Intel / video

Scaling to Long Horizons — Ross Taylor & Chengxi Taylor, General Reasoning

Source: youtube.com
Author: AI Engineer
Date: Published Jul 31, 2026

Context

Training an on lengthy tasks creates a feedback delay: the final outcome may arrive long after the decisions that caused it. General Reasoning's presentation describes using a second model, called a critic, to estimate how promising an unfinished attempt is. Those estimates can provide a learning signal before the entire task finishes, reducing the need to leave training hardware waiting for final results. The exchange is consequential: training can proceed sooner, but it now depends partly on a learned estimate that can be wrong. Longer tasks therefore require choices about when feedback arrives and how trustworthy it is, alongside enough memory to retain the agent's working .

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.

Key quotes

“So again, that also shows you how powerful RL is. You can have a sota base model, but that is not enough.”
Ross Taylor

“Like a 1 billion parameter model with RLHF was outperforming 175 billion models. So two orders of magnitude fewer parameters, but getting better results.”
Ross Taylor

“it was just like the bitter lesson, like the most purest form of bitter lesson possible. Like, better base models, more RL computes, bigger context windows, and that's all you need for this kind of emergent behavior.”
Ross Taylor

“Long horizon task is not just an engineering problem. It is a mindset.”
Chengxi Taylor

“we gave all the frontier models a 100K to start. All of them lost”
Chengxi Taylor

Read the source www.youtube.com