Ampersand Training Revealed: Simulation Feedback Over Rewards, Scaling RL Long-Horizon
gharik · x · 2026-10-21
Humansand shared how its Ampersand model was trained: unlike assistant models trained with verifiable or preference rewards, Ampersand learns the long-term consequences of its actions through simulation feedback.
Team members parth29, saurabhshah2, and joeyxhong handled the RL scaling work. This diverges from mainstream RLHF/RLVR pipelines and underpins Ampersand's positioning as a long-horizon collaborator that "learns with us."
More from Models
- Musk: Grok Bot is the fastest-growing product he's ever seen, enterprise rollout in weeks — elonmusk · 2026-10-21
- Google: Gemini 4 Argon processed ~25 trillion tokens in first 72 hours — OfficialLoganK · 2026-10-21
- Leaked GPT-6.1-Sol reasoning traces show familiar GPT caveman CoT once spaces are restored — teortaxesTex · 2026-10-21
- Gemini 4 Scores Just 78% on FrontierMath Tier 4 as OpenAI Models Saturate It — haider1 · 2026-10-21
- Hamel Husain: public benchmark quality is abysmal — broken graders, sloppy harnesses — HamelHusain · 2026-10-21
- 163MB open-source speech model Bark 3.0 beats Whisper (57x bigger) and NVIDIA Parakeet (15x bigger) — Scobleizer · 2026-10-21