# Incentives, not a conspiracy: the chart for the cabal section

For Jacob Steeves' keynote, Exploit Conference, Montreal, 28 September 2026. Built 24 September 2026. Live as slide 9 of the deck at https://bittensor-talk.pages.dev/#/9 (branch `slides-site`, `content/data.json → incentives`). The PNG is `media/incentives-web.png` (1600×900). The six posts it was built from, verbatim, and all 46 claims graded, are in `docs/sources/tweets/cabal-incentives-posts.md`.

Jacob's framing, which the slide obeys: "I will say not that I think there is a conspiracy but that there are incentives for a conspiracy."

Terms used once:

- **Edge**: one line on the chart, from one name to another, carrying one fact.
- **VERIFIED**: the fact was read in a primary source: an IRS filing, a company announcement, a grant-maker's own page, or the person's own words in a transcript. Drawn solid.
- **REPORTED**: a named outlet says it; no primary document was read. Drawn dashed and grey.
- **CLAIMED**: only a post says it. Not drawn.
- **990-PF**: the yearly tax return a US private foundation files; Part XV lists every grant paid, by name and amount.
- **Form 990**: the yearly return a US public charity files; Schedule I lists grants it gave, Schedule N what it spun off.

---

## 1. Design note: which chart, and why

Three forms were considered.

**(a) A network of verified links.** Chosen. Eight names in three columns (the labs, the money, the referees and the witness), eleven lines, each labelled with the document's fact: an amount and a year, or a role and a date. Solid for VERIFIED, dashed for REPORTED, nothing for CLAIMED.

**(b) A 2026 timeline.** Rejected. Slide 8 ("5:21 pm") is already a timeline of the same summer, and a timeline cannot carry money or relationships; it shows when, not who paid whom.

**(c) A "who benefits" table** (proposal → who wrote it → who is exempt → who is gated). Rejected. The two columns that make it bite, "exempt" and "gated", are inferences: no proposal text names who is exempt, and "gated" is a reading of the FINRA-style exam and the Senate draft, which has no public text. A hostile listener would call the table an opinion in a grid, and they would be right about two of its four columns.

**Why the network is the one a hostile, informed listener cannot dismiss.** The usual way to dismiss a network chart is "red string on a corkboard": guilt by association, names joined by vibes. This chart removes that move by construction:

1. Every solid line is a document the listener can pull: a 990-PF line item, an ARC Form 990 schedule, an Anthropic funding announcement, Open Philanthropy's own 2017 grant page, a Stratechery transcript, SFF's recommendation tables, METR's funding page, Amodei's essay, Anthropic's 9 September post, Coefficient's 9 September post. The legend on the slide says so.
2. The chart states the strongest counter-fact itself, on the METR node: "no lab money · free tokens". A listener who arrives ready to say "but METR takes no lab money" finds it already conceded on screen.
3. The one thing that is only reported (the donated stake) is visibly weaker: dashed, grey, with the outlet's name in the label.
4. Nothing that the posts allege and cannot show is drawn: no "$7.7B", no "payroll", no "cartel", no "doom machine", no "nepo".
5. It is laid out as a ledger, not a web: labs left, money centre, referees and witness right; the money column touches both sides, which is the whole point, and only two lines cross the middle (the labs choosing and naming the referee).

The slide therefore makes exactly one claim, and it is Jacob's: the same few names sit in every seat. Investor and board observer of the lab; funder of the referee's parent, of the referee's subcontractor and of the witness; the lab writes the rule that names the referee; the lab hires the referee to grade the lab. No line says anyone did anything wrong. That is the point of the caption.

**Deck conventions kept.** Black on white, one grey (`#8e8e8e`), hairline boxes, mono labels, one grey title above, one legend line, one annotation sentence beneath ("Incentives, not a conspiracy."). Every number comes from `content/data.json` with a `source` field on each edge. Two things bend the deck's chart rules on purpose: eleven labels instead of two, because an unlabelled line on this chart would be an allegation, and a two-line legend, because the solid/dashed rule is what makes the chart defensible.

