How to make System 1 LLMs show System 2 competencies?

Based on 6 sampled comments from this thread. Analyzed 2026-10-17.

Commenters spent far less time on techniques for eliciting deliberate reasoning than on whether the System 1/System 2 label fits language models at all. The strongest objection reverses the article’s premise: autoregressive models produce tokens one after another, can articulate intermediate reasoning, and therefore resemble a mechanized version of slow, verbal deliberation more than fast intuition. From that view, the major capability gap is not teaching a reflexive model to reason, but giving text-only systems the embodied, emotional, and sensorimotor machinery associated with instinctive judgment and perhaps consciousness. Others argued that mapping a human psychological metaphor onto transformer internals is simply category error. Sequential output may resemble one aspect of deliberation, but it does not establish that model computation has the same organization as human cognition; proposed parallels between holistic processing and diffusion models were similarly tentative. Several readers also attacked the intellectual foundation of the terminology, saying the popularized dual-process account rests on contested psychology and has become AI-industry jargon rather than a rigorous technical framework. The only real convergence was that architecture and training should be discussed concretely, not smuggled into an anthropomorphic two-system story. No commenter supplied evidence that a recurrent internal-language approach would reliably create the desired capabilities, and one explicitly warned against it.

The central fight is whether LLMs are fast System 1 engines that need added System 2 deliberation, or already resemble System 2-style sequential reasoning while lacking embodied intuition. Defenders of the reversal point to autoregressive token-by-token generation and verbal reasoning; critics say the human dual-process metaphor is too scientifically shaky and mechanically unlike transformers to support either classification.

• Concern: The System 1/System 2 framing was widely treated as a flimsy psychological metaphor, so using it to prescribe LLM architecture risks replacing mechanism with fashionable jargon.
• Concern: Autoregressive next-token generation looks more like serial verbal deliberation than instinct, undermining the article’s premise that present-day LLMs are fundamentally System 1 systems.
• Concern: Text-only models may be missing embodied, emotional, and sensorimotor intelligence rather than an extra reasoning loop, leaving a deeper capability gap than chain-of-thought-style deliberation.
• Concern: The suggested recurrent internal-language route drew no support and was explicitly discouraged, with readers offering no reason to expect it to yield robust reasoning gains.