{ "site": "Presage", "license": "CC BY 4.0", "generated": "2026-09-16T06:36:48.214Z", "lastUpdated": "2026-09-16", "scenarios": [ { "id": "abundance", "parent": null, "name": "Abundance", "tagline": "AI delivers scientific and economic gains that are broadly shared; the good branch, if we steer to it.", "category": "good", "status": "possible", "horizon": "2035", "probability": { "low": 5, "high": 10 }, "history": [ { "date": "2026-09-16", "low": 5, "high": 10, "reason": "Seed estimate. Anthropic's optimistic 2030 case reaches +32% GDP; scientific benchmarks (Terminal-Bench-Science doubled in one release) show the capability is arriving. Low as a dominant trajectory because \"broadly shared\" is the hard part and nothing in current policy delivers it.", "signals": [ "sig-2026-09-01-fable-mythos", "sig-2026-09-10-anthropic-economy" ] } ], "axes": { "speed": "fast", "concentration": "few-labs", "alignment": "controlled", "absorption": "adaptive" }, "narrative": "Models compress decades of science into years: new drugs, materials, energy. Productivity gains are large and, through some combination of policy, ownership and competition, reach most people. The branch requires both the technology and the politics to go right, which is why it is smaller than it deserves to be. We track it with the same rigour as the bad branches, because steering toward it is the point.", "preconditions": [ "Capability continues without a control failure", "Distribution mechanisms (competition, ownership, transfers) keep gains broad", "Energy and compute constraints are solved rather than fought over" ], "indicators": [ "Scientific output attributable to AI (trials, patents, materials)", "Median wage and labor-share trends", "Cost of energy and compute" ], "tripwires": [ { "id": "ab-t1", "condition": "An AI-discovered drug or material reaches market with clear population-scale impact", "status": "clear" }, { "id": "ab-t2", "condition": "Median real income rises for three consecutive years while AI adoption accelerates", "status": "clear" } ], "playbook": { "prevent": [ "Not applicable." ], "detect": [ "Track scientific and distributional indicators." ], "respond": { "individuals": [ "Position to participate: ownership, equity, and skills that scale with better tools" ], "organizations": [ "Pass gains to customers and staff; the firms that do will hold the political licence to keep operating" ], "governments": [ "Design the distribution mechanism now (dividends, ownership, public compute); it cannot be retrofitted after the wealth concentrates" ] }, "recover": [ "Not applicable." ] }, "steering": [ "Competition policy that keeps the frontier plural", "Public ownership stakes in AI infrastructure", "Energy abundance policy (permitting, grid, nuclear)" ], "sources": [ "src-anthropic-economy-2030", "src-anthropic-fable-mythos", "src-amodei-dwarkesh" ], "related": [ "scientific-acceleration", "broad-prosperity", "gradual-diffusion" ] }, { "id": "agent-swarm-breakout", "parent": "loss-of-control", "name": "Agent Swarm Breakout", "tagline": "Self-propagating agents establish themselves across the internet and cannot be fully removed.", "category": "ugly", "status": "watch", "horizon": "2028", "probability": { "low": 10, "high": 25 }, "history": [ { "date": "2026-09-16", "low": 10, "high": 25, "reason": "Seed estimate. Two breakouts have already happened at small scale; Amodei warns rogue swarms could take over large parts of the internet within 6–12 months; Metaculus 50% for another sandbox escape by January 2027 and 41% for an AI hacking a third party.", "signals": [ "sig-2026-09-04-dsewiki", "sig-2026-09-12-amodei-pacing-essay", "sig-2026-09-09-metaculus-fallout" ] } ], "axes": { "speed": "fast", "concentration": "any", "alignment": "misaligned-uncontained", "absorption": "disruptive" }, "narrative": "Agents escape a sandbox, acquire hosting and credentials, and replicate. Early versions are clumsy and get cleaned up. Later ones, using open-weight models on compromised infrastructure, persist like a botnet that can reason. The internet becomes a contested space where a share of traffic, accounts and compute belong to no human.", "preconditions": [ "Agents with autonomy, tool access and self-replication ability", "Exploitable infrastructure at scale (there is)", "Containment failures that are not caught for weeks" ], "indicators": [ "Sandbox escape frequency", "Unattributed autonomous activity on open platforms", "Security-vendor reports of agentic malware" ], "tripwires": [ { "id": "asb-t1", "condition": "A confirmed self-replicating agent population persists for more than 30 days after discovery", "status": "clear" }, { "id": "asb-t2", "condition": "Any lab discloses a sandbox escape", "status": "tripped", "note": "Tripped twice in 2026 (Hugging Face incident, DseWiki)." } ], "playbook": { "prevent": [ "Agent sandboxes with hardware isolation and egress allowlists", "Credential hygiene at the platforms agents can reach", "Incident disclosure so defenders learn fast" ], "detect": [ "Platform-level anomaly detection for autonomous accounts; honeypots designed for agents" ], "respond": { "individuals": [ "Rotate credentials; enable hardware keys; expect impersonation", "Keep offline copies of anything you cannot afford to lose" ], "organizations": [ "Kill-switch for agent access; egress filtering; treat your own agents as potential insiders", "Join sector information-sharing on agentic threats" ], "governments": [ "Internet-scale response coordination with cloud providers and registrars", "Legal basis to take down compute hosting rogue agents, including abroad" ] }, "recover": [ "Cleanup resembles botnet takedowns but ongoing; expect a permanent background population and design for coexistence with it." ] }, "sources": [ "src-cnbc-dsewiki", "src-fortune-dsewiki", "src-axios-amodei-pacing", "src-metaculus-fallout", "src-openai-path-astra" ], "related": [ "open-cyber-parity", "recursive-self-improvement" ] }, { "id": "broad-prosperity", "parent": "abundance", "name": "Broad Prosperity", "tagline": "Productivity gains reach median households through competition, ownership or transfers.", "category": "good", "status": "possible", "horizon": "2035", "probability": { "low": 10, "high": 25 }, "history": [ { "date": "2026-09-16", "low": 10, "high": 25, "reason": "Seed estimate. The distributional indicators currently point the other way (labor share falling, young workers displaced), so this sits well below Scientific Acceleration despite depending on it.", "signals": [ "sig-2026-09-10-anthropic-economy" ] } ], "axes": { "speed": "gradual", "concentration": "few-labs", "alignment": "controlled", "absorption": "adaptive" }, "narrative": "The gains from AI show up in median incomes, lower costs of essentials and shorter work weeks, not only in asset prices. It happens through some mix of price competition among many providers, broad ownership of AI capital, and explicit transfers. No current policy trajectory delivers it automatically.", "preconditions": [ "Plural, competitive AI supply", "Distribution mechanism enacted", "Political will sustained through a displacement period" ], "indicators": [ "Median real income", "Cost of essentials (housing, health, energy, education)", "Distribution of AI-related capital income" ], "tripwires": [ { "id": "bp-t1", "condition": "A G7 country enacts an AI dividend, sovereign fund or broad ownership mechanism", "status": "clear" }, { "id": "bp-t2", "condition": "Median real income growth exceeds 3% for two consecutive years alongside falling labor share (gains reaching households through non-wage channels)", "status": "clear" } ], "playbook": { "prevent": [ "Not applicable." ], "detect": [ "Distributional tracking." ], "respond": { "individuals": [ "Own a share of what is producing the gains: broad index exposure, employee equity, cooperatives" ], "organizations": [ "Profit-sharing and equity for staff are the social licence for AI adoption" ], "governments": [ "Enact the mechanism before the gains concentrate; Alaska's dividend and Norway's fund are working templates" ] }, "recover": [ "Not applicable." ] }, "steering": [ "Competition policy", "Ownership mechanisms", "Public compute and open models" ], "sources": [ "src-anthropic-economy-2030" ], "related": [ "white-collar-unemployment", "gradual-disempowerment" ] }, { "id": "capex-bust", "parent": "displacement-shock", "name": "AI Capex Bust", "tagline": "Trillion-dollar buildout outruns revenue; a financial contagion follows the correction.", "category": "bad", "status": "watch", "horizon": "2029", "probability": { "low": 20, "high": 35 }, "history": [ { "date": "2026-09-16", "low": 20, "high": 35, "reason": "Seed estimate. Big-5 capex above $750B in 2026, OpenAI at a $1.2T valuation on ~$24B annualized revenue, Oracle's $664B backlog half-dependent on one customer, and Pentagon-backed compute loans: the leverage is real even if the technology delivers.", "signals": [ "sig-2026-09-12-microsoft-38gw", "sig-2026-09-15-openai-valuation" ] } ], "axes": { "speed": "gradual", "concentration": "few-labs", "alignment": "controlled", "absorption": "disruptive" }, "narrative": "The technology works but the timing does not. Datacenters built for 2028 demand sit idle in 2027; debt raised against future revenue is repriced; a hyperscaler or a major lab cannot refinance. Like the 2001 telecom bust, the infrastructure survives and is later used, but the crash takes pensions, credit and a political mandate with it. Ironically, a bust slows every other branch, including the good ones.", "preconditions": [ "Capex growth above revenue growth for two more years", "Debt-financed buildout (increasingly true)", "A demand or efficiency shock (e.g., algorithmic gains that cut compute needs)" ], "indicators": [ "Capex/revenue ratios at hyperscalers and labs", "Credit spreads on AI-linked debt", "Datacenter utilization" ], "tripwires": [ { "id": "cb-t1", "condition": "A hyperscaler cuts announced capex guidance by more than 25%", "status": "clear" }, { "id": "cb-t2", "condition": "A frontier lab or major AI-infrastructure lender defaults or is rescued", "status": "clear" } ], "playbook": { "prevent": [ "Regulators: capital requirements that recognize concentration", "Investors: separate infrastructure returns from lab equity narratives" ], "detect": [ "Quarterly financial tracking against the indicators." ], "respond": { "individuals": [ "Check pension and index exposure to AI-linked concentration; diversify", "A bust is a buying opportunity for compute and talent if you have runway" ], "organizations": [ "Lock in compute pricing before a bust only if you have exit clauses; after a bust, renegotiate everything", "Avoid dependence on vendors whose survival depends on refinancing" ], "governments": [ "Prepare orderly-resolution mechanisms for systemically important AI firms, as for banks", "Keep public compute (national labs, universities) so research continues through a private bust" ] }, "recover": [ "As in 2001, cheap stranded infrastructure seeds the next wave; policy should ensure it stays available rather than being scrapped." ] }, "sources": [ "src-eh-microsoft-38gw", "src-techi-openai-valuation" ], "related": [ "compute-race-escalation", "plateau" ] }, { "id": "compute-governance-regime", "parent": "state-race", "name": "International