{"path":"research/schmachtenberger-and-the-ai-race.md","content":"# Two attractors and a trap: Schmachtenberger read into the AI-risk material (2026-10-01)\n\n**Status**: research, nothing ruled. The proposals are rows U8–U9 of ledger Arc U\n(`session26-open-rulings.md`). **Origin**: the founder, while the AI-risk material of 2026-10-01 was\nbeing read: *\"Schmachtenberger is also always relevant for so much.\"* This doc reads Schmachtenberger's\nframework into that material:\n- [ai-risk-for-mathematicians-and-the-coordination-game.md](ai-risk-for-mathematicians-and-the-coordination-game.md);\n- [four-cuts-of-the-ai-debate.md](four-cuts-of-the-ai-debate.md);\n- the governance map, [mapping-a-structured-position-on-ai-governance.md](mapping-a-structured-position-on-ai-governance.md).\n\n**What the corpus already holds, and what this adds.** Two earlier docs:\n- [zizek-schmachtenberger-connections.md](zizek-schmachtenberger-connections.md) (March 2026) covers the\n  metacrisis, sensemaking as the upstream problem, Game B, the epistemic commons, the Consilience Project's\n  lessons, and multipolar traps.\n- [ai-augmented-cooperation-infrastructure.md](ai-augmented-cooperation-infrastructure.md) § 6 (April\n  2026) summarises the generator functions and the three attractors.\n\nNeither had met the AI-race debate. That debate is where the framework has the most to say, and the game\ntheory found on 2026-10-01 gives one of Schmachtenberger's oldest claims a mechanism.\n\n**Read depth** (revised 2026-10-03, when the excerpts were replaced by full readings):\n- **In full**: *Joe Rogan Experience* #1736 with Tristan Harris (November 2021). The transcript has no\n  speaker labels, so attributions follow its internal cross-references: Harris named in the third person,\n  or a hand-back to \"Daniel\". Attributions that rest on continuity alone are marked *probably*.\n- **In full**: *The Great Simplification* #71, *Artificial Intelligence and the Superorganism* (May 2023),\n  and #132, *Silicon Dreams and Carbon Nightmares* (July 2024), both with Nate Hagens as host.\n- **In full**: *The Great Simplification* #42 (recorded 3 October 2022, Part 4 of *Bend Not Break*, also\n  with Hagens), read 2026-10-03. It is about the economy and the biosphere; AI appears only in passing, and\n  the word *sensemaking* does not occur.\n- **In full**: *Theories of Everything* with Curt Jaimungal (May 2023), from its automatic transcript, which\n  has no timestamps or speaker labels.\n- **Not read**: *Misalignment, AI & Moloch* with Liv Boeree; the February 2026 *Lifeworlds* conversation\n  (*Soulfire Sessions I*).\n\n**What the full reading corrected (2026-10-03).**\n1. The quote first used for the remedy, *\"agreements that get out of the race to the bottom type\n   dynamics\"*, answers a question about pricing resources, not about the AI race. The AI-race passage is a\n   different one, and it names enforcement and verification as the problems (§ 2).\n2. *\"Voluntary\"* is #42's word. In #71 and #132 the remedy is an agreement among those who could win. It\n   is hard because nobody can verify the other side and nothing enforces it internationally (§ 2).\n3. #71's closing is not titled after McGilchrist. McGilchrist is the second of two closing citations,\n   *\"Wisdom is the master and intelligence is the emissary?\"* is the host's question, and *wide* and\n   *narrow boundary* are the host's vocabulary (§ 5).\n4. The ecological objection does not need much capability. Schmachtenberger's Jevons argument runs on\n   *\"marginal efficiency gains that don't require fundamental breakthroughs\"*, and it comes in three\n   circles, not two (§ 4).\n5. An *\"FDA for AI\"* judgment was first paraphrased loosely (*needed and nowhere near enough*), then\n   removed as appearing in no transcript, before the 2023 interview had been read in full. The full\n   transcript has it: an FDA for AI is *\"both true and a profoundly inadequate analogy\"*. It is restored in\n   his words (§ 3).\n6. The image of two gutters on a bowling alley is Harris's, in the same episode. The two-attractor\n   sentence quoted in § 1 is Schmachtenberger's.\n\n---\n\n## 0. In short\n\n- **Schmachtenberger's two attractors are the axis the debate charts lack.