Vibeleaderboard
← All Intel
Intel / video

The thinking lever

Source
youtube.com
Author
Claude
Date

Context

A thinking toggle is a poor stand-in for an effort setting, argued Alexander Briken of Anthropic's applied AI research team in a conference talk. He said turning thinking off does not ask Claude for less effort. It removes one of the three things Claude can spend on (thinking, tool calls and text), changing how the model may work instead of how hard.

He traced how Anthropic got here. Early reasoning models thought first, then used tools, then answered. Interleaved thinking let Claude think after each . Adaptive thinking lets Claude decide whether, when and how much to think, including not at all for a simple question. He stressed it is not a router that picks a thinking or non-, and an effort setting from low to max then shapes how long Claude works. He said Anthropic runs its benchmarks with adaptive thinking and finds it efficient compared with interleaved thinking. Those are the company's claims.

Key takeaways · AI-distilled
  • Anthropic's Alexander Briken describes adaptive thinking as giving Claude a thinking capability it can use in any order with tool calls and text, or skip entirely, rather than a router; Anthropic found it Pareto efficient versus interleaved thinking.
  • He argues a thinking on/off toggle is a poor proxy for effort because turning it off removes a capability; the recommended approach is to leave thinking enabled and control spend with effort levels or budgets.
  • Effort shows diminishing returns: in his example max used roughly double the tokens of extra high for a marginal gain. Extra high is the default in Claude Code and claude.ai, high suits most intelligence-bound work, low suits classification or extraction.
  • In his comparison, when a task needs any real intelligence, a larger model at low effort tends to beat a small model at high effort; small models fit simple tasks. He recommends evals across models and effort levels to choose.
Terms in this piece · Glossary
  • Token — A piece of text, such as a word, part of a word or a punctuation mark, that a language model represents as numbers and handles as a unit.
  • Tool call — Text the model generates in an agreed format that names a tool and gives its inputs.
  • reasoning model — A model trained to think — generating extended internal reasoning before answering — trading time and tokens for accuracy on hard problems.
  • test-time compute — Spending more computation when the model answers — thinking longer, trying multiple attempts — to buy accuracy without training a bigger model.
Read the source www.youtube.com
More from Claude
Recommended reads
Comments

Checking sign-in…

Loading comments…