Scenarios · Scorecard · v0.1
The Scaling Question.
Five outcomes. One bet, three threads. Four binaries. Where named voices on the AI-100 cohort have placed their bets — and where they disagree about which strand of the wire is most likely to fail. Refreshed quarterly.
Of the AI-100 cohort: 11 voices engage the scaling-question framework — 3 full-path · 5 partial · 3 light.
Other AI-100 voices work in adjacent topic spaces (compute infrastructure, alignment, policy) and engage this framework's questions only sparsely or not at all.
§ 01
The cascade.
Substrate density · scrub the cascade
missingness-robustness · 4 states
↑ Holds the prediction
Disputes the prediction ↓
§ 01
The cascade.
§ 01 · The cascade
S1
THE THREE CONDITIONS
All three hold?
- Q1a · science — does the curve hold?
- Q1b · inputs — can the build-out clear?
- Q1c · agents — does long-horizon work?
S2
LABS DURABLE?
Labs durable?
conditional on: S1 ✓
↑ HOLDS → continue ↓ FAILS → Outcome 4 · Distributed
S3
POSITION CAPTURED?
Position captured?
conditional on: S2 ✓
↑ HOLDS → continue ↓ FAILS → Outcome 3 · Commoditized
S4
PERMISSION GRANTED?
Permission granted?
conditional on: S3 ✓
↑ HOLDS → continue ↓ FAILS → Outcome 2 · Constrained
Bull dream realized.
Concentrated + constrained.
Labs commoditized.
Distributed, post-paradigm.
Bubble corrects.
Other.
Voice matrix · scroll →
Voice Q1aQ1bQ1cQ2Q3Q4
Altman H H H H H H
Amodei H H H H h d
Hassabis H h H H H h
Sutskever H H h H — —
Sutton H d H — — —
+ 6 more voices · drag to scroll
§ 02
The five outcomes.
H = holds QH = qualified hold QD = qualified dispute D = disputes — = early termination OFF = off-cascade
| Code | Outcome | Lab valuations | Infra absorption | Labor displacement |
|---|---|---|---|---|
| HHHH | Bull dream realized Outcome 1 S | Frontier labs trade through current valuations as capability gains compound into revenue. | Hyperscaler buildout fully absorbed; utilization holds; depreciation schedules survive contact. | Long-horizon agent work substitutes broadly; labor-share compression visible in coverage industries. |
| HHH·QD | Concentrated + constrained Outcome 2 XL | Top-of-cohort lab values hold; downstream multiple compression as societal permission narrows the market. | Buildout absorbed but at lower utilization; long-horizon contracts re-priced. | Displacement contested at the regulation layer; rollout sectoral and slowed. |
| HH·D— | Labs commoditized Outcome 3 XS | Lab moat erodes; capability becomes table stakes; valuation drift toward applications & distribution. | Compute monetizes broadly through inference economics rather than training capex. | Substitution real but distributed; no single lab captures the labor wedge. |
| H·D—— | Distributed, post-paradigm Outcome 4 M | Lab leadership decoupled from market leadership; valuations re-anchor to product distribution and data rights. | Compute footprint absorbed via non-LLM workloads; new architecture lines pull demand. | Augmentation > substitution; agent economics modest; coverage industries adjust without rupture. |
| D——— | Bubble corrects Outcome 5 M | Capability gains fail to compound; lab valuations compress materially within the refresh window. | Excess buildout stranded or repurposed; depreciation accelerates; secondary market for compute soft. | Displacement thesis postponed; net effect indistinguishable from prior automation waves. |
| OFF | Other Outcome 6 L | Off-cascade voices land here. The framework does not predict where they route; the matrix names them. | Reserved for positions that decline the question rather than answer it. | Tracked separately in the methodology page; not modelled within this cascade. |
§ 03
The reading. Who said what.
S1A · SCIENCE
↑
Scaling-law extrapolations through 2025 continue to predict frontier benchmark gains; loss curves on next-token prediction remain monotonic past 10^26 FLOP. Architecture innovation contributes second-order improvements on top.
Current reading
Capability gains: rising trend at frontier scale.
Source · Cottier et al., arXiv:2511.23455 — update 2026-Q1
S1B · INPUTS
→
Compute supply growth tracking 4.1× / yr through Q1 2026; high-quality token supply tightening but synthetic data substitution offsetting. Capital availability uncorrelated with public-market AI sentiment in this cycle.
Current reading
Inputs scaling: holding, with substitution.
Source · Epoch AI compute-index Q1 2026; SemiAnalysis
S1C · AGENTS
↑
Long-horizon agent task success rates on METR / OSWorld continue to rise; tool-use composition shows compounding behaviour. Real-world deployment failure modes remain present but trending down quarter over quarter.
Current reading
Agent compounding: rising, with caveats.
