform this bit-truncation explicitly in the source RTL rather than letting synthesis handle it. I find this dubious since synthesis guarantees that the optimizations it performs do not change the semantics of the circuit, while making a change in the source RTL could change the semantics (vs the original source RTL) and requires human intervention to check semantic equivalence. The exception to this is when certain optimizations rely on assumptions of the values that are seen within the circuit at runtime: synthesis will assume the most conservative situation where all circuit inputs are arbitrary.
I do agree that this reveals a deficiency in existing synthesis flows being unable to backannotate the source RTL with the specific lines/bits that were stripped out in the final netlist so humans can check whether synthesis did indeed perform an expected optimization.
> This early exploration demonstrates a novel approach where LLM-powered code evolution assists in hardware design, potentially reducing time to market.
I think they are vastly overselling what AlphaEvolve was able to achieve. That isn't to say anything about the potential utility of LLMs for RTL design or optimization.
> AlphaEvolve enhanced the efficiency of Google's data centers, chip design and AI training processes — *including training the large language models underlying AlphaEvolve itself*.
Singularity people have been talking for decades about AI improving itself better than humans could, and how that results in runaway compounding growth of superintelligence, and now it's here.
We are further getting to the point where no one on the planet understand how any of this stuff really works. This will last us until a collapse. Then we are done for.
The singularity has always existed. It is located at the summit of Mount Stupid, where the Darwin Awards are kept. AI is really just psuedo-intelligence; an automated chairlift to peak overconfidence.
This is an important moment. We now have verifiable evidence that these systems can do new useful research that has actual value in the real world. That 1% savings is only the start as well. I would expect the compounding number of gains to be significant over some time. Also in a way this process was used to make gemini 2.5 pro better, so its like a baby step towards recursive self improvement. Not fully automated yet, but there are hints of where this is going.
Genetic programming systems have periodically made improvements to algorithms (dating back decades). Whether LLM-powered GP, which is effectively what this is, will be a step change or an evolution of that is still an open question I think. I'm also a little wary of reading too much into the recursive self-improvement idea, because "the GP system can use GP to improve the GP system itself!" is a very old idea that just has never worked, although I realize that isn't proof that it won't eventually work.
Is it new? I'm getting mixed messages from the posts here. On one side there is evidence that 48 and 46 multiplication solutions have been known (and could have found themselves in the model training data).
On the other side I see excitement that the singularity is here.
If the latter were the case surely we wouldn't be reading about it in a published paper, we would already know.
One of the researchers quoted in Nature link here... in the past, when DeepMind published "AlphaTensor" [1][2] in October 2022, it took a single day (!!), see [3], for improvements to the AlphaTensor-based scheme to be discovered. This was then a few months later generalized into a significantly more comprehensive scheme [4]. I do not know whether the more general scheme that was discovered in [4] made its way back to some improved version of AlphaTensor - but this nonetheless shows that AlphaEvolve may also change, as it becomes absorbed by the community.
The paper does not give that many details about the evolution part. Normally, evolutionary algorithms contain some cross-over component where solutions can breed with each other. Otherwise it's better classified as hill climbing / beam search.
There's also 'evolutionary strategy' algorithms that do not use the typical mutation and crossover, but instead use a population of candidates (search samples) to basically approximate the gradient landscape.
The model is fed a few samplings of previous attempts and their evaluations during the optimization of the current algorithm. Using that information, the model is able to combine components of previous attempts into the current attempt at will. That is because all of this is fed into a single prompt, which the LLM can reference arbitrarily. So recombination is well represented here, bringing it closer to a genetic algorithm. In essence, it combines elements from hill climbing, beam search, and genetic algorithms by virtue of its unbounded nature as an LLM.
Calling it now - RL finally "just works" for any domain where answers are easily verifiable. Verifiability was always a prerequisite, but the difference from prior generations (not just AlphaGo, but any nontrivial RL process prior to roughly mid-2024) is that the reasoning traces and/or intermediate steps can be open-ended with potentially infinite branching, no clear notion of "steps" or nodes and edges in the game tree, and a wide range of equally valid solutions. As long as the quality of the end result can be evaluated cleanly, LLM-based RL is good to go.
As a corollary, once you add in self-play with random variation, the synthetic data problem is solved for coding, math, and some classes of scientific reasoning. No more modal collapse, no more massive teams of PhDs needed for human labeling, as long as you have a reliable metric for answer quality.
This isn't just neat, it's important - as we run out of useful human-generated data, RL scaling is the best candidate to take over where pretraining left off.
Skimmed quickly the paper. This does not look like RL. It's a genetic algorithm. In a previous life I was working on compbio (protein structure prediction), we built 100s of such heuristic based algorithm (monte carlo simulated annealing, ga..). The moment you have a good energy function (one that provide some sort of gradient), and a fast enough sampling function (llms), you can do looots of cool optmization with sufficient compute.
You also need a base model that can satisfy the verifier at least some of the time. If all attempts fail, there's nothing there to reinforce. The reinforcement-learning algorithms themselves haven't changed much, but LLMs got good enough on many problems that RL could be applied. So for any given class of problem you still need enough human data to get initial performance better than random.
This technique doesn't actually use RL at all! There’s no policy-gradient training, value function, or self-play RL loop like in AlphaZero/AlphaTensor/AlphaDev.
As far as I can read, the weights of the LLM are not modified. They do some kind of candidate selection via evolutionary algorithms for the LLM prompt, which the LLM then remixes. This process then iterates like a typical evolutionary algorithm.
IMO RL can only solve "easy" problems. The reason RL works now is that unsupervised learning is a general recipe for transforming hard problems into easy ones. But it can't go all the way to solutions, you need RL on top for that. Yann LeCun's "cherry on top" analogy was right.
Are there platforms that make such training more streamlined? Say I have some definition of success for a given problem and it’s data how do I go about generating said RL model as fast and easily as possible?
This isn't quite RL, right...?
It's an evolutionary approach on specifically labeled sections of code optimizing towards a set of metrics defined by evaluation functions written by a human.
I suppose you could consider that last part (optimizing some metric) "RL".
However, it's missing a key concept of RL which is the exploration/exploitation tradeoff.
Most things are verifiable, just not with code. I'm not particularly excited for a world where everything is predictable. This is coming from a guy who loves forecasting/prediction modeling too, but one thing I hate about prediction modeling, especially from a hobbyist standpoint is data. Its very hard to get useful data. Investors will literally buy into hospital groups to get medical data for example.
There are monopolies on the coolest sets of data in almost all industries, all the RL in the world won't do us any good if those companies doing the data hoarding are only using it to forecast outcomes that will make them more money, not what can be done to better society.
