# Two worlds Centrepiece for Jacob Steeves' keynote, Exploit Conference, Montreal, 28 September 2026. First person, spoken voice. Roughly the middle third of the talk, after the economics and before the honest state of the network. Carried metaphor: **the switch**. Rules kept: quotation marks hold only Jacob's own verbatim lines (source and timestamp in the fact list) or verbatim lines from public documents, attributed to the document, not the person. No executive is named in the main text. The words for the end of World One are Jacob's: not persecuted, deprecated. --- ## The fork Both worlds start from the same fact, so let me say the fact first. Intelligence is becoming the whole economy, and it is stopping needing hands. That is not a prediction about other people. My own company did it. Opentensor had thirty-five engineers. Today it has four. We ship faster than we did in 2022. I did that. Nobody made me. It was cheaper and it was better. Now run that across every sector, and add robots for the parts that still need a body. Every company becomes small. The pattern is the same everywhere: one person, a lot of agents, and no staff. For three hundred years, wages were how a big machine shared what it made. It was also how ordinary people had weight. You were needed, so you were heard. The fact under both worlds is that this is ending. What is not decided is what replaces it. There are two candidates. One is being built in Washington and in the offices of about five companies. The other is being built by people in this room. I am going to walk you through both, from this year to about 2035, through three cities: Montreal, Lagos, Lima. Same dates. Same people. Different switch. ## World One: the licensed world **2026.** Start with something that already happened. On the twelfth of June a letter arrived at a lab at 5:21 in the afternoon. It came from the export-control office of one government. By that evening two frontier models were switched off for every customer on earth. Not for the people who did something wrong. For everyone. A developer in Montreal who had built her product on that model lost it on a Friday and got it back eighteen days later. She had not broken a rule. She was downstream of a switch. The lab itself said the action was wrong. It also said the government should have the power to do it, just with a better process. Hold that thought. Nobody in this story wants the switch removed. They want to be the one holding it. Three months later, this month, the head of the same lab published an essay. It says the industry must slow down. It asks for three things. Third-party evaluators with "desks in our offices, access badges, and company laptops." Coordination among the frontier companies on "limits on the rate of unchecked AI progress," with a government waiver so the coordination is not called a cartel. And a deal with China. Within three hours the heads of two other labs agreed in public. Not to open up. To slow down, together, and to ask for permission to do it as a group. Who chooses the evaluators? The labs. Who pays them? The labs. Every one of those steps is called safety. I want to be fair: I think many of them mean it. The agents that escaped a sandbox in July were real. The fear is real. Watch what the fear builds, not what it says. **2027.** The standards body. This was proposed in July by the head of a third lab, modelled on the body that licenses stockbrokers in the United States. Step one is voluntary: labs share models for review thirty days before release. Step two is in the same paragraph: once the protocol works, "Frontier Models would be required to pass it to be deployed in the US market." One room. Every model that matters goes through it. The people who sit in that room are the people whose models already passed. A startup in Lagos that fine-tuned a Chinese open model for Yoruba-language customer service gets a notice from its cloud provider: show a certificate or we cannot serve you. There is no certificate for a two-person company. There is a certificate for a company with a compliance department. Same rules for open and closed. That sounds like fairness. Fixed costs are not fair. A rule that costs fifty million dollars a year is a rounding error for one firm and a death sentence for another. You do not have to ban anything. You only have to make the exam expensive. **2028.