Intel / article

【開催レポート】Sakana AIシンポジウム：Jürgen Schmidhuberが語るAIのこれまでとこれから、日本から挑む言語モデル・世界モデル・Physical AI

Source: unknown
Author: Sakana AI
Date: Published Nov 1, 2026

Why it matters

Sakana AI says it will pretrain its own language models, with a small model (Sawara Small, provisional name) targeted for spring 2027. This signals a new Japan-based model lab and a research direction in world models and autonomous discovery.

Key takeaways · AI-distilled

Sakana AI says Sawara will be pretrained from scratch rather than adapted from another model, and built as -native for tool-integrated reasoning, coding and long-horizon tasks. of Sawara Small is set to start this December.
Chief Scientist Takuya Akiba, citing Epoch AI estimates, argued development is mostly a research contest: final training runs at OpenAI are estimated at about 10% of R&D compute, so research efficiency matters as much as GPU count.
In Sakana's Physical AI design, a small vision-language model proposes plans, a world model predicts and scores them several steps ahead, and only a few survivors run on real hardware, with results used to improve the VLM.
To limit reward hacking in its self-improvement loop, Sakana says it uses a frozen gate that accepts changes only with repeated gains across evaluations, statistical evidence, and gains on held-out tasks.
Schmidhuber argued artificial mathematicians improved fastest because math is deductive from given axioms, while finding a new physical law requires induction from raw observations and is much harder.

Terms in this piece · Glossary

LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
pretraining — The first, biggest phase of building a model: training it on enormous amounts of text so it learns language, facts, and reasoning in general.

Source link unavailable.unknown