본문으로 건너뛰기
AI NEWS JOURNAL
The SequenceAI NEWS / JOURNAL

The Sequence Knowledge - Issue #973: Learning RSI: World Models as Training Grounds for Self Improving Agents

SIMA 2 and the challenge of generating useful experience

번역 서비스가 일시 중지되어 원문을 표시합니다.

SIMA 2 and the challenge of generating useful experience TLDR: Practice has an inconvenient dependency: somebody must build the warehouse, then another warehouse, before an agent can learn to navigate it. World models make that dependency interesting. A learned simulator can generate the consequences of an agent’s actions, and when it is combined with a system that proposes tasks and evaluates attempts, the training environment becomes part of the improvement machinery. DeepMind’s SIMA 2 and Genie 3 provide an early demonstration of this arrangement. The central question is whether generated experience teaches capabilities that survive outside the worlds that produced them. We cover how a world model supplies consequences without automatically supplying a curriculum or a trustworthy definition of success, how SIMA 2 uses Gemini to set tasks and rate attempts, why a useful curriculum is a branching search rather than a staircase, and how a convincing world can still teach the wrong lesson. We also look at the economic cost of generating and judging experience, and close with the measurements that would show the improvement is recursive: that better practice for the next generation, not simply more compute, is what moved the needle. Imagine teaching an agent to navigate a warehouse. You can collect more demonstrations, write more instructions, or let it practice. Practice has an inconvenient dependency: somebody must build the warehouse. Then another warehouse, with different shelves, lighting, obstacles, and mistakes waiting to happen. An agent can exhaust your training environments long before it exhausts its capacity to learn. World models make that dependency interesting. A learned simulator can generate the consequences of an agent’s actions, creating experience on demand. Combine it with a system that proposes tasks and evaluates attempts, and the training environment becomes part of the improvement machinery. DeepMind’s SIMA 2 and Genie 3 provide an early demonstration of this arrangement. The central question is whether generated experience teaches capabilities that survive outside the worlds that produced them. A world model supplies consequences This post is for paid subscribers

자료 출처

원문 읽기 ↗

관련 주제

← 전체 소식 이어 읽기