# Technique example slides: one slide per move

Exploit Conference, Montreal, 28 September 2026. 45 minutes. Speaker: Jacob Steeves (Const).

Eight slides, one for each technique in `docs/research/model-keynotes.md`. Each slide is built to drop into `docs/slides-draft-v2.md` in the format of v1: one `##` per slide, **On screen** in one line, **Notes** as beats, not lines.

Design: Bittensor theme. White background, black text, or the inverse for emphasis slides. One line or one image per slide. No bullet lists on screen.

How to read each section:
- **Technique**: the move and its one-line definition.
- **Model**: the talk it comes from and the verbatim quote. Timestamps are from the transcripts in `docs/sources/talks/`; the time is the position in the recording named in each transcript's header.
- **Slide**: title, what is on screen, speaker notes in Jacob's voice.
- **Where it sits**: the minute in the 45-minute shape proposed in `model-keynotes.md`, and the v1 slide it replaces or joins.
- **Why it works here**: one sentence about this room.

[VERIFY] marks a number from a secondary source or one that moves daily. Sources: `docs/research/top-subnets-2026.md`, `economics-of-ai-and-labor-power.md`, `lab-closure-and-us-politics-sep-2026.md`, `state-of-bittensor-2026.md`, `const-tweet-themes.md`. Tweet lines are verbatim from the archive; dates are given so they can be pulled up on stage.

---

## Technique 1. Announce the shape in one breath, then keep the promise

**Definition:** Tell the room the map of the talk in one sentence in the first two minutes, then follow it exactly. The room relaxes when it knows the shape.

**Model:** Steve Jobs, Stanford commencement, 2005 (`jobs-stanford-commencement-2005.md`).

> [00:45] Today, I want to tell you three stories from my life. [00:50] That's it, no big deal, just three stories.

Second model: Bret Victor maps four ideas at [05:26] ("The four ideas I want to talk about today") and recaps them at 27:48.

### Slide: Three things

**On screen:** White. Three short lines, stacked, black: "What is happening." "What the fear is." "What we built."

**Notes**
- Three things tonight. What is actually happening. What the real fear is. What we built. That is it.
- Say what I will not do: I will not list 128 subnets. I will not tell you why we are better than a lab. If you came for that, you can leave now and read a dashboard.
- If you want it in my own words: "Software is eating the world. AI is eating software. Incentives will devour AI." (1 Mar 2025.) Same map, nine words. Part one is the first line, part two is the second, part three is the third.
- Come back to this slide twice. At about minute 10: "That was the first thing." At about minute 21: "Second thing done. Now the thing we built." The room should be able to say where we are without looking at a clock.
- If time runs short, this slide tells the room what got cut. A named gap is better than a mystery.
- Do not read the three lines. Look at the room and say them.

**Where it sits:** 0:00–2:00, "Open in the room." After the feeling and before the exit is removed. In v2: a new slide between v1 slide 2 ("The feeling") and v1 slide 3 ("The builders are telling you").

**Why it works here:** Holders, builders, AI people, and the Montreal crowd each came for a different part; naming three parts tells every group where its part is, so nobody checks a phone while waiting for it.

---

## Technique 2. Concede the weakness early and pair it with a date

**Definition:** Say the thing you are worst at before anyone else can, and in the same breath give the date by which it changes. A concession without a date is an apology; with a date it is a plan.

**Model:** John F. Kennedy, Rice University, 1962 (`kennedy-rice-moon-speech-1962.md`).

> [11:39] To be sure, to be sure we are behind and will be behind for some time in manned flight, but we do not intend to stay behind and in this decade we shall make up and move ahead.

Second model: Ilya Sutskever reads his own 2014 slides and grades them, NeurIPS 2024, [27:42]–[27:51] "Was it wise to pipeline? As we now know, pipelining is not wise. But we were not as wise back then."

