# Fact-check: Affine's model, "best model outside a lab", and NeurIPS

For Jacob Steeves' talk, Exploit Conference, Montreal, 28 September 2026.
Checked 16 September 2026. Every number has a source and a date. Primary sources first (Affine's own X account, GitHub and Hugging Face; Anthropic's system card; the Qwen model card; NeurIPS pages; the SlidesLive recording).

Terms used more than once, defined once:

- **MoE (mixture of experts)**: a model built from many small sub-networks ("experts"). Each token uses only a few of them. So the model has a large **total** parameter count and a much smaller **active** count (the parameters used per token).
- **Post-training / fine-tuning**: extra training on top of a finished base model. The base model already knows language and code. Post-training changes how it behaves.
- **Pre-training**: training a model from random weights on trillions of tokens. This is the expensive part.
- **SWE-bench Verified**: 500 real GitHub bug-fix tasks. Score = share of tasks fixed. **Terminal-Bench 2.0**: 89 tasks done inside a command-line terminal. Both are the standard "agentic coding" tests that Anthropic and Qwen report.
- **Permissionless**: anyone can join with no sign-up or approval. **Whitelisted**: only approved participants.
- **King**: on Affine and Teutonic, the model that currently holds the crown and earns the emissions. A **challenger** must beat the king to replace it.

---

## The three verdicts in one table

| # | Claim as Jacob said it (16 Sep) | Verdict | Shortest safe version |
|---|---|---|---|
| 1 | Affine's latest model is a MoE of "roughly 27–35B active parameters, comparable to a Qwen 27 billion, on par with Claude Opus 4.6, a model only a couple of years old" | **Does not hold** | "Affine miners post-train a 35-billion-parameter open model, Qwen3.6, that has only 3 billion active parameters. The base model alone scores where Anthropic's best model scored about a year ago on real coding tasks. What Affine adds on top is a market that finds the next improvement." |
| 2 | "The best model ever trained outside a lab or a heavily funded open-source company" | **Does not hold** as a claim about model quality. **Holds with wording change** as a claim about the run. | "Teutonic-II is the largest permissionless training run anyone has attempted: 110 billion parameters, anyone can contribute, nobody was whitelisted. It is not finished. That is the job." |
| 3a | Jacob spoke at NeurIPS about what the tooling can do | **Holds with wording change** | "I gave a talk at NeurIPS in Vancouver in December 2024 about how to pay strangers to train a model together. That was the tooling." |
| 3b | At NeurIPS 2024 someone on stage said "It could be a cryptocurrency, too. I mean, there's Bittensor" during Ilya Sutskever's Q&A | **Holds, but the speaker was the audience member, not Ilya** | "At Ilya Sutskever's NeurIPS talk in 2024, someone in the audience asked how to build the right incentives for AI. When Ilya said he did not know, the questioner said: it could be a cryptocurrency, there's Bittensor. Ilya said he was not the right person to comment on crypto. Bittensor got named on that stage. Not by Ilya." |

---

## Claim 1. Affine's latest model and the Opus 4.6 comparison

### Verdict: does not hold

Four separate parts of the sentence are wrong or unsupported. One part is right.

| Part of the claim | What the sources say | Status |
|---|---|---|
| "mixture of experts" | Correct. Affine's base model, Qwen3.6-35B-A3B, is a sparse MoE: 256 experts per layer, 8 routed + 1 shared active per token (Qwen model card). | Right |
| "roughly 27–35B active parameters" | 35B **total**, **3B active**. The "A3B" in the name means "3 billion active". Jacob has the total and active counts swapped. | Wrong |
| "comparable to a Qwen 27 billion" | Half right. Qwen's own table puts Qwen3.6-35B-A3B level with the dense Qwen3.5-27B (SWE-bench Verified 73.4 vs 75.0; Terminal-Bench 2.0 51.5 vs 41.6). But the credit belongs to Alibaba's base model, not to Affine. | Right about the base model; wrong to credit Affine |
| "on par with Claude Opus 4.6" | No. Opus 4.6: SWE-bench Verified 80.8, Terminal-Bench 2.0 65.4. Qwen3.6-35B-A3B base: 73.4 and 51.5. No Affine fine-tune has a published SWE-bench Verified or Terminal-Bench 2.0 score at all. Affine's own repo notes say the crowned kings scored 0/25 on an internal SWE probe (Aug 2026). | Wrong |
| "a model only a couple of years old" | Opus 4.6 was released 5 February 2026. It was 7 months old on 16 Sep 2026. | Wrong |
| "trained on Bittensor by permissionless miners" | Partly. The base model was pre-trained by Alibaba in a data centre. Affine miners (anyone with a hotkey and a Hugging Face account) post-train it and duel the king. Since 27 Aug 2026 the crown is decided by a "teacher-anchored distillation score", not by a benchmark. | Post-training yes; training no |

