Sakana AI launches Fugu, an LLM orchestrator using GRPO to dynamically route queries across external models

read more: https://www.alphaxiv.org/abs/2606.21228

@askalphaxiv https://github.com/trotsky1997/OpenFugu

The picture shows regular Fugu’s fast routing mechanism: it reads the user request, but does not answer it itself.

The key part is the “lightweight head,” a small extra decision layer attached to the language model.

That layer looks at the model’s hidden state, meaning its internal summary of what the request is about.

Then it gives one score to each available worker model, and the highest score decides which outside LLM gets the task.

The red diagonal marks a small tuning trick, where they adjust only a tiny part of the model’s internal weights so it gets better at choosing the right worker.

@askalphaxiv this is massive because it's not about training one larger model, it's about recognizing that a single agent can't keep up with the complexity of user queries, and dynamically combining them on demand for better performance

@askalphaxiv @hardmaru Sorry this model is weaker than Opus 4.8. I cancelled my subscription

@askalphaxiv Don’t see the part of the technical report where you give them $200 and they give you nothing of use. @SakanaAILabs’s tech is good. It’s quite good. It’s real. Their product is complete fraud. Complete highway robbery.

@askalphaxiv basically made a fantasy football team out of AI models

and somehow that beats just signing one good player lmao

@askalphaxiv One small prompt = 50% usage on lite

6 prompt and reached weekly usage in a day and half

@askalphaxiv stop building massive models for simple shit

@askalphaxiv corticothalmic loop as a SaaS offering

@askalphaxiv 🔆⏫

@askalphaxiv 技術レポート楽しみです。オーケストレーションが実運用でどう広がるのか注目しています。

@askalphaxiv @SakanaAILabs I actually found that it’s results weren’t any better than a SOTA model for a specific kind of task.

@askalphaxiv @SakanaAILabs the big providers have been doing this for quite a while

@askalphaxiv fascinating approach – fugu’s speed routing could revolutionize ai workflows!

@askalphaxiv @SakanaAILabs This was interesting to read but I guess it's more useful if the underlying models are much lighter than frontier models (I know it's doable but I'd like to see performance reports)

@askalphaxiv @SakanaAILabs https://www.hiddenleafnetworks.com/research/bridge-mode

@askalphaxiv @askalphaxiv wonder how Fugu handles latency with all those models. sounds cool but could get slow?

@rohanpaul_ai http://Z.ai and Heavenly sword. Two different Chinese companies announced they have their own Mythos. It looks like Fable is coming back on Tuesday June 29