NVIDIA представила SoL-Pi Систему, которая автоматически улучшает обвязку ИИ-агентов и сокращает расход токенов почти вдвое без заметной потери качества. Вместо ручной настройки исследователи запускают циклы автоматических экспериментов над самой обвязкой в разных репозиториях и средах с проверяемым результатом. В итоговую версию попадают только механизмы, которые стабильно проходят отбор. Таких механизмов осталось четыре: объединение действий, сжатие контекста прямо во время работы, более эффективная обработка наблюдений и сокращение делегированного чтения с сохранением важных доказательств. На EdgeBench из 51 задачи SoL-Pi показала качество на уровне базовой системы с GPT-5.6 Sol и Opus 5, при этом сократила трафик токенов почти на 50%. По оценке авторов, это уменьшает расходы на API примерно на треть. Экономия составляет около $8,75–13,50 в час по сравнению со штатными системами Codex и Claude Code и $4,36–5,71 в час по сравнению с базовой обвязкой. Код опубликован на GitHub в NVlabs. 🤖 Статья: https://arxiv.org/abs/2609.20519
From the original source:
Telegram — Вайб-кодинг
Read full article at Telegram — Вайб-кодингThis material was compiled from public sources. If its publication affects your rights → Request removal
NVIDIA introduced SoL-Pi
NVIDIA introduced the SoL-Pi System, which automatically improves the binding of AI agents and reduces token consumption by almost half without a noticeable loss of quality. Instead of manual configuration, researchers run cycles of automated experiments on the harness itself in different repositories and...
AI Summary & Key Takeaways
- NVIDIA introduced the SoL-Pi System, which automatically improves the binding of AI agents and reduces token consumption by almost half without a noticeable loss of quality
- Instead of manual configuration, researchers run cycles of automated experiments on the harness itself in different repositories and...
NVIDIA introduced SoL-Pi
Related stories
Judge dismisses Sam Altman’s sister’s sexual abuse lawsuit
On Tuesday, US District Judge Zachary Bluestone dismissed the lawsuit Ann Altman filed last year alleging that her brother, OpenAI CEO Sam Altman, had sexually abused her throughout her childhood. As reported earlier by The Guardian, the judge wrote that over the last year, Ann…
Anthropic pitches AI safety bill that may block a model if 50+ people could be killed
The proposed policy seems to offer protections against AI run amok but actually forbids federal agencies from blocking a model's release without the permission of a judge.
Hugging Face CEO warns that Netflix’s OpenAI hack documentary isn’t accurate
Before you share the Netflix documentary "AI Gone Rogue" take note. Clem Delangue, Hugging Face's co-founder who is featured in the film, says it went "full doomer mode" and isn't accurate.
Stop paying for AI, run everything locally (here’s how)
Big labs are not going to talk about this but continual learning the idea of improving the models for your [music] own use cases for your own workflows. Big labs have no incentive of providing users with that [music] because they're training their models for the normal…