** One failure is catastrophe\n  from power spread with no binding order. The other is dystopia from the top-down control built to stop\n  that. Asking every position *which failure do you fear more, and what do you accept from the other?*\n  sorts the material better than *instrument or agent*, and it yields a move, which is the corpus's test for\n  a borrowed category.\n- **The trap's type turns on beliefs, and Schmachtenberger sorts the racers the same way.** In 2021 he put\n  shared sensemaking under every issue: *\"if you can't make shared sense of the world, you can't act\n  effectively on issues\"*. The game theory read on 2026-10-01 says the AI race is a dilemma or a trust\n  problem depending on what its players believe. His own sorting is by a belief about *winning*. Agreement\n  is possible *\"if everybody can realize that their likelihood of winning the race is low\"*, and those who\n  could bind the race *\"don't want to because they believe they can win\"*.\n- **The trust problem's hard part is verification, and he states it without offering a mechanism**: *\"how\n  do we know the other guy's keeping the agreement\"*. Controlling capability once it has spread would take,\n  in his words, *\"universal surveillance on everyone's use of their home computer\"*. The verification\n  literature narrows ledger U9 accordingly: creation can be verified at chokepoints without watching people,\n  and spread capability cannot (`shared-sensemaking-what-deliberus-holds-and-owes.md` § 5).\n\n## 1. The missing axis: which failure each position fears\n\n**The framing, in Schmachtenberger's own words** (*Joe Rogan Experience* #1736, 2021): *\"either you\ncentralize the power and you get dystopias or it's decentralized and you get catastrophes. And right now,\nthe future looks like one of those two attractor states, most likely. Catastrophes or dystopias. We want a\nthird attractor. How do you have a world that has exponential tech that doesn't go catastrophic where the\ncontrol mechanisms to keep it from going catastrophic aren't dystopic?\"* Earlier in the same conversation\nHarris introduces the picture as a bowling alley with two gutters.\n\n**#71 states the horns in AI terms.** On one side, *\"a centralized AI system runs the world\"*: *\"a central\npoint of failure capture and corruption and no checks and balances on the power\"*. On the other, *\"separate\ncompeting groups that don't have totalizing power, in which case they're caught in multipolar traps\"*. And\nthen: *\"to control for all the catastrophes orients towards very centralized power, which gives dystopia.\nSo what we need is a third attractor that is neither of those\"*.\n\n**#42 gives the fullest statement of the third attractor**, about exponential technology and\nexternalities in general rather than AI. There are *\"two attractor states for the future, meaning states\nthat are most likely based on the kind of general dynamics we're on\"*, and *\"The catastrophes largely\nresult from not having more coordinated governance.\"* Preventing them needs *\"some powerful control\nsystems because incentives alone won't do it. You will need deterrents\"*. Then the other horn: *\"if you\ncreate powerful enough control systems to check all the exponential tech and to check all of the\nexternalities, how does that not become dystopic?\"*, and *\"If you make a system that's centralized, even\nif the initial governance of it is good, isn't it capturable by negative forces eventually?\"* So *\"the third\nattractor we're looking for is able to prevent all the catastrophes and have the coordination system that\ndoes that not be corruptible or capturable and that is threading the eye of a needle. And I don't think it\nis impossible and it's also not easy.\"* What he asks for is *\"effective global governance, which does not\nhave to be a government, it can be a decentralized process\"*, because some underrate *\"how bad a one world\ngovernment would be from any type of government we've ever had so far\"*. Near the end the question stays\nopen: *\"how do we create that is both enforceable and non dystopic?\"*\n\n**The 2023 interview names what control would take.