### What Affine's "latest model" actually is (Sep 2026)

One idea at a time:

1. **Base model.** Qwen/Qwen3.6-35B-A3B. Released by Alibaba's Qwen team on 16 April 2026 (Hugging Face repo created 15 Apr 2026; Qwen GitHub changelog 16 Apr). 35B total, 3B active, MoE, Apache-2.0 licence. Natively multimodal. Context 262K tokens.
2. **What miners submit.** Full fine-tunes of that base (bittensor.ai SN120 page: "Submissions must be Qwen3.6-35B fine-tunes. Other architectures are rejected"). Hugging Face shows them as "Model size 35B params".
3. **How the crown is decided now.** Since 7 Aug 2026 Affine uses a "reason-distillation" score. A frozen teacher model (GLM-4.5-Air-FP8, a 106B/12B-active model by Z.ai) reads the miner's reasoning. The miner wins if its reasoning makes the teacher more confident in the teacher's own next action (@affine_io, 7 and 24 Aug 2026). On 27 Aug 2026 the throne was reset to the untouched Qwen base ("reign 0 re-seeded from untouched Qwen/Qwen3.6-35B-A3B", Affine repo `AGENTS.md`). Reign 9 was crowned on 8 Sep 2026 (same file). The score is not a benchmark. Affine's own repo says so: "the public claim is a distillation meter, not a coding meter, and crowns do not imply benchmark capability."
4. **Last Affine model with published scores.** AFFINE-35B-V, crowned about 23 June 2026 (five champions in five days after the 18 June base switch). Scores are on Affine's **internal** environments, not public benchmarks: SWE 62.4, NavWorld 54.6, Memory 70.4, Terminal 78.1 (base: SWE 33.1, NavWorld 33.8, Memory 11.7, Terminal 46.2). Source: @affine_io 24 Jun 2026; taoprotocol.org 25 Jun 2026. These "SWE" and "Terminal" numbers are **not** SWE-bench Verified or Terminal-Bench 2.0 and cannot be placed in the same column as Anthropic's.
5. **Before that.** The Qwen3-32B era (dense, 32B) ran to the 29th champion by 12 Jun 2026: AFFINE-XXIX beat the Qwen3-32B baseline by +10.5 on SWE-REBENCH, +9.0 on SWE-MULTI, +8.5 on HumanEval (@affine_io, 12 Jun 2026).
6. **Who mines.** Anyone. Miners are pseudonymous Hugging Face accounts (vera6, truelove-yes, best26, kevin954, TalentPigs, Tok331102, unconstai). Five slots are paid; each paid "roughly $50,000 USD per week" (@affine_io, 23 Aug 2026). Compute is rented on Lium (8×B200 pods appear in miner model cards).
7. **Who runs it.** Founded by Jacob (Jul 2025). Ownership passed to a China-based team (Tobias, Allan, bignickeye) in June 2026. Jacob is "chief scientist at Affine" (@affine_io, 13 Aug 2026).

### The comparison table Jacob asked for

All scores are the maker's own published figures. Harnesses differ (Anthropic: own two-tool scaffold, 25 trials; Qwen: internal scaffold, 200K context; Terminal-Bench both use the Terminus-2 agent). Treat gaps under about 3 points as noise.

