Anthropic reasoning
Claude's chain-of-thought tokens are worked computations, not inner thoughts.
- Filed under
- Tech
- Type
- Background entry
- Sources
- 10
Image: anthropic.com
Summary
The strongest supported account treats reasoning as a computational capability rather than a human-like inner monologue. Claude’s thinking mode gives the model more opportunity to restate problems, test approaches, and check intermediate results, which can help with mathematics, coding, analysis, and long-running tasks. Anthropic’s own studies also show that displayed chain-of-thought is only variably faithful, while the anthropic principle is best understood as an observation-selection framework rather than a single cosmological proof.
- CapabilityAdditional inference-time computation can improve performance on complex tasks, but its benefits depend on the task and configuration.
- InterpretabilityReadable reasoning can provide useful evidence for evaluation without serving as a guaranteed window into the model’s actual causal process.
- CosmologyObserver-selection effects constrain what kinds of universes or histories can be observed, but do not by themselves settle every fine-tuning explanation.
- Claude’s thinking mode allows the same model to spend more time and effort on difficult problems rather than switching to a separate model.
- The API describes thinking as intermediate work that can include restating the task, trying approaches, checking results, and abandoning failed paths.
- Anthropic’s experiments found substantial variation in how strongly models rely on their stated chain of thought, with less faithful reasoning in many tasks as models become larger and more capable.
Deeper Insight
The sharper dissent focuses on category errors: intermediate tokens may be useful learned computation without being “thoughts,” and longer reasoning traces may sometimes overthink, mislead, or fail systematically. In anthropic philosophy, critics examine whether self-location principles remain coherent in infinite worlds or interact badly with decision theory, including Dutch-book vulnerabilities. These critiques do not erase practical value, but they challenge strong claims about what reasoning traces or observer-selection arguments prove.
- Model languageCalling generated intermediate tokens reasoning may anthropomorphize an optimization process and encourage unjustified confidence in explanations.
- ReliabilityMore inference-time effort is not uniformly beneficial; some controlled puzzle results report overthinking and complexity-dependent failure.
- Anthropic theorySelf-sampling, self-indication, and related principles can produce divergent or internally costly conclusions in unusual world models.
- A prominent critique argues that intermediate tokens should not automatically be treated as reasoning traces, with “reasoning” potentially functioning as shorthand for learned prompt augmentation.
- A review of controlled puzzle experiments reports that reasoning-enhanced models can underperform on low-complexity tasks and continue exploring after finding a correct path.
- Anthropic reasoning in infinite worlds can create problems for the self-indication assumption, motivating alternative formulations such as distribution-sensitive approaches.
Sources · 10
Featured · Analysis · mkodama.org
Anthropic and Decision-Theoretical Dutch Books
Anthropic and Decision-Theoretical Dutch Books ## Introduction Here's a fun fact—some combinations of anthropic assumptions and decision theories expose you to Dutch books. [1] If you are a causal decision theorist who makes the self sampling assumption or an evidential decision theorist who makes t
Read sourceAnalysis · anthropic.com
Reasoning models don't always say what they think \ Anthropic
Reasoning models don't always say what they think \ Anthropic # Reasoning models don't always say what they think - Published Apr 3, 2025 - Team Alignment - Topics - Reasoning - Alignment - Reward hacking - Safety evaluations - Claude Sonnet - Training See more... Since late last year, “reasoning mo
Read sourceAnalysis · news.ycombinator.com
Stop Anthropomorphizing Intermediate Tokens as Reasoning/Thinking Traces | Hacker News
Stop Anthropomorphizing Intermediate Tokens as Reasoning/Thinking Traces (2025) | Hacker News Stop Anthropomorphizing Intermediate Tokens as Reasoning/Thinking Traces (2025) | Hacker News||[](https://news.ycombinator.com)|**[Hacker News](news)**[new](newest) | [past](front) | [comments](
Read source