Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Sakana Fugu

sakana.ai

Post reply on HN
Print view View original

41–50 of 141 posts

Re: Sakana Fugu

#41

Post by chvid »

This would have been much more interesting and impactful if it had relied on open source models rather than commercial models that are only availble via an API.

The reasoning chains could have been used, and the resulting combined model could easily and effectively have been distilled.

Re: Sakana Fugu

#44

Post by totetsu »

chenzhekl wrote: post #43

I probably will never pay to Sakana, as they are involved in military contracts. https://japannews.yomiuri.co.jp/politics/defense-security/20...

Yeah, I was trying to parse their "defense policy" https://sakana.ai/company-info/defense-policy.html?lang=en But it seems like lot of words to say we have no policy and we'll just go along with the powers that be. Like they rely on deferring to the Pacifist constitution, which the current administration if moving mountains to try and change. And when it it you can bet they will not want to give up their defense contracts.

Re: Sakana Fugu

#45

Post by chenzhekl »

david_shi wrote: post #22
puttycat wrote: post #17

Can someone explain this in layman terms? I don't understand any of it

It's similar to this: https://openrouter.ai/blog/announcements/fusion-beats-fronti... Basically, if you combine a bunch of near-frontier models (like GPT 5.5, etc) you can get performance that sometimes surpasses top line models like Claude's Fable. Sakana seems to have a separate approach using a domain specific model to perform the model routing step.

But it's priced the same as frontier models. Why do I not directly pay for frontier models?

Re: Sakana Fugu

#46

Post by sofixa »

holistio wrote: post #26
rvz wrote: post #25

Earlier quoted context omitted.

Pay $0 to run a local model or even a cheap DeepSeek V4 model via their API which is close to free per million tokens. These prices are just going to get raced to $0.

Maybe. But for now it's fascinating how $200/month has kind of become a normal tier. It's similar to how AirPods normalised all of us having $300+ headphones. All of us would have scoffed at the idea a decade ago.

The Sony WH-1000XM series and the Bose QC35 were the standard quality headphones years before AirPods were a thing, and both retailed at $300+.

Re: Sakana Fugu

#47

Post by sofixa »

kijin wrote: post #29
rvz wrote: post #25

Earlier quoted context omitted.

Pay $0 to run a local model or even a cheap DeepSeek V4 model via their API which is close to free per million tokens. These prices are just going to get raced to $0.

Not while the hardware required to run a local model at an acceptable speed costs way more than $200. Guess what, the big players are hoarding all the RAM and GPUs so that other people can't afford decent hardware. It's working out beautifully for them!

> Not while the hardware required to run a local model at an acceptable speed costs way more than $200

It's $200/month. You have to take into account energy costs and all the rest of a system, but if you break even within 1-2 years ($2400-$4800) it'd be a pretty good deal. And $4000 buys you a pretty decent system.

Re: Sakana Fugu

#48

Post by NitpickLawyer »

GolfPopper wrote: post #15

This is a joke, right?

Not necessarily. There were some tests last year-ish from hf that showed that simply alternating (randomly) between claude and gpt (whatever their versions were at the time) on a task produced better results than either of them individually. So during a task, the first call was sent to one, then the other and so on.

There's also the concept of "smart routing" requests based on some heuristics / embeddings. You'd get "simple" tasks handled by smaller (cheaper) models and use a bigger model to curate / sort / merge the results.

There's a lot of things to try here. I wouldn't personally pay for this service, but I don't think it's "a joke"...

Re: Sakana Fugu

#49

Post by nickandbro »

chenzhekl wrote: post #43

I probably will never pay to Sakana, as they are involved in military contracts. https://japannews.yomiuri.co.jp/politics/defense-security/20...

I imagine if it was Deepseek partnering with the CCP it would be different?

Re: Sakana Fugu

#50

Post by njoyablpnting »

Looking at the technical report I'm a bit confused. The improvement from using their orchestrator models seems minimal (in some cases lower than just the model which I'm assuming is in the orchestrator's pool?). Maybe it's sort of acting as an additional reasoning step upfront? Sort of like how if you asked Claude to create a plan for how best to prompt itself, you would probably end up with a better result than just the base prompt.

Also, from the technical report, looks like they're training on the output of Claude Code, etc. I'm guessing this doesn't violate TOS because they're technically not a directly competing model. This brings me to what I see as the main risk with this service, which is that it seems like an easy thing for a frontier lab to make obsolete, either by models beginning to converge in terms of strengths or by improving their own harnesses to include more of this meta-reasoning.

Post reply on HN