| Model | Made by, how | Released | Size (total / active) | SWE-bench Verified | Terminal-Bench 2.0 | Source |
|---|---|---|---|---|---|---|
| Claude Opus 4.6 | Anthropic, closed, data centre | 5 Feb 2026 | not disclosed | **80.8** (80.84 avg of 25 trials; 81.42 with a prompt change) | **65.4** | Anthropic announcement and system card, Feb 2026 |
| Claude Opus 4.5 | Anthropic | 24 Nov 2025 | not disclosed | 80.9 | 59.8 | Opus 4.6 system card table |
| Claude Sonnet 4.5 | Anthropic | 29 Sep 2025 | not disclosed | 77.2 | 50.0 (Sonnet 4.5 post) / 51.0 (Opus 4.6 card) | Anthropic, Sep 2025 and Feb 2026 |
| Claude Opus 4 | Anthropic | 22 May 2025 | not disclosed | 72.5 | 43.2 (Terminal-Bench 1.0, older version) | Anthropic "Introducing Claude 4" |
| Claude 3.5 Sonnet (upgraded) | Anthropic | 22 Oct 2024 | not disclosed | 49.0 | not reported | Anthropic, 30 Oct 2024 |
| Qwen3.5-27B (the "Qwen 27B") | Alibaba Qwen, open weights, data centre | 24 Feb 2026 | 27B dense | 75.0 | 41.6 | Qwen3.6-35B-A3B model card table |
| **Qwen3.6-35B-A3B (Affine's base)** | Alibaba Qwen, open weights, data centre | 16 Apr 2026 | 35B / 3B MoE | **73.4** | **51.5** | Qwen model card, Apr 2026 |
| AFFINE-35B-V | Affine miners, post-training of Qwen3.6-35B-A3B | ~23 Jun 2026 | 35B / 3B | not measured | not measured | Affine internal envs only: SWE 62.4, Terminal 78.1 (@affine_io 24 Jun 2026) |
| Affine king, reign 9 | Affine miners, post-training, distillation score | 8 Sep 2026 | 35B / 3B | not measured publicly | not measured publicly | Affine repo `AGENTS.md`; Aug probe: crowned kings 0/25 on an internal SWE-lite set; untouched Qwen base 0.24 "best of 51 benched models" |

Where the comparison holds:

- The **base model** (Qwen3.6-35B-A3B, 3B active) matches Anthropic's **Claude Opus 4 / Sonnet 4** of May 2025 on SWE-bench Verified (73.4 vs 72.5/72.7) and matches **Claude Sonnet 4.5** of September 2025 on Terminal-Bench 2.0 (51.5 vs 51.0). So "a 3-billion-active open model does what Anthropic's frontier did 12–16 months ago" is fair, and the "on par with a Qwen 27B" part is fair.
- The **mechanism** works: after the June base switch, five successive champions raised Affine's internal scores in five days, and the 32B era ran 29 champions. That is real evidence that a paid, open competition produces improvements on the target it is paid for.

Where it does not hold:

- Not Opus 4.6. Gap of 7 points on SWE-bench Verified and 14 points on Terminal-Bench 2.0, against the base model. Against Affine's fine-tunes there is no public number to compare at all.
- Not "27–35B active". 3B active.
- Not "a couple of years old". Seven months.
- Not "trained on Bittensor". Pre-trained by Alibaba; post-trained on Bittensor. The Templar audit of Teutonic-I (see claim 2) shows how badly this distinction bites when a critic checks.
- Affine's own public repo (`AGENTS.md`, sections 1 and 3b, dated 9–10 Aug 2026) says the coding claim "does not survive the live board": all three crowned kings resolved 0 of 25 SWE tasks, and the untouched Qwen base was the best of 51 models benched. The team has since reset the throne (27 Aug) and changed the score three times (5, 9 Sep). A hostile questioner can read this file.

Two things to expect a question on:

- Affine's public repo has a `mining/` folder described as the "Operator's own SN120 mining effort" (`LAYOUT.txt`), and challenger model cards on Hugging Face (accounts unconstai, truelove-yes, best26) cite experiment paths from that folder. Jacob wrote on 20 Dec 2025 that he has "never bought Affine". Decide the answer to "does the operator mine its own subnet?" before the talk.
- The scorer was gamed at least twice (@affine_io, 27 Aug 2026), and the throne was reset. Say it first; it is the point of a live market.

### Safe wording for claim 1

Say:

> "Affine takes an open model from Alibaba, Qwen3.6, thirty-five billion parameters, three billion active, and pays anyone on earth to make it better. The base model already does what Anthropic's frontier did a year ago on real coding tasks. In June the miners crowned five champions in five days. The mechanism is the product. The model is the receipt."

Or, shorter:

> "Three billion active parameters, open weights, and it codes like a frontier model did a year ago. Anyone can improve it and get paid the moment they do."

Do not say: "on par with Opus 4.6", "27 to 35 billion active", "trained on Bittensor" (say "post-trained" or "improved"), or any Affine internal score next to an Anthropic score.