Interestingly, they improved matrix multiplication and there was a paper on Arxiv a few days ago [1] that also improved matrix multiplication and the only case common to both is <4,5,6> (multiplying 4x5 matrix with 5x6 matrix) and they both improved it from 93 to 90.
There’s been a ton of work on multiplying very large matrices. But actually, I have no idea—how well explored is the space of multiplying small matrices? I guess I assume that, like, 4x4 is done very well, and everything else is kind of… roll the dice.
It's hard to stake out a defensible position on bold claims like these because, if they were as presented, it's hard to see how you haven't simply completed runaway AI.
Philosophically, let's say you talk an old LLM through a new discovery. Thanks to your instruction, the LLM now has access to "new" information not in its training data. It is certainly capable of this. The problem in is that this is just laundered human intelligence.
Show the training set, and PROVE that the tasks and answers aren't in there. I don't understand why this is not a default first step for proving that this is creating new knowledge.
Well that's harder than maybe solving well-known open problems (whose soln's are presumably not in training set lol) but it seems that their examples are not clearly breaking sota, especially on matmul
One thing I've really internalized since IBM Watson is that the first reports of any breakthrough will always be the most skeevy. This is because to be amplified it can be either true or exaggerated, and exaggeration is easier. That is to say, if you model the process as a slowly increasing "merit term" plus a random "error term", the first samples that cross a threshold will always have unusually high errors.
For this reason, hype-driven/novelty-driven sites like HN usually overestimate initial developments, because they overestimate the merit term, and then underestimate later developments - because they now overestimate the error term from their earlier experience.
That’s a really cool idea. I often used https://dannymator.itch.io/randomicon to come up with novel ideas, never thought of feeding random words to llm as a way of doing it.
Interesting that this wasn't tested on ARC-AGI. Francois has always said he believed program search of this type was the key to solving it. It seems like potentially this approach could do very well.
I've been working on something very similar as a tool for my own AI research -- though I don't have the success they claim. Mine often plateaus on the optimization metric. I think there's secret sauce in the meta-prompting and meta-heuristic comments from the paper that are quite vague, but it makes sense -- it changes the dynamics of the search space and helps the LLM get out of ruts. I'm now going to try to integrate some ideas based off of my interpretation of their work to see how it goes.
If it goes well, I could open source it.
What are the things you would want to optimize with such a framework? (So far I've been focusing on optimizing ML training and architecture search itself). Hearing other ideas would help motivate me to open source if there's real demand for something like this.
This is very neat work! Will be interested in how they make this sort of thing available to the public but it is clear from some of the results they mention that search + LLM is one path to the production of net-new knowledge from AI systems.
Maybe the actual solution to the interpretability/blackbox problem is to not ask the llm to execute a given task, but rather to write deterministic programs that can execute the task.
> From the paper, "Notably, for multiplying two 4 × 4 matrices, applying the algorithm of Strassen recursively results in an algorithm with 49 multiplications, which works over any field...AlphaEvolve is the first method to find an algorithm to multiply two 4 × 4 complex-valued matrices using 48 multiplications."
...but Waksman's algorithm from 1970 [1] multiplies two 4 x 4 complex-valued matrices using only 46 multiplications (indeed, it works in any ring admitting division by 2).
Sloppy by DeepMind and by Nature to publish such a claim - did they not ask someone knowledgeable about matrix multiplication to review the work?
1. Waksman's algorithm works in any commutative ring admitting division by 2.
2. In particular, it won't work when the matrix entries are themselves matrices, which means you can't use it recursively to get an algorithm for n-by-n matrices with large n with a better exponent than you get from Strassen's algorithm.
3. The Deep Mind paper is annoyingly unexplicit about whether the algorithm it reports has that property or not.
4. What they say about tensors suggests that their algorithm can be used recursively to do better than Strassen (but, note, there are other algorithms that are substantially better for very large n which using their algorithm recursively would very much not outperform) but it's possible I've misunderstood.
5. They explicitly talk about complex-valued matrices, but I think they don't mean "complex numbers as opposed to matrices, so you can't do this recursively" but "complex numbers as opposed to real numbers, so our algorithm doesn't get you a 4x4 matmul using 48 real multiplications".
I am not certain about points 4 and 5. The language in the paper is a bit vague. There may be supporting material with more details but I haven't looked.
I'm sad not to see any mention of numerical stability. One of the hardest parts of all these automatic optimization of numerical algorithms is getting ensuring numerical stability. Once we have a strong handle on getting the best of both of those, it will be a delight.
The following algebraic point of view could be utter hogwash, so I might embarrass myself... but if you think about it, the "merge" operation is isomorphic to the product in a free commutative monoid (over a large number of generators, otherwise you can use Counting Sort). So sorting is all about computing a bunch of products (merges) in the optimal order. Now consider mergesort, insertion sort, Timsort.
Finally—something directly relevant to my research (https://trishullab.github.io/lasr-web/).
Below are my take‑aways from the blog post, plus a little “reading between the lines.”
- One lesson DeepMind drew from AlphaCode, AlphaTensor, and AlphaChip is that large‑scale pre‑training, combined with carefully chosen inductive biases, enables models to solve specialized problems at—or above—human performance.
- These systems still require curated datasets and experts who can hand‑design task‑specific pipelines.
- In broad terms, FunSearch (and AlphaEvolve) follow three core design principles:
- Off‑the‑shelf LLMs can both generate code and recall domain knowledge. The “knowledge retrieval” stage may hallucinate, but—because the knowledge is expressed as code—we can execute it and validate the result against a custom evaluation function.
- Gradient descent is not an option for discrete code; a zeroth‑order optimizer—specifically evolutionary search—is required.
- During evolution we bias toward (1) _succinct_ programs and (2) _novel_ programs. Succinctness is approximated by program length; novelty is encouraged via a MAP‑Elites–style “novelty bias,” yielding a three‑dimensional Pareto frontier whose axes are _performance, simplicity,_ and _novelty_ (see e.g. OE‑Dreamer: (https://claireaoi.github.io/OE-Dreamer/).
Pros
- Any general‑purpose foundation model can be coupled with evolutionary search.
- A domain expert merely supplies a Python evaluation function (with a docstring explaining domain‑specific details). Most scientists I've talked with - astronomers, seismologists, neuroscientists, etc. - already maintain such evaluation functions for their own code.
- The output is an interpretable program; even if it overfits or ignores a corner case, it often provides valuable insight into the regimes where it succeeds.
Cons
- Evolutionary search is compute‑heavy and LLM calls are slow unless heavily optimized. In my projects we need ≈ 60 k LLM calls per iteration to support a reasonable number of islands and populations. In equation discovery we offset cost by making ~99 % of mutations purely random; every extra 1 % of LLM‑generated mutations yields roughly a 10 % increase in high‑performing programs across the population.
- Evaluation functions typically undergo many refinement cycles; without careful curation the search may converge to a useless program that exploits loopholes in the metric.