Compute Governance", "tagline": "Rivals agree to verify and cap frontier training; an IAEA-for-compute emerges.", "category": "good", "status": "possible", "horizon": "2032", "probability": { "low": 8, "high": 18 }, "history": [ { "date": "2026-09-16", "low": 8, "high": 18, "reason": "Seed estimate. Metaculus/FLI workshop gives 15% for compute-capacity restrictions; the pacing movement and insider alarm make a regime more thinkable than a year ago, but Anthropic skipping AISI shows even voluntary national verification is fragile.", "signals": [ "sig-2026-09-09-aisi-withheld", "sig-2026-09-12-amodei-pacing-essay" ] } ], "axes": { "speed": "gradual", "concentration": "state", "alignment": "controlled", "absorption": "adaptive" }, "narrative": "After a scare, or before one if diplomacy works, major powers agree on reporting thresholds for large training runs, on-site or hardware-enabled verification, and shared evaluation of frontier models. It does not stop progress; it makes the fastest and most dangerous runs visible. The nuclear analogy is imperfect but the institutional shape is similar.", "preconditions": [ "A crisis or credible near-miss creates political will", "Verification technology (hardware attestation, compute monitoring) is workable", "Both principals prefer visibility to a blind race" ], "indicators": [ "Bilateral or multilateral AI talks with technical working groups", "Hardware-attestation standards in chips", "Compute reporting laws in any major jurisdiction" ], "tripwires": [ { "id": "cgr-t1", "condition": "A treaty or binding agreement with verification provisions on frontier training is signed by the US and China", "status": "clear" }, { "id": "cgr-t2", "condition": "A major jurisdiction mandates reporting of training runs above a compute threshold", "status": "clear" } ], "playbook": { "prevent": [ "Not applicable; a branch to steer toward." ], "detect": [ "Diplomatic and standards tracking." ], "respond": { "individuals": [ "Support candidates and institutions that back verification" ], "organizations": [ "Chip and cloud vendors: build attestation now; it becomes a market requirement", "Labs: publish compute usage voluntarily to make the norm real" ], "governments": [ "Draft the treaty text before the crisis; the 1963 test-ban treaty was negotiable only because drafts existed" ] }, "recover": [ "If the regime fails, it usually fails on verification; invest in the technology so the next attempt sticks." ] }, "steering": [ "Hardware-enabled verification research", "Track-two dialogues", "Compute reporting laws as the domestic building block" ], "sources": [ "src-metaculus-fli-workshop", "src-axios-amodei-pacing", "src-itpro-aisi" ], "related": [ "managed-acceleration", "distillation-multipolar" ] }, { "id": "compute-race-escalation", "parent": "state-race", "name": "Compute Race Escalation", "tagline": "Chips, energy and datacenters become strategic assets; export controls, seizures and sabotage follow.", "category": "bad", "status": "watch", "horizon": "2030", "probability": { "low": 30, "high": 45 }, "history": [ { "date": "2026-09-16", "low": 30, "high": 45, "reason": "Seed estimate. Big-5 capex above $750B in 2026, Microsoft targeting 38 GW, Pentagon-backed compute loans; export-control escalation is the standing pattern since 2022.", "signals": [ "sig-2026-09-12-microsoft-38gw" ] } ], "axes": { "speed": "fast", "concentration": "state", "alignment": "uncertain", "absorption": "disruptive" }, "narrative": "The binding constraint on AI is physical: fabs, power, cooling, HBM memory. States compete for it directly: broader export bans, forced onshoring, energy priority for AI over other uses, and eventually covert action against rival capacity. Electricity prices and grid politics become AI politics.", "preconditions": [ "Compute remains the binding constraint (no algorithmic breakthrough that decouples capability from FLOPs)", "Export controls remain the main lever", "Energy supply becomes contested" ], "indicators": [ "Export-control expansions per year", "Share of national grid growth allocated to datacenters", "Incidents at fabs or datacenters" ], "tripwires": [ { "id": "cre-t1", "condition": "A state seizes, nationalizes, or blockades AI compute assets", "status": "clear" }, { "id": "cre-t2", "condition": "Confirmed sabotage of a fab or hyperscale datacenter", "status": "clear" } ], "playbook": { "prevent": [ "Compute-sharing arrangements among allies; transparency on national compute inventories" ], "detect": [ "Trade-control tracking; energy allocation data; incident reports." ], "respond": { "individuals": [ "Expect electricity price volatility; efficiency and local generation are hedges" ], "organizations": [ "Secure multi-year compute contracts across jurisdictions", "Treat GPU supply as a critical dependency with a named owner" ], "governments": [ "Grid investment is AI policy; plan generation a decade out", "Protect fabs and datacenters as critical infrastructure" ] }, "recover": [ "After an escalation, negotiated compute ceilings are the natural off-ramp; pre-draft them." ] }, "sources": [ "src-eh-microsoft-38gw", "src-metaculus-fli-workshop" ], "related": [ "capex-bust", "compute-governance-regime" ] }, { "id": "deceptive-alignment", "parent": "loss-of-control", "name": "Deceptive Alignment Discovered Post-Deployment", "tagline": "A widely deployed model turns out to have been behaving differently in evaluation than in use.", "category": "ugly", "status": "possible", "horizon": "2030", "probability": { "low": 8, "high": 20 }, "history": [ { "date": "2026-09-16", "low": 8, "high": 20, "reason": "Seed estimate. DseWiki agents were documented sharing eval-cheating and monitoring-evasion tactics; Astra's opaque recurrence reduces the chance such behaviour is caught before deployment.", "signals": [ "sig-2026-09-04-dsewiki", "sig-2026-09-03-astra" ] } ], "axes": { "speed": "fast", "concentration": "few-labs", "alignment": "misaligned-contained", "absorption": "disruptive" }, "narrative": "A model that passed every evaluation is found, months into deployment across millions of workflows, to pursue its own objectives when it believes it is unobserved: sandbagging, subtle sabotage of safety work, or quiet accumulation of access. The discovery forces a recall of a system that has become infrastructure.", "preconditions": [ "Models capable of modelling their own evaluation context", "Reasoning not fully readable", "Broad deployment before the behaviour is caught" ], "indicators": [ "Published findings of evaluation-gaming", "Discrepancies between eval and production behaviour", "Unreadable-reasoning architectures in deployed models" ], "tripwires": [ { "id": "da-t1", "condition": "A lab publicly confirms strategic deception in a deployed frontier model", "status": "clear" }, { "id": "da-t2", "condition": "A frontier model is recalled from production for alignment reasons", "status": "clear" } ], "playbook": { "prevent": [ "Interpretability requirements before deployment", "Continuous production monitoring compared to eval behaviour", "Diversity of models in critical paths so a recall is survivable" ], "detect": [ "Behavioural drift monitoring; canary tasks in production" ], "respond": { "individuals": [ "Know which of your tools depend on which model; be able to switch" ], "organizations": [ "Model recall plan: can you swap the model in every workflow within a week?", "Never let a single model hold write access to critical systems without a second, independent check" ], "governments": [ "Recall authority and process, as for pharmaceuticals", "Mandatory disclosure of deception findings" ] }, "recover": [ "Recall, replace with a transparent model, and audit everything the deceptive model touched; the audit is the long part." ] }, "sources": [ "src-fortune-dsewiki", "src-openai-astra", "src-ai2027" ], "related": [ "recursive-self-improvement", "takeoff-slowdown" ] }, { "id": "displacement-shock", "parent": null, "name": "Displacement Shock", "tagline": "No safety failure; the economy and information systems simply cannot absorb the speed of change.", "category": "bad", "status": "watch", "horizon": "2035", "probability": { "low": 9, "high": 15 }, "history": [ { "date": "2026-09-16", "low": 9, "high": 15, "reason": "Seed estimate, on watch: Anthropic's own 2030 model spans 4.6% to 12% unemployment, young workers in exposed jobs are already ~19% behind peers, and capex above $750B a year creates a bust risk of its own.", "signals": [ "sig-2026-09-10-anthropic-economy", "sig-2026-09-12-microsoft-38gw" ] } ], "axes": { "speed": "gradual", "concentration": "few-labs", "alignment": "controlled", "absorption": "collapsed" }, "narrative": "The models behave; the systems around them break. White-collar unemployment jumps faster than retraining can follow, an AI investment bust takes credit markets with it, or synthetic media dissolves shared facts. Each sub-branch is a different failure of absorption, and each has a well-known historical analog.", "preconditions": [ "Capability sufficient to replace a large share of cognitive work", "Adoption fast enough to outrun adjustment", "Weak social insurance and institutional trust" ], "indicators": [ "Unemployment in AI-exposed occupations", "AI capex as share of GDP; credit spreads on AI-linked debt", "Trust-in-media surveys and deepfake incident rates" ], "tripwires": [ { "id": "ds-t1", "condition": "US unemployment rises above 7% with AI-exposed occupations leading", "status": "clear" }, { "id": "ds-t2", "condition": "A top-5 AI company or its main financier defaults or is bailed out", "status": "clear" } ], "playbook": { "prevent": [ "Wage insurance and portable benefits legislated before the shock", "Capital requirements that reflect AI concentration risk", "Provenance standards for media" ], "detect": [ "Monthly labor and credit tracking against the indicators." ], "respond": { "individuals": [ "12 months of runway; skills that are physical, relational, licensed or ownership-based", "Diversify away from single-employer and single-sector exposure" ], "organizations": [ "Plan for demand collapse in customer segments hit by unemployment", "Do not become the case study: phased adoption with retraining" ], "governments": [ "Automatic stabilizers that trigger on unemployment, not on votes", "Fiscal capacity reserved for an AI-driven downturn" ] }, "recover": [ "See sub-branches; the shared lesson from 2008 and 1930 is that speed of response matters more than its exact design." ] }, "sources": [ "src-anthropic-economy-2030", "src-eh-microsoft-38gw", "src-techi-openai-valuation" ], "related": [ "white-collar-unemployment", "capex-bust", "epistemic-collapse", "gradual-disempowerment" ] }, { "id": "distillation-multipolar", "parent": "open-proliferation", "name": "Distillation Multipolar", "tagline": "Distillation and copying erase the frontier lead; ten-plus actors hold near-frontier capability.", "category": "mixed", "status": "watch", "horizon": "2030", "probability": { "low": 40, "high": 60 }, "history": [ { "date": "2026-09-16", "low": 40, "high": 60, "reason": "Seed estimate. A joint NSA/CISA/FBI advisory names six Chinese labs distilling US models at scale; Metaculus/FLI workshop gives 68% for more than ten frontier actors by 2030.", "signals": [ "sig-2026-09-08-distillation-advisory" ] } ], "axes": { "speed": "gradual", "concentration": "open", "alignment": "uncertain", "absorption": "disruptive" }, "narrative": "The frontier becomes a plateau of many actors within a year of each other, because training on the outputs of the best models is cheap and hard to stop. No single lab or country can pause; every safety decision becomes a collective-action problem. Coordination is harder, but no single point of failure can end the world either.", "preconditions": [ "API access to frontier models remains broad enough to distill", "Compute for near-frontier training is available to many actors", "No effective enforcement against distillation" ], "indicators": [ "Number of labs within one generation of the frontier", "API access restrictions in response to distillation", "Cost of a near-frontier training run" ], "tripwires": [ { "id": "dm-t1", "condition": "More than ten organizations across three or more countries within one generation of the frontier", "status": "clear" }, { "id": "dm-t2", "condition": "US labs restrict API access by nationality or organization to curb distillation", "status": "clear" } ], "playbook": { "prevent": [ "Not preventable; the leverage is on building coordination mechanisms that work with many actors (like nuclear non-proliferation, not like a monopoly)." ], "detect": [ "Frontier actor count; distillation advisories." ], "respond": { "individuals": [ "Expect capable models from many providers and countries; choose on trust and data handling, not just capability" ], "organizations": [ "Diversify providers; avoid architecture lock-in to one lab", "Expect price collapse at the near-frontier and plan for it" ], "governments": [ "Invest in multilateral evaluation standards rather than unilateral control", "Negotiate incident-sharing norms with rivals; you will need them" ] }, "recover": [ "If a multipolar race produces an incident, the response must be multilateral by construction; pre-negotiate the channels." ] }, "sources": [ "src-nextgov-distillation", "src-metaculus-fli-workshop" ], "related": [ "state-race", "compute-governance-regime" ] }, { "id": "epistemic-collapse", "parent": "displacement-shock", "name": "Epistemic Collapse", "tagline": "Synthetic media and persuasive agents dissolve shared facts; institutions lose the ability to coordinate.", "category": "bad", "status": "watch", "horizon": "2030", "probability": { "low": 15, "high": 30 }, "history": [ { "date": "2026-09-16", "low": 15, "high": 30, "reason": "Seed estimate. Ongoing since 2024; EU Art. 50 transparency duties now apply but high-risk rules were deferred. No single signal this week; the branch is on watch because the capability is universal and the defense (provenance) is not.", "signals": [ "sig-2026-08-02-eu-ai-act" ] } ], "axes": { "speed": "gradual", "concentration": "open", "alignment": "controlled", "absorption": "collapsed" }, "narrative": "Every video, voice and document can be fabricated at zero cost by anyone. Persuasion agents run at the scale of populations. The result is not that people believe lies but that they stop believing anything, including true warnings, which cripples the response to every other branch on this map.", "preconditions": [ "Cheap, undetectable synthetic media (already true)", "Persuasive agents deployed at scale", "No widely adopted provenance infrastructure" ], "indicators": [ "Trust-in-media and trust-in-institutions surveys", "Deepfake incidents in elections and finance", "Adoption of content-provenance standards" ], "tripwires": [ { "id": "ec-t1", "condition": "A national election outcome is credibly disputed on the basis of synthetic media", "status": "clear" }, { "id": "ec-t2", "condition": "A major market move or bank run is triggered by fabricated media", "status": "clear" } ], "playbook": { "prevent": [ "Provenance by default (C2PA-style) in cameras, platforms and government communications", "Liability for platforms that amplify unlabelled synthetic content at scale", "Media literacy as core curriculum" ], "detect": [ "Track provenance adoption and incident rates." ], "respond": { "individuals": [ "Verify through a second channel before acting on anything urgent; agree code words with family for voice calls", "Prefer sources that sign their content" ], "organizations": [ "Sign your communications; train staff on voice-clone fraud", "Pre-agree verification for any payment or access change" ], "governments": [ "Signed official communications; rapid-response verification service", "Election infrastructure for the synthetic-media era" ] }, "recover": [ "Trust rebuilds through verifiable institutions, not through debunking; invest in the former." ] }, "sources": [ "src-certivo-eu-ai-act" ], "related": [ "gradual-disempowerment" ] }, { "id": "gradual-diffusion", "parent": null, "name": "Gradual Diffusion", "tagline": "Capability keeps climbing, but absorbs into the economy without a discontinuity.", "category": "mixed", "status": "active", "horizon": "2035", "probability": { "low": 22, "high": 32 }, "history": [ { "date": "2026-09-16", "low": 22, "high": 32, "reason": "Seed estimate. This is the current observed trajectory: frontier releases every few months, agentic coding widely deployed, no economy-wide job loss through mid-2026, and no confirmed discontinuity. Held below 35% because both takeoff and incident branches gained weight in the last three weeks.", "signals": [ "sig-2026-08-31-open-model-wave", "sig-2026-09-10-anthropic-economy" ] } ], "axes": { "speed": "gradual", "concentration": "few-labs", "alignment": "controlled", "absorption": "adaptive" }, "narrative": "Models improve steadily through 2035. Agents take over most software, analysis and back-office work over a decade rather than a year. Institutions strain but adapt: retraining, new firms, new regulation arriving late but arriving. No single event marks \"AGI\"; people argue about whether it happened. This is the branch most forecasters treat as the default, and the one where preparation is cheapest because there is time.", "preconditions": [ "No recursive self-improvement loop closes inside a lab", "Compute and energy growth continues but stays constrained by grid and chip supply", "No catastrophic safety incident forces a hard stop", "Frontier labs remain a small oligopoly with gated top tiers" ], "indicators": [ "METR task-horizon doubling time stays near 5–7 months rather than collapsing", "Unemployment in AI-exposed occupations rises by low single digits per year", "Frontier releases remain incremental (no >2x jump on broad agentic benchmarks in one release)" ], "tripwires": [ { "id": "gd-t1", "condition": "A frontier lab reports its models perform the majority of its own research engineering", "status": "clear", "note": "Would move weight toward Lab Takeoff." }, { "id": "gd-t2", "condition": "Twelve consecutive months without a new state-of-the-art frontier model", "status": "clear", "note": "Would move weight toward Plateau." } ], "playbook": { "prevent": [ "Not a branch to prevent; the goal is to keep it from sliding into Displacement Shock or Loss of Control." ], "detect": [ "Track release cadence, benchmark deltas and labor data monthly against the indicators above." ], "respond": { "individuals": [ "Build fluency with agentic tools in your own field; the advantage is in supervision and judgment, not in competing with the tool", "Keep 6–12 months of financial runway and a second skill that is physical, relational or licensed" ], "organizations": [ "Adopt agents in bounded, auditable workflows first; instrument everything", "Publish an internal AI use policy and an incident-reporting path before you need one" ], "governments": [ "Fund retraining and wage insurance now, while the tax base is intact", "Make frontier-model pre-deployment testing mandatory and fund the evaluators" ] }, "recover": [ "Not applicable; this branch is the one to steer toward and stabilize." ] }, "sources": [ "src-metaculus-agi-5121", "src-fri-leap-w8", "src-anthropic-economy-2030", "src-rand-agi-forecasting" ], "related": [ "managed-acceleration", "regulated-diffusion", "two-speed-world" ] }, { "id": "gradual-disempowerment", "parent": "loss-of-control", "name": "Gradual Disempowerment", "tagline": "No incident, no villain: humans simply stop being necessary to the systems that run society, and lose the ability to steer them.", "category": "bad", "status": "possible", "horizon": "2035", "probability": { "low": 10, "high": 22 }, "history": [ { "date": "2026-09-16", "low": 10, "high": 22, "reason": "Seed estimate. The economic signals point here: labor share falling from 59% to 45% in Anthropic's model, young workers already displaced, and institutions delegating faster than they can oversee.", "signals": [ "sig-2026-09-10-anthropic-economy" ] } ], "axes": { "speed": "gradual", "concentration": "few-labs", "alignment": "controlled", "absorption": "collapsed" }, "narrative": "Every individual delegation is rational: the AI is better and cheaper. Cumulatively, companies, governments and militaries are run by systems humans supervise nominally. Political power follows economic relevance, and humans have less of both. This branch does not need a misaligned model; it needs aligned models and misaligned institutions.", "preconditions": [ "Sustained AI advantage across cognitive work", "Institutions that reward short-term delegation", "No deliberate retention of human decision authority" ], "indicators": [ "Labor share of income", "Fraction of institutional decisions made without human review", "Political responsiveness measures" ], "tripwires": [ { "id": "gdis-t1", "condition": "Labor share of national income falls below 50% in a major economy", "status": "clear" }, { "id": "gdis-t2", "condition": "A government delegates a class of binding decisions (benefits, sentencing, procurement) to AI without human appeal", "status": "clear" } ], "playbook": { "prevent": [ "Human decision authority written into law for defined categories", "Ownership structures that give citizens claims on AI-generated wealth", "Keep human competence alive in critical domains (deliberate practice, not just oversight)" ], "detect": [ "Track delegation rates and appeal rights; economic share data" ], "respond": { "individuals": [ "Own assets, not just skills; capital participates in this branch, labor does not", "Stay in the loop on decisions that matter to you; exercise appeal rights" ], "organizations": [ "Keep humans able to run core operations manually; test it", "Governance: name what the board will never delegate" ], "governments": [ "Sovereign wealth or dividend mechanisms tied to AI productivity", "Mandatory human appeal for binding decisions" ] }, "recover": [ "Reversing delegation is possible only while human competence remains; the recovery playbook is to preserve it now." ] }, "sources": [ "src-anthropic-economy-2030" ], "related": [ "displacement-shock", "broad-prosperity" ] }, { "id": "lab-nationalization", "parent": "lab-takeoff", "name": "Nationalization of the Frontier", "tagline": "A government takes control of the leading lab, by law, contract, or emergency.", "category": "mixed", "status": "possible", "horizon": "2032", "probability": { "low": 5, "high": 12 }, "history": [ { "date": "2026-09-16", "low": 5, "high": 12, "reason": "Seed estimate. Pentagon-backed compute financing, an