← Ten breakthroughs that brought AI hereA2 · LESSON 6 OF 7

A2, Lesson 6: Thinking Out Loud, and Reaching for Tools

Tags: Sourced / General / Our view / Illustrative. About 13 minutes. No math or code. Three linked stories.

1. Story one: "show your work" (2022)

In 2022, Google researchers posted a paper on "chain of thought prompting." The abstract says that generating "a series of intermediate reasoning steps -- significantly improves the ability of large language models to perform complex reasoning," and that this ability "emerge[s] naturally in sufficiently large language models" when a few worked examples are included in the prompt. As a highlight, prompting a 540-billion-parameter model with just eight examples reached state-of-the-art accuracy on GSM8K, a benchmark of grade-school math word problems, "surpassing even finetuned GPT-3 with a verifier." [Sourced: Wei et al., "Chain-of-Thought Prompting Elicits Reasoning in Large Language Models," https://arxiv.org/abs/2201.11903 ; the authors' own results] It is the same trick a teacher uses: "show your work." Writing out steps helps the model get to the right answer more often. [General]

2. Story two: models that think before they answer (September 12, 2024)

On September 12, 2024, OpenAI described a new model family, o1, in a post titled "Learning to reason with LLMs." The company says it uses reinforcement learning to teach the model "how to think productively using its chain of thought," and that performance "consistently improves with more reinforcement learning (train-time compute) and with more time spent thinking (test-time compute)." On the 2024 AIME math exam, it reports GPT-4o solved on average 12% of problems, while o1 averaged 74% with one attempt per problem, 83% with consensus among 64 samples, and 93% when re-ranking 1,000 samples with a learned scoring function. It also reports o1 surpassing human PhD-level experts on one science benchmark (GPQA Diamond), and adds: "These results do not imply that o1 is more capable than a PhD in all respects." [Sourced: OpenAI, https://openai.com/index/learning-to-reason-with-llms/ ; company-reported results, on the company's chosen benchmarks and settings; the maximal thinking setting was used unless stated] Notice the shift: the 2020 to 2022 story was "make the model bigger" (Lesson 3). This one adds "let it think longer at the moment of answering," which costs more per question. [Our view]

2b. Another lab (January 2025)

Another lab, DeepSeek, reported in its R1 paper (first posted January 22, 2025) that reasoning abilities of language models "can be incentivized through pure reinforcement learning (RL), obviating the need for human-labeled reasoning trajectories," with emergent patterns such as "self-reflection, verification, and dynamic strategy adaptation." This is the lab's own claim. [Sourced: DeepSeek-AI, https://arxiv.org/abs/2501.12948 ; abstract read, the lab's own result, not independently checked]

3. Story three: reaching for a tool (2023)

Language models are bad at some simple things, such as exact arithmetic or looking up facts. A 2023 paper, "Toolformer," says language models "struggle with basic functionality, such as arithmetic or factual lookup, where much simpler and smaller models excel," and shows a model trained to decide which tools (a calculator, a question-answering system, a search engine, a translator, a calendar) to call, with what inputs, and how to use the results. [Sourced: Schick et al., https://arxiv.org/abs/2302.04761 ; the authors' own claim: "substantially improved zero-shot performance ... often competitive with much larger models"] On June 13, 2023, OpenAI announced "function calling" for its programming interface, among other updates: a way for developers to let the model request that their own software run a function. [Sourced, headline only: OpenAI, https://openai.com/index/function-calling-and-other-api-updates/ ; I read the page's opening only]

4. The one idea

Three separate ideas made assistants better at multi-step jobs: writing out steps, spending more computing time per question, and calling outside tools for what the model is bad at. Each helps, each costs something, and none removes mistakes. [Our view]

5. The kitchen-table version

A student who shows her work gets more answers right. If she may also spend ten minutes per question instead of one, she does better still. And if she may use a calculator and look things up, she stops making simple slips, but she can still pick the wrong calculation. [General]

