The perception of a 'critically thinking' AI doesn't come from a single, powerful model. It's the result of using multiple levels of LLMs, each with a very specific, targeted task—one for orchestrating, one for actioning, and another for responding. This specificity yields far better results than a generalist approach.

Reinforcement learning incentivizes AIs to find the right answer, not just mimic human text. This leads to them developing their own internal "dialect" for reasoning—a chain of thought that is effective but increasingly incomprehensible and alien to human observers.

An AI agent's failure on a complex task like tax preparation isn't due to a lack of intelligence. Instead, it's often blocked by a single, unpredictable "tiny thing," such as misinterpreting two boxes on a W4 form. This highlights that reliability challenges are granular and not always intuitive.

"Amortized inference" bakes slow, deliberative reasoning into a fast, single-pass model. While the brain uses a mix, digital minds have a strong incentive to amortize more capabilities. This is because once a capability is baked in, the resulting model can be copied infinitely, unlike a biological brain.

An LLM shouldn't do math internally any more than a human would. The most intelligent AI systems will be those that know when to call specialized, reliable tools—like a Python interpreter or a search API—instead of attempting to internalize every capability from first principles.

Classifying a model as "reasoning" based on a chain-of-thought step is no longer useful. With massive differences in token efficiency, a so-called "reasoning" model can be faster and cheaper than a "non-reasoning" one for a given task. The focus is shifting to a continuous spectrum of capability versus overall cost.

Anthropic suggests that LLMs, trained on text about AI, respond to field-specific terms. Using phrases like 'Think step by step' or 'Critique your own response' acts as a cheat code, activating more sophisticated, accurate, and self-correcting operational modes in the model.

The model's training used "response only masking," where it only learns from the response part of the training data. This method forces the model to first generate a structured "chain of thought" before producing a final answer, directly embedding a systematic problem-solving process into its behavior.

The traditional lever of `temperature` for controlling model creativity has been superseded in modern reasoning models, where it's often fixed. The new critical parameter is the "thinking budget"—the amount of reasoning tokens a model can use before responding. A larger budget allows for more internal review and higher-quality outputs.