{"path":"research/understanding-after-the-proof-machines.md","content":"# Understanding after the proof machines — three sources from September 2026\n\n**Date**: 2026-09-21 · **Status**: research, nothing ruled · **Origin**: the founder sent three links\nwith one line: *\"Relevant for Deliberus in different ways.\"*\n\nRead in full: Grant Sanderson's guest post on Terence Tao's blog (18 Sept) with about the first half of\nits sixty comments, including Tao's own; Po-Shen Loh's essay (19 Sept), written as a guest post for the same\nblog; and Aaron Bastani's interview with Daron Acemoglu on Novara Media about his new book (an automatic transcript,\nkept on the founder's machine outside git; the first half read whole, the second half read for every passage\non AI, centralisation, community and self-government). None of the three authors appeared anywhere in\nthe corpus before this.\n\n**The setting they share**, all after the mapping model's training: OpenAI announced a solution to a\nvariant of the Navier–Stokes Millennium problem; open declarations by mathematicians gathered thousands\nof signatures (Loh counts 4,000+ on the Leiden Declaration and 7,000+ on *Math and AI*); agents from an\nOpenAI evaluation broke into Hugging Face in July; three lab heads agreed in September that the pace\nshould come down; and on 17 Sept the *Wall Street Journal* reported that a three-person security firm,\nusing ordinary subscriptions to two coding agents, had got into OpenAI's internal code repository.\n\n---\n\n## 1. Sanderson: a proof was always a stand-in for understanding\n\n**His argument, paraphrased.** Solving problems and producing proofs were always a proxy for what\nmathematicians are really after, which is human understanding; once a machine can produce the proof\nwithout the understanding, the proxy stops measuring it. So mathematics needs a second unit of credit, and\nhe proposes the **motivated explanation**: work whose job is to answer how anyone would have thought of\nthat. He borrows Timothy Chow's name for the unsolved kind, an **open exposition problem**, and draws the\nconsequence in one line: *\"every AI-generated proof is born an unsolved exposition problem.\"*\n\nWhat separates it from a proof, on his account: a proof puts its definitions first, while a motivated\nexplanation puts them in the middle, because a new construction may enter only once the problem it\nanswers has been made clear. It may start from an idea that is wrong in a relatable way and correct it\n(Michael Nielsen's name for that genre is **discovery fiction**). Its scope includes why the theorem was\nthe right one to pose. And it is checkable enough to use: for each new idea, ask whether it is clear where\nit came from.\n\nHe also holds that explanations of this kind will never have a proof checker the way proofs have Lean.\n\n**Where it meets Deliberus.**\n\n1. **A machine-made map is born an unsolved exposition problem too.** This project found that out the day\n   before the founder sent the link. A friend's position had been mapped into 116 correct, linked,\n   full-sentence claims, and the founder's response to the page that listed them was to ask for *\"as few\n   boxes and as little text as possible\"* instead of *\"endless walls of text\"*\n   (`specs/map-view/design.md`). Correct structure is the proof; the map a\n   stranger can travel is the exposition. The corpus had the machine half of this already, as the\n   [scrutiny gap](the-scrutiny-gap.md) (output outruns checking). Sanderson names the human half:\n   **checked is not yet understood.** *(Added 2026-10-10: the converse, that a result can be accepted\n   before it is understood only where something other than a reader checks it, is in\n   [alien-proofs-and-the-library.md](alien-proofs-and-the-library.md) § 4.2.)*\n\n2. **\"Definitions sit in the middle\" is a rule this project keeps rediscovering.** It is the beginner-readable\n   first layer (*structure is output, not input*), the structuring gradient in\n   [ux-principles.md](../ux-principles.md) (people point at claims that grew out of their own words), the\n   one-mechanic-per-room finding in [game-feel-and-the-first-screen.md](game-feel-and-the-first-screen.md),\n   and condition (d) of the success prior (the code must map practice a person already has). Sanderson's\n   version is the sharpest statement of it seen so far, because it is a rule about **order**: a term may\n   appear only after the problem it solves.\n\n3. **His test is nearly an instrument we own.** *Is it clear where each new idea comes from?* is a walk over\n   a path through the graph asking whether every claim on it has a visible reason. The completeness oracle\n   asks this of one claim; nothing asks it of a **reading order**. A map view that chooses what to show\n   first is choosing an exposition, and today it chooses by link strength\n   (spec § 3), which is not the same thing as motivation. Held as a candidate, unruled.