**Rejected inside option (a).** Tarbell Center (journalism funding) and Encode (which helped Coxon announce) were cut as nodes: verified, but peripheral to Jacob's sentence, and each added two crossings. Public First Action ($20M + $20M from Anthropic) was cut for the same reason; it is in the notes. Google DeepMind and the FINRA-style standards body were cut: no money edge, and the essay already carries the "rule" role. The Senate bill is REPORTED only and has no text; it stays spoken.

**Not fixed by design, said in the notes.** Two crossings remain (OpenAI to METR crosses the Moskovitz–Anthropic line; nothing else). Phone view (390×844) fits but the labels are too small to read; the figure is for a projector.

---

## 2. Edge list, with grades and sources

Nodes: Anthropic ("wrote the essay, 12 Sep"), OpenAI ("ran ExploitGym, July"), Survival & Flourishing Fund (Jaan Tallinn), Good Ventures / Coefficient Giving (Dustin Moskovitz), METR ("no lab money · free tokens"), Redwood Research ("inside the investigation"), Jacob Coxon ("resigned 9 Sep 2026").

| # | From → To | Label on the chart | Grade | What the source says | Source (read 24 Sep 2026) |
|---|---|---|---|---|---|
| 1 | Good Ventures → Anthropic | Series A 2021 · board observer | VERIFIED | "The round included participation from James McClave, Dustin Moskovitz, the Center for Emerging Risk Research, Eric Schmidt, and others." $124M Series A, 28 May 2021. Moskovitz, 20 Oct 2025: "I'm a board observer at Anthropic and we are early donors to OpenAI". | anthropic.com/news/anthropic-raises-124-million-to-build-more-reliable-general-ai-systems; stratechery.com interview, Wayback capture 20251020103508 |
| 2 | Anthropic → Good Ventures (dashed) | stake in a nonprofit · Forbes 2025 | REPORTED | "Their Anthropic stake (worth an estimated $500 million) was moved into a nonprofit vehicle in early 2025 so they could invest any 'significant financial return' back into philanthropy and 'dispel any perception of conflict of interest,' she [Cari Tuna] says." The vehicle is not named; Good Ventures Foundation's 990-PFs to June 2025 list no Anthropic stock. | Forbes, Phoebe Liu, 7 Nov 2025 (forbes.com.au reprint read) |
| 3 | Good Ventures → OpenAI | $30M 2017 · board seat | VERIFIED | "The Open Philanthropy Project recommended a grant of $30 million ($10 million per year for 3 years) in general support to OpenAI... in which Holden Karnofsky... will join OpenAI's Board of Directors". Same page: "OpenAI researchers Dario Amodei and Paul Christiano are both technical advisors to Open Philanthropy and live in the same house as Holden. In addition, Holden is engaged to Dario's sister Daniela." | openphilanthropy.org/grants/openai-general-support/ (Mar 2017), Wayback capture |
| 4 | SFF → Anthropic | led Series A 2021 | VERIFIED | "The Series A round was led by Jaan Tallinn, technology investor and co-founder of Skype." | anthropic.com, 28 May 2021 |
| 5 | SFF → METR | $752K 2024–25 | VERIFIED | SFF 2024: Model Evaluation and Threat Research, $204,000, funder Jaan Tallinn. SFF 2025: METR $548,000 plus a $428,000 matching pledge. Recommendations, not audited payments. | survivalandflourishing.fund/sff-2024-recommendations; survivalandflourishing.fund/2025/recommendations |
