{
  "data": {
    "author": "Vlad Cealicu",
    "faq": null,
    "kicker": "Blog · Opinion · 10 November 2026",
    "lede": "Robert Sapolsky opens his Stanford course on human behaviour with a multiple-choice question about a man who punches somebody at work. I think most of what he says in that hour about people is true of a model, and that with a model we can run the experiments he can't.",
    "readingMinutes": 9,
    "words": 2096
  },
  "kind": "anchor.page",
  "links": {
    "api": "https://www.anchorterminal.com/api/v1/index.json",
    "html": "https://www.anchorterminal.com/blog/sapolsky-first-lecture-decision-models",
    "json": "https://www.anchorterminal.com/blog/sapolsky-first-lecture-decision-models.json",
    "llms": "https://www.anchorterminal.com/llms.txt",
    "markdown": "https://www.anchorterminal.com/blog/sapolsky-first-lecture-decision-models.md",
    "mcp": "https://www.anchorterminal.com/mcp",
    "slim": "https://www.anchorterminal.com/blog/sapolsky-first-lecture-decision-models.min.md"
  },
  "markdown": "By Vlad Cealicu, 10 November 2026. 2,096 words, about 9 min to read.\n\n- 26: decision models we list\n- 17: of them with at least one model built on a Qwen base\n- 60%: how often two models that were both wrong had picked the same wrong answer, against 33% by chance\n\nDecision models in the directory on 10 November 2026, and how alike models are when they're wrong. Sources [3] and [7].\n\nRobert Sapolsky starts his Stanford course on human behavioural biology with a multiple-choice question. A man of 40, married for 15 years, punches somebody at work, has an affair with a 16-year-old who works on a supermarket checkout, embezzles money and disappears, and the hall gets three explanations to pick from. He's a creep, he's having a very immature midlife crisis, or he has a mutation in one gene. Then Sapolsky tells them there is a neurological disease that looks like this, and that it comes from one gene [[2]](#ref-2). Stanford put a 2024 recording of the lecture on YouTube on 12 March [[1]](#ref-1), and the details here are from the 2010 recording, which has a published transcript [[2]](#ref-2).\n\nI read that as a decision request, because it has a state, a question and a fixed list of answers, which is what an agent sends a [decision model](/categories/decision-models), a model that gives back a probability for each answer you allowed and no text [[3]](#ref-3). There were fourteen of them when I [first wrote about them](/blog/decision-models-everyone-wants-your-decisions) on 9 October and we list 26 today [[3]](#ref-3). The rest of the hour is Sapolsky telling a hall of students how a question like that goes wrong and how to think about the answer, and I think this is exciting, because most of what he says about a person is true of a model, and with a model we can run the experiments he can't.\n\nThe third answer is only on Sapolsky's list because somebody found the gene. Take it off and send the other two to a decision model, and it spreads its probability over a creep and a midlife crisis, because those are the answers it was given, and the right one gets nothing because it isn't there. When I described these models in October I wrote \"no fourth option the model made up\" and I meant it as a good thing. I still think it is for most of what an agent decides (which queue, which tool, is this done), and the lecture opens with the case where the right answer is the one nobody wrote down.\n\nThe fix is one more answer on the list, none of these, and then watching how much probability lands on it over a week, because that number tells you how much of what's coming in your list doesn't cover. [kettei](/tools/sqwish-kettei) has a version of it as a setting, an abstain policy below a probability you choose [[3]](#ref-3).\n\n## What we do to the number\n\nSapolsky's first big point is that thinking in categories does three things to you. Two things in the same category look more alike than they are, two things on either side of a line look more different than they are, and while you're watching the lines you miss the whole picture. His example of the second is two exam scores, 65 and 66, with the pass mark between them [[2]](#ref-2).