@aaronjmars.eth
cool article by https://x.com/thinkymachines on forecasting using AI
> Mantic used Tinker to RL fine-tune gpt-oss-120b on ~10k binary "will event happen by date?" questions, lifting its forecasting score from 38.6 to 45.8 - matching frontier models like Gemini 3 Pro and GPT-5.
> The fine-tuned model also adds real diversity to model ensembles, earning ~40% weight in the optimal 5-sample ensemble alongside Grok 4, GPT-5, and Gemini 3 Pro.
1 reply
0 recast
16 reactions