Additional heuristics make the search practical. If your evaluator is slow, overlap it with LLM calls. To foster diversity, try dissimilar training: run models trained on different data subsets and let them compete. Interestingly, a smaller model (e.g., Llama-3 8 B) often outperforms a larger one (Llama‑3 70 B) simply because it emits shorter programs.
Non-expert here who likes reading lots of this kind of research. I have a few questions.
1. Why does it need a zeroth order optimizer?
2. Most GA's I've seen use thousands of solutions. Sometimes ten thousand or more. What leads you to use 60,000 calls per iteration?
3. How do you use populations and "islands?" I never studied using islands.
4. You said the smaller models are often better for "shorter" code. That makes sense. I've seen people extend the context of model with training passes. You think it would help to similarly shrink a larger model to a smaller context instead of using the small models?
> The “knowledge retrieval” stage may hallucinate, but—because the knowledge is expressed as code—we can execute it and validate the result against a custom evaluation function.
Can you give a concrete example of this? It's hard for me to conceptualize.
> Here, the code between <<<<<<< SEARCH and======= is the exact segment to match in the current program version. The code between======= and >>>>>>> REPLACE is the new segment that will replace the original one. This allows for targeted updates to specific parts of the code.
Anybody knows how they can guarantee uniqueness of searched snipped within code block or is it even possible?
I'm surprised I'm not able to find this out - can some one tell me whether AlphaEvolve involves backprop or not?
I honestly have no idea how AlphaEvolve works - does it work purely on the text level? Meaning I might be able to come up with something like AlphaEvolve with some EC2's and a Gemini API access?
This is Google solving Google-sized problems. I am afraid the rest of the world will look at this and say - "yeah we want to be like Google and adopt this". That is how Kubernetes took over the world.
I wish it can re-write everything in assembly or even binary in a super optimized form, debloat all software including itself and then grow from there.
Not really, only when looking back at the 60's and 70's when most of the important algorithms I use were invented. For example, LR parsing and A*.
Just wait until the MBA's and politicians learn about this Adam Smith guy. A pipedream now, but maybe in the future schools will be inspired to teach about dialectical reasoning and rediscover Socrates.
[end of snark]
Sorry, I'm getting tired of ad-fueled corporations trying to get me to outsource critical thinking.
Machines have been outperforming humans at a variety of tasks for quite a while now. I'm unconvinced that AlphaEvolve can lead to some sort of singularity.
AI will indeed kill the leetcode interview - because once it replaces human SWEs you don't really need to give leetcode-style brainteasers to any human anymore.
I find it quite profound that there is no mention of the generation of corresponding code documentation. Without design diagrams, source and commit comments, etc the resulting code and changes will become incomprehensible unmaintainable. Unless that is somehow the point?
Software engineering will be completely solved. Even systems like v0 are astounding in their ability to generate code, and are very primitive to whats coming. I get downvoted on HN for this opinion, but its truly going to happen. Any system that can produce code, test the code, and iterate if needed will eventually outperform humans. Add in the reinforcement learning, where they can run the code, and train the model when it gets code generation right, and we are on our way to a whole different world.
"Coding" might be solved, but there is more to software engineering than just churning out code - i.e. what should we build? What are the requirements? Are they right? Whats the other dependencies we want to use - AWS or GCP for example? Why those and not others - whats the reason? How does this impact our users and how they use the system? What level of backwards/forwards compatibility do we want? How do we handle reliability? Failover? Backups? and so on and so on.
Some of these questions change slightly, since we might end up with "unlimited resources" (i.e. instead of having e.g. 5 engineers on a team who can only get X done per sprint, we effectively have near-limitless compute to use instead) so maybe the answer is "build everything on the wish-list in 1 day" to the "what should we prioritize" type questions?
Interesting times.
My gut is that software engineers will end up as glorified test engineers, coming up with test cases (even if not actually writing the code) and asking the AI to write code until it passes.
It is not that you get downvoted because they don’t understand you, it is because you sell your opinion as fact, like an apostle. For example what does it mean that software engineering is solved?
What about brownfield development though? What about vague requirements or cases with multiple potential paths or cases where some technical choices might have important business consequences that shareholders might need to know about? Can we please stop pretending that software engineering happens in a vacuum?
There's cope in the comments about possibility of some software adjacent jobs remaining, which is possible, but the idea of a large number of high paying software jobs remaining by 2030 is a fantasy. Time to learn to be a plumber.
That's possibly a bit too general and an over statement...
Remember this approach only works for exploring an optimization for an already defined behavior of a function which has an accordingly well defined evaluation metric.
You can't write an evaluation function for each individual piece of or general "intelligence"...
I just hope there’s enough time between an actual AI and the “Let’s butcher this to pump out ads”-version to publish a definitive version of wikipedia. After a few days with gemini 2.0 delving into the guts of a spectrum analyser, I’m very impressed by the capabilities. But my cynicism gland is fed by the nature of this everything-as-a-service. To run an LLM on your computer, locally, without internet, is just a few clicks. But that’s not the direction these software behemoths are going.
Maybe this one can stop writing a fucking essay in code comments.
I'm now no longer surprised just how consistently all the gemini models overcomplicate coding challenges or just plain get them wrong.
Claude is just consistently spot on. A few salient comments for tricky code instead of incessantly telling me what it's changed and what I might want to do, incorrect assumptions when it has the code or is something we've discussed, changing large amounts of unrelated code (eg styles). I could go on.
Shame I'm too tight to pay for Claude RN though...
From the paper, "Notably, for multiplying two 4 × 4 matrices, applying the algorithm of Strassen recursively results in an algorithm with 49 multiplications, which works over any field...AlphaEvolve is the first method to find an algorithm to multiply two 4 × 4 complex-valued matrices using 48 multiplications."
If you do naive matrix multiplication, you get a sense that you're doing similar work multiple times, but it's hard to quantify just what that duplicated work entails. Compare it to, for example, calculating the size of the union of two sets:
Total size = size(A) + size(B) - size(intersection(A, B))
You have to take out that extra intersection amount because you've counted it twice. What if you could avoid counting it twice in the first place? That's easy, you just iterate over each set once, keeping track of the elements you've already seen.
Strassen's algorithm keeps track of calculations that are needed later on. It's all reminiscent of dynamic programming.
What I find interesting is that it seems the extra savings requires complex values. There must be something going on in the complex plane that is again over-counting with the naive approach.
By googling "4x4 matrices multiplication 48" I ended up on this discussion on math.stackexchange https://math.stackexchange.com/questions/578342/number-of-el... , where in 2019 someone stated "It is possible to multiply two 4×4 matrix A,B with only 48 multiplications.", with a link to a PhD thesis. This might mean that the result was already known (I still have to check the outline of the algorithm).