\n\nI think a decision model is better at this than we are, because it gives you 0.51 and we're the ones who write the next line.\n\n```go\n// the usual line\nif p \u003e 0.5 {\n\tapprove()\n}\n\n// keeping the number a bit longer\nswitch {\ncase p \u003e= 0.9:\n\tapprove()\ncase p \u003c= 0.1:\n\trefuse()\ndefault:\n\taskAPerson(p)\n}\n```\n\nWith the first one, 0.51 and 0.99 go down the same branch and 0.49 and 0.51 go down different ones, which is his first two problems in one line of Go. The second is still categories (three of them), but the middle one says I don't know, which the first version throws away.\n\nThe number has to mean something for that to work, and that's calibration, whether the things a model gives 0.9 come true about nine times in ten. Chuan Guo and colleagues found in 2017 that the neural networks of the day were poorly calibrated, and that a fix with one parameter (temperature scaling) worked on most of the datasets they tried [[4]](#ref-4). [Laya](/tools/convai-laya)'s model card says its checkpoints ship over-confident and tells you to fit a temperature on your own data [[3]](#ref-3), which is the same fix, and calibration is in the test we've written for this category, along with how often an answer given 0.9 or more is wrong [[3]](#ref-3).\n\nHis third problem is the whole picture, and for an agent the whole picture is the chain, because an agent makes one decision after another and what you care about is whether the whole run comes out right. If each is right 95% of the time and the mistakes are independent, which is the simple case, ten in a row are all right 60% of the time and fifty in a row 8%.\n\nChance that every decision in a chain is right when each is right 95% of the time, in per cent\n\n| Decisions in a row | All right |\n| --- | --- |\n| 1 | 95 |\n| 5 | 77 |\n| 10 | 60 |\n| 20 | 36 |\n| 50 | 8 |\n\nEach series has a zero-based scale.\n\nSapolsky gives the students James Gleick's Chaos to read, for the argument that taking a thing apart and putting it back together is how you fix a clock, and a behaviour is closer to a cloud [[2]](#ref-2). A fixed set of labelled decisions is the clock test, and it's the one we've written [[3]](#ref-3). An agent chaining decisions for a week is the cloud, and in our benchmark that's Task success, which is pending [[5]](#ref-5).\n\n## Why did it say yes?\n\nHis second idea is the buckets. If you want to know why the man threw the punch you can ask what his neurons did a second before, what he saw or heard just before that, what his hormones had been doing for hours, what happened to him before he was born, which genes he carries and what evolution was selecting for millions of years ago, and each of those belongs to a different science that tends to think its bucket is the explanation. The course argues that none of them is, and that each one is only a convenient place to describe everything that came before it [[2]](#ref-2).\n\nWhy a decision model said yes, going back in time\n\n1. **The second before** (you can test it). The question, its wording and the order of the answers\n2. **Just before that** (you can test it). The state the agent passed in, and what it left out\n3. **The hours before** (you can test it). How busy the server was, which can change the answer to the same request\n4. **Before it was released** (on the card, if they say). The fine-tune, which decisions it was trained on and who labelled them\n5. **Before that** (on the card, if they say). The base model, and 17 of the 26 we list have one built on Qwen\n6. **Long before** (hard to see from outside). What the base model's makers were training it to be good at\n\nI think this is where it gets better than biology, because a biologist can't run the man again with one hormone changed, and we can send the same request 500 times, shuffle the answers, change one line of the state, and for the open ones put the same training on a different base and compare. [Decider](/tools/decider) and [Strands Decider](/tools/strands-decider) each ship the same family on Qwen bases and on Gemma bases [[3]](#ref-3), which is close to the experiment nobody can do on people, the same upbringing on different genes. (The server's load is the odd bucket, and I went through it in [Unreliable compared to what?](/blog/unreliable-compared-to-what) on 4 November. On 3 November I compared [editing a model's weights](/blog/editing-the-weights) with gene editing, so the genes can be changed as well.)