** Catastrophic technologies are *\"way easier to build\n[than] the nukes and way harder to control. When you have many actors that have access to many different\ntypes of catastrophic technology that can't be monitored, you don't get mutually assured destruction and\nthose types of safeties\"*. He then voices the control side, *\"we need ubiquitous surveillance to know what\neverybody's doing in their basement\"*, and judges it in the next breath: *\"So we can see catastrophes or\ndystopias, right? Because most versions of ubiquitous surveillance are pretty terrible.\"* The same interview\nputs both horns in one sentence: *\"If you don't have some coordination, you get these problems of\ncoordination failure. If your coordination is imposed, you end up getting oppression dynamics. So how do you\nget coordination that is global but that is emergent\"* (*Theories of Everything*, automatic transcript).\n\nIn this doc's summary: catastrophe where power is spread with no binding order, dystopia where it is\nbound by top-down control.\n\n**Read across the material on that axis:**\n\n| Position | Fears most | Accepts from the other side | Source |\n|---|---|---|---|\n| Xiaoyu He's essay | catastrophe (takeover, \"magic\" that cannot be contained) | *\"enforcing a global AI pause or ban\"* (S19). Concentrated power is set aside (*\"evil CEOs or soulless corporations\"*) | essay S19, § 4.5 |\n| The governance map | both, explicitly | — its thesis 4, as rendered in the map, is *\"The risks of distributed capability and of concentrated control must be modelled together\"*, and thesis 5 refuses to count the institutions' own harm as zero | map § 1 |\n| Erik Hoel | catastrophe, and dystopia in one specific form: *\"the concentration of all power into one machine\"*, which he reads as Madison's tyranny taken to the extreme | a ban written into national law, justified as time for voters to understand | *Humanity's Flight 93 Moment*, read in full |\n| The many-agent comment | catastrophe from spread-out agents claiming rights | nothing; it expects coordination to fail | the He essay doc, § 4a |\n| Misuse critique (as a quadrant) | partly the dystopia side: corporate concentration and harm (a reading of the quadrant, not of each name on it) | — | the chart, four-cuts § 3 |\n\n**What the axis shows.** He's essay is a single-attractor frame. The governance map was a two-attractor\nframe before anyone here had named the two attractors. Hoel's position names one form of dystopia, the\nconcentration of power in the machine, and answers it with the oldest anti-dystopia instrument there is:\nlaw made by voters.\n\n**The question that falls out, offered as a critical question for every position in the cluster:**\n***which failure do you fear more, and what do you accept from the other?*** A borrowed category earns its\nplace here by yielding a move (`what-belongs-in-the-ontology.md` § 6b). This one yields a question a\nreader can put to any claim, and the answer locates a disagreement that the instrument-or-agent axis\ncannot see. (Proposal, ledger U8.)\n\n## 2. The trap's type depends on what its players can see together\n\n**The claim** (JRE #1736): *\"we have decreasing ability to make shared sense of the world. And in any kind\nof democratic society, if you can't make shared sense of the world, you can't act effectively on issues.\"*\nThen: *\"underneath the AI issue, the CRISPR issue, the US-China issue, the how do we regulate markets issue\n… is, can we make sense of anything collectively, adequately, to be able to make choices effectively in the\nenvironment we're in?\"* For years this was an intuition: sensemaking comes first because coordination\nneeds it. By 2023 he names a different generator, *\"Narrow goal achieving is the underlying generative\ndynamic\"* (#71). So sensemaking is the bottleneck to correcting the trap, not its cause.\n\n**The mechanism.** The game theory read on 2026-10-01 makes the intuition precise:\n- The AI race is a prisoner's dilemma for players who believe the risk is low.\n- It is a stag hunt, a trust problem, for players who believe it is high enough. Under one simple model,\n  10% suffices if doom is at least nine times as bad as winning is good.