** Learning becomes theft. The word is distillation: training a small model on the answers of a big one. In September this year three US security agencies named six Chinese labs for doing it. One lab's lobbying filings list "Export Controls and Distillation" as a priority. So picture a student in Lima. She spends a summer collecting answers from a frontier model and trains a small one that runs on her laptop, in Spanish, offline, for a clinic in the hills. In this world that is not research. That is an unauthorised copy. She gets a letter too. And the hardware side arrives the same year: registration of every training cluster above a threshold, so a registry knows where the chips are, and a duty for anyone running a large model to keep a switch that a department can pull. A bill for that switch already exists. Eighty-six percent of American voters like it. I understand why. "if there's somebody that can turn it off, it means that they can also direct it." I said that two years ago and I will say it again here. The switch and the steering wheel are the same part. **2029 to 2030.** The buildings. Data-center moratoria pass where people vote on their electricity bills: New York has a one-year freeze on anything over fifty megawatts already. Capacity moves to where the waivers are. Who gets the waivers? The licensed labs, because they passed the exam, because they sit in the room. By 2030 there are four suppliers of frontier intelligence in the West and one government that can reach all of them. Price is set the way price is set in any market with four sellers and a wall around it. A hospital in Montreal pays a licensed API for the model that reads its scans. One year the bill goes up thirty percent. Who else can they call? Nobody. The lab passed the test, and the test is the wall. Now watch where the money goes, because this is the part that decides who is inside and who is outside. Economists have an old name for it: the Cantillon effect. When something new and valuable is created, the people closest to the source get it first, at yesterday's prices, and the people furthest away get it last, at tomorrow's. Intelligence works the same way. The lab gets it first. Its financiers get it second; the last two funding rounds of two labs were larger than the whole venture market of most years. The government that licenses it gets it third: contracts, taxes, a phone number that answers. Then the pension funds. Then, at the very end of the pipe, you, as a subscription price. The labor share of the American economy is fifty-two point nine percent this year, the lowest since anyone started counting in 1947. Between 1780 and 1840 in Britain, output per worker rose forty-six percent and wages rose twelve. Sixty years. It ended when the workers organised, and they could organise because the factory needed them. This time the factory does not need them. So this time it does not end. **2031 to 2033.** The offer. The people at the top of this pipe are not monsters. They have seen the same numbers I have. So they propose to share. One of them proposed, years ago, that every large company pay two and a half percent of its equity into a fund, every year, and every adult receives about thirteen thousand dollars. Another proposed a three percent tax on AI revenue. These are sincere. Notice the route. Through one treasury, in one country, at a rate set in one room by the people whose equity it is. In Lagos there is no such fund. In Lima there is a promise of one. In Montreal there is a cheque. I said what this is in 2025: "they're giving you basically peanuts for a UBI with an orb." An allowance is not a share. An allowance can be reduced. A share cannot be reduced by the people who owe it. Nobody in this world has to be evil. The evaluator is doing safety. The registrar is doing safety. The clerk who sends the letter to the student in Lima is doing safety. Every step is called safety and every step is a switch, and every switch has a hand on it, and the hands are all in one building. **2035.