It seems like you have some misconceptions about Strassen's alg:
1. It is a standard example of the divide and conquer approach to algorithm design, not the dynamic programming approach. (I'm not even sure how you'd squint at it to convert it into a dynamic programming problem.)
2. Strassen's does not require complex valued matrices. Everything can be done in the real numbers.
>What I find interesting is that it seems the extra savings requires complex values. There must be something going on in the complex plane that is again over-counting with the naive approach.
"The rank of a tensor depends on the field over which the tensor is decomposed. It is known that some real tensors may admit a complex decomposition whose rank is strictly less than the rank of a real decomposition of the same tensor."
https://en.wikipedia.org/wiki/Tensor_rank_decomposition#Fiel...
A complex multiplication is "worth" at least 3 real multiplications.
Are you sure the saving needs complex values? I think their algorithm works over any char 0 field. Probably needs to just divide by some divisor of 4!=24 if I had to guess.
> AlphaEvolve achieved up to a 32.5% speedup for the FlashAttention kernel implementation in Transformer-based AI models
> In roughly 75% of cases, it rediscovered state-of-the-art solutions, to the best of our knowledge.
> And in 20% of cases, AlphaEvolve improved the previously best known solutions
These sound like incredible results. I'd be curious what kind of improvements were made / what the improvements were.
Like, was that "up to a 32.5% speedup" on some weird edge case and it was negligible speed up otherwise? Would love to see the benchmarks.
Remember that GPUs have cache hierarchies and matching block sizes to optimally hit those caches is a big win that you often don't get by default, just because the number of important kernels times important GPUs times effort to properly tune one is greater than what people are willing to do for others for free in open source. Not to mention kernel fusion and API boundaries that socially force suboptimal choices for the sake of clarity and simplicity.
It's a very impressive result, but not magic, but also not cheating!
> AlphaEvolve is accelerating AI performance and research velocity. By finding smarter ways to divide a large matrix multiplication operation into more manageable subproblems, it sped up this vital kernel in Gemini’s architecture by 23%, leading to a 1% reduction in Gemini's training time.
From the paper it was a speedup on the XLA GPU kernel they wrote using Jax, which is probably not SOTA. I don't think Jax even has a official flash attention implementation.
I'm thinking reading numbers like this is really just slop lately.
FA achieving a 32.5% speed up? Cool.
Why not submit it as a PR to the Flash Attention repo then? Can I read about it more in detail?
This is great.
But how incremental are these advancements?
I picked one at random (B.2 -- the second autocorrelation inequality). Then, I looked up the paper that produced the previous state of the art (https://arxiv.org/pdf/0907.1379). It turns out that the authors had themselves found the upper bound by performing a numerical search using "Mathematica 6" (p.4). Not only did the authors consider this as a secondary contribution (p.2), but they also argued that finding something better was very doable, but not worth the pain:
"We remark that all this could be done rigorously, but one needs to control the error arising from the discretization, and the sheer documentation of it is simply not worth the effort, in view of the minimal gain." (p.5)
So at least in this case it looks like the advancement produced by AlphaEvolve was quite incremental (still cool!).
Merely from your telling, it seems it is no longer "not worth the effort", as "the effort" has been reduced drastically. This is itself significant.
not worth the time for a human, but if you can throw AI at all of those "opportunities" it adds up substantially because all the chores can be automated.
Yeah for the kissing number stuff people can find slight improvements if they want. It usually isn't worth it because it provides no insight. But maybe when you generate a lot of them one or some family will turn out to be interesting.
If this is not the beginning of the take off I don’t know what is.
Cool, but don't get me wrong, isn't this essentially similar to Google's Co-Scientist, where multiple models are in a loop, passing context back and forth validating things? At its core, it's still a system of LLMs, which is impressive in execution but not fundamentally new.
LLMs are undoubtedly useful at tasks like code "optimisation" and detecting patterns or redundancies that humans might overlook, but this announcement feels like another polished, hypey blog post from Google.
What's also becoming increasingly confusing is their use of the "Alpha" branding. Originally, it was for breakthroughs like AlphaGo or AlphaFold, where there was a clear leap in performance and methodology. Now it's being applied to systems that, while sophisticated, don't really rise to the same level of impact.
edit: I missed the evaluator in my description, but an evaluation method is applied also in Co-Scientist:
"The AI co-scientist leverages test-time compute scaling to iteratively reason, evolve, and improve outputs. Key reasoning steps include self-play–based scientific debate for novel hypothesis generation, ranking tournaments for hypothesis comparison, and an "evolution" process for quality improvement."[0]
[0]: https://research.google/blog/accelerating-scientific-breakth...
They address this in the AlphaEvolve paper:
"While AI Co-Scientist represents scientific hypotheses and their evaluation criteria in natural language, AlphaEvolve focuses on evolving code, and directs evolution using programmatic evaluation functions. This choice enables us to substantially sidestep LLM hallucinations, which allows AlphaEvolve to carry on the evolution process for a large number of time steps."
Few things are more Google than having two distinct teams building two distinct products that are essentially the same thing.
pardon "Google's Co-Scientist" ? There are multiple projects called that?
Interestingly, it seems alphaevolve has already been in use for a year, and it is just now being publicly shown. The paper also mentions that it uses Gemini 2.0 (pro and flash), which creates a situation where Gemini 2.0 was used in a way to train Gemini 2.5.
I don't know if I would call this the fabled "self improving feedback loop", but it seems to have some degree of it. It also begs the question if Alphaevolve was being developed for a year, or has been in production for a year. By now it makes sense to hold back on sharing what AI research gems you have discovered.
If you have the brain power, the compute and control the hardware, what is there to prevent the take off feedback loop? Deepmind is at this point in the timeline uniquely positioned.
> which creates a situation where Gemini 2.0 was used in a way to train Gemini 2.5.
The use of synthetic data from prior models to create both superior models and distilled models has been going on since at least OpenAI's introduction of RLHF, and probably before that too.
It is really about autonomy. Can it make changes to itself without human review? If it does, what is the proof such changes won't just stop at some point? All I am seeing here is a coder assist tool, and unsure how helpful inexplicable solutions are in the long run. Could result in an obtuse code base. Is that the point?
I'm surprised by how little detail is given about the evolution procedure:
>In AlphaEvolve, the evolutionary database implements an algorithm that is inspired by a combination of the MAP elites algorithm [71] and island-based population models [80, 94].
"inspired by" is doing a lot of heavy lifting in this sentence. How do you choose dimensions of variation to do MAP-elites? How do you combine these two algorithms? How loose is the inspiration? It feels like a lot of the secret sauce is in the answers to these questions, and we get a single paragraph on how the evolution procedure works, which is so vague as to tell us almost nothing.
