BaseLabs open-sources mammoth-1 and elephant-1, top-ranked decision models
saranormous · x · 2026-10-22
BaseLabs released and open-sourced mammoth-1 (27B, the #1 ranked decision model) and elephant-1 (9B, best under 15B, beating Jev), with all weights public plus a technical report on training and data methods.
The surrounding discussion crystallizes a key distinction: recursive improvement is here via agentic workflows — models can readily iterate on models one rung down the capability ladder (an agentic LLM training a decision model) — but recursive self-improvement isn't: attempts to have LLMs improve themselves (building their own RL environments, self-training, curating mid-training data) still require heavy human involvement. This decision-model training loop is one of the first outer loops that actually worked.
Related event: BaseLabs Open-Sources Decision Models mammoth-1 and elephant-1(4 posts)→
More from AGI Musings
- Ethan Mollick: AI coding tools help, but product bottlenecks lie elsewhere — emollick · 2026-10-22
- Oxford study: the public is an AI shareholder and deserves a happiness dividend — SandraWachter5 · 2026-10-22
- Alignment discourse checked: 'spooky emergent behavior' traced to fixable setup choices — random_walker · 2026-10-22
- Anthropic Publishes Its First Annual Public Benefit Report With Five Pillars — AnthropicAI · 2026-10-22
- "Mechanomorphizing": describing people with machine concepts — PierceLilholt · 2026-10-22
- Musk tells Isaacson: millionfold-smarter AI means no human control, money goes extinct — TheMoonMidas · 2026-10-22