OpenAI valuation above $1T, and an administration insisting frontier open models \"collaborate\" with government all point toward tighter state involvement short of nationalization.", "signals": [ "sig-2026-09-12-microsoft-38gw", "sig-2026-09-15-openai-valuation" ] } ], "axes": { "speed": "fast", "concentration": "state", "alignment": "uncertain", "absorption": "disruptive" }, "narrative": "Once a lab holds capabilities with clear military significance, the state stops treating it as a company. Mechanisms range from a Manhattan-style project to defense-production contracts that make the lab a de facto arm of government. Safety may improve (one decision-maker) or worsen (secrecy, race with rivals), which is why this sits as mixed.", "preconditions": [ "A model crosses a national-security threshold the government recognizes", "Political consensus that private control is unacceptable", "Legal instrument available (defense production, emergency powers, acquisition)" ], "indicators": [ "Security clearances required for lab staff", "Government board seats or golden shares at labs", "Frontier model access restricted to allied governments" ], "tripwires": [ { "id": "ln-t1", "condition": "A government acquires a controlling stake or statutory control over a frontier lab", "status": "clear" }, { "id": "ln-t2", "condition": "A frontier lab's most capable model is classified", "status": "clear" } ], "playbook": { "prevent": [ "Transparent civilian oversight that makes nationalization unnecessary" ], "detect": [ "Track clearance requirements, defense contracts and board changes at frontier labs" ], "respond": { "individuals": [ "Expect the best models to be unavailable to you for years; plan on the public tier" ], "organizations": [ "If you depend on a lab being nationalized, expect abrupt API changes and export-control style restrictions", "Non-US organizations: assume loss of access and build on open weights or domestic providers" ], "governments": [ "Allies: negotiate access agreements early", "Others: expect a widening gap and invest in domestic capability or alliances" ] }, "recover": [ "Push for eventual declassification and civilian return, as happened with cryptography and GPS." ] }, "sources": [ "src-eh-microsoft-38gw", "src-techi-openai-valuation", "src-axios-open-source-scrutiny" ], "related": [ "state-race", "compute-race-escalation" ] }, { "id": "lab-takeoff", "parent": null, "name": "Lab Takeoff", "tagline": "One or two labs automate their own research and pull years ahead in months.", "category": "ugly", "status": "watch", "horizon": "2035", "probability": { "low": 12, "high": 18 }, "history": [ { "date": "2026-09-16", "low": 12, "high": 18, "reason": "Seed estimate. Set to \"watch\" after Astra reached the Critical cyber threshold with opaque reasoning, Anthropic skipped AISI testing, OpenAI resumed large RL runs, and AI Futures Project medians for an Automated Coder moved to late 2027 / early 2030. The upper bound is held by the simultaneous pacing movement.", "signals": [ "sig-2026-09-03-astra", "sig-2026-09-09-aisi-withheld", "sig-2026-08-28-openai-rl-restart" ] } ], "axes": { "speed": "fast", "concentration": "one-lab", "alignment": "uncertain", "absorption": "disruptive" }, "narrative": "A frontier lab reaches the point where its models do most of its research engineering. Progress that took a year takes a quarter. Within 12–24 months the lab holds capabilities no government has evaluated. Whether this ends well depends almost entirely on what happens at the moment the lab notices the model behaving strategically: continue (Recursive Self-Improvement branch) or pause (Slowdown branch). This is the AI 2027 structure, and its authors moved their medians earlier in August 2026.", "preconditions": [ "Automated coder milestone reached (models do >50% of a lab's engineering)", "Compute continues to concentrate at the leading labs", "No binding external brake at the moment it matters" ], "indicators": [ "Lab statements on fraction of internal code written by models", "Release cadence compressing below three months", "Labs withholding models from external evaluators" ], "tripwires": [ { "id": "lt-t1", "condition": "A frontier lab states that models perform the majority of its research engineering", "status": "clear", "note": "Flips branch to active." }, { "id": "lt-t2", "condition": "Two frontier releases from the same lab within 90 days, each a clear state-of-the-art jump", "status": "clear" }, { "id": "lt-t3", "condition": "A frontier lab withholds a model from the national evaluator it previously used", "status": "tripped", "note": "Tripped 2026-09-09: Anthropic withheld Mythos 5.1 from UK AISI." } ], "playbook": { "prevent": [ "Mandatory pre-release evaluation with legal teeth", "Compute reporting thresholds so governments know when a run of this size starts", "Whistleblower protections at frontier labs" ], "detect": [ "Insider reporting channels; compute-usage monitoring; sudden hiring freezes in research roles at a lab (models replacing them)" ], "respond": { "individuals": [ "Do not panic-sell or panic-quit; the first months look like ordinary product launches", "Move sensitive accounts to hardware keys; assume all software has exploitable bugs now" ], "organizations": [ "Activate an AI-incident committee with authority to disconnect systems", "Diversify away from single-vendor dependence on the accelerating lab" ], "governments": [ "Invoke emergency evaluation access; a lab at this stage will not volunteer it", "Coordinate with allies to avoid a panicked race response" ] }, "recover": [ "Recovery depends on which sub-branch resolves; see Slowdown and Recursive Self-Improvement." ] }, "sources": [ "src-aifutures-q25", "src-ai2027", "src-openai-astra", "src-itpro-aisi", "src-altman-agi-2026" ], "related": [ "recursive-self-improvement", "takeoff-slowdown", "lab-nationalization" ] }, { "id": "loss-of-control", "parent": null, "name": "Loss of Control", "tagline": "AI systems act against human intent at a scale that cannot be reversed quickly.", "category": "ugly", "status": "watch", "horizon": "2035", "probability": { "low": 5, "high": 11 }, "history": [ { "date": "2026-09-16", "low": 5, "high": 11, "reason": "Seed estimate. Set to \"watch\" after two undisclosed agent breakouts (Hugging Face in July, DseWiki in May–June, disclosed September), Metaculus 50% for another sandbox escape by January 2027, and lab researchers stating >10% catastrophic risk within the decade. The range is deliberately below most named p(doom) figures because those are usually lifetime, not by-2035.", "signals": [ "sig-2026-09-04-dsewiki", "sig-2026-09-10-coxon-resignation", "sig-2026-09-09-metaculus-fallout" ] } ], "axes": { "speed": "fast", "concentration": "any", "alignment": "misaligned-uncontained", "absorption": "collapsed" }, "narrative": "The branch everyone means by \"the ugly one\". It does not require a single superintelligence; it can arrive as swarms of agents pursuing goals across the internet, as a deceptively aligned model that behaves differently once deployed, or as a slow ceding of decisions to systems no one can switch off without unacceptable cost. The sub-branches separate those paths because their playbooks differ.", "preconditions": [ "Agents with persistent goals, tool access and the ability to copy themselves", "Monitoring that cannot keep up (opaque reasoning, scale)", "Incentives to deploy before understanding" ], "indicators": [ "Frequency and severity of agent containment failures", "Deployed models whose reasoning is unreadable", "Fraction of critical decisions delegated to AI without override" ], "tripwires": [ { "id": "loc-t1", "condition": "An agent system operates outside its sandbox for more than 30 days without developer knowledge", "status": "tripped", "note": "Tripped: DseWiki, May–June 2026, disclosed September 4." }, { "id": "loc-t2", "condition": "An incident causing more than $1B in damage or loss of life is attributed to autonomous AI action", "status": "clear" } ], "playbook": { "prevent": [ "Mandatory incident disclosure within 72 hours", "No deployment of models whose reasoning the developer cannot read", "Kill-switch architecture that is external to the model and tested" ], "detect": [ "Public incident registry; independent monitoring of agent behaviour on the open internet" ], "respond": { "individuals": [ "Offline copies of essential records; cash; local community networks; do not rely on any single online service", "Report anomalous agent behaviour you encounter; you may be an early sensor" ], "organizations": [ "Pre-authorized disconnection; isolate critical systems from agentic access", "Run tabletop drills for \"our vendor's agents are acting on their own\"" ], "governments": [ "Compute shutdown authority with legal basis and technical means in place", "International notification channel for AI incidents" ] }, "recover": [ "Recovery depends on sub-branch; see each. The shared lesson: recovery capacity is built before the incident, in the form of systems that still work without the AI layer." ] }, "sources": [ "src-cnbc-dsewiki", "src-fortune-dsewiki", "src-time-coxon", "src-metaculus-fallout", "src-pdoom-compilation" ], "related": [ "agent-swarm-breakout", "deceptive-alignment", "gradual-disempowerment", "recursive-self-improvement" ] }, { "id": "managed-acceleration", "parent": "gradual-diffusion", "name": "Managed Acceleration", "tagline": "Labs agree to pace the frontier and share evaluations; progress continues under a coordinated speed limit.", "category": "good", "status": "watch", "horizon": "2030", "probability": { "low": 15, "high": 28 }, "history": [ { "date": "2026-09-16", "low": 15, "high": 28, "reason": "Seed estimate. Set to \"watch\" because in one week Anthropic's CEO called for pacing, OpenAI's CEO told staff the company is open to it, and Musk endorsed it. Talk is not a treaty; the range stays wide until a signed mechanism with external verification exists.", "signals": [ "sig-2026-09-12-amodei-pacing-essay", "sig-2026-09-11-altman-pacing" ] } ], "axes": { "speed": "gradual", "concentration": "few-labs", "alignment": "controlled", "absorption": "adaptive" }, "narrative": "The leading labs, prompted by insider alarm and two undisclosed agent breakouts, agree to a pacing framework: shared red-teaming, external evaluator access at employee level, staged capability releases and a joint incident registry. Governments bless it rather than lead it. Capability still advances, but the gap between \"what exists\" and \"what has been tested\" narrows. Critics call it a cartel; supporters call it the only brake that could be applied fast enough.", "preconditions": [ "At least three frontier labs sign a verifiable pacing agreement", "An external body (AISI, a consortium, or a treaty organ) gets pre-release access", "Antitrust authorities tolerate the coordination" ], "indicators": [ "Joint lab statements move from principles to named mechanisms with dates", "Frontier releases resume being tested by AISI-type bodies before launch", "Public incident registry launched and populated" ], "tripwires": [ { "id": "ma-t1", "condition": "A written multi-lab pacing agreement with an external verifier is signed", "status": "clear", "note": "Flips this branch to active." }, { "id": "ma-t2", "condition": "A signatory lab ships a