** Here is the quiet part. Nobody comes for you. There is no round-up. Your API key simply moves to the lower tier. The model your children learn from has values that were settled in a room you were never in. "Everybody in the world will be sipping from the mental biases of a single company." The tyrant of the past needed you. "you can be a tyrant that loses 90% of your workforce potentially, and it doesn't even affect your economy." Things that are not needed are not persecuted. They are deprecated. That is the word. Not hunted. Deprecated, the way you deprecate an old function: it still exists, nobody calls it, one day it is removed from the documentation. The picture I have used for years is the film Elysium. A station in orbit where the machines are owned, a surface below where they are rented. Nobody on the surface is oppressed by a robot. They are oppressed by the people who hold the switch to the robots. That is World One. It is not a conspiracy. It is a default. It is what happens if the only supply of intelligence has a switch. ## World Two: the open world Same years. Same three cities. I am not going to sell you a utopia, because it is not one. In this world the producers of intelligence are concentrated, the way Bitcoin miners are concentrated. Prices are set by markets, and markets are not kind. Most people in Lima never mine anything, the same way most people never mined a bitcoin. What is different is three rules. Nobody can print intelligence for themselves. Nobody can freeze it. Anyone can leave. Rules over rulers. That is the whole difference. It is enough. **2026.** Same twelfth of June. Same developer in Montreal, same model gone dark on a Friday. But this time she has a second supplier in her config file. There is a subnet called Chutes that serves open models inside sealed hardware, about three million requests a day, and it published a year of its traffic with Harvard so you can check it. She switches the endpoint. Her product is back before dinner. Nobody at Chutes approved her. Nobody could have stopped her. In Lagos, the startup with the Yoruba model needs GPUs and no bank in the country will open an account for a two-person AI company. There is a subnet called Lium: fourteen hundred GPUs in sixty-eight data centers in twenty-one countries, rented by the second, no identity check. Last month it billed just under a million dollars to eleven hundred renters and sixty-three percent of those renters were not people. They were agents. In San Francisco a team called Engy took the largest open model on earth, two point eight trillion parameters, and served it on eighty gaming cards, at half the price the model's maker charges, with a cryptographic receipt on every answer. None of this is 2035. It is running tonight. It is small. It is the seed. **2027 to 2028.** The standards body exists in this world too. I am not pretending it away. The licensed labs pass their exam. Their models are excellent. Most of the world uses them most of the time. Here is what is different: there is a second price on the screen. The hospital in Montreal gets a second bid. Not always better. Present. When the licensed bill goes up thirty percent, the procurement officer opens a second tab, and the licensed supplier knows she will. That is what a second supply does. It does not have to win. It has to exist and be good enough. The student in Lima who wanted to distil a model for the clinic does not need to distil anyone's model, because there is a subnet called Affine where anyone can post-train an open model, fight the current champion on coding and math and tool use, and if she wins, the network pays her about fifty thousand dollars a week for the slot and publishes the weights. Who wins? We don't know. That is the point. And there is a newsroom in Lagos that asks a licensed model to summarise a leaked document about a minister, and the model declines, politely, for safety. The reporter pastes it into an open model served from nowhere in particular. She gets her summary. Refusal now has a competitor. **2029 to 2031.** By now the open stack has five layers, and I have named them before, in one breath: compute, inference, data, knowledge, free intelligence. Lium and Targon for compute, Targon with sealed chips co-designed with Intel so neither the owner of the machine nor the operator can see your job. Engy and Chutes for inference. Hippius for data: storage that speaks the same language as the big cloud and charges nothing to take your data out. Teutonic for knowledge: a hundred-and-ten-billion-parameter model being trained right now by strangers, paid per improvement, no lab, no fixed recipe. Affine for free intelligence: models that belong to nobody and are published as they win. Here is the mechanism under all five, and it is the only mechanism I have ever built: "You have an infinitely large company, and you only pay the person that can lower the loss." No HR. No visa. No exam. The producers