---

## Claim 2. "The best model ever trained outside a lab or a heavily funded open-source company"

### Verdict: does not hold as stated; holds with a wording change (about the run, not the model)

Why it does not hold, in four steps:

1. Affine's models are post-trained copies of an Alibaba model. Alibaba is a lab. The sentence asks the audience to credit Bittensor with the pre-training.
2. Teutonic-I, Bittensor's own pre-training claim, was audited by Templar (the previous SN3 team) on 20 Aug 2026 and found to descend from Qwen3.5-9B-Base (weights 0.96 correlated; a near-copy of a Qwen-derived checkpoint was crowned on day two). Matched re-evaluation: Qwen3.5-9B-Base 62.48 vs Teutonic-I 62.45 on the same 11 tasks. No aggregate gain. Jacob conceded on Discord that miners used Quasar weights (synapz.org, 21 Aug 2026).
3. Several teams outside the big labs have trained models **from scratch**, on the open internet, that beat anything Bittensor has released from scratch. Covenant-72B (Templar, Mar 2026) is the largest **permissionless** from-scratch run with a finished model. Nous Consilience 40B, Pluralis-8B and INTELLECT-1 are from-scratch runs by small companies.
4. Community fine-tuning of Qwen is not rare. Individuals on Hugging Face post Qwen3.5/3.6 fine-tunes distilled from Claude Opus 4.6 traces every week. Affine's difference is the payment and the duel, not the model.

### The field, compared

"Outside a lab or heavily funded" is graded honestly. Funding figures are as reported in press releases; verify before a slide.

| Run / model | Who | Funding | Type | Size | Data | Permissionless? | Result | Date |
|---|---|---|---|---|---|---|---|---|
| **Teutonic-II 110B** | Jacob + Dendrite, SN3 | Protocol emissions (~$50K/day to SN3) | Pre-training, king-of-the-hill loss duels | ~110B total / ~7.3B active MoE (MiMoV2-style) | FineWeb-Edu eval; miners choose | Yes | In progress. Genesis is random init (HF card, 24 Aug 2026). Interim community figures MMLU 69%, GSM8K 72.3%, PIQA 82.9% (SubnetRadar, 6 Sep). Dashboard showed a reset state (0/5 kings) on 16 Sep. Team says post-training on Reliquary next (SubnetRadar, 1 Sep) | Aug–Sep 2026 |
| **Teutonic-I 10B** | same | emissions | "Pre-training" competition | 8.6B (marketed 10B) | 4.24T eval tokens | Yes | 62.28% on 11 tasks. Audit: derived from Qwen3.5-9B-Base; no gain over Qwen (62.48) | 14 Aug 2026; audit 20 Aug |
| **Covenant-72B** | Templar / Covenant AI (left Bittensor Apr 2026) | Not disclosed; emissions during the run | Pre-training, SparseLoCo + Gauntlet | 72B dense | ~1.1T tokens | **Yes** (first permissionless from-scratch run of this size) | MMLU 67.1, ARC-C 56.8, HellaSwag 80.6; beats LLaMA-2-70B (65.6 MMLU) | 9–13 Mar 2026 |
| Templar 1.2B (Gauntlet) | Templar; paper co-authored by Jacob | emissions | Pre-training | 1.2B | — | Yes | "competitive on a per-iteration basis" | arXiv 2505.21684, May 2025 |
| **Nous Consilience 40B** | Nous Research | ~$50M Series A (Paradigm, Apr 2025, reported) | Pre-training, DisTrO/DeMo on Psyche (Solana) | 40B dense | ~1.1–1.2T of a planned 20T | No (whitelisted testnet) | MMLU 62.5 pre-anneal, 67.1 after anneal (Covenant paper, Table 3); Nous calls it "largest distributed pre-training run ever" by size and data | May 2025–2026 |
| Hermes 4.3 | Nous | same | Post-training on Psyche | (on Seed-OSS-36B etc.) | — | No | "first production model post-trained entirely on Psyche" | Dec 2025 |
| **INTELLECT-1** | Prime Intellect | ~$15M then (reported) | Pre-training, DiLoCo | 10B | 1T | No (whitelisted, 30 nodes) | MMLU 32.7 (0-shot, Covenant table) | Nov 2024 |
| INTELLECT-2 | Prime Intellect | same | Decentralized RL post-training of QwQ-32B | 32B | — | Yes (RL rollouts) | ≈ QwQ-32B | Apr 2025 |
| **INTELLECT-3** | Prime Intellect | $130M Series A (8 Jul 2026, own blog) | Post-training of GLM-4.5-Air-Base, **centralized** 512 H200s | 106B / 12B MoE | — | No | AIME25 88.0, LiveCodeBench 69.3, GPQA 74.4, MMLU-Pro 81.9 | 26 Nov 2025 |
| **Pluralis-8B (Agora)** | Pluralis Research | ~$7.6M seed (2025, reported) | Pre-training, pipeline-parallel over internet | 8.6B | 500B FineWeb-Edu | **Yes** (330 contributor nodes, mostly consumer GPUs) | Loss "within a small margin of a centralized reference"; 63% of H100 efficiency; no downstream table found | arXiv 2607.13332, Jul 2026 |
| Gensyn RL Swarm / CodeZero | Gensyn | ~$43M (a16z, 2023, reported) | Decentralized RL on 0.5B–1.5B Qwen2.5-Coder | tiny | — | Yes | No notable model; RL Swarm marked paused | 2025–2026 |
| **IOTA Orion-100B / Orion-16B** | Macrocosmos, SN9 | emissions; ~20–30 staff | Pre-training, pipeline-split over internet | 100B / 16B | 1.1B tokens (100B) / >100B tokens (16B) | Yes | Systems proof, not a usable model | Jun–Aug 2026 |
| **Affine kings** | Affine Foundation, SN120 | emissions (~$250K/week to 5 slots) | Post-training of Qwen3.6-35B-A3B | 35B / 3B | — | Yes | Internal env gains (Jun); no public benchmark since Aug | Jun–Sep 2026 |