### Slide: My own slide from 2024

**On screen:** Black. White text. Jacob's tweet of 22 Sep 2024, set plain, no screenshot chrome: "The chain operates on a Proof of Authority (PoA) mechanism, with plans to transition to Proof of Stake (PoS) by the end of next year." Beneath it, smaller: "Missed. New date: December 2027."

**Notes**
- Read it out loud. My words, September 2024. "End of next year" meant the end of 2025. It did not happen. I am not going to pretend it did, and I am not going to wait for the first question to bring it up.
- Why it slipped: we rebuilt the economics first. Dynamic TAO in February 2025. Three emission models in sixteen months. I chose to fix what pays before I fixed who governs. Right call then. Wrong call now.
- Where we are: the chain is proof of authority. One privileged key can upgrade or pause it, and it runs through me. In July 2024 we used it to halt the chain about 35 minutes into a wallet drain of roughly 32,000 TAO, about $8 million at the time. That saved people money. The same key could be used badly. [VERIFY the TAO figure; secondary]
- I said this in June 2024 and I will say it here: "we cannot pretend for a second that at our level of decentralized immutability we are comparable to Bitcoin." (19 Jun 2024.) We are behind Bitcoin on this and we will be for some time.
- We do not intend to stay behind. Nominated proof of stake, where holders choose who makes blocks, within about a year. [VERIFY the current target month] The governance pallet that replaces the key, proposed May 2026, three phases, the last one a vote to switch sudo off. [VERIFY status] Full decentralization and then ossification around December 2027. Screenshot this slide. Hold me to it.
- "The goal of power is to give it away. The art of power is its decentralization." (18 Nov 2023.) That was three years ago. This is the date attached to it.

**Where it sits:** 34:00–39:00, "I hold a key." Right after the plain confession (v1 slide 21) and in place of v1 slide 22 ("The path, with dates"), so the dates arrive on the same slide as the missed date.

**Why it works here:** Crypto holders have watched founders promise decentralization for a decade; putting your own missed date on the screen before they can quote it is the only thing that gives the new date any weight.

---

## Technique 3. Remove the option of opting out before you ask for anything

**Definition:** State early that the thing is going to happen with or without the room. After that, every choice you offer is a choice between versions, not a choice about whether.

**Model:** John F. Kennedy, Rice University, 1962 (`kennedy-rice-moon-speech-1962.md`).

> [04:42] The exploration of space will go ahead, whether we join in it or not, and it is one of the great adventures of all time

Second model: Ilya Sutskever, NeurIPS 2024, [30:28] "But pre-training, as we know it, will unquestionably end. Pre-training will end. Why will it end?" (Session-audio timestamp; subtract about 22:20 for the standalone video.)

### Slide: Whether we join or not

**On screen:** White. Quote, large: "AI isn't going away. Barring unprecedented global government coordination, it isn't even slowing down. The only real decision is whether to centralize power, or disperse it." Small attribution: Naval, 12 September 2026.

**Notes**
- Every AI talk right now carries a hidden option: maybe it stops. It will not. On 12 September the heads of three frontier labs agreed, in public, to slow down together and asked for an antitrust waiver to do it. Even the slowdown is a deal among a few companies. Nobody proposed stopping. Nobody can.
- Naval wrote this the same day. Not me. A man with no stake in this network said the only decision left is centralize or disperse.
- So I am not going to argue tonight about whether AI should exist. That decision has been made, by physics and by capital. The decision left is who owns it.
- Elon, 13 September: "Nothing can shut down open source." My reply: "No. But they can defund it." Opting out is not on the menu. Being starved is. That is the whole reason this network exists.
- Sixty-eight percent of American voters told a pollster on 9 September they support a pause on advanced AI. [VERIFY, Data for Progress] Fear is the median position. Fine. The question is whether the answer to that fear is five companies with a federal licence, or a network nobody can licence.
- From this slide on, when you hear me say "AI" and think "should we," replace it with "who."

**Where it sits:** 2:00–4:00, "Remove the exit." In v2: a new slide directly after "Three things" and before v1 slide 3 (the Amodei quote), so the first outside voice the room hears is the one that closes the exit.