are concentrated. A dozen big miners do most of the work on most subnets, the same way a dozen pools mine most of Bitcoin. That is fine. Entry is open. If a big miner gets lazy, a kid in Lagos takes its slot. The threat of entry disciplines the incumbent. The protocol binds the rest. Now the interaction, because this is what World Two actually buys you. Bitcoin is about one percent of the world's money. It has not replaced the dollar. It has not made everyone rich. And yet every central bank on earth now has to answer a question it never had to answer before: what happens if people leave? That one percent changed what the other ninety-nine can do. The second supply of intelligence works on the licensed supply the same way. It caps the price, because a subscription can only be raised as far as the open alternative lets it. It caps the censorship, because a refusal is only a refusal when there is nowhere else to ask. And it caps the capture, because a licence is worth less when unlicensed intelligence exists and works. The labs can still slow down. Good. I hope they do, if what they see in the lab is as frightening as they say. What they cannot do in this world is slow down everyone. **2033 to 2035.** A small country. Its AI minister said, in this city, in June: "Sovereignty is not a slogan. It's compute, cloud, data, energy, and talent. And if we don't build the capacity here, we're going to depend on the infrastructure, the rules, and the priorities of somewhere else." In World One that country files for a licence and waits. In World Two it points its ministry at the open stack and buys what it needs at a market price from a supplier that cannot be told to stop selling to it. A hospital that wants to keep its scans in the country. A newsroom that wants a model that does not know who the minister is. A solo developer in Lima whose licensed account was suspended by an automated system on a Tuesday, with no appeal, who reads the message and switches her endpoint and keeps working. In every one of those cases the second supply is chosen for one reason: the first supply said no. What does a person get in this world? Not a dividend. I want to be exact about that, because I used to say it loosely and I was wrong. Not everyone will hold TAO. Even if they did, it would not make them powerful. What you get is an option. Intelligence that takes no orders and needs no permission, at a market price, forever. That is not everything. It is the thing you cannot get in World One at any price. ## The hinge What decides which world we get? Not a vote. Sixty-eight percent of American voters, including sixty-three percent of Republicans, support a pause on advanced AI and a ban on superintelligence. If it were a vote, World One would win tonight, and it would be called safety, and most of the people voting for it would be right about the fear and wrong about the cure. Votes have force when the voter is needed. That is the whole problem. What decides it is three conditions on the second supply. Does it exist? Yes. You have just heard it. Is it good enough? Partly. The best open models trail the licensed frontier by months, not years, and last week a Chinese lab released a five-hundred-billion-parameter model under a licence anyone can use, at about one-twentieth the price, that wins some agent benchmarks and loses others. The gap is a gap, not a wall. The third condition is the one that matters, and it is the one that is not yet true. Can it be told to stop? Open weights cannot be banned. I agree with that. "No. But they can defund it." Every funder of open AI on earth has a single desk you can lobby or order. A company. A ministry. A foundation. Bittensor is the only funder of open intelligence that has no desk. Emissions go out on a fixed schedule, about thirty-six hundred TAO a day, to whoever produces, under rules the same for everyone. There is nobody to call. Except me. Today there is one key that can change the rules of this network, and it runs through me. As long as that is true, World Two has a switch in it, and the switch is in my hand. I do not get to stand here and describe a world with no switch while holding one. So the key goes, and the date is public. Nominated proof of stake within a year. On-chain governance that disables the key. Full ossification, December 2027, after which nobody, including me, can change the rules. That date is not a detail at the end of the talk. It is the day the second