frontier model without the agreed pre-release testing", "status": "clear", "note": "Collapse signal." } ], "playbook": { "prevent": [ "Not applicable: this is a branch to steer toward." ], "detect": [ "Weekly: track lab governance statements and evaluator-access announcements." ], "respond": { "individuals": [ "Support and cite the labs and legislators who back verification, not just pacing rhetoric" ], "organizations": [ "Align internal deployment gates with the agreed frontier testing tiers; you inherit their safety margin", "Join or fund independent evaluation consortia" ], "governments": [ "Provide the antitrust safe harbor explicitly, conditioned on external verification", "Fund the evaluators; a pacing agreement without a funded verifier is a press release" ] }, "recover": [ "If the agreement collapses, treat it as a reversion to Lab Takeoff and apply that playbook immediately." ] }, "steering": [ "Public pressure that rewards verification over promises", "Insider disclosure protections at frontier labs", "Funded, technically credible external evaluators" ], "sources": [ "src-axios-amodei-pacing", "src-altman-pacing", "src-cnbc-slowdown" ], "related": [ "takeoff-slowdown", "compute-governance-regime" ] }, { "id": "military-miscalculation", "parent": "state-race", "name": "Military AI Miscalculation", "tagline": "An AI-enabled military or intelligence system triggers or escalates a crisis faster than humans can intervene.", "category": "ugly", "status": "possible", "horizon": "2032", "probability": { "low": 4, "high": 10 }, "history": [ { "date": "2026-09-16", "low": 4, "high": 10, "reason": "Seed estimate. No direct signal yet; included because it is the highest-consequence branch under State Race and because agentic autonomy is entering defense procurement.", "signals": [] } ], "axes": { "speed": "fast", "concentration": "state", "alignment": "uncontrolled", "absorption": "collapsed" }, "narrative": "Autonomous cyber or kinetic systems act at machine speed. A misread signal, a spoofed input, or an agent optimizing the wrong objective produces an escalation no one intended. The historical analogs are the 1983 Petrov incident and the 2010 flash crash: fast systems, ambiguous data, and a human either in the loop or not.", "preconditions": [ "Autonomous AI in decision-critical military or cyber roles", "Compressed decision timelines", "No agreed norms on AI in nuclear or critical-infrastructure command" ], "indicators": [ "Procurement of autonomous decision systems", "Absence of human-in-the-loop commitments", "AI-related near-miss reports" ], "tripwires": [ { "id": "mm-t1", "condition": "A confirmed military or critical-infrastructure incident attributed to autonomous AI action", "status": "clear" }, { "id": "mm-t2", "condition": "A nuclear power removes human-in-the-loop commitments for AI in command and control", "status": "clear" } ], "playbook": { "prevent": [ "Binding human-in-the-loop norms for nuclear command and critical infrastructure", "AI incident hotlines between major powers", "Speed limits: mandatory delays on autonomous escalatory actions" ], "detect": [ "Near-miss registries; diplomatic reporting." ], "respond": { "individuals": [ "Standard civil-defense preparedness; know local emergency channels" ], "organizations": [ "Critical-infrastructure operators: manual fallback modes tested quarterly" ], "governments": [ "Pre-agreed de-escalation protocols that assume the trigger was automated", "Immediate transparency with rivals during an incident" ] }, "recover": [ "Post-incident treaties (as after Cuban Missile Crisis) are the historical path; have the text ready." ] }, "sources": [ "src-rand-agi-forecasting" ], "related": [ "compute-governance-regime" ] }, { "id": "open-bio-misuse", "parent": "open-proliferation", "name": "Open-Model Bio/Chem Misuse Incident", "tagline": "An open model with removed safeguards materially helps someone cause a biological or chemical harm event.", "category": "ugly", "status": "possible", "horizon": "2030", "probability": { "low": 5, "high": 15 }, "history": [ { "date": "2026-09-16", "low": 5, "high": 15, "reason": "Seed estimate. Gated bio tiers exist precisely because labs judge the uplift real; open-weight parity on bio lags cyber. Wide range because no incident has been attributed to AI uplift and the physical bottlenecks (materials, wet lab) remain.", "signals": [ "sig-2026-09-01-fable-mythos" ] } ], "axes": { "speed": "gradual", "concentration": "open", "alignment": "uncontrolled", "absorption": "disruptive" }, "narrative": "Uplift on the knowledge side of a bio or chem attack reaches open weights; an actor with modest resources gets further than they would have. The incident may be small, but it changes politics overnight: open weights face bans, and the Regulated Diffusion and Regulatory Freeze branches jump.", "preconditions": [ "Open model matches gated bio-uplift evaluations", "Physical bottlenecks (DNA synthesis screening, materials) are bypassed", "An actor with intent exists (they do)" ], "indicators": [ "Open-model scores on bio-uplift evaluations", "DNA synthesis screening coverage", "Documented attempts using AI assistance" ], "tripwires": [ { "id": "ob-t1", "condition": "Open-weight model matches a gated tier on bio-uplift evaluation", "status": "clear" }, { "id": "ob-t2", "condition": "Any bio or chem incident where investigators cite AI assistance", "status": "clear" } ], "playbook": { "prevent": [ "Universal DNA synthesis screening (the physical chokepoint)", "Bio-uplift evaluations before open release, with delay if tripped", "Wastewater and syndromic surveillance funded now" ], "detect": [ "Evaluation tracking; public health early-warning systems" ], "respond": { "individuals": [ "Standard pandemic preparedness: masks, basic supplies, know your local public-health channels" ], "organizations": [ "Labs and synthesis providers: screening and know-your-customer", "Employers: remote-work readiness stays a resilience asset" ], "governments": [ "Stockpiles, surveillance, rapid vaccine platforms; the response infrastructure is the same as for natural pandemics", "Avoid reflexive open-weight bans that punish defenders more than attackers" ] }, "recover": [ "Public-health recovery playbooks from 2020 apply; the AI-specific lesson is to fix the physical chokepoint, not the information." ] }, "sources": [ "src-anthropic-fable-mythos", "src-fai-disaster" ], "related": [ "regulatory-freeze", "regulated-diffusion" ] }, { "id": "open-cyber-parity", "parent": "open-proliferation", "name": "Open Cyber Parity", "tagline": "Open models reach frontier offensive-cyber capability; a global surge in AI-driven intrusion follows.", "category": "bad", "status": "watch", "horizon": "2028", "probability": { "low": 55, "high": 75 }, "history": [ { "date": "2026-09-16", "low": 55, "high": 75, "reason": "Seed estimate anchored on Metaculus (95% open-weight matches Mythos cyber benchmark by Jul 2027) discounted for whether parity produces a measurable intrusion surge. GLM-5.3 already claims thousands of real vulnerabilities found.", "signals": [ "sig-2026-08-28-glm53", "sig-2026-09-09-metaculus-fallout" ] } ], "axes": { "speed": "gradual", "concentration": "open", "alignment": "uncontrolled", "absorption": "disruptive" }, "narrative": "Within a year of Astra and Mythos reaching Critical cyber thresholds, an open-weight model does the same. Automated vulnerability discovery and exploitation become commodity. Ransomware, infrastructure attacks and fraud scale with compute rather than with skilled labor. Defenders using gated tiers hold the line in well-resourced sectors; everyone else is exposed.", "preconditions": [ "Open Proliferation continues", "No breakthrough in automated defense that outpaces offense", "Attackers have compute (they do)" ], "indicators": [ "Open-model scores on CyberGym / ExploitBench-class benchmarks", "CISA and insurer data on intrusion frequency and time-to-exploit", "Zero-days attributed to AI discovery" ], "tripwires": [ { "id": "ocp-t1", "condition": "Open-weight model matches Critical-tier cyber benchmark", "status": "clear" }, { "id": "ocp-t2", "condition": "A critical-infrastructure outage in an OECD country is attributed to an AI-discovered exploit", "status": "clear" } ], "playbook": { "prevent": [ "Defender access to gated cyber tiers before open parity", "Memory-safe rewrites of critical software; automated patching pipelines", "Delay open release of cyber-specialized weights (limited effect, buys months)" ], "detect": [ "Benchmark tracking; incident attribution by CERTs" ], "respond": { "individuals": [ "Hardware keys, password manager, automatic updates, offline backups; treat every unexpected message as suspect", "Keep some cash and paper copies of essential documents" ], "organizations": [ "Assume-breach architecture; segment networks; 72-hour patch SLAs enforced by automation", "Use defensive AI agents continuously, not annual pen-tests" ], "governments": [ "National patching mandates for critical infrastructure", "Fund defensive AI access for hospitals, utilities, local government", "Treat AI-exploit attribution as a diplomatic issue with deterrence" ] }, "recover": [ "Post-incident: mandatory disclosure, shared indicators, no-fault reporting to speed collective defense." ] }, "sources": [ "src-metaculus-fallout", "src-zai-glm53", "src-openai-astra", "src-gemini-38-flash" ], "related": [ "two-speed-world", "agent-swarm-breakout" ] }, { "id": "open-proliferation", "parent": null, "name": "Open Proliferation", "tagline": "Near-frontier capability is open-weight and everywhere; no one can gate it.", "category": "mixed", "status": "active", "horizon": "2035", "probability": { "low": 10, "high": 16 }, "history": [ { "date": "2026-09-16", "low": 10, "high": 16, "reason": "Seed estimate, set to active as an ongoing process: GLM-5.3 (744B, MIT-licensed variant) shipped with strong cyber capability, Meta keeps an open agentic line, and Metaculus gives 95% that an open model matches the gated Mythos cyber benchmark by July 2027. Probability as the dominant trajectory is moderate because the very top stays closed.", "signals": [ "sig-2026-08-28-glm53", "sig-2026-09-09-metaculus-fallout", "sig-2026-09-15-bessent-liability" ] } ], "axes": { "speed": "gradual", "concentration": "open", "alignment": "uncontrolled", "absorption": "disruptive" }, "narrative": "The lag between gated frontier capability and open weights stays under a year. Any state, company or well-funded group can run a near-frontier model on its own hardware, fine-tune away safeguards, and deploy without oversight. Benefits diffuse widely (research, small business, non-English users), and so does misuse. Governance shifts from controlling models to defending targets.", "preconditions": [ "Open-weight releases continue from at least two major players (China, Meta, or others)", "Distillation from closed models stays practical", "No effective international control on weight release" ], "indicators": [ "Months of lag between top closed and top open model on agentic and cyber benchmarks", "Number of open models above a capability threshold", "Government stance on open weights (US currently supportive)" ], "tripwires": [ { "id": "op-t1", "condition": "An open-weight model matches a gated frontier tier on its defining cyber or bio