35B Model Beats 1T: BigBang Achieves Self-Evolution with 100% AI-Synthesized Data
量子位 · wechat · 2026-08-07
The Endless Frontier team released BigBang-V1, the first foundation model natively trained via Recursive Self-Improvement (RSI). With 35B parameters and 100% AI-synthesized post-training data, it ranks first among 35B models in several hardcore scientific benchmarks, even outperforming 1T-scale models.
The core innovation is a self-evolving synthetic data pipeline:
- Dual-Agent Collaboration: A Generator Agent proposes and solves frontier scientific tasks, while a Critic Agent adversarially reviews and filters the data.
- Inner & Outer Loops: An inner loop enables AI self-play for data generation; an outer loop uses real training results to calibrate the Critic, preventing model collapse.
- Task Criteria: Uses scientific tasks that are both cutting-edge and verifiable as training vehicles, allowing the data production system to evolve alongside model capabilities.
This mechanism breaks the bottleneck of human-generated data, enabling AI to autonomously decide what the next generation of models should learn. It demonstrates strong multi-step evidence verification and self-correction in complex tasks like bioinformatics and paper code reproduction.
Related event: BigBang-V1: 35B Model Beats 1.6T Parameter Giants(3 posts)→
More from Models
- OpenAI stealth-testing mysterious internal model 'Aeon' in Codex desktop app and Dots — cedric_chee · 2026-10-16
- Argon edges out astra as daily driver, but Claude still wins heavy coding work — iruletheworldmo · 2026-10-16
- NVIDIA ships FP4-quantized Qwen3.8-Flash for local use, trending on Hugging Face — nvidia · 2026-10-16
- RoboVista benchmark: Gemini-4 Argon hits 81.9% on robotics understanding — berkeley_ai · 2026-10-16
- Dev reports decision model jev beats gpt-6-luna on quality in real training pipeline — genmon · 2026-10-16
- Google won't publish Argon 4 limits in AI Pro—users compare experiences — Consistent-Text-2662 · 2026-10-16