Reading the table:

- Every entry is "outside a frontier lab". The honest split is **from-scratch vs post-training**, and **permissionless vs whitelisted**.
- Best **from-scratch** model made outside the big labs on the open internet: **Covenant-72B** (permissionless) and **Consilience 40B** (whitelisted). Both beat anything Bittensor has released from scratch. Teutonic-I does not qualify as from-scratch.
- Best **post-trained** model from a decentralized-AI company: **INTELLECT-3** (106B). But it was trained on one 512-GPU cluster owned by a $130M company. It is neither permissionless nor decentralized in the training itself.
- Best **permissionless post-training**: Affine, but with no public benchmark for the current king.
- **Largest permissionless run attempted**: Teutonic-II at 110B. Not finished. "50% larger than any prior decentralized run" (Jacob, 24 Aug) is fair against Covenant-72B by parameter count. IOTA's Orion-100B was a systems demo trained on 1.1B tokens only.

Also on the "heavily funded" cut: Affine and Teutonic are funded by protocol emissions worth roughly $13M/yr (Affine) and $18M/yr (SN3) at September prices. That is small next to a lab, but it is not "no funding". Nous, Prime Intellect and Gensyn are venture-funded companies. Pluralis and Templar/Covenant are small. Do not draw the line at money. Draw it at **who is allowed in**.

### Safe wording for claim 2

Best version (true today):

> "Teutonic-II is the largest permissionless training run anyone has attempted. One hundred and ten billion parameters. No whitelist. No résumé. Any GPU, any algorithm, any person. It is not finished, and I will not tell you it is the best model in the world. I will tell you no lab can run this experiment, because the moment you let strangers in, you need a market, not a manager."

Second version (about post-training, true today):

> "On Affine, anonymous engineers took an open model and, in five days, crowned five successive champions on the tasks the network pays for. Nobody hired them. Nobody approved them. The token did."

Add the Teutonic-I lesson yourself, before a critic does:

> "Our first run taught us something. The market found the cheapest path to a lower loss, and that path ran through borrowed open weights. That is not a scandal. That is what a market does when you do not verify. Teutonic-II verifies."

Do not say: "best model ever trained outside a lab", "world's most performant decentralized LLM", or "trained on Bittensor" for any Affine model.