**Why it works here:** The AI people in this room have spent the week arguing about pacing the frontier versus regulatory capture; moving the question from "whether" to "who" stops both camps arguing with you and starts them listening.

---

## Technique 4. Make the enemy a pattern or a past self, not a person in the room

**Definition:** Name a structure, a recurring type, or your own past mistakes as the villain. Never name a rival company or a person who might be in the audience.

**Model:** Yuval Noah Harari, TED 2018 (`harari-why-fascism-is-so-tempting-ted-2018.md`).

> [08:29] And dictatorship now means that too much data is being concentrated in the hands of the government or of a small elite.

Second model: Bret Victor makes a dead man the villain, [04:15] "And von Neumann was furious at him." Third: Andreas Antonopoulos voices the type instead of naming it, [01:10] "See? See? It's not working. It's slow. It doesn't work so well."

### Slide: Too much intelligence in too few hands

**On screen:** Black. White text, one line: "Too much intelligence in too few hands." Small beneath: after Harari, 2018.

**Notes**
- I am not going to say a company's name tonight as an enemy. Not out of politeness. The company is not the problem. The pattern is.
- Harari, 2018: land, then machines, then data, each time "concentrated in the hands of the government or of a small elite." Put the word intelligence in that sentence. That is what I am afraid of.
- My version, October 2025: "The real risk of AI was never robots with a vendetta. It's that all the capital ownership concentrates down to a small elite set of technocrats in SF." Notice I did not say a name then either.
- The pattern has a name in economics. Stigler, 1971: "as a rule, regulation is acquired by the industry and is designed and operated primarily for its benefit." The railroads backed the first federal regulator in 1887. The phone company took federal oversight in 1913 and got a protected monopoly for seventy years. Say "incumbents often welcome the rules." Do not say "always write them." [VERIFY the railroad framing; historians dispute Kolko]
- Three years ago this room could not open a bank account. Twenty-three banks got letters from the FDIC telling them to "pause" anything touching crypto. We found out through a lawsuit. Same pattern. Different asset. [VERIFY count, Coinbase FOIA]
- What I want you to hear: the people inside those labs are sincere. The fear is real. The pattern does not need villains. It needs a state and a few firms holding the same asset. Call that state-corporate capture. Do not call it fascism unless you spend ninety seconds defining it first.
- The enemy in this room, if there is one, is my own past self: the founder who chose speed over decentralization and is still holding a key. I get to that at minute 34.

**Where it sits:** 10:00–14:00, "The real fear, defined." In v2: directly before v1 slide 9 (Elysium), so the definition comes first and the picture second.

**Why it works here:** There are lab employees, press, holders, and Montreal researchers in the same seats; a pattern with 140 years of examples lets all of them agree without having to attack or defend an employer.

---

## Technique 5. Teach the pattern three times before applying it to your thing

**Definition:** Show the same shape in three cases the room already accepts. Then show your case. The audience predicts the fourth, so it feels like their conclusion, not your pitch.

**Model:** Andreas Antonopoulos, Zurich 2016 (`antonopoulos-infrastructure-inversion-2016.md`).

> [02:53] The ones who succeed are the ones who continue, even though the rest of society tells them they are crazy, [03:00] to pursue this crazy idea, automobiles, electrification, the internet, and bitcoin. [03:10] Every time, crazy pioneers who were made fun of by everybody else in society persisted until everybody could see what they were doing was correct.

Then, after three worked examples: [21:12] "Then we get to Bitcoin." [21:40] "What we're doing is we're riding the automobile on the muddy roads of banking."

### Slide: 2011. 2016. 2026.

**On screen:** White. Three dates across the slide, one short phrase under each: "2011 · a joke" "2016 · for criminals" "2026 · the power of Poland, and an ETF."