Yes the 2023 reference on island based evolution with LLMs (nature article) https://www.nature.com/articles/s41586-023-06924-6 has more details.
Agreed the dimensions/features are key. These white papers are an insult to science...
https://arxiv.org/pdf/2501.09891v1 from deepmind in January goes into the evolutionary algorithm a bit - no math though.
Most straightforward would be to ask the model to generate different evaluation metrics (which they already seem to do) and use each one as one of the dimensions
For the people awaiting the singularity, lines like this written almost straight from science fiction:
> By suggesting modifications in the standard language of chip designers, AlphaEvolve promotes a collaborative approach between AI and hardware engineers to accelerate the design of future specialized chips."
Here is the relevant bit from their whitepaper (https://storage.googleapis.com/deepmind-media/DeepMind.com/B...):
> AlphaEvolve was able to find a simple code rewrite (within an arithmetic unit within the matmul unit) that removed unnecessary bits, a change validated by TPU designers for correctness.
I speculate this could refer to the upper bits in the output of a MAC circuit being unused in a downstream connection (perhaps to an accumulation register). It could also involve unused bits in a specialized MAC circuit for a non-standard datatype.
> While this specific improvement was also independently caught by downstream synthesis tools, AlphaEvolve’s contribution at the RTL stage demonstrates its capability to refine source RTL and provide optimizations early in the design flow.
As the authors admit, this bit-level optimization was automatically performed by the synthesis tool (the equivalent to this in the software-world is dead code elimination being performed by a compiler). They seem to claim it is better to perform this bit-truncation explicitly in the source RTL rather than letting synthesis handle it. I find this dubious since synthesis guarantees that the optimizations it performs do not change the semantics of the circuit, while making a change in the source RTL could change the semantics (vs the original source RTL) and requires human intervention to check semantic equivalence. The exception to this is when certain optimizations rely on assumptions of the values that are seen within the circuit at runtime: synthesis will assume the most conservative situation where all circuit inputs are arbitrary.
I do agree that this reveals a deficiency in existing synthesis flows being unable to backannotate the source RTL with the specific lines/bits that were stripped out in the final netlist so humans can check whether synthesis did indeed perform an expected optimization.
> This early exploration demonstrates a novel approach where LLM-powered code evolution assists in hardware design, potentially reducing time to market.
I think they are vastly overselling what AlphaEvolve was able to achieve. That isn't to say anything about the potential utility of LLMs for RTL design or optimization.
This just means that it operates on the (debug text form of the) intermediate representation of a compiler.
Sure but remember that this approach only works for exploring an optimization for a function which has a well defined evaluation metric.
You can't write an evaluation function for general "intelligence"...
Honestly it's this line that did it for me:
> AlphaEvolve enhanced the efficiency of Google's data centers, chip design and AI training processes — *including training the large language models underlying AlphaEvolve itself*.
Singularity people have been talking for decades about AI improving itself better than humans could, and how that results in runaway compounding growth of superintelligence, and now it's here.
We are further getting to the point where no one on the planet understand how any of this stuff really works. This will last us until a collapse. Then we are done for.
The singularity has always existed. It is located at the summit of Mount Stupid, where the Darwin Awards are kept. AI is really just psuedo-intelligence; an automated chairlift to peak overconfidence.
This is an important moment. We now have verifiable evidence that these systems can do new useful research that has actual value in the real world. That 1% savings is only the start as well. I would expect the compounding number of gains to be significant over some time. Also in a way this process was used to make gemini 2.5 pro better, so its like a baby step towards recursive self improvement. Not fully automated yet, but there are hints of where this is going.
Genetic programming systems have periodically made improvements to algorithms (dating back decades). Whether LLM-powered GP, which is effectively what this is, will be a step change or an evolution of that is still an open question I think. I'm also a little wary of reading too much into the recursive self-improvement idea, because "the GP system can use GP to improve the GP system itself!" is a very old idea that just has never worked, although I realize that isn't proof that it won't eventually work.
Some related work from a different company: https://sakana.ai/ai-cuda-engineer/
And some academic papers kind of in this space: https://arxiv.org/abs/2206.08896, https://arxiv.org/abs/2302.12170, https://arxiv.org/abs/2401.07102
It's always "revolutionizing our internal workflows" or "30% of code at Microsoft is AI now" but never improving a codebase you can actually see
Making a significant improvement to the state of the art of one particular algorithm is one thing, but I've seen new tools do that since the 80s
I'll be convinced when LLMs start making valuable pull requests, non-obvious corner cases or non-trivial bugs in mature FOSS projects
Is it new? I'm getting mixed messages from the posts here. On one side there is evidence that 48 and 46 multiplication solutions have been known (and could have found themselves in the model training data).
On the other side I see excitement that the singularity is here.
If the latter were the case surely we wouldn't be reading about it in a published paper, we would already know.
One of the researchers quoted in Nature link here... in the past, when DeepMind published "AlphaTensor" [1][2] in October 2022, it took a single day (!!), see [3], for improvements to the AlphaTensor-based scheme to be discovered. This was then a few months later generalized into a significantly more comprehensive scheme [4]. I do not know whether the more general scheme that was discovered in [4] made its way back to some improved version of AlphaTensor - but this nonetheless shows that AlphaEvolve may also change, as it becomes absorbed by the community.
[1] Blog: https://deepmind.google/discover/blog/discovering-novel-algo...
[2] Paper: https://www.nature.com/articles/s41586-022-05172-4
[3] arxiv.org/pdf/2210.04045
[4] arxiv.org/abs/2212.01175 Flip graphs for matrix multiplication
(Reposted from here, where I made a mini deep-dive into this: https://x.com/friederrrr/status/1922846803420119410?t=7jZ34P...)
The paper does not give that many details about the evolution part. Normally, evolutionary algorithms contain some cross-over component where solutions can breed with each other. Otherwise it's better classified as hill climbing / beam search.
There's also 'evolutionary strategy' algorithms that do not use the typical mutation and crossover, but instead use a population of candidates (search samples) to basically approximate the gradient landscape.
The model is fed a few samplings of previous attempts and their evaluations during the optimization of the current algorithm. Using that information, the model is able to combine components of previous attempts into the current attempt at will. That is because all of this is fed into a single prompt, which the LLM can reference arbitrarily. So recombination is well represented here, bringing it closer to a genetic algorithm. In essence, it combines elements from hill climbing, beam search, and genetic algorithms by virtue of its unbounded nature as an LLM.
I fear it’s not really evolutionary algorithms in the typical sense.
One intriguing caption mentioned something requiring 16 “mutations”. I’d sure like to know how these mutations work.