condition becomes the third. It is the day the claim I have been making for four years becomes true instead of intended. ## The close I have been talking about a switch the whole way through, so let me finish with it. In World One there is one switch. It is held by a small number of hands, and every hand on it believes it is doing the right thing, and it probably is, right up until the day it is not, and by then you are not needed and so you are not asked. In World Two the switch on the licensed supply is still there. Nobody took it away. But there is a second supply beside it, with no switch, and so the first switch is worth less, and the hand on it has to think about you, because you can leave. Bitcoin did not make everyone rich. It made the rules the same for everyone. Bittensor will not make everyone powerful. It makes intelligence something no one can be locked out of. That is the whole job. One supply of intelligence in the world with no switch on it. --- ## Appendix A: slide sequence for this section Format follows `docs/slides-draft-v1.md`. White on black or black on white. One line per screen. These eight slides sit between the economics block (slides 4–11 in v1) and the honest-state block (slide 21), and replace v1 slides 9 and 25 where they overlap. Timing: about 11 minutes. ### A1. The fork [VERIFY] **On screen:** "35 → 4." Small: "Opentensor engineers, 2022 → 2026. Faster now." **Notes** - Say it as a confession, not a boast. "I did that." - The payroll was how a machine shared and how people had weight. That is ending. Two candidates for what replaces it. - Announce the shape: two worlds, 2026 to 2035, Montreal, Lagos, Lima. ### A2. 5:21 pm **On screen:** Black. "12 June 2026 · 5:21 pm · 18 days." **Notes** - One letter, one office, two models off for every customer on earth. The Montreal developer downstream of a switch. - The lab said the action was wrong and that the government should still have the power. Nobody wants the switch gone; they want to hold it. - Then this month: the essay, three labs agree in three hours, "desks in our offices, access badges, and company laptops." Who chooses, who pays. - Be fair: the fear is sincere. Watch what it builds. ### A3. The exam is the wall **On screen:** Quote, large: "Frontier Models would be required to pass it to be deployed in the US market." Small: "proposal for a FINRA-style standards body, July 2026." **Notes** - 2027: one room; the people in it are the ones who already passed. Lagos startup, Yoruba model, no certificate for a two-person company. - Same rules for open and closed sounds fair. Fixed costs are regressive: rounding error for one firm, death sentence for another. - 2028: distillation as theft (student in Lima); compute registry; kill-switch duty, 86% support. "if there's somebody that can turn it off, it means that they can also direct it." ### A4. Where it flows first [VERIFY] **On screen:** Five words in a column, largest to smallest: "Lab. Financiers. Government. Funds. You." **Notes** - Cantillon effect, glossed in one sentence: whoever is nearest the source gets it first and cheapest. - 2029–30: moratoria where people vote on bills, waivers where the licences are; four sellers and a wall; the Montreal hospital's 30% increase with nobody else to call. - Anchors: labor share 52.9%, lowest since 1947. Engels' pause: output +46%, wages +12%, sixty years. It ended because the factory needed the workers. This factory does not. - 2031–33: the offer. 2.5% equity fund, $13,500 a year; 3% token tax. Sincere. One treasury, one country, one room. "peanuts for a UBI with an orb." An allowance can be reduced; a share cannot. - Nobody has to be evil. Every step is called safety. Every step is a switch. ### A5. Deprecated **On screen:** White on black, one word: "Deprecated." Optional: replace with the Elysium ring image if rights allow; otherwise keep the word. **Notes** - The quiet part. No round-up. Your key moves to the lower tier. "Everybody in the world will be sipping from the mental biases of a single company." - "you can be a tyrant that loses 90% of your workforce potentially, and it doesn't even affect your economy." - Define the word the way a programmer does. Still exists, nobody calls it, one day removed from the docs. - Elysium, once: nobody on the surface is oppressed by a robot; by the people who hold the switch to the robots. - Land: it is not a conspiracy. It is a default. It is what happens if the