The video discusses Sakana AI, a Japanese startup that has achieved a $265 billion valuation by taking a contrarian approach to AI development. Unlike large companies like OpenAI that focus on building massive, all-encompassing AI models, Sakana AI has developed 'Fugu,' a smaller AI system designed to manage a team of specialized AIs. Fugu's function is to receive a query, intelligently select the most appropriate AI model for each component of the task, and then synthesize their outputs into a final answer. This approach is significant because one of Sakana's founders was a co-author of the foundational research that underpins current leading AI models such as ChatGPT, Gemini, and Claude. The video also touches upon regulatory actions, noting that Anthropic had to restrict access to its powerful 'Fable 5' model in the US due to a government order. Sakana AI claims that Fugu can achieve comparable performance to Fable 5 at a significantly lower cost, starting at $20 per month. Furthermore, Fugu has reportedly outperformed Opus, Gemini, and GPT on challenging coding benchmarks. However, the video acknowledges mixed early reception, citing Perplexity's earlier release of a similar concept called 'Model Council' and quoting Wharton Professor Ethan Mollick, who described a comparable system as 'painfully slow' and less effective in practical application than Fable 5. The central question posed is whether Sakana AI's Fugu represents genuine innovation or is merely hype.

Matty | AI Models & Monetization

While America builds bigger AI, a tiny Japanese startup did the complete opposite and just became worth $265 billion.

A Japanese lab called Sakana AI built Fugu, and the idea is wild. Companies like OpenAI spend billions training one giant model to be good at everything. Fugu flips that. It's a small AI whose entire job is to manage a team of other AIs. You ask one question, it picks the best model for each part, then combines the answers.

And one of Sakana's founders co-wrote the original research paper that unlocked the tech powering Chat GPT, Gemini, and Claude today. Two weeks ago, Anthropic had to ban public access to its strongest model, Fable 5, on US government order. Sakana claims Fugu matches Fable on benchmarks starting at $20 a month. On brutal coding tests, Fugu beat Opus, Gemini, and GPT.

But early signs are mixed. Perplexity shipped a version months ago called Model Council. The scores are Sakana's own. And Wharton Professor Ethan Mollick said it was painfully slow and didn't match Fable 5 in real use.

So, is it hype or the real deal?

Small AI Fugu manages multiple models for better answers, starting at #AI #Tech

While America builds bigger AI, a tiny
Japanese startup did the complete
opposite and just became worth $265
billion.
A Japanese lab called Sakana AI built
Fugu, and the idea is wild. Companies
like OpenAI spend billions training one
giant model to be good at everything.
Fugu flips that. It's a small AI whose
entire job is to manage a team of other
AIs. You ask one question, it picks the
best model for each part, then combines
the answers. And one of Sakana's
founders co-wrote the original research
paper that unlocked the tech powering
Chat GPT, Gemini, and Claude today. Two
weeks ago, Anthropic had to ban public
access to its strongest model, Fable 5,
on US government order. Sakana claims
Fugu matches Fable on benchmarks
starting at $20 a month. On brutal
coding tests, Fugu beat Opus, Gemini,
and GPT. But early signs are mixed.
Perplexity shipped a version months ago
called Model Council. The scores are
Sakana's own. And Wharton Professor
Ethan Mollick said it was painfully slow
and didn't match Fable 5 in real use.
So, is it hype or the real deal?

Get Our Mobile App

Small AI Fugu manages multiple models for better answers, starting at #AI #Tech

Transcription