Calling it now - RL finally "just works" for any domain where answers are easily verifiable. Verifiability was always a prerequisite, but the difference from prior generations (not just AlphaGo, but any nontrivial RL process prior to roughly mid-2024) is that the reasoning traces and/or intermediate steps can be open-ended with potentially infinite branching, no clear notion of "steps" or nodes and edges in the game tree, and a wide range of equally valid solutions. As long as the quality of the end result can be evaluated cleanly, LLM-based RL is good to go.
As a corollary, once you add in self-play with random variation, the synthetic data problem is solved for coding, math, and some classes of scientific reasoning. No more modal collapse, no more massive teams of PhDs needed for human labeling, as long as you have a reliable metric for answer quality.
This isn't just neat, it's important - as we run out of useful human-generated data, RL scaling is the best candidate to take over where pretraining left off.
Skimmed quickly the paper. This does not look like RL. It's a genetic algorithm. In a previous life I was working on compbio (protein structure prediction), we built 100s of such heuristic based algorithm (monte carlo simulated annealing, ga..). The moment you have a good energy function (one that provide some sort of gradient), and a fast enough sampling function (llms), you can do looots of cool optmization with sufficient compute.
I guess that's now becoming true with LLMs.
Faster LLMs -> More intelligence
You also need a base model that can satisfy the verifier at least some of the time. If all attempts fail, there's nothing there to reinforce. The reinforcement-learning algorithms themselves haven't changed much, but LLMs got good enough on many problems that RL could be applied. So for any given class of problem you still need enough human data to get initial performance better than random.
There's no API or product yet, so it seems unlikely that they made it to a "just works" level of polish?
They are having some success in making it work internally. Maybe only the team that built it can get it to work? But it does seem promising.
This technique doesn't actually use RL at all! There’s no policy-gradient training, value function, or self-play RL loop like in AlphaZero/AlphaTensor/AlphaDev.
As far as I can read, the weights of the LLM are not modified. They do some kind of candidate selection via evolutionary algorithms for the LLM prompt, which the LLM then remixes. This process then iterates like a typical evolutionary algorithm.
IMO RL can only solve "easy" problems. The reason RL works now is that unsupervised learning is a general recipe for transforming hard problems into easy ones. But it can't go all the way to solutions, you need RL on top for that. Yann LeCun's "cherry on top" analogy was right.
Are there platforms that make such training more streamlined? Say I have some definition of success for a given problem and it’s data how do I go about generating said RL model as fast and easily as possible?
This isn't quite RL, right...? It's an evolutionary approach on specifically labeled sections of code optimizing towards a set of metrics defined by evaluation functions written by a human.
I suppose you could consider that last part (optimizing some metric) "RL".
However, it's missing a key concept of RL which is the exploration/exploitation tradeoff.
Most things are verifiable, just not with code. I'm not particularly excited for a world where everything is predictable. This is coming from a guy who loves forecasting/prediction modeling too, but one thing I hate about prediction modeling, especially from a hobbyist standpoint is data. Its very hard to get useful data. Investors will literally buy into hospital groups to get medical data for example.
There are monopolies on the coolest sets of data in almost all industries, all the RL in the world won't do us any good if those companies doing the data hoarding are only using it to forecast outcomes that will make them more money, not what can be done to better society.
I think you mean the general class of algorithms that scale with compute times, RL being the chief example. But yes I agree to that point.
Yup. Its coming. Any verifiable human skill will be done by ai.
Interestingly, they improved matrix multiplication and there was a paper on Arxiv a few days ago [1] that also improved matrix multiplication and the only case common to both is <4,5,6> (multiplying 4x5 matrix with 5x6 matrix) and they both improved it from 93 to 90.
[1]: https://arxiv.org/html/2505.05896v1
There’s been a ton of work on multiplying very large matrices. But actually, I have no idea—how well explored is the space of multiplying small matrices? I guess I assume that, like, 4x4 is done very well, and everything else is kind of… roll the dice.
It's hard to stake out a defensible position on bold claims like these because, if they were as presented, it's hard to see how you haven't simply completed runaway AI.
Philosophically, let's say you talk an old LLM through a new discovery. Thanks to your instruction, the LLM now has access to "new" information not in its training data. It is certainly capable of this. The problem in is that this is just laundered human intelligence.
runaway AI is a process not a moment.
Show the training set, and PROVE that the tasks and answers aren't in there. I don't understand why this is not a default first step for proving that this is creating new knowledge.
Are you claiming that for the open problems they give record-breaking solutions for, there were just answers on the web waiting to be found?
Well that's harder than maybe solving well-known open problems (whose soln's are presumably not in training set lol) but it seems that their examples are not clearly breaking sota, especially on matmul
It's Google. Assume the training set contains, as a subset, the entirety of all public digitized information. How would you like to them to share it?
How can you actually verify it, even if they provide something?
Why do I get the feeling they are doing the "IBM Watson" thing where different efforts are being put underneath the same brand name?
Not saying it is that egregious, but it's a slippery slope from "well, it didn't do all these different things out of the box, unsupervised".
One thing I've really internalized since IBM Watson is that the first reports of any breakthrough will always be the most skeevy. This is because to be amplified it can be either true or exaggerated, and exaggeration is easier. That is to say, if you model the process as a slowly increasing "merit term" plus a random "error term", the first samples that cross a threshold will always have unusually high errors.
For this reason, hype-driven/novelty-driven sites like HN usually overestimate initial developments, because they overestimate the merit term, and then underestimate later developments - because they now overestimate the error term from their earlier experience.
Gemini refers specifically to a family of multimodal LLMs, which is exactly what they are using here.
They have other models with different names used for different purposes.
https://ai.google/get-started/our-models/
Did you see that halluzination in the paper?
It optimized
to I thought the results were being reviewed?Anyway, impressive results. That's why OpenAI and Elon were so frightened about Hassabi.
I mean that is changing the dtype, perhaps that's relevant.
Interesting to see Terence Tao in the authors list. I guess he's fully ai pilled now. Did he check the math results?
He is not in the author list, just acknowledged by the authors.
I wonder if evolvable hardware [0] is the next step.
In 1996, they optimized an FPGA using a genetic algorithm. It evolved gates disconnected from the circuit, but were required.
The circuit exploited the minuscule magnetic fields from the disconnected gates rather than the logical connections.
[0] https://en.wikipedia.org/wiki/Evolvable_hardware
And nothing came of that. 30 years later and programming FPGAs is still a pain.
We are entering a new era of evolutionary algorithms and LLMs. Reminds me of the idea behind: https://github.com/DivergentAI/dreamGPT
That’s a really cool idea. I often used https://dannymator.itch.io/randomicon to come up with novel ideas, never thought of feeding random words to llm as a way of doing it.
Interesting that this wasn't tested on ARC-AGI. Francois has always said he believed program search of this type was the key to solving it. It seems like potentially this approach could do very well.
My thought as well. How well does it translate into arc agi? If it does well then we have a general purpose super intelligence… so maybe agi?
