Thinking Fast and Slow in AI: The 2021 Paper That Priced Deliberation
In 2021, an IBM-led team asked what Kahneman would build if he designed agents. Thinking Fast and Slow in AI proposes SOFAI: fast solvers that answer from experience in constant time, slow solvers that reason on demand, and a meta-cognitive module refereeing between them. Five years later, every team running agents relitigates this paper daily. Its core claim: deliberation is a purchase, and something should check the price first.
The architecture defaults to fast, like humans do. Every incoming problem hits a System 1 solver first, which reacts from past experience without reasoning about problem size. Slow solvers never self-start. They wait to be invoked, which keeps time-to-action minimal on everything routine.
The referee asks one economic question. The meta-cognitive module weighs the expected extra reward of slow reasoning against its cost in time and memory, using the fast solver’s own confidence plus a model of each solver’s track record. Think twice only when the arithmetic says the second thought earns its keep.
The check itself runs in two speeds, which is the detail I admire most. MC1 is a quick gate: enough resources left, and is fast-solver confidence high relative to expected reward? If yes, accept and move on. Only the doubtful cases reach MC2 for full accounting. Even the referee refuses to deliberate about whether to deliberate.
Slow is not assumed better. The paper states plainly that some tasks belong to fast solvers permanently, perception being the human example, and that experience should migrate problems from slow to fast the way reading in your native language once needed effort and now needs none. Complex arithmetic never migrates. A serious architecture tracks which is which instead of worshipping depth.
For sequences of decisions, the paper offers two clocks. Arbitrate per step for flexibility, letting each move pick its own solver, or arbitrate once for the whole plan when solvers carry deep domain knowledge. Their test instances span both ends: step-by-step grid navigation and full epistemic planning. The referee’s influence shrinks as solver expertise grows, which feels right.
You might think this is algorithm portfolios wearing Kahneman branding. The difference matters. Portfolios pick the best solver for the job; SOFAI assumes the fast one and demands justification for anything slower, gated on confidence and priced against reward. That default-fast posture is exactly how teams ship today: a cheap model drafts, a reasoning model checks only the uncertain parts. The 2021 paper aged into a design pattern.
My note: at publication this was architecture without results. The authors disclose the simplifications plainly, one fast and one slow solver, assumed experience on exactly the same problem, no reported numbers yet. Read it as a blueprint with unusual intellectual honesty about its own scaffolding, not as an empirical verdict.
The takeaway: If you run agent loops, steal the MC1 question verbatim. Set a confidence floor below which the cheap model must escalate, and price every slow call against the gain it is supposed to buy. Deliberation without a budget is just latency with confidence.
Related TMFNK Content
- Contrastive Language Models: Fast Decisions Without Generating a Word The same fast-slow trade inside the model itself, deciding without the expensive part.
- Is Grep All You Need? How Agent Harnesses Reshape Agentic Search What the referee looks like in production search stacks today, choosing tools per step.
- The Magic of Average: Why LLMs Make Simple the New Powerful The case for the fast default: average output is often good enough that deliberation rarely earns its price.
Crepi il lupo! 🐺