Thinking Machines выпустили ИИ-модель, которая самостоятельно обучила другую модель Бывшая команда OpenAI выпустила свою открытую ИИ-модель Inkling 2 Small, которой дали задачу - научить Qwen3.5-4B прогнозировать временные ряды. Дальше Inkling 2 Small действовала одна: написала собственный скрипт оценки, нашла баг в нём же, перебрала несколько подходов к обучению (SFT, затем попытку RL - та оказалась неудачной, модель сама это заметила и откатилась), и в итоге сдала рабочий чекпойнт. Заняло это 4,5 часа и $426 из выделенного бюджета в $500. Итоговая модель ошибается примерно на 15% меньше, чем необученная версия, на данных, которые видела впервые. Надо отметить, что это постановочный бенчмарк, но интересно к чему все идет. Сама модель Inkling 2 Small состоит из 594 млрд параметров, из которых 17 млрд активных, архитектура MoE. Веса откроют через неделю. Модель отстает от DeepSeek-V4.1-Flash.
From the original source:
Telegram — blockchainRF (блокчейн / мозг / space / WEB 3.0)
Read full article at Telegram — blockchainRF (блокчейн / мозг / space / WEB 3.0)Thinking Machines released an AI model that independently trained another model
Thinking Machines released an AI model that independently trained another model. The former OpenAI team released their open AI model Inkling 2 Small, which was given the task of teaching Qwen3.5-4B to forecast time series. Then Inkling 2 Small acted alone: she wrote...
AI Summary & Key Takeaways
- Thinking Machines released an AI model that independently trained another model
- The former OpenAI team released their open AI model Inkling 2 Small, which was given the task of teaching Qwen3.5-4B to forecast time series
- Then Inkling 2 Small acted alone: she wrote...
Thinking Machines released an AI model that independently trained another model
Related stories
Moonshot revealed the weights of Kimi K3.1 Pro
🚀 Moonshot has revealed the weights of the Kimi K3.1 Pro Chinese Moonshot AI has published the full weights of its new model Kimi K3.1 Pro on Hugging Face. Previously, it was only available through the API and application. What's interesting: ➖ 2.8 trillion parameters, of which 104 billion are active for each token. ➖…
GPT-6.2 Leaks Are Getting Seriously Interesting
OpenAI's next model is getting very close, and the leaks around it are getting harder and harder to ignore. Two mysterious model names appeared in OpenAI's official list of models for a few minutes and then vanished. A giant new model called Bell is reportedly sitting inside…
Astana Hub has created an IT ecosystem worth $6.7 billion in eight years: what does it include?
Astana Hub celebrates eight years. The total income of participants reached $6.7 billion, export of IT services - $2.4 billion. We tell you how many companies and startups are united by the ecosystem.
Anthropic has eliminated token overspending in Claude Code
✔️ Anthropic has eliminated token overspending in Claude Code Previously, Claude Code background subagents performing code reviews or tests updated their status every 30 seconds and each time the base model re-read the entire correspondence history. Users online noted that...