\n\nThe first bucket is already measured. Pouya Pezeshkpour and Estevam Hruschka reordered the answers on multiple-choice benchmarks in 2023 and found a performance gap of about 13% to 75%, depending on the benchmark [[6]](#ref-6). [Kev](/tools/jaredpalmer-kev)'s release notes list option order changing its answers among the things it does badly, and it has an endpoint that permutes the answers so you can check your own questions [[3]](#ref-3).\n\nThe base model is the bucket I'd look at first. 17 of our 26 listings have at least one model built on a Qwen base (one of them by way of a Xiaomi distil), going by their cards and files [[3]](#ref-3). [Drex Max](/tools/drex-max-v1)'s card named Qwen3.8-27B as its base until 30 October and now names none, while the files in the repository still do, and Kev's notes say its answers to knowledge questions are set by the base model [[3]](#ref-3). So if you ask three decision models and take the majority to be safe, look at what they're built on first, because three fine-tunes of one base are relatives. Elliot Kim and colleagues looked at more than 350 models in 2025, and on one leaderboard (71 models answering the same multiple-choice questions) two models that were both wrong had picked the same wrong answer about 60% of the time, against a third if they were guessing, and models from the same developer or on the same architecture agreed more [[7]](#ref-7).\n\n## Two quotes he reads out\n\nSapolsky reads out lines from three of the most influential scientists of the last century, as he puts it, each of them explaining behaviour from inside one bucket, and two of the three have a version in this market.\n\nThe first is John Watson, who founded behaviourism and said that given a child at birth and control of how it was raised he could make it into anything, a doctor, a lawyer, a beggar or a thief. Sapolsky's answer is that he couldn't, because what the child starts with counts (his example is protein malnutrition before birth) [[2]](#ref-2). [Part two](/blog/decision-models-train-your-own) of my decision model posts said to fine-tune a small open model on your own decisions, and read unkindly that's Watson's offer with a GPU. I still think it's the right way to keep your decisions in the building, and Sapolsky's answer to Watson applies to it. The base comes with the fine-tune, so pick it as carefully as the training data.\n\nThe second is Egas Moniz, who invented the lobotomy and, in the line Sapolsky reads, had seen \"cures and improvements but no failures\" [[2]](#ref-2). He got half of the 1949 Nobel Prize in Physiology or Medicine for it [[8]](#ref-8), and Sapolsky says the operation was done to tens and hundreds of thousands of people [[2]](#ref-2). A refund sent to the wrong queue isn't a lobotomy, and I'd still want to see the failures. Seven of our 26 listings note a model card or docs with no limitations section [[3]](#ref-3), and we're in that group for now, because Performance and Task success are both pending in our benchmark and the test for this category hasn't run, so the grades on those pages come from what's published and none of it from us watching a model decide [[5]](#ref-5) [[3]](#ref-3).\n\n- [Kev](https://www.anchorterminal.com/tools/jaredpalmer-kev.md), B. release notes that list what it does badly, option order among them\n- [Strands Decider 2B](https://www.anchorterminal.com/tools/strands-decider.md), C. measured figures for each limitation\n- [Jev](https://www.anchorterminal.com/tools/typesafe-jev.md), C. a known-limitations page\n\nThree that publish where they fail\n\n## Does a model decide?