**Notes**
- Do not say "we are early." Show a curve and let the room place us on it.
- Bitcoin. 2011: pizza money. A joke. 2016: "for criminals." That was the year Andreas stood in Zurich and said Bitcoin was a supercar on the muddy roads of banking. 2026: about a thousand exahashes a second, the electricity of Poland, a spot ETF. [VERIFY, Hashrate Index Q2 2026; Cambridge CBECI]
- Open source. 1991: a hobby. 2001: a CEO called it a cancer. Today it runs nearly every server on Earth. [VERIFY the 2001 quote and the server-share claim]
- Open weights. 10 September, four days before I wrote these notes: a 552-billion-parameter model shipped under an MIT licence at about one twentieth the token price of the leading closed model. You cannot pull the ladder up when the ladder is a torrent file. [VERIFY DeepSeek V4.1 Flash figures]
- Now let them say the fourth case before I do. A market for intelligence, five years old. Market cap two to three billion against five trillion. A founder key. Emissions bigger than revenue. This is what a market for intelligence looks like while it still runs on the labs' roads. "There is no way to compete with fiat-money tank engines in the center lane." (31 Dec 2025.) So we are not in the center lane.
- Predict the sequence, not the date: ignore, dismiss, buy from it, publish to it. Step three has started. Chutes is on OpenRouter. Grayscale and Bitwise have filed. I am not saying when step four comes. I am saying that it will.

**Where it sits:** 14:00–18:00, "Bitcoin: the proof." In v2: directly before v1 slide 12 ("Bitcoin's second invention"), so the pattern is taught before the mechanism is explained.

**Why it works here:** The crypto holders in this room lived through 2011 and 2016 and will finish the pattern themselves, which turns "Bittensor is early" from a founder's plea into the room's own conclusion.

---

## Technique 6. One metaphor, carried to the end

**Definition:** Choose one image for the danger and one for the cure. Return to each at least three times. Do not stack five images for the same idea.

**Model:** Martin Luther King Jr., 1963 (`king-i-have-a-dream-1963.md`).

> [03:02] Instead of honoring this sacred obligation, America has given the Negro people a bad check, a check which has come back marked insufficient funds. [...] [03:40] And so we've come to cash this check, a check that will give us upon demand the riches of freedom and the security of justice.

The check is set at 02:08, cashed at 03:40, and never replaced by another image for the same idea.

### Slide: A payroll with no employer

**On screen:** Black. White text, one line: "A payroll with no employer."

**Notes**
- The metaphor for the cure is payroll. It was set at minute 4 and this is where it gets cashed. Do not introduce anything new here.
- Minute 4, first pass: for a hundred years a frontier company shared its gains through payroll. GM, 1979, 853,000 people. [VERIFY, secondary] The lab today is worth about a trillion with a few thousand people and has no payroll to share through. "You will work for the machines that work for the elites — and you'll own none of it." (31 May 2025.)
- Minute 21, second pass: Bittensor is a payroll with no employer. About 3,600 TAO a day, paid to whoever produces, on a schedule nobody can change. "Dont ask permission. No HR department. Just bring compute and your mind." (22 Nov 2023.)
- Now, third pass, cash it: Bitcoin proved a network with no owner can outgrow any company. Bittensor is that network, pointed at intelligence, paying ownership to anyone who helps build it. It is the machine that pays you back.
- The metaphor for the danger is Elysium. Once, at minute 12, as a picture. Not Gaza. Not feudalism. Not fascism. Not corporatism. Two images for the whole night. If a third one comes to mind on stage, do not say it.
- Stop after the line. The next slide is the τ glyph and the URL. No thank-you slide.

**Where it sits:** 43:00–45:00, "Close." In v2: replaces v1 slide 25 ("Aim the fear correctly"), with one line added to v1 slide 5 (payroll, first pass) and one to v1 slide 15 (payroll, second pass) so the image appears three times.

**Why it works here:** Holders will post one sentence from this talk tonight; if the danger and the cure each have a single image, the sentence they post is yours and not a list of villains.