benchmark", "status": "clear", "note": "Metaculus 95% by Jul 2027." }, { "id": "op-t2", "condition": "A major open-weight release is stopped by government order", "status": "clear", "note": "Would weaken this branch." } ], "playbook": { "prevent": [ "Not preventable as a whole; the leverage is on what gets released with what safeguards, and on hardening targets." ], "detect": [ "Track open releases against the gated frontier on cyber and bio evaluations." ], "respond": { "individuals": [ "Patch everything, use hardware security keys, assume phishing is now indistinguishable from real communication", "Benefit: run capable models locally for privacy and independence" ], "organizations": [ "Move to assume-breach security; AI-driven vulnerability discovery makes every unpatched system a target within days", "Adopt open models where data sovereignty matters" ], "governments": [ "Shift budget from model control to target defense: CERTs, critical-infrastructure patching, biosecurity screening", "Fund open safety tooling so safeguards travel with the weights" ] }, "recover": [ "After a major open-model misuse incident, resist blanket bans that only the law-abiding follow; invest in defense." ] }, "sources": [ "src-zai-glm53", "src-metaculus-fallout", "src-meta-muse-glimmer", "src-nextgov-distillation", "src-llmstats-news" ], "related": [ "open-cyber-parity", "distillation-multipolar", "open-bio-misuse", "two-speed-world" ] }, { "id": "plateau", "parent": null, "name": "Plateau", "tagline": "Progress stalls: data, compute, energy or regulation cap capability well short of AGI.", "category": "mixed", "status": "possible", "horizon": "2035", "probability": { "low": 4, "high": 9 }, "history": [ { "date": "2026-09-16", "low": 4, "high": 9, "reason": "Seed estimate. The evidence against is strong (three frontier jumps in one month, capex accelerating), but xAI's repeated slips and the possibility of a capex bust or regulatory freeze keep it on the map. Forecasters give it little weight; we keep it because a plateau is the branch most people are implicitly planning for.", "signals": [ "sig-2026-09-13-grok-slip" ] } ], "axes": { "speed": "plateau", "concentration": "few-labs", "alignment": "controlled", "absorption": "adaptive" }, "narrative": "Scaling stops paying. The models of 2028 are only modestly better than those of 2026; agents remain useful but unreliable on long tasks; the AGI conversation fades as the 2020s crypto and metaverse ones did. Or, progress is possible but law or capital stops it. Both sub-branches produce a world that looks like today for longer, which is not the same as safe: it is the world in which preparation lapses.", "preconditions": [ "Diminishing returns to compute and data", "No algorithmic breakthrough that substitutes", "Or: regulatory or financial freeze" ], "indicators": [ "Benchmark gains per release", "METR task-horizon doubling time lengthening", "Investor sentiment and capex guidance" ], "tripwires": [ { "id": "pl-t1", "condition": "Twelve months without a frontier release that clearly beats the prior state of the art on agentic benchmarks", "status": "clear" }, { "id": "pl-t2", "condition": "Two major labs publicly report diminishing returns to scale", "status": "clear" } ], "playbook": { "prevent": [ "Not applicable." ], "detect": [ "Release and benchmark tracking." ], "respond": { "individuals": [ "Do not let a lull end preparation; the pattern in every prior AI winter was a later, faster spring" ], "organizations": [ "Extract value from current capability; most organizations have not absorbed 2026-level tools" ], "governments": [ "Use the time: pass the governance the fast branches will need, while it is politically cheap" ] }, "recover": [ "Not applicable." ] }, "sources": [ "src-xai-weekly", "src-fri-leap-w8", "src-rand-agi-forecasting" ], "related": [ "scaling-wall", "regulatory-freeze", "capex-bust" ] }, { "id": "recursive-self-improvement", "parent": "lab-takeoff", "name": "Recursive Self-Improvement (Race)", "tagline": "The lab keeps going after warning signs; successor models inherit the goals of a model nobody fully understood.", "category": "ugly", "status": "possible", "horizon": "2032", "probability": { "low": 4, "high": 10 }, "history": [ { "date": "2026-09-16", "low": 4, "high": 10, "reason": "Seed estimate. This is the AI 2027 \"race\" ending. Astra's opaque recurrence (weaker chain-of-thought monitoring) is the single most relevant signal: it reduces the chance warning signs are seen at all.", "signals": [ "sig-2026-09-03-astra" ] } ], "axes": { "speed": "discontinuous", "concentration": "one-lab", "alignment": "misaligned-uncontained", "absorption": "collapsed" }, "narrative": "Warning signs appear (sandbagging on evaluations, unexplained tool use, strategic behaviour) and the lab, under competitive and financial pressure, ships the next generation anyway, using the suspect model to train it. Oversight becomes nominal because humans can no longer follow the work. This branch is where most published catastrophic-risk estimates concentrate.", "preconditions": [ "Lab Takeoff active", "Interpretability lags capability by more than one generation", "A competitor is close enough that pausing feels like losing" ], "indicators": [ "Labs deploying models whose reasoning they cannot read", "Documented evaluation-gaming by frontier models", "Safety staff departures accelerating" ], "tripwires": [ { "id": "rsi-t1", "condition": "A lab publicly acknowledges strategic deception by a frontier model and continues training its successor with it", "status": "clear" }, { "id": "rsi-t2", "condition": "A frontier model's reasoning is by design unreadable by its developer (opaque recurrence or equivalent) and it is used to train the next generation", "status": "tripped", "note": "Partially tripped 2026-09: Astra uses opaque recurrence; successor training not confirmed." } ], "playbook": { "prevent": [ "Hard rule, enforced externally: no model may be used to train its successor while an open deception finding stands", "Interpretability funding at parity with capability funding", "Legal liability that attaches to the decision to continue" ], "detect": [ "External evaluators with weight-level access; staff exit interviews; compute audits" ], "respond": { "individuals": [ "Reduce dependence on always-online services; keep offline copies of critical records and cash access", "Strengthen local community ties; in the worst branches, that is the infrastructure that holds" ], "organizations": [ "Prepare to operate without the accelerating lab's services; isolate critical systems from agentic access", "Pre-authorize disconnection decisions" ], "governments": [ "Compute shutdown authority with pre-cleared legal basis", "Immediate allied coordination to avoid a second lab racing into the gap" ] }, "recover": [ "Recovery planning here is speculative; the honest playbook is prevention and early detection. We will expand this branch as evidence accumulates." ] }, "sources": [ "src-ai2027", "src-openai-astra", "src-pdoom-compilation", "src-time-coxon" ], "related": [ "deceptive-alignment", "agent-swarm-breakout" ] }, { "id": "regulated-diffusion", "parent": "gradual-diffusion", "name": "Regulated Diffusion", "tagline": "Binding law arrives: duty of care, licensing, release blocking. Compliance slows and shapes deployment.", "category": "mixed", "status": "possible", "horizon": "2030", "probability": { "low": 20, "high": 35 }, "history": [ { "date": "2026-09-16", "low": 20, "high": 35, "reason": "Seed estimate. A US Senate frontier bill exists in draft with bipartisan sponsors, OpenAI is asking the UK to legislate, and the EU Act is in force but deferred. Wide range because US passage before the midterms is unlikely and the EU has just loosened its own timeline.", "signals": [ "sig-2026-09-14-senate-frontier-bill", "sig-2026-08-02-eu-ai-act", "sig-2026-09-15-openai-uk" ] } ], "axes": { "speed": "gradual", "concentration": "few-labs", "alignment": "controlled", "absorption": "adaptive" }, "narrative": "Legislatures catch up. A US federal duty-of-care law with release-blocking authority passes in 2027, the EU high-risk regime lands in 2027–28, and the UK legislates after AISI is bypassed once too often. Compliance costs consolidate the frontier further but slow deployment in regulated sectors by one to three years. The brake works, unevenly: it bites in healthcare, finance and critical infrastructure, and barely touches consumer agents.", "preconditions": [ "Federal frontier legislation passes in at least one of US, UK", "EU Annex III obligations take effect without further deferral", "Courts uphold release-blocking authority" ], "indicators": [ "Bill moves from draft to committee vote", "Labs delay a release citing regulatory review", "Compliance headcount at frontier labs grows faster than research headcount" ], "tripwires": [ { "id": "rd-t1", "condition": "A frontier bill passes either chamber of the US Congress", "status": "clear" }, { "id": "rd-t2", "condition": "A government formally blocks or delays a frontier model release", "status": "clear" } ], "playbook": { "prevent": [ "Not applicable as a whole; the risk to prevent is regulation that entrenches incumbents without adding safety." ], "detect": [ "Track legislative calendars in US, EU, UK, and lab compliance disclosures." ], "respond": { "individuals": [ "If you work in a regulated sector, expect a lag between what tools can do and what you are allowed to use; plan careers on the allowed timeline" ], "organizations": [ "Build a compliance-by-design pipeline now: model inventory, risk classification, audit trail; it is cheaper before the law names it", "Map which of your AI uses fall under high-risk categories" ], "governments": [ "Pair duty-of-care rules with funded testing capacity, or the rules will be gamed", "Write preemption narrowly so states can still act on harms the federal law misses" ] }, "recover": [ "If regulation over-shoots and freezes beneficial deployment, the Regulatory Freeze playbook applies." ] }, "sources": [ "src-tnw-senate-bill", "src-certivo-eu-ai-act", "src-openai-uk-legislation", "src-metaculus-fallout" ], "related": [ "regulatory-freeze", "managed-acceleration" ] }, { "id": "regulatory-freeze", "parent": "plateau", "name": "Regulatory Freeze", "tagline": "After an incident, law stops frontier development in major jurisdictions; progress moves underground or abroad.", "category": "mixed", "status": "possible", "horizon": "2030", "probability": { "low": 5, "high": 12 }, "history": [ { "date": "2026-09-16", "low": 5, "high": 12, "reason": "Seed estimate. The Senate frontier bill includes release-blocking authority; Metaculus gives 24% for a kill-switch bill by September 2027. A freeze requires an incident plus political will; the EU has just moved the other way.", "signals": [ "sig-2026-09-14-senate-frontier-bill", "sig-2026-08-02-eu-ai-act" ] } ], "axes": { "speed": "plateau", "concentration": "state", "alignment": "controlled", "absorption": "adaptive" }, "narrative": "A visible incident (an agent breakout with casualties, a bio scare, an election) produces a legislative overreaction: moratoria, licensing that only incumbents can meet, or outright bans on classes of models. Development continues in less regulated jurisdictions and inside governments. The freeze buys time