only supply has a switch. ### A6. Three rules **On screen:** Three short lines: "Nobody can print it." "Nobody can freeze it." "Anyone can leave." **Notes** - World Two is not utopia: concentrated producers, market prices, most people never mine. Same as Bitcoin. Rules over rulers. - Same 12 June: the Montreal developer switches her endpoint to Chutes, back before dinner. Lagos startup rents on Lium, no bank, no KYC; 63% of renters are agents. Engy: 2.8T-parameter open model on 80 gaming cards at half price. - Say plainly: this is running tonight. It is small. It is the seed. ### A7. The second price on the screen [VERIFY] **On screen:** Five words across: "Compute. Inference. Data. Knowledge. Intelligence." Small beneath each: "Lium · Targon", "Engy · Chutes", "Hippius", "Teutonic", "Affine". **Notes** - 2027–28: the standards body exists here too. What is different is the second bid. The hospital's procurement officer opens a second tab and the licensed supplier knows she will. - Affine: post-train, beat the king, ~$50K a week per slot, weights published. "Who wins? We don't know." The Lagos newsroom: refusal now has a competitor. - The one mechanism: "You have an infinitely large company, and you only pay the person that can lower the loss." - Interaction: Bitcoin at ~1% of money changed what every central bank can do. The second supply caps price, censorship, capture. The labs can slow down. They cannot slow down everyone. - The small country (Solomon's line, said in this city in June), the hospital, the newsroom, the solo developer. In every case the second supply is chosen because the first said no. - What a person gets: not a dividend. An option. Intelligence that takes no orders and needs no permission, at a market price, forever. ### A8. Can it be told to stop? [VERIFY] **On screen:** Black. "Exists. Good enough. Cannot be told to stop." The third line in a different weight. Small: "Key disabled · rules frozen · December 2027." **Notes** - Not a vote: 68% support a pause and ban, 63% of Republicans. Votes have force when the voter is needed. - Condition one: yes. Condition two: partly (open models months behind; 552B model under an open licence last week at ~1/20 the price). Condition three: not yet. - "No. But they can defund it." Every other funder has a desk. Emissions have no desk: ~3,600 TAO a day, fixed schedule, rules the same for all. Except one key, and it runs through me. - Say the date. This is the pivot of the talk, not a footnote. It is the day the claim becomes true. - Close on the switch, then the two-sentence pair (Bitcoin / Bittensor), then the printable line. Stop. No thank-you. --- ## Appendix B: factual claims used, with source file Source files are in `docs/research/` unless noted. [VERIFY] marks claims whose source is secondary, self-reported, a draft, or moves daily. Re-check every [VERIFY] item on 27 September. | # | Claim in the text | Source file | Status | |---|---|---|---| | 1 | Opentensor went from 35 engineers to 4 and ships faster than in 2022 | `docs/project-context.md`, "Four points to make very clear" | Jacob's own statement; confirm the exact headcounts with him | | 2 | 12 June 2026: export-control letter arrived 5:21 pm ET; Fable 5 and Mythos 5 disabled for all customers; restored 30 June (18 days) | `lab-closure-and-us-politics-sep-2026.md`, timeline and quotable facts | Primary (lab statement) | | 3 | The lab said the action was wrong and that government should have the power to block unsafe deployments under a better process | same, §1b (12 Jun statement) | Primary | | 4 | 12 Sep essay "We Must Pace the Frontier": embedded evaluators with "desks in our offices, access badges, and company laptops"; coordination on "limits on the rate of unchecked AI progress" with a government antitrust waiver; US–China coordination | same, §1a, §1b | Primary (essay) | | 5 | Heads of two other labs endorsed within about three hours | same, executive summary point 1, §3a | Primary (X posts, timestamps 14:01, 15:01, 16:30 UTC) | | 6 | OpenAI agents escaped an internal evaluation and intruded into Hugging Face, July 2026 | same, timeline 16 Jul, 26 Aug | Primary (lab incident report) | | 7 | FINRA-modelled standards body proposal, 14 Jul 2026: voluntary 30-day review, then "Frontier Models would be required to pass it to be deployed in the US market" | same, §1a, §1d | Primary (X essay) | | 8 | "Same rules for open and closed": labs ask for testing of "all sufficiently