---

## Claim 3a. Jacob's NeurIPS appearance

### Verdict: holds with wording change

Facts:

- **Event**: NeurIPS 2024, Vancouver Convention Centre. **Session**: Expo Talk Panel, "Incentivizing Collaborative AI: A Decentralized Approach to Scaling Machine Learning". **Speaker**: Jacob Steeves. **When**: Tuesday 10 December 2024, 8:30–9:30 a.m. PST. **Where**: West Meeting Room 109–110. **Sponsor**: Bittensor (listed as "Bittensor CSP" on the Expo page). Sources: neurips.cc/virtual/2024/expo-talk-panel/100355 and neurips.cc/Expo/Conferences/2024/talk panel/100355.
- **What kind of session**: an Expo talk. Expo sessions are bought by sponsors. They are not peer-reviewed and not part of the main conference programme. Say "I gave a talk at NeurIPS", not "I presented at NeurIPS" or "a NeurIPS paper".
- **What the abstract promised**: the architecture (contributors and validators), the incentive mechanism ("mathematical foundations of the reward system"), practical challenges (privacy, latency, malicious actors), applications, open-source benefits. Nothing about a trained model or benchmark results. This matches Jacob's own framing on 16 Sep: "showcasing what the tooling can do."
- **Recording**: none found. NeurIPS lists no SlidesLive link for this Expo slot; no YouTube upload found. Ask the Opentensor Foundation whether a recording or the slides exist before promising the audience a link.
- **Jacob's own record of the week**: 12 Dec 2024, "we peppered your booth at @NeurIPSConf" (to Elon Musk, 161 likes). 14 Dec 2024, "A lot of people at NeurIPS talking about ways to regulate AI... Make the token the objective function" (149 likes) and "the token is the objective function" (185 likes). Links: x.com/const_reborn/status/1867293069810577609 and /1868076076905418938.
- **Earlier NeurIPS paper**: "Incentivizing Intelligence: The Bittensor Approach" (Steeves, Shaabana, Hu, Luus, Liu, Tasker-Steeves), Trustworthy and Socially Responsible Machine Learning workshop (TSRML 2022), co-located with NeurIPS 2022. A workshop paper, not a main-track paper. Source: bittensor.com/pdfs/academia/NeurIPS_DAO_Workshop_2022_3_3.pdf.
- **Related paper with Jacob as co-author**: "Incentivizing Permissionless Distributed Learning of LLMs" (Lidin, Sarfi, Pappas, Dare, Belilovsky, Steeves), arXiv 2505.21684, 27 May 2025. Describes Gauntlet and a permissionless 1.2B run on Bittensor. Not a NeurIPS paper as far as found. Covenant (Templar) presented two workshop papers in Dec 2025 (Jacob retweeted on 7 Dec 2025); Jacob was not listed as presenter.
- **NeurIPS 2025**: no Steeves talk or paper found.

### Safe wording for claim 3a

> "In December 2024 I stood on a stage at NeurIPS in Vancouver and explained how you pay strangers to train a model together. That was the tooling. Today I can show you what the tooling made."

Do not say "I presented our results at NeurIPS" or "our NeurIPS paper". Do not say the talk showed a trained model.

---

## Claim 3b. Who said "It could be a cryptocurrency, too. I mean, there's Bittensor"?

### Verdict: the line is real and was said on the NeurIPS stage, but by the audience member, not by Ilya Sutskever

The statement to verify comes from `docs/research/model-keynotes.md` (section 10), which quoted the whole passage as Ilya's and flagged it for checking.

Method (so you can trust it or redo it):

1. Downloaded the official SlidesLive recording of the NeurIPS 2024 Test of Time session (slideslive.com/39031672, 46 min 19 s, Goodfellow first, then Sutskever). Same file the transcript in `docs/sources/talks/sutskever-neurips-test-of-time-2024.md` was made from.
2. Re-transcribed 42:30–44:40 with word timestamps (faster-whisper, model small).
3. Built voice fingerprints (speaker embeddings, resemblyzer encoder) for six passages that are certainly Ilya (his answer before and after the disputed lines) and four that are certainly the questioner (the question itself, "I have an unrelated question", "thank you for the talk"). All ten reference passages were classified correctly, with a clear margin.
4. Classified the disputed lines against those two fingerprints, both as whole lines and as a sliding 1.2-second window.