\n- RAND states the condition outright: cooperation becomes possible when *\"a common shared assessment\n  emerges\"*, and needs mechanisms *\"to establish common knowledge and verify mutual commitments\"*.\n\n(`ai-risk-for-mathematicians-and-the-coordination-game.md` § 3.)\n\n**Schmachtenberger sorts the racers by a belief too.** #71: *\"Agreement is a way out of a multipolar trap.\nThe agreements are hard, because if everybody can realize that their likelihood of winning the race is\nlow, the race as a whole might mess everything up for everybody, and not doing the thing is better\"*. #132:\n*\"people who are at the front of an arms race, like who really can have a chance of winning, could bind it\n… But they don't want to because they believe they can win\"*, and *\"if I think I can win, I like that world\nbetter than the one where I have to share power in a multipolar way\"*. The belief that matters to him is\nabout the chance of winning; to Papenhausen it is about the chance of doom. Both are beliefs a shared\nassessment could move. Those who think they can win play something nearer a dilemma, and those who think\nthey cannot see the trust problem.\n\n**The critique this refines.** The corpus's March doc put a critique in Schmachtenberger's mouth: Deliberus\n*\"addresses the sensemaking layer but not the incentive layer\"*. That is the corpus's own construction of\nwhat he would say, not a quotation (`zizek-schmachtenberger-connections.md` § 3.4). The mechanism sharpens\nit into two cases:\n- **Where the race is a trust problem**, the shared assessment largely *is* the incentive. Each side\n  already prefers mutual restraint and needs assurance, not new payoffs.\n- **Where it is a true dilemma** (players who disbelieve the risk, or who expect to win), the critique holds\n  in full. Payoffs must change, through liability, law and mechanism design.\n\nThe material contains both halves: a governance mechanism (the friend's second document, held privately)\nand the sensemaking question. Mapped together, they would be the coupling the critique asks for.\n\n**The remedy, as he states it, is an agreement, and he names why it is hard.** #71, on the AI race\nspecifically:\n- Within a nation, a law can bind *\"where you have a monopoly of violence and backing up rule of law\"*.\n- Internationally: *\"But if you don't have monopoly of violence and rule of law, how do you enact it? … The\n  tricky thing is, how do we know the other guy's keeping the agreement, and that they're not defecting and\n  secretly doing it?\"*\n- #42 gives the shape of the agreement: parties accepting *\"constraints on their own behavior knowing that\n  the other ones will also accept constraints on their behavior, some kind of voluntary agreement\"*, because\n  *\"there is no international monopoly of force because how do you enforce a deterrent on countries that\n  have nukes in critical parts of your supply chain? It's very hard.\"* Hence *\"you can't force that\n  thing\"*. That is the stag hunt's assurance logic in his own words. *Voluntary* is not the whole of it:\n  the same episode says *\"incentives alone won't do it. You will need deterrents\"* (§ 1).\n- #42 also has a *voluntarily* line that reads, out of context, like a remedy for the race. It answers\n  Hagens's doubt about internalising externalities across the whole economy (*\"we're not going to voluntarily on\n  mass choose that\"*), not the AI race: *\"we do have to voluntarily at a minimum adequate scale, create a\n  new system that doesn't self terminate or we get the plausible deniability of saying we couldn't have\n  done anything else\"*, *\"which is why understanding what the alternatives look like is key.\"* The last\n  clause ties understanding to choice, the same two stages as the 2021 sensemaking sentence.\n- #132 insists it is possible: *\"What would it take to create real trust there? Could we do it? Yes, we\n  could do it. The argument that we can't and therefore we have to rush ahead to win and there is no safe\n  solution is actually gibberish.\"*\n\nHe is not confident it will happen: *\"it doesn't look good. To be honest, it doesn't look good\"* (#71).