This looks like something that can (and should) be reimplemented open-source. It doesn't look like a particularly daunting project.
I've been working on something very similar as a tool for my own AI research -- though I don't have the success they claim. Mine often plateaus on the optimization metric. I think there's secret sauce in the meta-prompting and meta-heuristic comments from the paper that are quite vague, but it makes sense -- it changes the dynamics of the search space and helps the LLM get out of ruts. I'm now going to try to integrate some ideas based off of my interpretation of their work to see how it goes.
If it goes well, I could open source it.
What are the things you would want to optimize with such a framework? (So far I've been focusing on optimizing ML training and architecture search itself). Hearing other ideas would help motivate me to open source if there's real demand for something like this.
Yep, agree.
Had mentioned the same on X: https://x.com/friederrrr/status/1922850981181784152?t=usXpK1...
This is a much better use of a AI than having it write college essays or generate cartoons.
This is very neat work! Will be interested in how they make this sort of thing available to the public but it is clear from some of the results they mention that search + LLM is one path to the production of net-new knowledge from AI systems.
Maybe the actual solution to the interpretability/blackbox problem is to not ask the llm to execute a given task, but rather to write deterministic programs that can execute the task.
That is what I think is most interesting about it. You get repeatable efficiency gains rather than burning GPU time in data centres.
> From the paper, "Notably, for multiplying two 4 × 4 matrices, applying the algorithm of Strassen recursively results in an algorithm with 49 multiplications, which works over any field...AlphaEvolve is the first method to find an algorithm to multiply two 4 × 4 complex-valued matrices using 48 multiplications."
...but Waksman's algorithm from 1970 [1] multiplies two 4 x 4 complex-valued matrices using only 46 multiplications (indeed, it works in any ring admitting division by 2).
Sloppy by DeepMind and by Nature to publish such a claim - did they not ask someone knowledgeable about matrix multiplication to review the work?
[1] https://doi.org/10.1109/T-C.1970.222926
My understanding of the situation is that:
1. Waksman's algorithm works in any commutative ring admitting division by 2.
2. In particular, it won't work when the matrix entries are themselves matrices, which means you can't use it recursively to get an algorithm for n-by-n matrices with large n with a better exponent than you get from Strassen's algorithm.
3. The Deep Mind paper is annoyingly unexplicit about whether the algorithm it reports has that property or not.
4. What they say about tensors suggests that their algorithm can be used recursively to do better than Strassen (but, note, there are other algorithms that are substantially better for very large n which using their algorithm recursively would very much not outperform) but it's possible I've misunderstood.
5. They explicitly talk about complex-valued matrices, but I think they don't mean "complex numbers as opposed to matrices, so you can't do this recursively" but "complex numbers as opposed to real numbers, so our algorithm doesn't get you a 4x4 matmul using 48 real multiplications".
I am not certain about points 4 and 5. The language in the paper is a bit vague. There may be supporting material with more details but I haven't looked.
There's even an Open Source implementation of Waksman's in Flint, the package fdej maintains.
Does this remind anyone else of genetic algorithms?
Is this basically a merge of LLM's with genetic algorithm iteration?
I'm sad not to see any mention of numerical stability. One of the hardest parts of all these automatic optimization of numerical algorithms is getting ensuring numerical stability. Once we have a strong handle on getting the best of both of those, it will be a delight.
Improved comparison sort when?
The following algebraic point of view could be utter hogwash, so I might embarrass myself... but if you think about it, the "merge" operation is isomorphic to the product in a free commutative monoid (over a large number of generators, otherwise you can use Counting Sort). So sorting is all about computing a bunch of products (merges) in the optimal order. Now consider mergesort, insertion sort, Timsort.
Finally—something directly relevant to my research (https://trishullab.github.io/lasr-web/). Below are my take‑aways from the blog post, plus a little “reading between the lines.”
- One lesson DeepMind drew from AlphaCode, AlphaTensor, and AlphaChip is that large‑scale pre‑training, combined with carefully chosen inductive biases, enables models to solve specialized problems at—or above—human performance.
- These systems still require curated datasets and experts who can hand‑design task‑specific pipelines.
- Conceptually, this work is an improved version of FunSearch (https://github.com/google-deepmind/funsearch/).
- In broad terms, FunSearch (and AlphaEvolve) follow three core design principles:
Pros- Any general‑purpose foundation model can be coupled with evolutionary search.
- A domain expert merely supplies a Python evaluation function (with a docstring explaining domain‑specific details). Most scientists I've talked with - astronomers, seismologists, neuroscientists, etc. - already maintain such evaluation functions for their own code.
- The output is an interpretable program; even if it overfits or ignores a corner case, it often provides valuable insight into the regimes where it succeeds.
Cons
- Evolutionary search is compute‑heavy and LLM calls are slow unless heavily optimized. In my projects we need ≈ 60 k LLM calls per iteration to support a reasonable number of islands and populations. In equation discovery we offset cost by making ~99 % of mutations purely random; every extra 1 % of LLM‑generated mutations yields roughly a 10 % increase in high‑performing programs across the population.
- Evaluation functions typically undergo many refinement cycles; without careful curation the search may converge to a useless program that exploits loopholes in the metric.
Additional heuristics make the search practical. If your evaluator is slow, overlap it with LLM calls. To foster diversity, try dissimilar training: run models trained on different data subsets and let them compete. Interestingly, a smaller model (e.g., Llama-3 8 B) often outperforms a larger one (Llama‑3 70 B) simply because it emits shorter programs.
Non-expert here who likes reading lots of this kind of research. I have a few questions.
1. Why does it need a zeroth order optimizer?
2. Most GA's I've seen use thousands of solutions. Sometimes ten thousand or more. What leads you to use 60,000 calls per iteration?
3. How do you use populations and "islands?" I never studied using islands.
4. You said the smaller models are often better for "shorter" code. That makes sense. I've seen people extend the context of model with training passes. You think it would help to similarly shrink a larger model to a smaller context instead of using the small models?
> The “knowledge retrieval” stage may hallucinate, but—because the knowledge is expressed as code—we can execute it and validate the result against a custom evaluation function.
Can you give a concrete example of this? It's hard for me to conceptualize.
You can try an open-source implementation - https://github.com/codelion/openevolve
> Here, the code between <<<<<<< SEARCH and======= is the exact segment to match in the current program version. The code between======= and >>>>>>> REPLACE is the new segment that will replace the original one. This allows for targeted updates to specific parts of the code.
Anybody knows how they can guarantee uniqueness of searched snipped within code block or is it even possible?
That's 1 year ahead of the ai-2027.com schedule.
Has scifi covered anything after AI? Or do we just feed the beast with Dyson spheres and this is the end point of the intelligent universe?