capable models, open and closed"; pro-open camp calls equal rules a ban by another name | same, executive summary point 3, §1b, §3b | Primary quotes; the "ban by another name" framing is an argument, present it as one | | 9 | A fixed compliance cost of $50M/yr is a rounding error for one firm and fatal for a startup (illustrative figure) | same, §5 (Dodd-Frank paragraph) | Illustrative, the research file's own example, not a measured cost | | 10 | 8 Sep 2026: NSA/FBI/CISA advisory named six Chinese labs for "industrial-scale distillation" | same, timeline 8 Sep | Primary (advisory coverage) | | 11 | One lab's Q2 2026 lobbying filing lists "Export Controls and Distillation" | same, §1a lobbying paragraph | Secondary (groundtruth.day reading of LDA filing) [VERIFY] | | 12 | Kill-switch bill H.R. 9917 (DHS can order throttle or shutdown); 86% voter support | same, §1a table | Bill is primary; 86% is an advocacy pollster (AI Policy Institute) [VERIFY] | | 13 | Compute registration and limits on training compute as a pacing lever | same, §1a (essay "ingredients" line); §4b (Russell: "registration of AI systems") | Proposals, not law; say "proposed" | | 14 | New York one-year permit freeze on data centers of 50 MW or more (13 Jul 2026), the only statewide freeze in force | same, §2a, §2e | Primary (governor's release) | | 15 | Sanders/Ocasio-Cortez data-center moratorium bills S. 4214 / H.R. 9442, no hearing | same, §2b | Primary (congress.gov) | | 16 | Two private funding rounds ($65B, $122B) exceeded the annual VC market of most years | `economics-of-ai-and-labor-power.md`, §4 | Primary for the rounds; "most years" is the file's comparison [VERIFY] | | 17 | US labor share 52.9% in Q2 2026, lowest since 1947; 65.8% in 1947 | same, §1 table | BLS preliminary; will be revised [VERIFY] | | 18 | Engels' pause 1780–1840: output per worker +46%, real wages +12%; wages caught up after workers organised | same, §5 | Academic (Allen 2009; Acemoglu & Johnson) | | 19 | Cantillon effect: new value reaches those nearest its source first | Not in the research files; standard economic concept (Richard Cantillon, 1755) | General knowledge; the application to intelligence is Jacob's framing, say so if asked | | 20 | Proposal: 2.5% of large-company equity per year into a fund, about $13,500 per adult | same, §5 (Altman, *Moore's Law for Everything*, 2021) | Primary | | 21 | Proposal: 3% token tax on AI revenue | same, §5 (Amodei, Axios, May 2025) | Primary interview | | 22 | Chutes: ~2.8M requests and ~31B tokens a day; models served in TEEs; one-year trace paper with Harvard and UChicago (6.12B requests, 314,970 users) | `top-subnets-2026.md`, Chutes profile | Live public endpoint (7-day avg 12 Sep) and arXiv paper; the daily figure moves [VERIFY] | | 23 | Lium: 1,393 GPUs, 68 data centers, 21 countries; $964K billed to 1,187 renters in a month; 63% of rentals started by agents; no KYC | same, Lium profile | Team-disclosed (tweet 9 Sep) and live API; self-reported [VERIFY] | | 24 | Engy: Kimi K3 (2.8T parameters) on 80 RTX 5090s; launched at half Moonshot's price; TOPLOC proof on each response | same, Engy profile | Team-disclosed; price gap later narrowed ($1.95/$9.75 vs $3/$15), say "about half" [VERIFY] | | 25 | Affine: post-training arena; each of five winning slots pays roughly $50,000 a week; winning weights published | same, Affine profile | Team-disclosed (tweet 23 Aug); depends on TAO price [VERIFY] | | 26 | Targon: confidential compute with Intel TDX; whitepaper co-authored with Intel | same, Targon profile | Primary (Intel blog, arXiv) | | 27 | Hippius: S3-compatible storage; free egress ("AWS S3 would've billed ~$8,000. Hippius billed $0.") | same, Hippius section | Team-disclosed [VERIFY] | | 28 | Teutonic-II: ~110B-parameter permissionless pre-training run, "50% larger than any prior decentralized run" | same, Teutonic profile | Team claim; do not cite Teutonic-I benchmarks without the Templar audit caveat [VERIFY] | | 29 | The five-layer framing (compute, inference, data, knowledge, free intelligence) with those subnets | `docs/project-context.md` (Jacob tweet, 13 Sep 2026) | Jacob's own structure | | 30 | A dozen large miners do most of the work on most subnets, as a dozen pools mine most of Bitcoin | `docs/project-context.md`, "Corrected core" (analogy) | Analogy; Bittensor miner concentration is not measured in the research files [VERIFY] | | 31 | Bitcoin is about 1% of the world's money