\n\n*(The sentence this section first quoted, *\"we talk about the need to be able to make agreements that get\nout of the race to the bottom type dynamics, the multipolar trap, the social trap\"*, is verbatim in #71. It\nanswers a question about countries that simplify first being out-competed, and the next sentence is about\npricing resources: *\"So either everybody has to or nobody can, right?\"*)*\n\n**His own protocol for a debate.** In the same 2021 conversation, probably Schmachtenberger speaking, a\nfacilitated debate is described:\n- *\"One of the things they agree to is, what would I need to change my mind about this? If the answer is\n  nothing, then you don't even engage in the debate.\"*\n- The debaters read each other's material first, start from what they agree on, *\"both around the values\n  that are relevant and the facts of the matter\"*.\n- *\"Then we try to formalize our disagreement\"*.\n- They then ask what it would take to settle it: *\"Do we need to do a new piece of science? Do we disagree\n  because of an intuitive hunch or a different value?\"*\n- The output informs institutions, *\"all of that be transparently oversighted\"*.\n\nThe transcript has no speaker labels; Harris's interjection *\"Partial truth, but not complete. Go on.\"* marks\nhim as the listener. This is the closest thing in these sources to Deliberus's own method:\n- the disagreement made explicit;\n- its basis typed, as an empirical gap, an intuition or a value;\n- the shared ground found first.\n\nIts entry condition, *what would change my mind*, is the slot the ontology lacks\n(`shared-sensemaking-what-deliberus-holds-and-owes.md` § 2.5).\n\n## 3. Where assurance meets the dystopia gutter\n\nThe trust problem's hard part, in Papenhausen's words, is *\"verification and assurance\"*. Schmachtenberger\nties controlling widely spread catastrophic capability to surveillance of everyone. Two places:\n- **JRE, probably Schmachtenberger, on cyber weapons**: *\"the only way to regulate that is universal\n  surveillance on everyone's use of their home computer. And we don't want that future.\"*\n- **The 2023 interview**, voicing the control side of the trap, *\"we need ubiquitous surveillance to know\n  what everybody's doing in their basement\"*, then rejecting it: *\"most versions of ubiquitous surveillance\n  are pretty terrible\"*.\n\nHe does not tie that to verifying a treaty. That join is this doc's inference. The verification literature\nsupports the inference for spread capability and refutes it for creation (§ 3a below). **So the escape\nroute the game theory points to runs toward the dystopia gutter once capability has spread**, and not\nbefore.\n\n**3a. Narrowed 2026-10-03.** Compute monitoring can verify frontier *training* by watching chips, data\ncentres and supply chains, while leaving consumer devices alone. Once a model exists, its weights *\"may from\nthen on be copied undetectably\"*, and policing their use *\"would require surveilling and policing individual\ncitizens' use of their personal computers\"* (Shavit 2023). The 2023 interview draws a related asymmetry,\nbetween developing a model and running it: monitoring all large GPU clusters *\"to some degree can be\ndone\"*, but *\"it takes a very large GPU cluster to develop an LLM. It takes a very small one to run that LLM\nafterwards\"*. It says nothing there about copying weights or policing personal computers. The argument and its sources are in `shared-sensemaking-what-deliberus-holds-and-owes.md`\n§ 5.\n\n**That makes the clearest crux in the material the two attractors themselves:**\n- **Openness**, which the governance map treats as dual-use while arguing against custody, buys\n  verification of what a model *is* without surveillance. In the map's own words, *\"open weights can enable\n  distributed, model-level error correction and safety improvement\"*. It pays with proliferation, which the\n  map also records (*\"open weights make a provider's guardrails removable\"*). Once weights are open,\n  verifying what people *do* with them is exactly the verification that needs watching everyone.\n- **Custody** (what the essay's S6–S7 would favour) buys containment. It pays with concentration, and\n  depends on the essay's own doubt (S7) that any lab can hold it.