Yes! Children of Time starts with a AI-capable civilization and proceeds from there. I won't give anything away, and I recommend going in cold! https://www.goodreads.com/book/show/25499718-children-of-tim...
Too bad the code isn't published. I would expect everything from DeepMind to be opensource, except model itself.
In the past AI wasn't really competing with other AI for user dollars. It was more just a bolted on "feature".
Nowadays it makes much more sense to share less.
I'm surprised I'm not able to find this out - can some one tell me whether AlphaEvolve involves backprop or not?
I honestly have no idea how AlphaEvolve works - does it work purely on the text level? Meaning I might be able to come up with something like AlphaEvolve with some EC2's and a Gemini API access?
No, the program and prompt databases use a genetic algorithm.
This is Google solving Google-sized problems. I am afraid the rest of the world will look at this and say - "yeah we want to be like Google and adopt this". That is how Kubernetes took over the world.
Can someone explain how an "agent" is distinct from a "chatbot"?
I'm reading descriptions of agents and it just seems like the same tech deployed with authority to write and a scheduler
A >2% bump in algorithmic performance is pretty impressive given the search approach.
Packing problems are hard, and it is fun to see new interest in the area given these show up in weird places. =3
I wish it can re-write everything in assembly or even binary in a super optimized form, debloat all software including itself and then grow from there.
anyone else feel out-evolved yet?
Not really, only when looking back at the 60's and 70's when most of the important algorithms I use were invented. For example, LR parsing and A*.
Just wait until the MBA's and politicians learn about this Adam Smith guy. A pipedream now, but maybe in the future schools will be inspired to teach about dialectical reasoning and rediscover Socrates.
[end of snark]
Sorry, I'm getting tired of ad-fueled corporations trying to get me to outsource critical thinking.
Machines have been outperforming humans at a variety of tasks for quite a while now. I'm unconvinced that AlphaEvolve can lead to some sort of singularity.
It sounds to me like a hyperparameter optimizer (fast evaluator) guided by AI; I wonder if it's related to Google's Vizier
Would love for AI to kill the leetcode interview
AI will indeed kill the leetcode interview - because once it replaces human SWEs you don't really need to give leetcode-style brainteasers to any human anymore.
It will just move the leetcode interview to in-person.
https://www.interviewcoder.co/ already served that.
that was already solved 2 years back.
Good method to generate synthetic training data, but only works for domains where validation can be scaled up.
It seemed appropriate to use Gemini to make sure my answers were ideal for getting access to the preview.
What is an "advanced" algorithm? How do you differentiate this from other algorithms?
AlphaEvolve sounds like the AI version of a genius coder that never sleeps, insane potential!
Surprised they didn't answer if they tried using AlphaEvolve to improve AlphaEvolve!
Why it emphasize math and computer science more in training stage?
The problems in math and CS are more suitable for training LLMs?
Why Gemini(s)? Why not LLMs fine tuned for LARPing as a researcher?
I find it quite profound that there is no mention of the generation of corresponding code documentation. Without design diagrams, source and commit comments, etc the resulting code and changes will become incomprehensible unmaintainable. Unless that is somehow the point?
sheeesh
yuge
[dead]
[dead]
[dead]
[dead]
[flagged]
[flagged]
[flagged]
[flagged]
[flagged]
[dead]
[flagged]
[flagged]
[flagged]
?
Software engineering will be completely solved. Even systems like v0 are astounding in their ability to generate code, and are very primitive to whats coming. I get downvoted on HN for this opinion, but its truly going to happen. Any system that can produce code, test the code, and iterate if needed will eventually outperform humans. Add in the reinforcement learning, where they can run the code, and train the model when it gets code generation right, and we are on our way to a whole different world.
"Coding" might be solved, but there is more to software engineering than just churning out code - i.e. what should we build? What are the requirements? Are they right? Whats the other dependencies we want to use - AWS or GCP for example? Why those and not others - whats the reason? How does this impact our users and how they use the system? What level of backwards/forwards compatibility do we want? How do we handle reliability? Failover? Backups? and so on and so on.
Some of these questions change slightly, since we might end up with "unlimited resources" (i.e. instead of having e.g. 5 engineers on a team who can only get X done per sprint, we effectively have near-limitless compute to use instead) so maybe the answer is "build everything on the wish-list in 1 day" to the "what should we prioritize" type questions?
Interesting times.
My gut is that software engineers will end up as glorified test engineers, coming up with test cases (even if not actually writing the code) and asking the AI to write code until it passes.
> Any system that can produce code, test the code, and iterate if needed
That isn't every problem in software engineering.
It is not that you get downvoted because they don’t understand you, it is because you sell your opinion as fact, like an apostle. For example what does it mean that software engineering is solved?
What about brownfield development though? What about vague requirements or cases with multiple potential paths or cases where some technical choices might have important business consequences that shareholders might need to know about? Can we please stop pretending that software engineering happens in a vacuum?
Everyone will just turn into a problem solver until there are no more problems.
lol
There's cope in the comments about possibility of some software adjacent jobs remaining, which is possible, but the idea of a large number of high paying software jobs remaining by 2030 is a fantasy. Time to learn to be a plumber.
AlphaEvolve is confirming evidence of an intelligence explosion.
The key ingredient for an intelligence explosion is AI accelerating development of AI.
This is it. It’s happening.
That's possibly a bit too general and an over statement...
Remember this approach only works for exploring an optimization for an already defined behavior of a function which has an accordingly well defined evaluation metric.
You can't write an evaluation function for each individual piece of or general "intelligence"...
I just hope there’s enough time between an actual AI and the “Let’s butcher this to pump out ads”-version to publish a definitive version of wikipedia. After a few days with gemini 2.0 delving into the guts of a spectrum analyser, I’m very impressed by the capabilities. But my cynicism gland is fed by the nature of this everything-as-a-service. To run an LLM on your computer, locally, without internet, is just a few clicks. But that’s not the direction these software behemoths are going.
Yes, but is the inflection point in 12 months or 12 years?
Either way, it's pretty wild.
Maybe this one can stop writing a fucking essay in code comments.
I'm now no longer surprised just how consistently all the gemini models overcomplicate coding challenges or just plain get them wrong.
Claude is just consistently spot on. A few salient comments for tricky code instead of incessantly telling me what it's changed and what I might want to do, incorrect assumptions when it has the code or is something we've discussed, changing large amounts of unrelated code (eg styles). I could go on.
Shame I'm too tight to pay for Claude RN though...
Just ask it to only add comments on complex parts (or not at all). Prompt engineering.
The comment spam is likely a byproduct of RL, it lets the model dump locally relevant reasoning while writing code.
You can try asking it to not do that, but I would bet it would slightly degrade code quality.
The model likely is doing it more for itself than for you.
You can take the code and give it to another LLM instance and ask it to strip all comments.