Capital Bets on AI Building AI: RSI Paradigm & Frontis-MA1's Solution
新智元 · wechat · 2026-08-01
As the potential for AI to participate in its own R&D emerges, Recursive Self-Improvement (RSI) is moving from sci-fi to industrial reality, attracting over $2.5 billion in funding. The article distinguishes between AI4AI, Meta-Evolution, and RSI, noting that true RSI requires an executable, verifiable engineering closed-loop.
Xianyuan Tech and Tsinghua University released the Frontis-MA1 model and the full-stack open-source suite OpenMLE to bridge this gap. The core innovation reconstructs linear context into an evolvable graph that allows crossover and backtracking, turning even failed attempts into training signals. In the MLE-Bench evaluation, the system pushed the effective submission rate to 71.21%, proving the effective stacking of model and search gains.
More from Venture
- The simplest agent money-making idea: let agents take gigs on Fiverr/Upwork — yangyi · 2026-10-16
- Solo creator swears by just two tables: one for 100K YouTube subs, one for $10K MRR — jackedAJ · 2026-10-16
- Proofpoint to buy Varonis for $6.3B in cash, a 52% premium, closing in 2027 — shashib · 2026-10-16
- Investor: Anthropic may never publish ARR again, was FCF-positive in Q3 with 75% compute on training — mjdramstead · 2026-10-16
- From $100K debt to $1K in 4 days: solo founder's AI website journey — gefei55 · 2026-10-16
- Barclays: Tesla–SpaceX merger could be announced before December, giving Musk 56% voting power — aakashgupta · 2026-10-16