\n\nSo the proposal is that every proposed escape from the race should say:\n- which verification it relies on;\n- whether that verification checks creation, use, or an artifact's properties;\n- which gutter it leans toward.\n\n(Ledger U9, revised.)\n\n**Law as restraint is shared ground, with two caveats from the full reading.** On #71: *\"you could think\nabout the Sabbath as an example of law binding a multipolar trap\"* and *\"almost all of law is about\nrestraint\"*. Three positions agree that law is the instrument:\n- the essay's pause or ban;\n- Hoel's ban, justified by democratic digestion time;\n- the governance map's insistence that extraordinary governance owe proofs of hazard, effectiveness,\n  systemic effects and legitimacy.\n\nThe caveats:\n1. His Sabbath is religious law with a death sanction inside a shared culture, not a pact between rivals.\n2. He separates law from wisdom at scale. Where a group grows past mutual acquaintance, *\"We start to get\n   this kind of intelligence rule-based law that is different than wisdom\"*.\n\nOn a regulator for AI he is specific. An *\"FDA for AI\"* is *\"both true and a profoundly inadequate\nanalogy because a single new chemical that comes out is not an agent\"*, and regulation kept national\nfails, so the ask is *\"some type of global governance, if not government\"* (2023 interview).\n\nSo *law as the shared ground* is this doc's synthesis. **The live question is the scale and the\nverification, not whether law should bind.** Can a law bind the race at the scale where the race runs, when\nthe international case has no monopoly of violence, without the verification it needs becoming the\ndystopia it was meant to prevent?\n\n## 4. Three ecological objections, not one\n\nO'Connor's five objections counted *\"ecological catastrophe\"* as one claim, compatible with slop\n(`four-cuts-of-the-ai-debate.md` § 4). In #132 Schmachtenberger draws it as three concentric circles:\n1. *\"The energy and material use of the AI\"*;\n2. *\"AI applied to extractive and polluting industries and their efficiencies\"*;\n3. *\"AI simply increasing the market as a whole, even where it's doing software and financial services\"*.\n\nThe second and third run through Jevons: *\"the gains in efficiency will lead to more total energy use,\nmaterial use, et cetera, not less\"*. #71 frames the same argument as first-, second- and nth-order effects,\nand in both episodes the host introduces Jevons first.\n\n**Corrected 2026-10-03: none of the three needs much capability.** This section first said the acceleration\nobjection *\"depends on the technology working\"* and needs a great deal of it. Schmachtenberger says the\nreverse. The mechanism runs on *\"marginal efficiency gains that don't require fundamental breakthroughs,\nthat's gonna happen\"*, and the climate uses on offer are *\"almost all very marginal efficiency gains that\nwill not change the curves very much and will lead to Jevons paradoxes\"*. The argument is conditional\ninstead on AI being *\"in service to companies\"* inside a market that seeks returns, and he proposes binding\nit by law and treaty.\n\nSo *slop* undercuts none of the three, and neither does *bubble* in the investment sense. O'Connor's single\nobjection is three, with different evidence for each, and the headline narrows again.\n\n## 5. Intelligence unbound by wisdom\n\n**Corrected 2026-10-03.** #71's closing movement starts at 02:46:25, when Schmachtenberger offers to *\"bring\nback a few threads\"*. It runs through:\n- AI as an accelerant: *\"AI is not a risk within the meta crisis. It is an accelerant to all of them\"*;\n- wisdom defined against any metric set;\n- then two citations, Bohm and McGilchrist.\n\n*\"Wisdom is the master and intelligence is the emissary?\"* is the host's question. Schmachtenberger's\nanswer is that AI takes *\"that part of us already not bound by wisdom and putting it on a completely\nunbound, recursive exponential curve\"*. *Wide* and *narrow boundary* are the host's terms; his are narrow\ngoals and wide values.\n\n**The claim is stronger than *wider boundary*.