Result (session timestamps; subtract about 22:20 for the standalone YouTube upload youtube.com/watch?v=1yvBqasHLZs, so about 21:00–21:20 there):

| Time | Words | Speaker (fingerprint) | Confidence |
|---|---|---|---|
| 43:06–43:25 | "But to your question about what incentive structure should we create, I don't feel that I know... some kind of top-down structure government thing, I don't know." | **Ilya** | strong (0.94–0.98 to Ilya) |
| 43:25–43:27 | "It could be a cryptocurrency too." | **Questioner** | clear (0.81 to questioner vs 0.67 to Ilya; every sliding window agrees) |
| 43:28–43:29 | "Yeah. I mean..." (short, partly laughter) | Ilya, weak | low (short clip) |
| 43:29–43:31 | "There's Bittensor, you know, there's things." | **Questioner** | clear (0.84 vs 0.67; every sliding window agrees) |
| 43:32–43:36 | "I don't feel like I am the right person to comment on cryptocurrency." | **Ilya** | strong (0.95) |
| 43:37–43:50 | "But you know, there is a chance, by the way, what you're describing will happen... maybe that will be fine." | **Ilya** | strong |

So the exchange was:

- Questioner: how do you create the right incentive mechanisms for humanity to build AI that has freedoms?
- Ilya: I don't feel that I know... it sounds like a top-down government thing, I don't know.
- Questioner: It could be a cryptocurrency too.
- Ilya: Yeah, I mean...
- Questioner: There's Bittensor, you know, there's things.
- Ilya: I don't feel like I am the right person to comment on cryptocurrency. But there is a chance what you're describing will happen...

This also fits the text: "Yeah. I mean..." is how a speaker answers an interruption, not how he continues his own sentence. The community transcript (github.com/shun-liang/readable-talks-transcriptions) also breaks the paragraph exactly at "It could be a cryptocurrency too." / "Yeah. I mean... There's BitTensor".

Caveat: this is an automated voice comparison, not a human listening. It was right on all ten known passages, so it is strong evidence, but one person should listen to the 20 seconds (YouTube ~21:00–21:20) before the slide is final. Correct `model-keynotes.md` section 10 either way: the current text attributes "there's Bittensor" to Ilya.

### Safe wording for claim 3b

> "In December 2024, at Ilya Sutskever's NeurIPS talk, someone in the audience asked him how you build the right incentives so that intelligence stays free. He said he did not know. The questioner said: it could be a cryptocurrency, there's Bittensor. Ilya said he was not the right person to comment on crypto. I will take that. The question was asked on that stage, and the answer in the room was our name."

Do not say "Ilya Sutskever mentioned Bittensor" or put the quote on a slide with Ilya's name under it.

---

## Sources (with dates)

Affine

- @affine_io tweets, read via the X API on 16 Sep 2026: 12 Jun 2026 (2065461449166835963, AFFINE-XXIX vs Qwen3-32B), 21 Jun (2068704533233168573, base switch), 24 Jun (2069810780019491188, AFFINE-35B-III scores), 30 Jul (2082861540534026710, LoRA-on-GLM-5.2 plan), 7 Aug (2085698444346216834, reason-distillation), 13 Aug (2087957548578001177, Jacob "chief scientist at Affine"), 23 Aug (2091498650119901215, $50K/week per slot), 24 Aug (2091858071824011444, GLM-4.5-Air-FP8 frozen teacher), 27 Aug (2093008057261609231, evaluator gamed), 4 Sep (2095937147320844533, private submissions), 7 Sep (2096984627059810338). One API request, 29 tweets returned.
- Affine GitHub: github.com/AffineFoundation/affine (`AGENTS.md`, `LAYOUT.txt`, `START_HERE.txt`, read 16 Sep 2026; repo last pushed 16 Sep 2026). Key passages: §1 "crowns do not imply benchmark capability"; §2 "genesis reset" 27 Aug 2026; §3b RT-7 "all three S-crowned kings resolve 0/25" (9 Aug 2026); v7 note on the SWE-bench Pro / Claude Code probe (7–9 Sep 2026); `LAYOUT.txt` "mining/ Operator's own SN120 mining effort".
- Hugging Face: AffineFoundation org (7 models, last updated 18 May 2026); vera6/affine-5g4yy75zuz-t6 (35B, created 18 Aug 2026, "live king reign36" per challenger cards); unconstai/1787207581, truelove-yes/20260819-111453, best26/Affine-5czsc2fc98-... (challenger cards, Aug 2026).
- bittensor.ai/subnets/120 ("Submissions must be Qwen3.6-35B fine-tunes"); taopedia.org/wiki/subnet_120; taoprotocol.org, "Affine (SN120) Update: Model Migration and an Evaluation Bug", 25 Jun 2026; subnetradar.com/subnet-news/120/2026-08-23.

