Agents
Outlook where frontier AI is headed next 18 months: The AI reasoning training + harness loop works if you can produce enough data and reason…
Outlook where frontier AI is headed next 18 months: The AI reasoning training + harness loop works if you can produce enough data and reasoning traces (via verifiers). Proven with code and math result
Outlook where frontier AI is headed next 18 months: The AI reasoning training + harness loop works if you can produce enough data and reasoning traces (via verifiers). Proven with code and math results. Frontier labs desire to scale this horizontally to many more domains as increasing generality isn't emerging from the NN substrate through data/scale. But it's very expensive and slow to do this training loop manually. Enter "RSI" discourse. Labs want to automate this training loop they now know works. How? By leveraging coding agents to build world models aka symbolic verifiers. The bet is automated horizontal domain scaling through automatic symbolic world modeling which offers the critical feedback into post-training. This is important because many domains are intolerant of learning "online". Symbolic world models offer a path to learning/training "offline". It's safe for an AI system to test hypotheses against a git repo with tests with no consequences. It's not safe for AI to test against online systems eg SaaS databases, tax filings, manufacturing spin up, etc. A couple other trends to overlay. We're seeing incredible capabilities emerge from better symbolic harnesses around frontier reasoning models. As time goes forward these harnesses produce the necessary training traces required to teach the models to emulate the harness, obviating the need for the fat harness. The other related push is towards multi-agent scaling. Many problems are search constrained (we see this in math right now) where a single linear CoT is not the optimal way to find a solution when you're optimizing for time.
Related
- We're teaming up with Hugging Face to make open agent traces the fuel for the next generation of open coding models. You can now upload @Dro…
- Production traffic from frontier models is a golden data asset. If you can efficiently mine the traces, filter for quality, and fine-tune sm…
- This 10-min read from @Vtrivedy10 changes how you build AI agents. Most people are stuck in the same loop; switching models when agents brea…
- Most people training agentic LLMs with RL right now have a silently broken training loop and have no idea. Here's the trap: single-turn RL w…
Source: Francois Chollet (X) | 2026-08-06