** *\"The difference between what the weighted metrics set of\nall the identified important metrics and the optimization function on it says you should do, the difference\nbetween that and what you should actually do is wisdom. And it requires being able to attune to more than\njust the known metrics\"*. And: *\"The metric set that is definable is not the right metric set.\"*\n\nThe corpus holds the same structure in its own terms: analysis and attunement as two poles neither of which\nmay eat the other, also McGilchrist-derived (`../analysis-and-attunement.md`). Schmachtenberger's word for\nwhat exceeds the metrics is *attune*. Three outside echoes this week fit it:\n- Gómez-Emilsson's gap between knowing what is good and being moved to pursue it;\n- the Fields Medalists' insistence, which Aaronson endorsed, that human understanding stay at the centre;\n- Hoel's worry that AI moves the world *\"from under the supervision of consciousness to unconsciousness\"*.\n\n**One consequence for Deliberus.** A typed claim graph with computed strengths is a definable metric set.\nOn Schmachtenberger's definition no such set is wisdom, which is one more reason the map shows a reading and\nnever chooses.\n\nThis is a resonance, not a finding, and it is recorded so that it is not rediscovered as one.\n\n## 6. A test the material makes possible\n\nThe metacrisis claim, that the races in AI, climate, arms and markets are one dynamic, is already queued\nhere as an empirical question about the graph. That is the three-scales auto-connect experiment\n(`fractal-scales-and-temporal-frame.md` § Where the metacrisis claim becomes testable; `TODO.md`). Once the\nAI-risk material is mapped, it is a ready test. Do its race claims (S17, the trust-problem reading, law as\nrestraint) share premises or structure with race claims already in the graph from other domains? If\nauto-connect finds shared structure, the connection is demonstrated rather than asserted. If it finds\nnothing, the claim loses a little, and that is what makes the test worth running.\n\n## 7. Cautions\n\n- **The framework is a lens, not a finding.** Never write *\"Schmachtenberger shows that…\"* in a funding\n  document. The rule is the one earned on assembly theory: do not borrow its authority. The two attractors\n  earn their place here by yielding a question, not by being true.\n- **The corpus has put words in Schmachtenberger's mouth before, and this doc did too.** The March doc's\n  § 3.4 critique is labelled there as what he *would* argue. This doc's first version made a quote carry a\n  context it did not have and paraphrased a judgment loosely. Its first correction then removed that\n  judgment as absent, before the one transcript that holds it had been read. Quote only from a transcript\n  you have read around, and never call a quote absent from a source you have not read in full.\n- **The game-theory vocabulary is ours, not his.** *Stag hunt* and *voluntary* do not occur in #71, #132,\n  JRE #1736 or the 2023 interview, and *prisoner's dilemma* occurs only in the host's primer there. #42 has\n  *\"some kind of voluntary agreement\"*.\n- **His remedy comes with his own pessimism**: *\"it doesn't look good\"*, and, in #132, *\"How much goes into\n  really trying to figure out how to bind arms races? Almost none.\"*\n- **The most recent AI-specific statements read date from 2024.** What Schmachtenberger has said in 2026,\n  after the summer's incidents, is unread. The February 2026 conversation is the first place to look.\n\n**Sources**:\n[JRE #1736 transcript](https://podscripts.co/podcasts/the-joe-rogan-experience/1736-tristan-harris-daniel-schmachtenberger) ·\n[TGS #71](https://www.thegreatsimplification.com/episode/71-daniel-schmachtenberger) ·\n[TGS #42 transcript](https://www.thegreatsimplification.com/wp-content/uploads/2022/10/TGS42DanielSchmactenberger04Transcript.docx.pdf) ·\n[TGS #132](https://www.thegreatsimplification.com/episode/132-daniel-schmachtenberger) ·\n[TOE, *AI Risks & the Metacrisis*](https://www.youtube.com/watch?v=g7WtcTATa2U) (read in full, automatic transcript) ·\n[*Misalignment, AI & Moloch*, with Liv Boeree](https://civilizationemerging.com/media/misalignment-ai-moloch/) (not read) ·\n[*Soulfire Sessions I*, Lifeworlds, February 2026](https://shows.acast.com/lifeworlds/episodes/soulfire-sessions-with-daniel-schmachtenberger) (not read).\n"}