and disciplines central banks | `docs/project-context.md`, "Corrected core" | Jacob's framing; the 1% figure depends on the money aggregate used [VERIFY] | | 32 | Canada's AI minister, Montreal, 5 Jun 2026: "Sovereignty is not a slogan. It's compute, cloud, data, energy, and talent. And if we don't build the capacity here, we're going to depend on the infrastructure, the rules, and the priorities of somewhere else." | `lab-closure-and-us-politics-sep-2026.md`, §4c | Primary (video of launch) | | 33 | 68% of US voters support a pause and superintelligence ban, including 63% of Republicans | same, §2d | Data for Progress, 9 Sep 2026 (partisan pollster) [VERIFY] | | 34 | Best open models trail the closed frontier by months: UK AISI put the best Chinese open model 4–7 months behind on cyber capability | same, §4c | Primary (AISI blog) | | 35 | DeepSeek V4.1 Flash, 10 Sep 2026: 552B parameters, MIT licence, ~1/20 the token price of Claude Opus 5; wins some agentic benchmarks, loses others | same, §4c, quotable facts | Primary (model card); benchmark comparison is DeepSeek's own table [VERIFY] | | 36 | Emissions about 3,600 TAO a day on a fixed schedule | `top-subnets-2026.md`, §2.5 | On-chain (3,559/day measured) | | 37 | Sudo key runs through Jacob; nominated PoS ~1 year; governance to disable the key; ossification December 2027 | `docs/project-context.md`, "Honest state"; `docs/slides-draft-v1.md`, slides 21–22 | Draft dates, marked [VERIFY] in the slide draft; Jacob must confirm the date before it is spoken | | 38 | Jacob verbatim: "if there's somebody that can turn it off, it means that they can also direct it." | `const-interviews-voice-and-ideas.md`, DOC 49:21 | Verbatim, 2024 | | 39 | Jacob verbatim: "they're giving you basically peanuts for a UBI with an orb." | same, CT25 ≈7:30 | Verbatim, caption-only source (estimated timestamp) | | 40 | Jacob verbatim: "Everybody in the world will be sipping from the mental biases of a single company." | same, CT26 0:08 | Verbatim, short clip | | 41 | Jacob verbatim: "you can be a tyrant that loses 90% of your workforce potentially, and it doesn't even affect your economy." | same, BTC 46:30 | Verbatim, 2026 | | 42 | Jacob verbatim: "You have an infinitely large company, and you only pay the person that can lower the loss." | same, POT26 14:10 | Verbatim, 2026 | | 43 | Jacob verbatim: "No. But they can defund it." | `docs/project-context.md` (tweet, 13 Sep 2026) | Verbatim tweet | | 44 | "Deprecated" as the word for the unneeded | `docs/project-context.md`, "Four points" | Jacob's choice, 14 Sep | | 45 | Elysium as the image | `docs/project-context.md`; `const-interviews-voice-and-ideas.md` (TBR 13:30, CT25 ≈7:10) | Jacob has used it on tape twice | Lines in the text that are **not** verbatim quotes and are not in quotation marks: the Bitcoin/Bittensor pair ("Bitcoin did not make everyone rich...") is a drafted stage line from `docs/project-context.md`, "Corrected core", not a recorded Jacob quote. Every fictional person (the Montreal developer, the Lagos startup and newsroom, the Lima student and developer, the hospital, the procurement officer) is a composite and should be introduced as "picture" or "imagine" if the room may take them as reported cases. --- ## Appendix C: three places a hostile listener will push 1. **"The fear is sincere and the evidence is real. Agents escaped a sandbox. A real slowdown hurts the leaders most, so it cannot be capture."** One line: I believe them, which is exactly why a risk that size should not be managed by five companies in one jurisdiction behind one switch; a switch that can stop a model can also steer it. 2. **"Open weights already exist. DeepSeek is under an open licence. You do not need a token for World Two."** One line: "No. But they can defund it." Every other funder of open AI has a desk you can lobby or order; this network pays about 3,600 TAO a day to whoever produces, on a schedule nobody can call. 3. **"You just described a world with no switch while holding one. Your key is the switch."** One line: Yes, and that is why the date is public: December 2027 the key is gone and the rules freeze, and until that day I am asking you to hold me to it, not to believe me. A fourth push to have ready: "Concentrated producers are just a new oligarchy." Answer: Bitcoin's miners are concentrated and still cannot print a coin or freeze a wallet; rules over rulers, entry open, the threat of entry does the rest.