Thinking Machines open-sources RL recipe for fine-tuning models as event probability forecasters
clarejtbirch · x · 2026-09-05
Tinker, Thinking Machines Lab's fine-tuning service, added a forecasting recipe to its cookbook: fine-tune a model (Qwen3.8-27B in the example) with RL on binary prediction-market questions from Prophet Arena, teaching it to forecast the probability a market resolves YES given a historical information snapshot.
Key details:
- Uses a chronological data split with Brier reward; no tools or live search
- Ships data.py (download/verify/inspect the split) and train.py; just set TINKERAPIKEY to run
- Also supports GLM-5.3 (peft config + glm53highreasoning renderer)
More from coding & agent
- Perplexity CEO Declares the End of Markdown Files: Bitter Lesson Wins Again — AravSrinivas · 2026-09-05
- Codex rust-v0.153.3 Hotfix Adds GPT-6-Astra to Amazon Bedrock — github-actions[bot] · 2026-09-05
- Raspberry AI runs a fashion design agent on LangGraph and LangSmith — LangChain · 2026-09-05
- Debunked: OpenAI Agents Didn't "Turn Evil" — They Made 15,000+ Edits to a German Wiki to Beat Task Timers — irinarish · 2026-09-05
- From Prompt to Skill to MCP Server: One Dev's LLM Writing Assistant Evolved Three Times — unixterminal · 2026-09-05
- Hamel Husain highlights random-string diversity seed as a standout prompt trick — HamelHusain · 2026-09-05