Source · METR long-task suite v0.4; OSWorld 2026-Q1
Q2 / Q3 / Q4
⊘
Lab durability, position capture, and societal permission are governance + market questions, not research ones. Evidence here is dispositional rather than empirical. Indicators below the wire are inconclusive at v0.1.
Current reading
Below-the-wire: dispositional rather than empirical at v0.1. The framework's downstream questions are governance and market judgments; their substrate matures on a different cadence.
Source · deslop synthesis; see methodology §3b
HOLDS
QUAL HOLD
not engaged
QUAL DISP
DISPUTES
rejects frame
| Voice | Q1a · Sci | Q1b · Inp | Q1c · Agt | Q2 · Dur | Q3 · Pos | Q4 · Soc | Distribution |
|---|---|---|---|---|---|---|---|
| Altman OpenAI | HOLD [L1] | HOLD [L2] | HOLD [L3] | HOLD [L4] | HOLD [L5] | HOLD [L6] | |
| Amodei Anthropic | HOLD [L7] | HOLD [L8] | HOLD [L9] | HOLD [L10] | Q·HD [L11] | Q·DS · | |
| Hassabis Google DeepMind | HOLD [L12] | Q·HD [L13] | HOLD [L14] | HOLD · | HOLD · | Q·HD [L15] | |
| Sutskever SSI | HOLD [L16] | HOLD [L17] | Q·HD · | HOLD · | — | — | |
| Sutton Keen / U. Alberta | HOLD · | Q·DS [L18] | HOLD · | — | — | — | |
| Hinton U. Toronto | HOLD [L19] | Q·HD · | Q·HD [L20] | Q·DS · | DISP · | DISP [L21] | |
| LeCun Meta FAIR | Q·DS [L22] | — | DISP [L23] | — | — | — | |
| Marcus NYU / indep. | DISP [L24] | DISP [L25] | Q·DS [L26] | — | — | — | |
| Chollet Ndea / ARC Prize | DISP [L27] | — | Q·DS [L28] | — | — | — | |
| Tegmark MIT / FLI | Q·HD · | — | — | Q·DS · | — | DISP · | |
| Hooker Adaption Labs | Q·DS [L29] | Q·DS [L30] | — | Q·DS [L31] | DISP [L32] | — |
Mitchell, Acemoglu, Bender, and Karpathy do not appear in this matrix at v0.1. They engage adjacent topic spaces (alignment, governance, narrow-AI critique) where their on-record positions live; the scaling-question cascade is not where they take their stance. Their positions on related questions are engaged in subsequent scorecards.
§ 04
Methodology.
How we verify these positions. Full standard at /scenarios/methodology/. Per-voice sources at § 05 below.
SOURCES · Where the stances come from
One primary citation per active-stance cell in the voice matrix above. 32 cells sourced; 13 active-stance cells render dimmed pending substrate verification. Quotes are byte-exact from primary substrate; read the full method →
- OpenAIQ1a · Sci HOLDS
“In three words: deep learning worked”
The Intelligence Age September 23, 2024 checked 2026-05-24 - OpenAIQ1b · Inp HOLDS
“The only responsible way to meet [demand] is to build more compute, faster”
OpenAI compute infrastructure post April 29, 2026 checked 2026-05-24 - OpenAIQ1c · Agt HOLDS
“In 2025 the first agents may 'join the workforce' and materially change company output”
Reflections January 6, 2025 checked 2026-05-24 - OpenAIQ2 · Dur HOLDS
“OpenAI's giant infrastructure spend is the right timing bet”
Stratechery October 8, 2025 checked 2026-05-24 - OpenAIQ3 · Pos HOLDS
“Destination site logic”
Altman to Stratechery, March 20, 2025 (on OpenAI's product strategy checked 2026-05-24 - OpenAIQ4 · Soc HOLDS
“AI as a 'few thousand days' to abundance”
The Intelligence Age framing of Q4-permitted-flywheel checked 2026-05-24 - AnthropicQ1a · Sci HOLDS
“Scaling up the training of AI systems leads to smoothly better results”
On DeepSeek and Export Controls January 29, 2025 checked 2026-05-24 - AnthropicQ1b · Inp HOLDS
“Export controls are one of our most powerful tools for preventing this”
On DeepSeek and Export Controls January 2025 checked 2026-05-24 - AnthropicQ1c · Agt HOLDS
“It could come as early as 2026”
Machines of Loving Grace, opening framework section, on powerful AI checked 2026-05-24 - AnthropicQ2 · Dur HOLDS
“Open-weights models present additional dangers in that guardrails can be simply stripped away”
The Urgency of Interpretability April 2025 checked 2026-05-24 - AnthropicQ3 · Pos QUAL HOLD
“Anthropic ARR has gone roughly 0 → $100M → $1B → $9-10B across 2023-2025”
Dwarkesh, February 2026 (operational proof of position-capture at the enterprise layer checked 2026-05-24 - Google DeepMindQ1a · Sci HOLDS