\n\nSapolsky also asks the hall for a show of hands on a list of beliefs, and free will is on it [[2]](#ref-2). In 2023 he published Determined, a book arguing that we don't have any, and that genes, experience and environment settle what we do (philosophers were mostly critical of it) [[9]](#ref-9). I don't know if he's right about people. For a model it's the description of how it works, the weights, the request and a random number, and nobody loses sleep over it. When people say a model doesn't decide anything and only computes, I think the fair answer is that on Sapolsky's account neither did the man who threw the punch, and the difference I care about is that the model's buckets can be opened and the man's mostly can't.\n\nIf the man's case came in as a ticket today, with the two answers most of us would have written, I'd want a none-of-these answer so the model has somewhere to put the disease nobody listed, the number kept as a number until the last step, and a second opinion from a model that isn't a relative of the first.\n\nWe haven't run any of this ourselves. The test for the category is written and hasn't run [[3]](#ref-3), so the 17 and the bases come from the cards and the rest is other people's papers, and when it runs I want the none-of-these answer and the same-base pairs in it. The experiment I'd most like to see is already sitting there, the same family on Qwen and on Gemma from two different vendors, asked the same questions, and if somebody runs it before we do I'd like to read it.\n\n---\n\nMore in [Opinion](https://www.anchorterminal.com/blog/opinion/index.md) · Older opinion [Just in time society](https://www.anchorterminal.com/blog/just-in-time-society.md)\n\n- Older: [A machine that builds the benchmark](https://www.anchorterminal.com/blog/a-machine-that-builds-the-benchmark.md)\n- Newer: [Google made four TPUs this year, and the rest of us get an interest form](https://www.anchorterminal.com/blog/google-tpu-8-four-chips-and-an-interest-form.md)\n- All posts: https://www.anchorterminal.com/blog/index.md\n\nAnchor Terminal grades the tools AI agents run on, with a source for every score. [The top tools](https://www.anchorterminal.com/top/index.md) · [Categories](https://www.anchorterminal.com/categories/index.md) · [Compare two tools](https://www.anchorterminal.com/compare/index.md) · [Prices](https://www.anchorterminal.com/prices/index.md) · [Sunsets](https://www.anchorterminal.com/sunsets/index.md) · [API, MCP server and CLI](https://www.anchorterminal.com/agents/index.md)\n\n## References\n\nRead on 10 November 2026. The lecture details are from the transcript of the 2010 recording, and the 2024 recording may differ from it in small ways. The quote from the lecture is as transcribed.\n\n[1] Stanford School of Humanities and Sciences, \"1. Human Behavioral Biology, Introduction by Robert Sapolsky\", a spring 2024 lecture posted on 12 March 2026. The lecture this post is about. https://www.youtube.com/watch?v=mZI2GQlNdJA\n\n[2] The Singju Post, full transcript of Robert Sapolsky's introduction to Human Behavioral Biology, recorded at Stanford on 29 March 2010. Source for the opening case and its three explanations, the three problems with thinking in categories and the exam scores of 65 and 66, the buckets, the lines from John Watson and Egas Moniz and what Sapolsky says of each, the show of hands on free will, and James Gleick's Chaos with the clock and the cloud. https://singjupost.com/stanford-professor-robert-sapolsky-on-introduction-to-human-behavioral-biology-full-transcript/\n\n[3] Anchor Terminal, decision model listings and the category page, 10 November 2026. Source for what a decision model is, the 26 listings, the test for the category and that it hasn't run, the 17 with a model on a Qwen base (Clef, Decider, Decision 3.0, Drex Max, Drex 1.5 by way of Xiaomi's MiMo distil, elephant-1, GLiClass, JEV-27B-VL, joey-1, Kev, mammoth-1, Microsoft-Decision-1, pplx-decider, pplx-decider-flash, SPEV, Strands Decider and Vela), Decider's and Strands Decider's Qwen and Gemma bases, Drex Max's card naming its base until 30 October, Kev's release notes and its permute endpoint, kettei's abstain policy, Laya's note on over-confident checkpoints, the seven listings that note no limitations section (Drex Max, elephant-1, GLiClass, joey-1, Liquid d1, mammoth-1 and ScaleDown), and the limitations published by Kev, Strands Decider and Jev. Each listing links the pages it was graded