---

## Technique 7. Anchor every number in the room

**Definition:** Give each big number one physical comparison the audience can check in its head, and where you can, make the comparison local. A number without an anchor is forgotten by the next slide.

**Model:** John F. Kennedy, Rice University, 1962 (`kennedy-rice-moon-speech-1962.md`).

> [14:07] That budget now stands at $5,400 million [14:13] a year, a staggering sum, though somewhat less than we pay for cigarettes and cigars [14:21] every year.

Also [10:32] "as long as two lengths of this field" and [11:16] "dropping it in this stadium between the 40-yard lines."

### Slide: Four hours

**On screen:** White. Large: "4 hours." Small beneath: "What the four largest cloud companies spend on data centers in the time it takes this network to pay out a year of TAO."

**Notes**
- Every number tonight gets one anchor. One. Where I can, the anchor is in this city.
- The network pays out about 3,600 TAO a day. At today's price that is about $300 million a year. [VERIFY on the day; $234/TAO on 12 Sep] The four largest cloud companies guide $720 to $745 billion of data-center spend this year. [VERIFY, earnings calls] That is about $2 billion a day. They spend our entire annual issuance in about four hours. That is the gap. Now you know it and I do not have to hide it.
- Bitcoin draws about as much electricity as Poland, 155 to 180 terawatt-hours a year. [VERIFY, Cambridge] In this room: close to what Hydro-Québec generates in a year. [VERIFY, Hydro-Québec annual output] Nobody planned it. Incentives did.
- The lab: about 2,300 people at the end of last year. [VERIFY, Fortune, Jun 2026] The Bell Centre holds about 21,000. [VERIFY] The whole company would fill one seat in ten. GM in 1979: 853,000 people, close to half the people living on the island of Montreal. [VERIFY both]
- Revenue: subnets bill outside customers for somewhere between $15 and $35 million a year depending on who counts. [VERIFY; $28–35M is a consultancy figure, $17–18M is checkable on-chain] That is about a tenth of what the protocol pays out. Nine months ago I said no subnet had serious revenue. I was right. Today three or four do.
- One anchor you can check right now: Lium billed $964,000 in a month to 1,187 renters and bought back and burned a million dollars of its token with the proceeds in one transaction, 9 September. Extrinsic 9032919-0005. Look it up on your phone while I talk. [Team-disclosed; the burn is on-chain]
- Rule for the whole talk: one anchor per number, said once. If I hear myself give a second comparison, stop.

**Where it sits:** The payroll anchors belong at 4:00–10:00 ("The firm without a payroll," v1 slides 5 and 6). The example slide itself sits at about 33:00, the end of "What has been built," directly before the scale admission at the bottom of v1 slide 20 ("Proof of mechanism, not yet a counterweight").

**Why it works here:** AI people think in capex and crypto people think in emissions, and the two groups have no shared unit; hours, seats, and a river the whole room drove past are the only units everyone present can check in their heads.

---

## Technique 8. Split the ask by audience and land on a printable sentence

**Definition:** Address each group in the room by name and give each one a single action. Then end on one short sentence that could be printed on a wall, and say it more than once.

**Model:** Yuval Noah Harari, TED 2018 (`harari-why-fascism-is-so-tempting-ted-2018.md`).

> [11:44] If you are an engineer, [11:46] then find ways to prevent too much data [11:50] from being concentrated in too few hands. [11:53] and find ways to make sure the distributed data processing [11:59] is at least as efficient as centralized data processing. [12:04] This will be the best safeguard for democracy. [12:07] As for the rest of us who are not engineers, [12:11] the number one question facing us [12:14] is how not to allow ourselves to be manipulated [12:19] by those who control the data.

For the printable sentence: Steve Jobs, Stanford 2005, [14:15] "Stay hungry, stay foolish. And I've always wished that for myself. And now, as [14:24] you graduate to begin anew, I wish that for you. Stay hungry, stay foolish."