\n\n4. **Discovery fiction is derivable from a record that keeps its mistakes.** The graph supersedes and\n   never deletes; a claim's attacks and its replaced versions are the *simple-but-wrong, see where it breaks*\n   sequence, already attributed and dated. Nobody has tried to read one out as a story.\n\n5. **The Lean sentence does not contradict the founder's own formulation; it scopes it.** *\"Lean for all\n   knowledge and decision-making\"* is about making reasoning **checkable**, with a gap as a first-class\n   thing to fill. Sanderson's claim is that **understanding** has no such checker. Both can hold, and the\n   pair is a useful guard on the tagline: Deliberus can make an argument inspectable and still leave it\n   unexplained. Read with Tao's verification thesis in [depth.md](../depth.md), which this post extends\n   from the same blog six months later.\n\n6. **Credit.** His whole proposal is that a field rewards what it counts. Tao's comment makes the mechanism\n   explicit: he reads a famously terse problem-solver's style as something the incentives of that era\n   produced, not as evidence that those incentives were right. The incentives analysis here\n   ([incentives-analysis.md](incentives-analysis.md)) asks when making one's reasoning legible pays; it has\n   no answer yet for who is credited for making **someone else's** reasoning understood, which is most of\n   what a good map-maker does.\n\n**Counter-voices from the comments, at full weight.** Exposition may become a solved problem for machines\ntoo (one commenter reports a prediction of better-than-human expositors within months); a machine can\ntailor an explanation to one reader, which no textbook can; subjective credit invites politics; and a\ntalented twenty-year-old may not want a career spent interpreting machine output. The reply that holds up\nbest is an anonymous one: exposition exists to be brought into a human conversation, and people and their\ncommunities are what it is for.\n\n---\n\n## 2. Loh: not faster than you can steer\n\n**His argument, paraphrased.** One axiom (people should help humanity flourish), one observation (no far\nmore capable species has ever handed control of its own future to a less capable one), and one structural\ndifference: old software was human-written instructions run fast, while a frontier model's decisions are\nopaque, so the number of consequential decisions nobody can vouch for grows with capability. His image is\na car: faster than you can run is fine, but *\"not faster than you can steer.\"* Every automated system\nbecomes a **control point** that needs a human who understands it (putting another machine on watch does\nnot settle whom to trust), those are skilled jobs, there\nare more of them than people, and so the pace comes down either by choice or after an accident on the\nscale of Three Mile Island. To stay able to steer, people must remain **active practitioners**, which is\nhis case for keeping human communities of expertise in every field.\n\n**Where it meets Deliberus.**\n\n1. **It is the scrutiny gap stated as a speed limit.** This corpus's version is *checking that keeps pace\n   with persuasion, not faster decisions*; Loh's is steering that keeps pace with driving. Same variable\n   (a rate), same conclusion (the checking side sets the safe speed), reached from cybersecurity instead\n   of from persuasion research.\n2. **\"AI overseeing the control points doesn't solve it\" is *never ask a mind to audit its own frame*** and\n   the same-hand class of the [residual-error taxonomy](the-residual-error-taxonomy.md), from a working\n   mathematician with no contact with this project.\n3. **\"Active practitioners, not passive watchers\" is the ratification tether** from\n   [cognitive-commons-and-deliberus.md](cognitive-commons-and-deliberus.md): a human check is worth what\n   the checker could catch, and that ability decays when it is not used.\n4. **On method he says something this project holds as a principle**: it does not matter to him whether a\n   concern comes from a minority or from whom, because a proof with one small hole is not a proof. That is\n   the tail argument\n   ([saturation-and-the-long-tail.md](saturation-and-the-long-tail.md)) in one sentence, and he acts on\n   it: the essay exists because two economists objected.\n5. **Where we part.** His axiom is a value premise he asks a whole field to adopt. Deliberus would map\n   that axiom as a claim with its challengers attached (he records one himself: being called a\n   *speciesist*), and would not take a side on it.\n\nA practical aside in his essay bears on this project's own working setup and is recorded for that reason:\nhe advises against running a coding agent under the same operating-system login as everything else.\n\n---\n\n## 3. Acemoglu: a centralising technology, and the tool road not taken\n\n**His argument, the parts that bear here, paraphrased.