Unfamiliar, multi-day tasks completed without human rescue
01
The four judgments
The minimum viable briefing. Select one to see the decision implication and the evidence that would reverse it.
Inference economics plus durable workflow switching costs
Weights released promptly after training, with reproducible evals
Unscripted task success, field uptime and paid repeat use
Shared thresholds, incident disclosure and observable consequences
02
Where they stand
Switch the question, drag the year, select a face. Paths show movement; the side panel keeps stated beliefs separate from committed actions.
later · constrainsooner · constrainlater · adaptsooner · adaptlong runwayshort runway →ship / adaptconstrain / verify →
03
Words, then wagers
The timeline is the evidence. Each entry preserves the exact public language and places the next observable action directly beneath it.
“AI is a General Purpose Technology with numerous use cases.”
Revealed next: Builds education, application companies and deployment practice—not a frontier lab.
Andrew Ng / U.S. Senate · Nov 2023 ↗“For humanoid robots, the biggest question is how to integrate with AI models.”
Revealed next: Ships robot bodies while treating general-purpose software as unresolved.
TMT / ChinaTalk translation · Apr 2024 ↗“We will not change to closed source.”
Revealed next: DeepSeek releases efficient open models and uses price as ecosystem pressure.
Waves / ChinaTalk translation · Jul 2024 ↗“I think it could come as early as 2026, though there are also ways it could take much longer.”
Revealed next: Threshold governance and a multi-gigawatt compute plan back the short-run case.
Dario Amodei · Oct 2024 ↗“AI is a once in a lifetime technology.”
Revealed next: Alphabet integrates Gemini across Search, Cloud, Workspace, Android and custom silicon.
Google · Feb 2025 ↗“I am launching a new non-profit AI safety research organization called LawZero.”
Revealed next: A research and institution-level pivot toward non-agentic Scientist AI.
Yoshua Bengio · Jun 2025 ↗“We believe in putting this power in people’s hands to direct it towards what they value.”
Revealed next: Meta spends for a concentrated superintelligence team while making openness conditional.
Meta · Jul 2025 ↗“We remain committed to open-sourcing Qwen and shaping it into the operating system of the AI era.”
Revealed next: $53B of infrastructure turns ecosystem language into a cloud-distribution bet.
Alibaba Group · Sep 2025 ↗“We are squarely an ‘age of research’ company.”
Revealed next: SSI declines near-term products and organizes around reliable generalization.
Dwarkesh Patel · Nov 2025 ↗“We are not building an ill-defined and ethereal superintelligence.”
Revealed next: Microsoft AI forms a bounded, domain-oriented superintelligence team.
Microsoft AI · Nov 2025 ↗“In ten more years, I believe we are almost certain to build superintelligence.”
Revealed next: OpenAI’s practical program is infrastructure, distribution and iterative deployment.
OpenAI · Dec 2025 ↗“The agentic AI inflection point has arrived.”
Revealed next: Seven production chips and rack-scale systems target every phase of the workload.
NVIDIA · Mar 2026 ↗“This partnership accelerates our capacity to build AI that people can shape and make their own.”
Revealed next: Thinking Machines secures at least one gigawatt and later releases open weights.
Thinking Machines Lab · Mar 2026 ↗“We will resume releasing some open source models soon.”
Revealed next: The recommitment follows a private-preview phase, showing access policy remains strategic.
Meta · Aug 2026 ↗04
The twenty
A deliberately bounded field: frontier labs, platforms, infrastructure, embodied systems, China, governance and research dissent.
20 / 20 shown
05
Who is useful to listen to
A prediction can be sincere and still be useless. This score rewards testability, proximity, costly backing and calibration—not fame.
Discount datesDecision-useful
How forecast utility is scored
Specificity 25 · concrete object, mechanism and window. Falsifiability 25 · observable resolution. Proximity 20 · privileged access. Costly backing 20 · capital, career, structure or product. Record 10 · stable definitions and deadline discipline. Editorial and provisional; utility is not honesty.
06
The games underneath
Public positions are moves inside races for capability, distribution and bargaining power. Select a game to see the equilibrium and the intervention that could change it.
Likely equilibrium
Every player invests because unilateral restraint transfers option value to rivals. Public safety language rises with capability, but spending does not slow.
Observed: Multi-gigawatt commitments are arriving simultaneously across competing stacks.
What changes the game: Shared evaluations, incident disclosure, verifiable thresholds and liability that make unsafe speed expensive.
Why restraint losesPayoff to A / B
A ↓ / B →RestrainAccelerateRestrain3 / 3fragile restraint1 / 5sucker / leadAccelerate5 / 1lead / sucker2 / 2stable race
07
Evidence + machine layer
The human view stays compressed. The full register and a stable JSON payload remain available for audit, retrieval and a future conversational layer.
LLM-ready
Open context.json ↗Structured context, not scraped prose.
Players, dated positions, exact quotes, revealed bets, conviction signals, games, reversals and source URLs in one versioned payload.
Open the 42-source evidence register
Ten years ↗OpenAI · Dec 2025
Altman puts superintelligence inside a ten-year window.
Built to benefit everyone: our plan ↗OpenAI · Jun 2026
Distribution, affordability and public oversight become the public frame.
Building the compute infrastructure for the Intelligence Age ↗OpenAI · Apr 2026
Stargate is described as the long-run compute base for broad deployment.
Machines of Loving Grace ↗Dario Amodei · Oct 2024
Defines powerful systems, gives an earliest-arrival case and maps scientific upside.
Responsible Scaling Policy v3.4 ↗Anthropic · Jul 2026
Capability thresholds are tied to security, evaluation and reporting requirements.