from. https://www.anchorterminal.com/categories/decision-models\n\n[4] Chuan Guo, Geoff Pleiss, Yu Sun and Kilian Q. Weinberger, 2017, \"On Calibration of Modern Neural Networks\", Proceedings of the 34th International Conference on Machine Learning. Source for modern neural networks being poorly calibrated, and for temperature scaling, a single-parameter fix that was effective on most of the datasets tested. https://proceedings.mlr.press/v70/guo17a.html\n\n[5] Anchor Terminal, the Agent Tool Benchmark, methodology 0.7. Source for Performance and Task success being pending in this run, with no score and nothing added to the total. https://www.anchorterminal.com/benchmark/\n\n[6] Pouya Pezeshkpour and Estevam Hruschka, August 2023, \"Large Language Models Sensitivity to The Order of Options in Multiple-Choice Questions\", later in Findings of NAACL 2024. Source for a performance gap of approximately 13% to 75% across benchmarks when the answer options are reordered. https://arxiv.org/abs/2308.11483\n\n[7] Elliot Myunghoon Kim, Avi Garg, Kenny Peng and Nikhil Garg, 2025, \"Correlated Errors in Large Language Models\", Proceedings of the 42nd International Conference on Machine Learning. Source for more than 350 models evaluated, 71 of them on Helm's multiple-choice questions, where two models that are both wrong agree about 60% of the time against a third for random guessing among the wrong answers, and for models from the same developer or on the same architecture agreeing more. https://arxiv.org/html/2506.07962v1\n\n[8] The Nobel Foundation, Egas Moniz, facts. Source for his half share of the Nobel Prize in Physiology or Medicine 1949, awarded for the discovery of the therapeutic value of leucotomy in certain psychoses. https://www.nobelprize.org/prizes/medicine/1949/moniz/facts/\n\n[9] Wikipedia, \"Determined, A Science of Life Without Free Will\". Source for the book's author, its publication by Penguin Press in 2023, its argument that genes, experience and environment settle our choices, and philosophers' mostly critical reviews. https://en.wikipedia.org/wiki/Determined:_A_Science_of_Life_Without_Free_Will\n",
  "meta": {
    "attribution": "Anchor Terminal (https://www.anchorterminal.com)",
    "attributionRequired": true,
    "credit": "CC BY 4.0, credit Anchor Terminal and link the page or API address you took it from (https://www.anchorterminal.com/terms/#data)",
    "docs": "https://www.anchorterminal.com/docs/",
    "generatedAt": "2026-11-10",
    "license": "CC-BY-4.0",
    "method": "https://www.anchorterminal.com/benchmark/",
    "methodology": "0.7",
    "openapi": "https://www.anchorterminal.com/openapi.json",
    "preview": false,
    "run": "2026-10-01",
    "runLabel": "October 2026 research run",
    "terms": "https://www.anchorterminal.com/terms/#data"
  },
  "page": {
    "breadcrumbs": [
      {
        "name": "Home",
        "url": "https://www.anchorterminal.com/"
      },
      {
        "name": "Blog",
        "url": "https://www.anchorterminal.com/blog/"
      },
      {
        "name": "Opinion",
        "url": "https://www.anchorterminal.com/blog/opinion/"
      },
      {
        "name": "Sapolsky's first lecture",
        "url": ""
      }
    ],
    "description": "Vlad Cealicu reads the first lecture of Robert Sapolsky's Human Behavioral Biology course as a guide to decision models. The answer nobody wrote, what a threshold does to a probability, the buckets behind a model's answer, 17 of the 26 decision models we list sharing Qwen ancestry, and why a model is the case where the experiment can be run.",
    "facts": [
      "Vlad Cealicu",
      "10 November 2026"
    ],
    "h1": "What Sapolsky's first lecture means for LLMs and decision models",
    "image": "https://www.anchorterminal.com/assets/og/blog-sapolsky-decision-models.png",
    "path": "/blog/sapolsky-first-lecture-decision-models",
    "published": "2026-11-10",
    "section": "blog",
    "title": "Sapolsky's first lecture, LLMs and decision models | Anchor Terminal",
    "toc": null,
    "updated": "2026-11-10",
    "url": "https://www.anchorterminal.com/blog/sapolsky-first-lecture-decision-models"
  },
  "tokens": {
    "markdown": 4600,
    "slim": 2080
  },
  "version": 1
}