### Slide: If you hold. If you build. If you are new.

**On screen:** White. Three lines, spaced: "If you hold." "If you build." "If you are new."

**Notes**
- If you hold TAO: you own a piece of the only AI capital on Earth that was never allocated to anyone before launch. Say it that way. About 600,000 TAO were mined in the open and later sold to funds at about $18; nobody was handed anything before block one. [VERIFY figures] Your validator now curates a basket of subnets. Watch what they curate. Stake where the allocation is good. Hold through the year it takes to remove the key.
- Ask my own side for discipline before I ask anything of anyone else: do not chase emissions, do not dump on the newcomers in this room. "nothing free, nothing given, everything fought for." (26 Mar 2026.)
- If you build: about 128 markets, and there is room for yours. [VERIFY count] Registration cost is falling toward the price of a transaction. Fourteen teams now buy back their own token with customer money; those are the ones that will still be here in two years. Build for a customer, not for emissions. "Don't submit a resume. Submit a model." (29 Jan 2024.)
- If you are an engineer, and this is Montreal so many of you are: Harari asked in 2018 for distributed processing at least as efficient as centralized. That is a spec, not a wish. One subnet runs a 2.8-trillion-parameter open model on 80 gaming cards at half the maker's price. [VERIFY, team-disclosed, Engy SN53] Come make it faster.
- If you are new: "You don't need no visa to mine a permissionless network." (21 Sep 2025.) There is no gate. Come and see if it is real.
- Then the sentence. Say it once here, slowly: "Everyone else funds open AI until someone tells them to stop. We built the one that nobody can tell." Say it again at the very end. Then stop talking.

**Where it sits:** 39:00–43:00, "The ask, three audiences." In v2: replaces v1 slide 24 ("Hold. Build. Come."). The printable sentence returns on the final slide at 44:30.

**Why it works here:** Four audiences share one room, and a sentence addressed to "you" only lands after each group has heard its own name; the printable line is what gets posted from the seats to X while you are still on stage.

---

## Drop-in map for v2

| Technique | Slide title | Minute | Replaces or joins in v1 |
|---|---|---|---|
| 1. Announce the shape | Three things | 0:00–2:00 | New, between slides 2 and 3 |
| 3. Remove the exit | Whether we join or not | 2:00–4:00 | New, directly after "Three things" |
| 7. Anchor the numbers | Four hours | ~33:00 (anchors also feed slides 5–6) | Before the scale admission on slide 20 |
| 4. Enemy as pattern | Too much intelligence in too few hands | 10:00–14:00 | New, before slide 9 (Elysium) |
| 5. Pattern three times | 2011. 2016. 2026. | 14:00–18:00 | New, before slide 12 |
| 2. Concede with a date | My own slide from 2024 | 34:00–39:00 | Replaces slide 22 |
| 8. Split the ask | If you hold. If you build. If you are new. | 39:00–43:00 | Replaces slide 24 |
| 6. One metaphor | A payroll with no employer | 43:00–45:00 | Replaces slide 25; add one line each to slides 5 and 15 |

Net effect on the slide count: five new slides, three replacements. If v2 keeps the fifteen-slide budget from `model-keynotes.md`, the candidates to cut are v1 slides 13 (the hash string), 14 (nature), and 18 (subnet count), whose content moves into speaker notes on neighbouring slides.

## Production notes

- Every [VERIFY] above re-checks on 27 September against the research files named at the top. The TAO price, market cap, subnet count, and the four-hour arithmetic move daily.
- Montreal anchors (Hydro-Québec output, Bell Centre capacity, island population) are from general knowledge, not from the research files. Confirm each before it goes on a slide or in Jacob's mouth.
- The Kennedy transcript reads "To be sure, to be sure we are behind"; the repeat is in the recording. Quote it either way.
- The Sutskever timestamps are from the full session audio. Subtract about 22:20 for the standalone video.
- Technique 2's slide reproduces a real tweet. Set it as plain text in the deck typeface, not as a screenshot, so it matches the theme.
- On-screen word count for the eight slides together is under 120 words.
