Sakana AI's Recursive Self-Improvement (RSI) Lab

by hardmaru • 2 months ago

29 22 2 months ago

I think the most impressive thing about Sakana.ai is their relentless pursuit of whatever is hype right now.

Genuinely it take a lot of work and talent to be this hype-motivated and completely ignore anything except what is popular on X at any given time.

Note: RSI is an incredibly important topic -- I just don't care to listen to Sakana on this matter -- they are the epitome of "hypebeast" https://www.urbandictionary.com/define.php?term=hypebeast

(Thanks for sharing hardmaru)

aleph__one• 2 months ago

On the contrary, I find them to be one of the least hypey companies. For instance, a cursory familiarity with David Ha's work would inform you that the team has been doing this kind of stuff for quite a long time.

laughingcurve8• 2 months ago

OpenAI is not Sam Altman Anthropic is not Dario Amodei and Sakana is not David Ha

Organizations, especially businesses, are not individuals. If the implication is that David Ha has always been doing this, and will always be doing this, and that Sakana is David Ha ... then that's a far worse insult to the employees at Sakana than my little tweaking.

mlmonkey2• 2 months ago

Being a Hypebeast leads to a rich acquisition.

laughingcurve8• 2 months ago

TRUE

granitepail• 2 months ago

A great deal of their research has been focused on zigging where others zag. Their paper "Continuous Thought Machines" (https://arxiv.org/abs/2505.05522, presented at NeurIPS) was posed specifically under the framing of there needing to be more fundamental research beyond squeezing as much as we can out of relatively vanilla transformer stacks. It is very biologically inspired and unique.

Now that models are getting stronger at agentic work, it is very natural that many labs are chasing some form of auto-research.

laughingcurve8• 2 months ago

It's natural for literally any AI lab to end up doing auto-research, since 'auto-research' is literally just 'autonomous AI' which is the whole darn point of all of this. I'm not going to hand out genius brownie points to folks working on RSI because of course its powerful. How about we hand brownie points to the folks who do things that are not hype and end up being important/powerful?

was posed specifically under the framing of there needing to be more fundamental research beyond squeezing as much as we can out of relatively vanilla transformer stacks.

Not to be contentious, but this is so broad of a description that it could include literally thousands of papers in the last year or two. I'm imagining double digits or more if we go back the full decade.

I'm saving brownie points for people who deserve them

artninja1988• 2 months ago

I think sakanas papers are one of the more creative, not just gunning for incremental benchmark improvements. But yeah I agree that they can be a bit (or very) hypey. But regardless, I want to see more of their kind of research than endless benchmark chasing. All the best to David Ha and the team!

laughingcurve8• 2 months ago

Can you explain how 'recursive self-improvement' functions without 'endless benchmark chasing'? I mean, RSI is literally that.

What do you think they're improving on? How would a model self-improve without some metric/data of some kind to check? When you have metrics+data, that is a benchmark. And yes, simulations and or soft-verification like LLM judges are still a kind of benchmarking. Maybe its not a static benchmark they can easily hack.

Folks -- RSI does not mean the self-improvement is them going to therapy and seeking inner peace to overcome trauma.

ansk• 2 months ago

Kind of a backwards take, both in your interpretation of this announcement and the company in general.

The point of this announcement is to draw attention to the fact that the currently hyped topic is what they have been working on since their inception. If anything, it gives off a Schmidhuber-esque 'actually it was me who invented that' vibe. But trying to retroactively claim credit for the hype is nowhere near the same as following the hype.

As for your impression that the company is more generally hype-chasing, I'm really not sure how you would come to that conclusion. At the time of their founding, chatbots were the hype on the product side and model scaling was the hype on the research side -- topics they have largely eschewed. They instead were founded with a focus on evolutionary and collective intelligence and have maintained a fairly cohesive research direction ever since.

laughingcurve8• 2 months ago

counter point: https://sakana.ai/asal/

"For the past 300,000 years, Earth has had only one form of advanced intelligence on it: humans. With the recent advent of AI foundation models, some believe we are at the dawn of a new kind of intelligence. As AI continues to evolve, we may witness the proliferation of diverse intelligent lifeforms coexisting with us."

I am not even going to link more than one thing I think I've made my point

dominotw• 2 months ago

recursive self-hype

pjjpo• 2 months ago

I've never considered them a hypebeast, just a scam to get investor money as easily as possible. They seem to be quite successful really.

laughingcurve8• 1 month ago

Dang, I thought I was being harsh I wouldn't say scam but maybe all AI is a scam to some.

hmmmmmmmmmmmmmm• 2 months ago

Fortunately a rational society like Japan is not as interested in outsourcing their capacity to curve fitting models as other societies.

tokioyoyo• 2 months ago

What?

mlmonkey2• 2 months ago

I don't know how RSI aligns with DPI ("Data Processing Inequality", which states that, basically, unless you have an infinite supply of real data, you will suffer from model collapse). Models can't keep improving themselves infinitely. See, for example, https://arxiv.org/html/2601.05280v2

tmaly• 2 months ago

I heard about the ShinkaEvolve on a podcast where the guest had used it to evolve an agent harness for a less capable model.

I ended up borrowing the ideas from it for one of my own personal projects.

algo_trader• 2 months ago

Do you actually understand why evolution methods are beneficial?

SGD generates a stronger learning signal,is more efficient, and scales better. Using it end-to-end makes it stronger yet.

Yet somehow mixing in a weaker blunt evolution stage improves the result?

Sakana AI's Recursive Self-Improvement (RSI) Lab

Related Stories