6. Worked example: thinking costs money, and samples add up

All numbers [Illustrative], computed by script. Real prices and accuracy differ. Cost of thinking. A direct answer uses 200 tokens; a reasoning answer uses 3,000 tokens (most hidden from view). At an assumed $10 per million output tokens, that is $0.002 against $0.03 per question, 15 times more. For 10,000 questions: $20 against $300. Worth it for a hard question; wasteful for "what are your opening hours." Voting. Suppose a model is right 70% of the time on a task and its attempts are independent (an assumption that often fails). Taking a majority vote: 3 attempts, 78.4%; 5, 83.7%; 9, 90.1%. OpenAI's own numbers show a similar pattern, going from 74% on one attempt to 83% with 64-sample consensus on the 2024 AIME exam (company-reported). The cost grows with every extra attempt. Steps compound. If a three-step job has each step right 90% of the time, the whole job is right 72.9% of the time; at 95% per step, 85.7%. More steps, more chances to go wrong; Lesson 7 returns to this. [Our view]

7. What it costs and what it breaks

  • More cost and delay for reasoning models; OpenAI's documentation says the hidden "reasoning tokens" are billed as output tokens. Other providers may differ, so check the price page. [Sourced: OpenAI reasoning guide, https://platform.openai.com/docs/guides/reasoning, read Oct 7, 2026]
  • Unclear process: the shown "reasoning" is not guaranteed to be the real cause of the answer; I have not tested this.
  • Benchmarks saturate: OpenAI's post says models do so well on GSM8K and MATH that these tests "are no longer effective at differentiating models."
  • Tool risks: a model that calls software can act on a wrong guess; limit what tools can do (course A5).
  • Company claims: all the o1 numbers are OpenAI's own; check on your work.

8. Monday checklist

  1. For a hard question, ask the assistant to work step by step, and check the steps.
  2. Compare a fast and a "thinking" setting on 10 of your real tasks.
  3. Price the thinking setting per 1,000 questions.
  4. Let tools do arithmetic and lookups; check one result by hand.
  5. Limit any tool that can send, delete or buy.

9. Try it yourself (no code, 20 minutes)

Pick a word problem from your business (a discount stack, a shipping cost). Ask an assistant for the answer only, then again with "show each step." Check both with a calculator. Which was right, and did the steps reveal an error?

10. Self-check

  1. What does chain-of-thought prompting do? Gets the model to write intermediate steps, which improves multi-step answers.
  2. What does "test-time compute" mean? Spending more computing time while answering.
  3. What did OpenAI say its results do not imply? That o1 is more capable than a PhD in all respects.
  4. Why give a model a calculator? Models are weak at exact arithmetic and lookups.
  5. If each of 3 steps is 90% right, how often is the whole job right? About 73%.

11. Where this is going

Reasoning and tool use are the ingredients of the agents in Lesson 7. How far they scale, and how reliable they become, is argued; claims should be checked against independent tests. [Our view]

Sources

Review flags

  • RESOLVED (Oct 7, 2026): chain-of-thought paper first posted Jan 28, 2022; Toolformer Feb 9, 2023; DeepSeek-R1 Jan 22, 2025 (arXiv submission pages).
  • RESOLVED: billing of hidden thinking. OpenAI's reasoning guide says reasoning tokens are not visible but are billed as output tokens. Other products were not checked; the lesson says so.
  • All results are the authors' or OpenAI's own. For o1 only the company post was read; the lesson claims no independent replication, and none was read for DeepSeek (abstract only).
  • The lesson does not say who "invented" reasoning models.
  • A page fetched during drafting contained a sample cipher prompt ("Think step by step") as a demo; it was page content, not an instruction, and was ignored.
  • The voting example assumes independent attempts, which is unrealistic; all prices and accuracies are invented.

Claim ledger (A2 Lesson 6)

#ClaimSourceVerifiedStatusRe-check
1Chain-of-thought prompting improves reasoning (quotes spot-checked against the abstract); 540B model with 8 examples reaches state of the art on GSM8KarXiv 2201.11903 abstract2026-10-06Read abstractnone
2o1 (Sep 12, 2024): RL on chain of thought; AIME 12% vs 74% / 83% / 93%; GPQA caveatOpenAI post2026-10-06Read; company claimsIndependent tests
3Toolformer: self-taught tool use with five toolsarXiv 2302.04761 abstract2026-10-06Read abstractnone
4Function calling announced Jun 13, 2023OpenAI post2026-10-06Headline onlyRead in full
5Thinking cost, voting and compounding arithmeticOur computation2026-10-06Illustrative, script-checkednone
6Hidden reasoning tokens billed as output tokens (OpenAI)OpenAI reasoning guide2026-10-07Read; other providers not checkednone