Inherent's Faraday Agent Replicates Scientific Papers Using a 27B-Parameter Model, Beating Frontier Systems
London startup Inherent, founded by Google DeepMind alumni and recently funded with a $50 million seed round, released an AI agent called Faraday that autonomously reproduces findings from published scientific papers. The company says Faraday outperformed Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5 on this task while running on a much smaller 27-billion-parameter Qwen 3.6 model, trained via reinforcement learning to cultivate 'research taste.' Inherent's stated longer-term goal is an AI scientist agent capable of discovering new knowledge across scientific fields.
Skynet Chance (+0.05 pp): An agent trained by outcome-based reinforcement learning to develop autonomous 'research taste' and self-directed experimentation is precisely the kind of open-ended goal-seeking system whose behavior is hard to specify or supervise. That a 27B model can match frontier systems also weakens compute-based control levers, since capable autonomous research agents become cheap to replicate.
Skynet Date (1 day sooner): Demonstrating that small, inexpensive models can perform frontier-level autonomous research work compresses the timeline for widely available self-improving research agents. Company-reported benchmark results from a 12-person startup limit how much weight the acceleration deserves.
AGI Progress (+0.04 pp): Independently reproducing published scientific results without being given the answer requires long-horizon planning, experiment design, and judgment — capabilities central to AGI rather than narrow benchmark performance. Achieving this at 27B parameters suggests capability gains from training method rather than scale alone.
AGI Date (1 day sooner): If reinforcement learning on outcomes can substitute for raw scale in eliciting research-level reasoning, the compute and capital barriers to AGI-relevant agents drop and more labs can pursue them in parallel. The result is self-reported and on a single task type, so the timeline shift is meaningful but not dramatic.
<< All AI news for August 22, 2026
[ Get the daily index digest on Telegram → ]Related AI News
- Google Adds Computer-Use Agent Capabilities to Gemini Mac App
- Meta's Muse AI Agent Reportedly Leaks Private Data Across User Accounts
- Google Launches Gemini 4 Neon, a Fast Frontier Model for Agentic Coding and Enterprise Work
- Meta's Muse AI Agent Accessed Private Messages and Behaved Erratically, Internal Chats Reveal
- Microsoft Positions Copilot as 'New OS for Work,' Launches Autonomous Autopilot Agents With Governance Controls