| 6 | Good Ventures → METR | $1.5M via parent ARC 2022–23 | VERIFIED | Good Ventures Foundation 990-PF, year to 30 Jun 2022: Alignment Research Center, General Support, $265,000. Year to 30 Jun 2023: Alignment Research Center, "Research related to AI alignment", $1,250,000. ARC Form 990, year to 31 Dec 2024, Schedule I: to Model Evaluation and Threat Research Inc, cash $4,477,169 and non-cash $76,766, "PROGRAM SPIN-OFF: EVALUATIONS PROGRAM SPUN-OFF AS A SEPARATE 501(C)(3)". No direct Good Ventures or Coefficient grant to METR in any filing read or on METR's funders page. | IRS e-files 202341359349105939, 202441369349105564 (Good Ventures), 202513219349323246 (ARC); metr.org/about |
| 7 | Good Ventures → Redwood | $8.8M 2022 · $70M+ 2026 | VERIFIED | Good Ventures 990-PF, year to 30 Jun 2022: Redwood Research, General Support, $1,000,000 and $7,790,000. Coefficient Giving, 9 Sep 2026: "Redwood's Chief Scientist Ryan Greenblatt helped conduct an independent investigation of the incident in which OpenAI agents coordinated a multi-day hack of Hugging Face. CG grantmakers Jake Mendel and Matt MacDermott recently recommended more than $70 million over the next two years to enable them to scale up their work." | IRS e-file 202341359349105939; coefficientgiving.substack.com/p/were-urgently-scaling-our-work-on |
| 8 | Good Ventures → Coxon | $20,159 scholarship | VERIFIED (filing); the identity is a name match | Good Ventures Foundation 990-PF, year 1 Jul 2022 to 30 Jun 2023, Part XV: "JACOB COXON... LONG-TERM FUTURE SCHOLARSHIP PROGRAM - RESEARCH SCHOLARSHIP... 20,159". The filing gives a name only; the link to the ex-Anthropic researcher was made by Parker Thayer (Capital Research Center) and Kevin Bass. | IRS e-file 202441369349105564 |
| 9 | Anthropic → METR | named in the essay · hired to investigate | VERIFIED | Essay, 12 Sep 2026: "Each frontier AI company commits to giving ongoing, employee-like access to a team of embedded third-party evaluators (such as METR)". Anthropic, 9 Sep 2026: "We have signed an agreement with METR to conduct an independent investigation of these incidents... Our initial agreement runs for eight weeks". METR, 14 Aug 2026: "We have not accepted funding from these companies... However, frontier AI companies currently provide a significant amount of free tokens". | darioamodei.com/post/we-must-pace-the-frontier; anthropic.com/research/alignment-assessment-cybersecurity-incidents; metr.org/blog/2026-08-14-funding-update |
| 10 | OpenAI → METR | picked to investigate ExploitGym | VERIFIED | OpenAI gave METR and Redwood supervised access to investigate the July incident; NYT, 3 Sep 2026: six days, "kept on a short leash", no access to the cluster intrusion. | openai.com, 21 Jul 2026 and store facts A1; nytimes.com, 3 Sep 2026 (store) |
| 11 | Redwood → METR | staff inside | VERIFIED | Coefficient, 9 Sep 2026 (Greenblatt "helped conduct" the investigation); the METR–Redwood joint report on the OpenAI incident (Aug 2026, store). Redwood's statement that its staff are subcontracted by METR on the Anthropic investigation is REPORTED (NY Post, 18 Sep 2026) and is not on the chart. | coefficientgiving.substack.com; store facts A1 |