Anthropic and Amazon expand compute collaboration ↗Anthropic · Apr 2026
A ten-year, $100B commitment secures up to 5GW of capacity.
Our position on open-weights models ↗Anthropic · Jul 2026
Supports lower-risk open weights while drawing a line at dangerous capabilities.
About Google DeepMind ↗Google DeepMind · Aug 2026
The lab states a responsibility-led path to general systems and scientific benefit.
AlphaFold: five years of impact ↗Google DeepMind · Nov 2025
A concrete record of the science-first thesis, including global research use.
2025 Q4 earnings call ↗Alphabet Investor Relations · Feb 2026
Alphabet guides to $175–185B of 2026 capital expenditure and describes model, product and cloud demand.
We’re moving from the age of scaling to the age of research ↗Dwarkesh Patel · Nov 2025
Sutskever identifies weak generalization as the bottleneck and changes the scaling thesis.
Safe Superintelligence ↗SSI · Jun 2024
A single-product company organized around safe superintelligence rather than near-term releases.
A path towards autonomous machine intelligence ↗Yann LeCun · Jun 2022
The technical program behind learned world models, memory, planning and hierarchical control.
LeCun leaves Meta to pursue advanced machine intelligence ↗Associated Press · Nov 2025
A career-level commitment to the non-LLM path after twelve years at Meta.
Open source AI is the path forward ↗Meta · Jul 2024
Zuckerberg argues that open models will become the industry standard.
Introducing Muse Spark ↗Meta · Apr 2026
The first new superintelligence-lab model launches in private preview, with future openness left conditional.
The future is for everyone ↗Meta · Aug 2026
Meta renews its case for personal, distributed and partly open superintelligence.
FY2026 Q3 earnings call ↗Microsoft Investor Relations · Apr 2026
Microsoft expects roughly $190B of calendar-2026 capital expenditure and reports large-scale agent management.
Towards Humanist Superintelligence ↗Microsoft AI · Nov 2025
Suleyman forms a superintelligence team around bounded, human-directed capability.
Vera Rubin opens the agentic frontier ↗NVIDIA · Mar 2026
Seven production chips span training, post-training and agent inference.
Global robotics leaders take physical systems to production ↗NVIDIA · Mar 2026
A full-stack robotics ecosystem backs the physical-world thesis.
$20B Series E ↗xAI · Jan 2026
xAI reports more than one million H100-equivalent GPUs at year-end 2025.
Musk predicts systems smarter than the smartest human by 2025 or 2026 ↗Reuters · Apr 2024
A short, testable deadline that can be compared with later statements.
Musk says systems may exceed any human as soon as next year ↗Associated Press · May 2026
The near-term deadline is renewed after the earlier window passes without a shared milestone.
A new frontier: spatial intelligence ↗World Labs · Sep 2024
Li’s move from language-centric systems toward persistent, controllable 3D worlds.
Generating bigger and better worlds ↗World Labs · Sep 2025
Marble turns the spatial thesis into a public, navigable product.
AI as normal technology ↗Arvind Narayanan & Sayash Kapoor · Apr 2025
A slow-diffusion counter-thesis centered on institutions rather than autonomous systems.
AGI is not a milestone ↗AI Snake Oil · May 2025
Argues that capability thresholds do not imply sudden economic or institutional change.
2026 AI Index: Economy ↗Stanford HAI · Apr 2026
Adoption is broad, agent deployment remains early, and infrastructure spending is at record levels.
Frontier risk report: Feb–Mar 2026 ↗METR · May 2026
Coding agents reach multi-day task horizons on the measured suite, with saturation caveats.
Trends in artificial intelligence ↗Epoch AI · Feb 2026
Tracks compute stock, training compute and data-center build constraints.
Statement to the U.S. Senate AI Insight Forum ↗Andrew Ng / U.S. Senate · Nov 2023
Frames the technology as general-purpose and argues for application-led diffusion.
DeepSeek CEO interview: The Quiet Giant ↗Waves / ChinaTalk translation · Jul 2024
Liang Wenfeng on open weights, original research, organizational design and DeepSeek’s AGI objective.
Alibaba Cloud’s Apsara Conference 2025 ↗Alibaba Group · Sep 2025
Eddie Wu ties Qwen openness to a $53B infrastructure commitment and full-stack distribution.
The AI Action Summit: A golden age of innovation ↗Google · Feb 2025
Pichai frames the shift as general-purpose, global and distribution-led.
AI at Scale Starts Here ↗AMD · Jun 2025
AMD’s open, heterogeneous-compute thesis, backed by an annual accelerator roadmap.
Thinking Machines and NVIDIA strategic partnership ↗Thinking Machines Lab · Mar 2026
Murati backs customizable systems with a multi-year, gigawatt-scale compute commitment.
Meta invests $14.3B in Scale AI and recruits its CEO ↗Associated Press · Jun 2025
The revealed bet behind Alexandr Wang’s authority inside Meta Superintelligence Labs.
Personal Superintelligence for Everyone ↗Meta · Jul 2025
Zuckerberg’s first-party statement of the personal-superintelligence and access thesis.
Introducing LawZero ↗Yoshua Bengio · Jun 2025
A career and institution-level pivot toward non-agentic, safe-by-design systems.
Unitree CEO on China’s robot revolution ↗TMT / ChinaTalk translation · Apr 2024
Wang Xingxing separates robotics hardware progress from the unresolved general-policy bottleneck.
With spatial intelligence, AI will understand the real world ↗TED · Apr 2024
Li’s public case for spatial intelligence as the bridge from seeing to acting.