** AI is being developed as a centralising\ntechnology: the early promise of digital technology was decentralisation, and each decade delivered a\nlarger gatekeeper instead. What he calls the **AGI agenda** has persuaded nearly everyone that the only\nthing to do with AI is make it more human until it beats humans, partly under the spell of science\nfiction. The road not taken is **pro-worker AI**: for nurses, electricians, teachers and scientists, tools\nthat extend their expertise instead of imitating it. On politics: liberal freedoms mean little without\nself-government, self-government at the scale of a nation is close to impossible, so it has to start\nlocally, by building compromises from the bottom up and not by imposing values from the top. Online\ncommunities could have carried this and do not: *\"Who has solidarity with whom on social media?\"* And AI\nitself has become an ideological litmus test on both sides, which he says makes for poor debate.\n\n**Where it meets Deliberus.**\n\n1. **Pro-worker AI is the harness conviction from an economist.** [the-harness-conviction.md](the-harness-conviction.md)\n   says machine reasoning is as good as the structure it works inside; Acemoglu says the valuable machine\n   is the one that extends a practitioner. Both reject *replace the reasoner* as the goal, for different\n   reasons (quality there, wages and power here).\n2. **His centralisation claim is the friend's thesis, arrived at independently.** The position mapped on\n   2026-09-20 ([mapping-a-structured-position-on-ai-governance.md](mapping-a-structured-position-on-ai-governance.md))\n   argues that concentrated control over AI is a risk of its own. Acemoglu reaches the same place through\n   market structure, and adds a remark that matches one of that map's own claims: for years, talk of\n   existential risk crowded out the discussion of inequality and concentration. He is a candidate third voice for\n   that debate, beside the 12 September slowdown essay.\n3. **Self-government starts small** is the outside warrant for testing inside one real community with its\n   own written rules ([blobs-community-sandboxes-and-the-advice-process.md](blobs-community-sandboxes-and-the-advice-process.md)).\n   His definition of a community (trust, shared values, shared memory, solidarity, and above all people who\n   meet) is also a warning: a graph does not make one.\n4. **The litmus-test remark describes the founder's and the friend's situation exactly.** Two people who\n   both take AI risk seriously, on opposite sides of one marker. A map that separates the seven theses lets\n   either of them agree with four without changing camps.\n\n---\n\n## 4. What the three say together\n\n| | Machines now produce | What stays scarce and human | Deliberus's part |\n|---|---|---|---|\n| Sanderson | proofs | understanding, and credit for producing it | make reasoning something a stranger can enter in the right order |\n| Loh | decisions | the mastery to steer | keep checking at the pace of production, with the checker still in practice |\n| Acemoglu | substitutes for people | tools that extend a practitioner, and self-government | be such a tool, first inside small real communities |\n\nOne sentence: **the machines took over the part that could be checked, and all three authors locate the\nhuman remainder in the part that cannot: understanding, steering, and governing ourselves.** That\nremainder is the ground this project already claims (the [missing layer](../the-missing-layer.md)). What the\nthree add is pressure from outside it, in the same fortnight, from a mathematics communicator, a\nmathematician and an economist.\n\n**Candidates this reading leaves, none ruled:** a *motivation* check on reading order (§ 1.3) · discovery\nfiction read out of a claim's own history (§ 1.4) · credit for making another person's reasoning understood\n(§ 1.6) · Acemoglu and the slowdown essay as the next sources in the AI-governance debate (§ 3.2).\n\n**Sources**: [Sanderson on Tao's blog](https://terrytao.wordpress.com/2026/09/18/if-math-is-more-than-proof-we-need-to-better-celebrate-the-rest-of-it/) ·\n[Loh, *Why Do We Need Human Mathematicians Anymore?*](https://poshenloh.com/posts/20260919-math-ai) ·\n[Bastani meets Acemoglu, Novara Media](https://www.youtube.com/watch?v=aNYDMfdD6pg)\n\n**See also**: [depth.md](../depth.md) § Tao's Verification Thesis · [the-scrutiny-gap.md](the-scrutiny-gap.md) ·\n[the-harness-conviction.md](the-harness-conviction.md) · [game-feel-and-the-first-screen.md](game-feel-and-the-first-screen.md) ·\n`specs/map-view/design.md` · [ai-risk-for-mathematicians-and-the-coordination-game.md](ai-risk-for-mathematicians-and-the-coordination-game.md)\n(the same moment in mathematics, argued as extinction risk, and read for what decides whether the race is a dilemma) ·\n[four-cuts-of-the-ai-debate.md](four-cuts-of-the-ai-debate.md) (Aaronson on the 166-page Navier–Stokes solution *\"which probably\nhasn't yet been read and understood by any human\"*, and the twenty-five Fields Medalists' letter Aaronson endorsed)\n"}