“To get all the way to something like AGI may require one or two more new breakthroughs”
Big Technology Podcast / Google I/O (with Sergey Brin) May 21, 2025 checked 2026-05-24 - Google DeepMindQ1b · Inp QUAL HOLD
“Goldfish brain... forget it after the session”
Big Technology / Davos, Jan 21, 2026. (Spoken about memory, not directly inputs, but reflects the substrate-not-purely-compute worldview checked 2026-05-24 - Google DeepMindQ1c · Agt HOLDS
“You actually want a system to... go and complete tasks”
Axios interview Dec 11, 2024 checked 2026-05-24 - Google DeepMindQ4 · Soc QUAL HOLD
“Race to the bottom on safety”
recurrent warning frame; he has used this 2024–2026 to argue for industry-wide coordination on capability deployment checked 2026-05-24 - SSIQ1a · Sci HOLDS
“We have the compute, we have the team, and we know what to do”
X post July 3, 2025 checked 2026-05-24 - SSIQ1b · Inp HOLDS
“Data is the fossil fuel of AI... we have but one internet”
NeurIPS 2024 "Test of Time" talk, Dec 13, 2024. (Data sub-input — bearish on natural data; not yet bearish on synthetic/agentic/multimodal substitution.) checked 2026-05-24 - Keen / U. AlbertaQ1b · Inp QUAL DISP
“People get locked into the human knowledge approach... get their lunch eaten by the methods that are truly scalable”
Dwarkesh Patel Sept 26, 2025 checked 2026-05-24 - U. TorontoQ1a · Sci HOLDS
“I thought it was 30 to 50 years or even longer away. Obviously, I no longer think that”
Euronews coverage of NYT interview, May 2023. Direct Q1a revision in the affirmative direction checked 2026-05-24 - U. TorontoQ1c · Agt QUAL HOLD
“[AI] got better at doing things like reasoning and also at things like deceiving people”
CNN State of the Union, December 2025. Combined Q1a + Q1c affirmation checked 2026-05-24 - U. TorontoQ4 · Soc DISPUTES
“10% to 20% risk that artificial intelligence will eventually take control from humans”
CBS coverage, 2024. The canonical Q4 dispute language checked 2026-05-24 - Meta FAIRQ1a · Sci QUAL DISP
“LLMs will never reach human-level intelligence — unless you change the architecture”
Meta's AI chief, TNW April 2024 checked 2026-05-24 - Meta FAIRQ1c · Agt DISPUTES
“I cannot imagine building agentic systems without an ability to predict… the consequences of their actions”
Fortune, Davos 2026 checked 2026-05-24 - NYU / indep.Q1a · Sci DISPUTES
“No single system will solve more than 4 of the AI 2027 Marcus-Brundage tasks by end of 2025”
Marcus on X, January 1, 2025 (a concrete forecast directly testing scaling continuity checked 2026-05-24 - NYU / indep.Q1b · Inp DISPUTES
“AI is basically an arms race... that can no longer be won”
Marcus's framing on Substack, January 2025 (wasteful-competition critique checked 2026-05-24 - NYU / indep.Q1c · Agt QUAL DISP
“AI agents are wildly premature technology that is being rolled out way too fast”
Dario Amodei, hype, AI safety, and the explosion of vibe-coded AI disasters April 27, 2026 checked 2026-05-24 - Ndea / ARC PrizeQ1a · Sci DISPUTES
“Skill-acquisition efficiency”
On the Measure of Intelligence, arxiv.org/abs/1911.01547 (2019 formal definition checked 2026-05-24 - Ndea / ARC PrizeQ1c · Agt QUAL DISP
“Significant breakthrough”
Chollet on o3, X December 20, 2024 checked 2026-05-24 - Adaption LabsQ1a · Sci QUAL DISP
“While the last decade was about compute, the next will be shaped by the efficiency of adaptation”
X post Oct 22, 2025 (sarahookr/status/1981127919025213691 checked 2026-05-24 - Adaption LabsQ1b · Inp QUAL DISP
“The most costly compute is pretraining compute… With inference compute, you get way more bang for [each unit of computing power]”
Fortune Feb 4, 2026 (Jeremy Kahn checked 2026-05-24 - Adaption LabsQ2 · Dur QUAL DISP
“This changes who gets to shape AI — and who AI ultimately serves”
IMP.NEWS coverage of AutoScientist (May 13 2026 checked 2026-05-24 - Adaption LabsQ3 · Pos DISPUTES
“How do you update a model without touching the weights? There's really interesting innovation in the architecture space”
Fortune Feb 4 2026 checked 2026-05-24
Originally published in Issue 1 · The Scaling Bet