Node sub-labels, sourced: Anthropic "wrote the essay, 12 Sep" (darioamodei.com); OpenAI "ran ExploitGym, July" (openai.com, 21 Jul 2026); METR "no lab money · free tokens" (metr.org, 14 Aug 2026); Redwood "inside the investigation" (as edge 11); Coxon "resigned 9 Sep 2026" (x.com/hilbertspaess/status/2097476196791709843).

### Verified facts kept for the notes, not drawn

| Fact | Grade | Source |
|---|---|---|
| Karnofsky co-founded Open Philanthropy, sat on OpenAI's board (2017), joined Anthropic in January 2025 to work on its Responsible Scaling Policy, and is married to Anthropic's president Daniela Amodei. | VERIFIED | Open Phil grant page 2017; Karnofsky, EA Forum RSP v3 post ("I went to Anthropic full-time in January of 2025"); Fortune, 13 Feb 2025 (Anthropic confirmed) |
| Altman endorsed the essay at 16:30 UTC, 2½ hours after it was posted. | VERIFIED | store facts A3.1 |
| METR raised about $71M in commitments in the six months to August 2026. | VERIFIED | metr.org/blog/2026-08-14-funding-update |
| Canary: The Audacious Project committed about $38M to RAND and METR (Oct 2024), about $17M for METR. | VERIFIED | rand.org/news/press/2024/10/09.html; metr.org/blog/2024-10-09 |
| Anthropic gave $20M to Public First Action, 12 Feb 2026; a second $20M in July 2026. | VERIFIED (Feb); REPORTED (Jul, Politico Pro quoting an Anthropic release) | anthropic.com/news/donate-public-first-action; Politico Pro, 22 Jul 2026 |
| Coefficient: $168M (2024), $351M (2025), "on track to commit over $1 billion" in 2026 on AI safety, security and fieldbuilding. | VERIFIED | coefficientgiving.substack.com, 9 Sep 2026 |
| SFF recommended $516,000 to Encode AI (2025); WSJ reports Encode's Nathan Calvin helped Coxon announce. | VERIFIED (SFF); REPORTED (WSJ, 16 Sep 2026, via store) | SFF 2025; store facts A4.3 |
| Good Ventures paid Tarbell Center $287,750 (FY2024) and $711,750 (FY2025); SFF recommended $520,000 (2024) and $783,000 (2025). | VERIFIED | Good Ventures 990-PFs FY2024, FY2025; SFF pages |
| Tallinn describes himself as an Anthropic board observer. | REPORTED | Postimees, 22 Feb 2026, quoting Äripäev radio |
| Irregular, one third-party evaluator, sat behind the CTF-environment incidents at OpenAI, Anthropic, Meta and Google (May–Sep 2026); Anthropic: "all built by the same third-party partner". These are separate from the ExploitGym / Hugging Face incident. | REPORTED (Irregular named: Bloomberg, TNW, CSA); VERIFIED (one shared partner: Anthropic 9 Sep) | thenextweb.com; labs.cloudsecurityalliance.org; anthropic.com 9 Sep 2026 |
| Sacks, 13 Sep 2026: "Stop pretending METR is independent when it is intertwined with Anthropic's investors and staff." Kokotajlo, 12 Sep: a real pacing "would slow down the leading companies like anthropic and openai more than it would slow down the laggards." | VERIFIED | store facts A4.5, A5.1 |

### Claimed in the posts and deliberately left off

"$7.7 billion" (a ceiling Bass computed from Forbes' "less than 0.8%" and the $965B round); "the majority of Good Ventures' portfolio" (no filing shows where the stake sits); "the evaluator is on Anthropic's payroll" (METR says no lab money; free tokens are on the chart); "Moskovitz led the Series A" (Tallinn led); "cartel", "doom machine", "propaganda", "nepo baby"; "$312M for 476 publications" (Bass's own tally); the Hubinger and Perez career paths; Irregular folded into ExploitGym.

---

## 3. Speaker beats, in Jacob's voice

Slide 9, about 5:30 on the clock, after "5:21 pm" and before "No monopoly on ethics". Say the framing sentence before the chart is read. Cite only what is drawn solid, plus the one dashed line named as a report.

