Sakana AI to pretrain its own LLM 'Sawara', first model due spring 2027
kaixhin · x · 2026-11-02
Sakana AI hosted the 'From World Models to Real-World AI' symposium in Tokyo (400+ attendees), headlined by new Chief Scientific Advisor Jürgen Schmidhuber, who co-wrote the 2018 'World Models' paper with CEO David Ha.
Key announcement: Chief Scientist Takuya Akiba unveiled the Sakana Sawara LLM series (tentative name), pretrained from scratch in-house. The first small model, 'Sawara Small', is slated for release in spring 2027 if things go well.
Schmidhuber's keynote traced 1990s world models, artificial curiosity, pretraining and distillation, arguing generative AI has realized only about half of his AI vision—autonomous scientist agents and Physical AI haven't scaled yet, an opportunity open outside big labs, with Japan a promising home. The event also covered open-ended autonomous discovery and world models & Physical AI.
Related event: Sakana AI to Release Homegrown LLM Sawara in Spring 2027(2 posts)→
More from AGI Musings
- Building AI hardware design loops while fearing models will one-shot it — samiramanabi · 2026-11-02
- Jobs focused on "fixing AI" mistakes surge over 85% — alejandroll10 · 2026-11-02
- eigenrobot on operating systems he doesn't fully understand — at work and in society — eigenrobot · 2026-11-02
- AI tutor asks how to spell 'seventeen': great lift in theory, implementation matters — Pascallisch · 2026-11-02
- Non-coder ships OneShot Redux, an open-source local screenshot-card Chrome extension — SuB8u · 2026-11-02
- Answering "Human Slop Is 1000x Worse": AI's Attention Edge Is the Bull Case for Ongoing Innovation — noahsolomon · 2026-11-02