sakana's run over the past 2 months has to be studied. since this post, it's been new research basically every few days: multi-agent orchestration (their big bet right now): Fugu - multi-agent orchestration packaged as a foundation model Conductor - orchestrate a team of agents using plain natural language Trinity - an evolved LLM "coordinator" that learns how to route work between models new architectures: DiffusionBlocks - train a neural net one block at a time instead of all at once TWELL - sparser, faster, lighter transformer language models KAME - real-time speech-to-speech AI that can pull in fresh knowledge mid-conversation Digital Ecosystems - multi-agent neural cellular automata (basically artificial life) methods + products: String Seed of Thought - prompting trick for diverse, distribution-faithful generation Marlin - an "ultra deep research" assistant (in beta) applied / deployments: SMBC - auto-generate proposals with a team of agents gov't (MIC) - map social-media space and flag disinformation DEEP DIVE - partnership for AI-powered intel analysis a 20-person lab moving at this pace is crazy, japan is back