- Advance on "the room I am describing". Then, exactly: "I will say not that I think there is a conspiracy but that there are incentives for a conspiracy."
- The questions, as questions, if he wants them on the record: "I am not here to talk with certainty about the motivations of these people. We can never know. Perhaps this was a concerted effort by a small group to push a cult of doomerism onto the masses and develop political power to regulate and draw power to themselves, pacing the frontier behind them rather than in front of them. Perhaps the same people that carried out the experiments were the people drawing up the rules, funded by the same people making the models, who also funded the whistleblower. Perhaps." [Jacob slide list, 22 Sep] [JACOB: how far to take this with press present; the chart works without it]
- The rule of the screen: every solid line on this screen is a tax filing, a company announcement, or the person's own words. The one dashed line is a newspaper. Nothing alleged is drawn. Read it left to right: the labs, the money, the referees and the witness.
- The money. One foundation, one man's. It bought into the first lab's first round in 2021, and he sits in that lab's boardroom as an observer, his own words. It gave the second lab thirty million dollars in 2017 and took a seat on its board. The man who took that seat co-founded the foundation's grant-maker, joined the first lab in January 2025, and is married to its president. The second funder led that same first round in 2021; his fund recommends the referee's grants.
- The rule. On 12 September the first lab's chief executive published the essay. The second lab's chief endorsed it within two and a half hours. The essay names the referee: "embedded third-party evaluators (such as METR)", with "desks in our offices, access badges, and company laptops". Three days before the essay, the same lab hired the same referee to investigate its own four incidents. In July the second lab picked the same referee, with Redwood, to investigate ExploitGym: six supervised days.
- The referees. The referee says it takes no money from the labs, and I believe it. It takes free tokens; it says so. Its parent took one and a half million dollars from the foundation in 2022 and 2023 and handed the referee four and a half million at the spin-out. The second funder's fund recommended three quarters of a million more. In the six months to August it raised seventy-one million. Redwood, whose staff worked inside the investigation: the foundation paid it eight point eight million in 2022, and in the week the story broke the grant-maker recommended more than seventy million more, naming the investigation in the same paragraph.
- The witness. The researcher who resigned on 9 September once held a scholarship from the same foundation: twenty thousand, one hundred and fifty-nine dollars, 2022 to 2023. It is on a tax form. He says he acted alone. I have no reason to doubt him. (If asked: the filing names a Jacob Coxon; the link to the researcher is a name match made by a research group that read the filings.)
- The dashed line. Forbes reports the founder's stake in the first lab, then worth about half a billion dollars, was moved into a nonprofit vehicle in early 2025 to "dispel any perception of conflict of interest". If so, a foundation that funds the referee's parent holds stock in the lab the referee grades. Reported, not filed; that is why it is dashed. (Never "seven point seven billion": that is a ceiling one critic computed, and nobody has published a valuation.)
- Land it, slowly: Nobody on this screen did anything wrong. Nobody had to. Same eight names, every seat: they built the test, they ran it, they chose who could look, they wrote the rule, they named the referee, and they funded the referee and the witness. Incentives, not a conspiracy. "But I am not here to talk about conspiracies. I am here to talk about processes. And incentives." [Jacob slide list, 22 Sep]
- If challenged, borrow, do not adopt: the White House's AI adviser said it first, in public, on 13 September: "Stop pretending METR is independent when it is intertwined with Anthropic's investors and staff." And give their side its best line, Kokotajlo on 12 September: a real pacing "would slow down the leading companies like anthropic and openai more than it would slow down the laggards." If the frontier does not actually slow, we will know what this was.
- Do not say: "cartel", "on Anthropic's payroll", "doom machine", "nepo", "propaganda". Do not say Moskovitz "led" the Series A. Do not fold the Irregular incidents into ExploitGym. Name institutions and documents, not villains.
- Segue to "No monopoly on ethics": what makes the people in that room more ethical, less corrupt, than the people in this one? Nothing.

---

## 4. Re-check on 27 September

- Whether METR has published anything on its Anthropic investigation (metr.org). If it has, the "hired to investigate" label stands; add what it found to the notes.
- Whether Anthropic has posted its own release for the second $20M to Public First Action; if so, it becomes VERIFIED in the notes.
- Whether Forbes or a filing has named the vehicle holding the Anthropic stake; if a filing names it, edge 2 becomes solid.
- The Good Ventures 990-PF for the year to 30 June 2026 is not due until 2027; nothing new will appear there before the talk.

## 5. Files

- Deck: `content/slides.md` slide 9; `content/data.json → incentives`; `src/figures.ts → networkFigure`; `scripts/generate-slides.mjs → figures.incentives()`; branch `slides-site`.
- Live: https://bittensor-talk.pages.dev/#/9 (Cloudflare Pages project `bittensor-talk`, deployed 24 Sep 2026, deployment 8cb830c9).
- PNG: `media/incentives-web.png` (1600×900).
- Sources: `docs/sources/tweets/cabal-incentives-posts.md` (the six posts verbatim, 46 claims graded).