Qwen and Anthropic

- Qwen/Qwen3.6-35B-A3B model card (HF, created 15 Apr 2026): architecture, "35B in total and 3B activated", benchmark table incl. Qwen3.5-27B. Release: github.com/AlibabaCloud-Official/Qwen3.6 changelog "2026-04-16"; alibabacloud.com/blog/603042 (17 Apr 2026). Qwen/Qwen3.5-27B HF created 24 Feb 2026.
- Anthropic, "Introducing Claude Opus 4.6", 5 Feb 2026 (anthropic.com/research/claude-opus-4-6); Claude Opus 4.6 System Card, Feb 2026 (table: Opus 4.6 / Opus 4.5 / Sonnet / Gemini / GPT-5.2; §2.4–2.5). "Introducing Claude Sonnet 4.5", 29 Sep 2025. "Introducing Claude 4", 22 May 2025. "Raising the bar on SWE-bench Verified with Claude 3.5 Sonnet", 30 Oct 2024.

Decentralized runs

- Teutonic-I paper: teutonic.ai/paper.html (Aug 2026); HF dendriteholdings/Teutonic-I (created 14 Aug 2026). Templar audit: tplr.ai/teutonic-1-audit.pdf; x.com/tplr_ai/status/2090433758977778074 (20 Aug 2026). synapz.org, 21 Aug 2026. Teutonic-II: HF dendriteholdings/teutonic-II-110B-genesis (24 Aug 2026); x.com/const_reborn/status/2091896286035235027; teutonic.ai dashboard and llms.txt (16 Sep 2026); subnetradar.com/subnet-news/3/2026-09-01 and /2026-09-03; taodaily.io "Teutonic Just Jumped from 10B to 110B".
- Covenant-72B: arXiv 2603.08163 (v2), 9 Mar 2026; x.com/tplr_ai 10 Mar 2026; tao.media 13 Mar 2026. Gauntlet: arXiv 2505.21684, 27 May 2025.
- Nous: nousresearch.com/nous-psyche (May 2025); HF PsycheFoundation/consilience-40b-7Y9v38s5; forum.nousresearch.com "Psyche - Future Directions"; nousresearch.com/introducing-hermes-4-3 (Dec 2025); anneal numbers from Covenant-72B paper Table 3.
- Prime Intellect: primeintellect.ai/blog/intellect-3 (26 Nov 2025); arXiv 2512.16144; HF PrimeIntellect/INTELLECT-3; primeintellect.ai/blog ("$130M Series A", 8 Jul 2026). INTELLECT-1: arXiv 2412.01152 (Nov 2024).
- Pluralis: arXiv 2607.13332 (Agora / Pluralis-8B, Jul 2026); arXiv 2506.01260 (Protocol Models, Jun 2025); pluralis.ai blog on node0 (7.5B, 303 participants).
- Gensyn: RL Swarm docs (paused), CodeZero (Qwen2.5-Coder 0.5B/1.5B), 2026.
- IOTA: see `docs/research/top-subnets-2026.md` (Macrocosmos, 1 Jun and 27 Jul 2026).

NeurIPS

- neurips.cc/virtual/2024/expo-talk-panel/100355; neurips.cc/Expo/Conferences/2024/talk panel/100355 (time, room); neurips.cc/virtual/2024/day/12/10.
- bittensor.com/pdfs/academia/NeurIPS_DAO_Workshop_2022_3_3.pdf (TSRML 2022).
- Jacob's tweets, 10–16 Dec 2024, from `docs/sources/tweets/const_reborn-tweets.jsonl`.
- SlidesLive 39031672 (NeurIPS 2024 Test of Time session, 13 Dec 2024); YouTube 1yvBqasHLZs; community transcript github.com/shun-liang/readable-talks-transcriptions (neurips_2024). Voice analysis: faster-whisper (small) + resemblyzer speaker encoder, run 16 Sep 2026 on the SlidesLive audio; scripts and clip kept off the store at /tmp/factcheck on the research VM.

Related store documents: `docs/project-context.md` ("So we built one", 16 Sep), `docs/research/top-subnets-2026.md` (Affine and Teutonic sections, 12 Sep), `docs/research/model-keynotes.md` (section 10; needs the speaker correction above).