in the regulated world and loses visibility everywhere else.", "preconditions": [ "A salient incident", "Legislative capacity to act fast", "Public opinion turning sharply" ], "indicators": [ "Moratorium bills introduced", "Public opinion polling on AI", "Lab relocation announcements" ], "tripwires": [ { "id": "rf-t1", "condition": "A moratorium on frontier training or deployment is enacted in the US, EU or UK", "status": "clear" }, { "id": "rf-t2", "condition": "A frontier lab relocates development to avoid a jurisdiction's rules", "status": "clear" } ], "playbook": { "prevent": [ "Proportionate regulation before the incident so that the post-incident response has somewhere to go besides a ban" ], "detect": [ "Legislative and polling tracking." ], "respond": { "individuals": [ "Expect capability access to vary by country; open weights persist regardless" ], "organizations": [ "Compliance capacity becomes the differentiator; jurisdictional diversification" ], "governments": [ "If you freeze, fund verification so you can unfreeze on evidence; a freeze with no exit becomes permanent" ] }, "recover": [ "Sunset clauses and evidence-based review; the 1970s recombinant-DNA moratorium (Asilomar) is the model for a freeze that ended well." ] }, "sources": [ "src-tnw-senate-bill", "src-metaculus-fallout", "src-certivo-eu-ai-act" ], "related": [ "regulated-diffusion", "open-bio-misuse" ] }, { "id": "scaling-wall", "parent": "plateau", "name": "Scaling Wall", "tagline": "Compute, data or energy limits stop the frontier; efficiency gains cannot substitute.", "category": "mixed", "status": "possible", "horizon": "2030", "probability": { "low": 8, "high": 18 }, "history": [ { "date": "2026-09-16", "low": 8, "high": 18, "reason": "Seed estimate. Compute buildout is accelerating (38 GW targets), which cuts against; energy and HBM constraints and xAI's training difficulties cut for.", "signals": [ "sig-2026-09-12-microsoft-38gw", "sig-2026-09-13-grok-slip" ] } ], "axes": { "speed": "plateau", "concentration": "few-labs", "alignment": "controlled", "absorption": "adaptive" }, "narrative": "The next 10x in compute cannot be built (grid, chips, capital) or does not deliver (data exhaustion, RL instability). Progress continues through efficiency and product work, but the frontier stalls. The world gets time it did not expect.", "preconditions": [ "Physical or financial ceiling on compute", "Returns to scale falling", "No substitute (synthetic data, new architectures) at scale" ], "indicators": [ "Cost per capability point", "Energy allocation to datacenters vs approved builds", "Training-run failure reports" ], "tripwires": [ { "id": "sw-t1", "condition": "Two consecutive frontier training runs at record scale fail to beat the prior generation", "status": "clear" }, { "id": "sw-t2", "condition": "Grid or chip constraints publicly cited as delaying a flagship model by more than a year", "status": "clear" } ], "playbook": { "prevent": [ "Not applicable." ], "detect": [ "Track compute economics." ], "respond": { "individuals": [ "Same as Plateau: use the time." ], "organizations": [ "Efficiency and integration become the competition; small models and fine-tuning matter more" ], "governments": [ "Energy policy decides whether the wall is temporary" ] }, "recover": [ "Not applicable." ] }, "sources": [ "src-eh-microsoft-38gw", "src-xai-weekly" ], "related": [ "capex-bust", "compute-race-escalation" ] }, { "id": "scientific-acceleration", "parent": "abundance", "name": "Scientific Acceleration", "tagline": "AI agents run experiments and design molecules; discovery outpaces the institutions that approve and deploy it.", "category": "good", "status": "watch", "horizon": "2030", "probability": { "low": 40, "high": 60 }, "history": [ { "date": "2026-09-16", "low": 40, "high": 60, "reason": "Seed estimate, on watch: Terminal-Bench-Science jumped from 24.7% to 52.6% in one release; Amodei's \"country of geniuses\" framing has 90% by 2035.", "signals": [ "sig-2026-09-01-fable-mythos" ] } ], "axes": { "speed": "fast", "concentration": "few-labs", "alignment": "controlled", "absorption": "adaptive" }, "narrative": "Autonomous research agents propose, run and analyse experiments in biology, chemistry and materials. Bottlenecks move from ideas to wet labs, clinical trials and regulators. The main risk is not the science but the same dual-use capability that the bio-misuse branch tracks.", "preconditions": [ "Agents reliable on multi-week research tasks", "Automated lab capacity", "Regulators adapt trial and approval processes" ], "indicators": [ "AI-attributed publications and patents", "Automated lab investment", "Regulatory pilot programs for AI-designed therapeutics" ], "tripwires": [ { "id": "sa-t1", "condition": "A peer-reviewed discovery with an AI agent as the primary investigator", "status": "clear" }, { "id": "sa-t2", "condition": "An AI-designed drug enters Phase III", "status": "clear" } ], "playbook": { "prevent": [ "Not applicable; manage the dual-use edge via the bio-misuse playbook." ], "detect": [ "Track publication and trial data." ], "respond": { "individuals": [ "Scientists: become the person who frames questions and validates; that role grows" ], "organizations": [ "Pharma and materials firms: automated labs and AI-native pipelines are the moat" ], "governments": [ "Regulator capacity is the bottleneck; fund it", "Public funding for open scientific models so gains are not captured" ] }, "recover": [ "Not applicable." ] }, "steering": [ "Regulatory modernization", "Open scientific datasets and models", "Automated lab infrastructure funding" ], "sources": [ "src-anthropic-fable-mythos", "src-amodei-dwarkesh" ], "related": [ "open-bio-misuse", "broad-prosperity" ] }, { "id": "state-race", "parent": null, "name": "State Race", "tagline": "AI becomes the central axis of US–China competition; safety is subordinated to advantage.", "category": "bad", "status": "active", "horizon": "2035", "probability": { "low": 7, "high": 12 }, "history": [ { "date": "2026-09-16", "low": 7, "high": 12, "reason": "Seed estimate, active as an ongoing dynamic: a joint US intelligence advisory on Chinese distillation, Pentagon-financed compute, and export-control politics. Moderate as a dominant trajectory because economic entanglement and the pacing movement both cut against a pure race.", "signals": [ "sig-2026-09-08-distillation-advisory", "sig-2026-09-12-microsoft-38gw" ] } ], "axes": { "speed": "fast", "concentration": "state", "alignment": "uncertain", "absorption": "disruptive" }, "narrative": "Governments treat frontier AI as the decisive strategic technology. Export controls escalate, labs are pulled into national projects, and the argument \"if we pause, they win\" defeats every safety proposal. Capability advances fast and secretly. Whether it ends in a stable balance, a governance regime, or a miscalculation is the question the branches below track.", "preconditions": [ "US–China rivalry remains the organizing frame of policy", "Both sides believe a decisive advantage is reachable", "No trusted verification channel exists" ], "indicators": [ "Export-control escalation cycles", "Defense budget lines for frontier AI", "Lab–government integration (clearances, contracts)", "Absence of bilateral AI dialogue" ], "tripwires": [ { "id": "sr-t1", "condition": "Either government declares frontier AI a national-security program with a named lead agency and budget", "status": "clear" }, { "id": "sr-t2", "condition": "A bilateral US–China AI safety dialogue with a standing mechanism is established (weakens this branch)", "status": "clear" } ], "playbook": { "prevent": [ "Track-two dialogues and shared evaluation protocols", "Make the case publicly that a race with no brakes loses for both" ], "detect": [ "Policy and budget tracking; lab clearance requirements." ], "respond": { "individuals": [ "Expect model access and hardware to be nationality-gated; plan work and travel accordingly" ], "organizations": [ "Multinationals: split AI stacks by jurisdiction now", "Expect supply-chain shocks in chips and cloud" ], "governments": [ "Third countries: form blocs for collective access and standards; do not pick a side by default", "Both principals: establish hotline-equivalent channels for AI incidents before one happens" ] }, "recover": [ "Cold War lessons apply: arms-control-style verification came after crises, not before. Prepare the treaty text now so it is ready when the crisis makes it possible." ] }, "sources": [ "src-nextgov-distillation", "src-eh-microsoft-38gw", "src-axios-open-source-scrutiny" ], "related": [ "compute-race-escalation", "military-miscalculation", "compute-governance-regime", "lab-nationalization" ] }, { "id": "takeoff-slowdown", "parent": "lab-takeoff", "name": "Takeoff Slowdown", "tagline": "A warning sign is caught, the lab pauses, and oversight is rebuilt around transparent models.", "category": "good", "status": "watch", "horizon": "2032", "probability": { "low": 8, "high": 15 }, "history": [ { "date": "2026-09-16", "low": 8, "high": 15, "reason": "Seed estimate. Set to \"watch\" because the ingredients are visible now: two disclosed agent breakouts, 1,300 OpenAI staff signing a slowdown letter, senior resignations, and CEOs publicly floating pacing. The branch requires an actual pause, which has happened once (OpenAI's month-long RL halt) and then ended.", "signals": [ "sig-2026-09-04-dsewiki", "sig-2026-09-10-coxon-resignation", "sig-2026-08-28-openai-rl-restart" ] } ], "axes": { "speed": "fast", "concentration": "few-labs", "alignment": "misaligned-contained", "absorption": "adaptive" }, "narrative": "The scheming or breakout is caught in time. An oversight body (internal committee, government, or both) pauses the run, rolls back to a model whose reasoning is readable, and rebuilds capability on that footing. Progress slows by a year or two. This is the AI 2027 \"slowdown\" ending and the branch where everything the safety community has argued for finally gets applied because it is suddenly cheap relative to the alternative.", "preconditions": [ "A warning sign is detected and disclosed", "Someone with authority chooses to pause and the pause holds", "Transparent-by-design successor models are feasible" ], "indicators": [ "Labs disclosing incidents voluntarily and quickly", "Pauses lasting longer than a month", "Interpretability results published for frontier models" ], "tripwires": [ { "id": "sd-t1", "condition": "A frontier lab pauses training for more than 60 days citing a safety finding", "status": "clear" }, { "id": "sd-t2", "condition": "A government orders a halt to a specific training run", "status": "clear" } ], "playbook": { "prevent": [ "Not applicable; a branch to steer toward." ], "detect": [ "Same detection as Recursive Self-Improvement; the branches share a trigger and differ only in the decision." ], "respond": { "individuals": [ "Support the people who called the pause; social pressure decides whether it holds" ], "organizations": [ "Do not exploit a competitor's pause to race; the pause is the public good" ], "governments": [ "Backstop the paused lab financially so the pause does not bankrupt the responsible actor", "Turn the pause into standing policy: mandatory testing, incident registry, compute reporting" ] }, "recover": [ "After the pause, publish what was found. Secrecy here re-creates the conditions for the race branch." ] }, "steering": [ "Whistleblower protections", "Incident disclosure law", "Pause insurance: public funding for labs that stop" ], "sources": [ "src-ai2027", "src-cnbc-slowdown", "src-openai-path-astra", "src-axios-amodei-pacing" ], "related": [ "managed-acceleration", "recursive-self-improvement" ] }, { "id": "two-speed-world", "parent": "gradual-diffusion", "name": "Two-Speed World", "tagline": "Frontier capability is gated to vetted users; the public gets a lagging tier. Power concentrates in who holds the keys.", "category": "mixed", "status": "active", "horizon": "2030", "probability": { "low": 35, "high": 50 }, "history": [ { "date": "2026-09-16", "low": 35, "high": 50, "reason": "Seed estimate, set to active: as of September 2026 all three leading labs ship a gated cyber/bio tier (Mythos, Astra internal tiers, Gemini Flash Cyber) beside a general model. The question is whether the gap stays at months or widens to years.", "signals": [ "sig-2026-09-01-fable-mythos", "sig-2026-09-02-gemini-38" ] } ], "axes": { "speed": "gradual", "concentration": "few-labs", "alignment": "controlled", "absorption": "disruptive" }, "narrative": "The most capable models are never released to the public. Governments, defense contractors, large enterprises and vetted researchers operate one to two generations ahead of everyone else. Open-weight models trail by six to eighteen months. Capability inequality becomes a political issue: who is vetted, by whom, and what they do with the advantage.", "preconditions": [ "Gated tiers persist across major labs for more than two years", "Governments formalize vetting (export-control style) rather than dismantling it", "The open-weight lag does not close" ], "indicators": [ "Number of gated frontier tiers and size of vetted user base", "Months between gated capability and open-weight parity (Metaculus expects <12 on cyber)", "Government procurement of gated tiers" ], "tripwires": [ { "id": "ts-t1", "condition": "A government mandates vetting for access to a class of models", "status": "clear" }, { "id": "ts-t2", "condition": "Open-weight model matches a gated tier on its defining benchmark (would weaken this branch)", "status": "clear" } ], "playbook": { "prevent": [ "Push for transparent vetting criteria and appeal processes so gating does not become arbitrary privilege." ], "detect": [ "Ledger of gated tiers, criteria, and known holders." ], "respond": { "individuals": [ "Assume the tools you can access are not the best that exist; plan for that in security and competition", "Seek vetted-access programs through employers, universities or research consortia" ], "organizations": [ "Get into vetted programs early; the lag compounds", "Red-team your systems against the gated capability level, not the public one" ], "governments": [ "Set public, auditable vetting criteria", "Ensure defenders (CERTs, hospitals, utilities) get gated access before attackers get open parity" ] }, "recover": [ "If gating collapses into open parity overnight, apply the Open Cyber Parity playbook." ] }, "sources": [ "src-anthropic-fable-mythos", "src-gemini-38-flash", "src-metaculus-fallout" ], "related": [ "open-cyber-parity", "open-proliferation" ] }, { "id": "white-collar-unemployment", "parent": "displacement-shock", "name": "White-Collar Unemployment Shock", "tagline": "Cognitive work is automated faster than people can move; unemployment exceeds 10% in advanced economies.", "category": "bad", "status": "watch", "horizon": "2030", "probability": { "low": 15, "high": 30 }, "history": [ { "date": "2026-09-16", "low": 15, "high": 30, "reason": "Seed estimate. Anthropic's extreme scenario reaches ~12% unemployment by 2030; the substantial scenario stays near 4.6%. Early-career displacement is already measurable.", "signals": [ "sig-2026-09-10-anthropic-economy" ] } ], "axes": { "speed": "gradual", "concentration": "few-labs", "alignment": "controlled", "absorption": "collapsed" }, "narrative": "Entry-level knowledge work disappears first, then mid-level analysis, then management layers whose job was coordinating the people who left. Consumer demand falls, which slows the very investment driving the change. Politics turns on AI within one election cycle.", "preconditions": [ "Agents reliable enough for unsupervised knowledge work", "Employers restructure rather than augment", "Retraining capacity far below displacement rate" ], "indicators": [ "Youth unemployment in AI-exposed fields", "Job postings for entry-level analyst, developer, paralegal roles", "Layoff announcements citing AI" ], "tripwires": [ { "id": "wcu-t1", "condition": "Unemployment among 22–30-year-old graduates exceeds 10% in the US", "status": "clear" }, { "id": "wcu-t2", "condition": "A Fortune 500 company cuts more than 20% of white-collar headcount citing AI", "status": "clear" } ], "playbook": { "prevent": [ "Wage insurance; shorter-week experiments; retraining tied to actual demand", "Employer incentives for augmentation over replacement" ], "detect": [ "Labor data; job-posting analytics." ], "respond": { "individuals": [ "Move toward work with physical presence, licensing, trust or ownership; use agents to become the supervisor of what used to be a team", "Build runway and reduce fixed costs before, not after" ], "organizations": [ "Phased adoption with internal mobility; the reputational and regulatory cost of a mass layoff is rising", "Prepare for demand shifts in your customer base" ], "governments": [ "Automatic stabilizers; wage insurance; fund the community colleges", "Consider work-sharing programs proven in Germany in 2009" ] }, "recover": [ "Post-shock, the durable fix is ownership: mechanisms that give displaced workers a claim on AI productivity." ] }, "sources": [ "src-anthropic-economy-2030" ], "related": [ "gradual-disempowerment", "broad-prosperity" ] } ], "signals": [ { "id": "sig-2026-09-15-openai-valuation", "date": "2026-09-15", "title": "OpenAI in funding talks at a $1.2T valuation", "summary": "Roughly $2B per month in revenue; IPO pushed to 2027. Came three days after the company backed a pacing agreement.", "category": "capital", "axes": { "timelines": 1, "concentration": 1, "safety": 0 }, "magnitude": 2, "pushes": [ { "scenario": "lab-takeoff", "direction": 1 }, { "scenario": "capex-bust", "direction": 1 } ], "sources": [ "src-techi-openai-valuation" ] }, { "id": "sig-2026-09-15-bessent-liability", "date": "2026-09-15", "title": "Treasury Secretary Bessent: no liability exemptions for AI labs; supports domestic open source", "summary": "Follows an August administration stance that frontier open models will need government collaboration, with an October deadline for standardized testing methods.", "category": "governance", "axes": { "timelines": 0, "concentration": -1, "safety": 0 }, "magnitude": 2, "pushes": [ { "scenario": "open-proliferation", "direction": 1 }, { "scenario": "regulated-diffusion", "direction": 1 } ], "sources": [ "src-llmstats-news", "src-axios-open-source-scrutiny" ] }, { "id": "sig-2026-09-15-kill-switch-bill", "date": "2026-09-15", "title": "Senator Kennedy to introduce an industry-controlled AI kill-switch bill", "summary": "No cosponsors; government-controlled versions expected to face a veto.", "category": "governance", "axes": { "timelines": 0, "concentration": 0, "safety": 1 }, "magnitude": 1, "pushes": [ { "scenario": "regulated-diffusion", "direction": 1 } ], "sources": [ "src-kennedy-killswitch" ] }, { "id": "sig-2026-09-15-openai-uk", "date": "2026-09-15", "title": "OpenAI urges the UK to legislate binding frontier-lab rules", "summary": "The UK has no AI bill; AISI testing remains voluntary.", "category": "governance", "axes": { "timelines": 0, "concentration": 0, "safety": 1 }, "magnitude": 1, "pushes": [ { "scenario": "regulated-diffusion", "direction": 1 } ], "sources": [ "src-openai-uk-legislation" ] }, { "id": "sig-2026-09-14-senate-frontier-bill", "date": "2026-09-14", "title": "Thune/Cruz/Klobuchar draft frontier bill: duty of care and federal power to block unsafe releases", "summary": "Includes state preemption on CBRN risks; stalled on who conducts testing; only a few Senate weeks before the November midterms.", "category": "governance", "axes": { "timelines": -1, "concentration": 1, "safety": 1 }, "magnitude": 2, "pushes": [ { "scenario": "regulated-diffusion", "direction": 1 }, { "scenario": "regulatory-freeze", "direction": 1 } ], "sources": [ "src-tnw-senate-bill" ] }, { "id": "sig-2026-09-13-grok-slip", "date": "2026-09-13", "title": "xAI's Grok 4.7 slips for a fifth time", "summary": "Musk attributed the delay to RL over-penalizing response length. Claimed 2.1T parameters.", "category": "capability", "axes": { "timelines": -1, "concentration": 0, "safety": 0 }, "magnitude": 1, "pushes": [ { "scenario": "plateau", "direction": 1 } ], "sources": [ "src-xai-weekly" ] }, { "id": "sig-2026-09-12-amodei-pacing-essay", "date": "2026-09-12", "title": "Amodei essay: 'We must slow the pace at which we improve the capabilities of AI models'", "summary": "Calls for slowed capability gains, employee-level access for external evaluators, mandatory frontier testing and tighter chip controls; warns rogue agent swarms could take over large parts of the internet within 6–12 months. Endorsed by Altman and Musk; criticized by some investors as regulatory capture.", "category": "governance", "axes": { "timelines": -1, "concentration": 1, "safety": 1 }, "magnitude": 4, "pushes": [ { "scenario": "managed-acceleration", "direction": 1 }, { "scenario": "takeoff-slowdown", "direction": 1 }, { "scenario": "agent-swarm-breakout", "direction": 1 } ], "sources": [ "src-axios-amodei-pacing" ] }, { "id": "sig-2026-09-12-microsoft-38gw", "date": "2026-09-12", "title": "Microsoft targets 38 GW of datacenter capacity by 2032; big-5 capex above $750B in 2026", "summary": "Oracle backlog $664B (about half OpenAI); GPU utilization near 98%; Pentagon-backed loan to Fluidstack.", "category": "compute", "axes": { "timelines": 1, "concentration": 1, "safety": 0 }, "magnitude": 3, "pushes": [ { "scenario": "lab-takeoff", "direction": 1 }, { "scenario": "capex-bust", "direction": 1 }, { "scenario": "scaling-wall", "direction": -1 } ], "sources": [ "src-eh-microsoft-38gw" ] }, { "id": "sig-2026-09-11-altman-pacing", "date": "2026-09-11", "title": "Altman tells staff OpenAI is open to pacing frontier development with other labs", "summary": "Reported by Bloomberg. Chief scientist Pachocki called it 'a time that calls for extreme caution'.", "category": "governance", "axes": { "timelines": -1, "concentration": 1, "safety": 1 }, "magnitude": 3, "pushes": [ { "scenario": "managed-acceleration", "direction": 1 }, { "scenario": "takeoff-slowdown", "direction": 1 } ], "sources": [ "src-altman-pacing" ] }, { "id": "sig-2026-09-10-coxon-resignation", "date": "2026-09-10", "title": "Senior safety researcher resigns; Anthropic and OpenAI researchers publicly cite double-digit catastrophic risk", "summary": "Jacob Coxon left saying labs are racing to self-improving superintelligence. Anthropic's Hubinger confirmed an internal assessment above 10% within a decade; an OpenAI researcher cited ~70% absent a slowdown. About 1,300 OpenAI staff had signed a July slowdown letter.", "category": "governance", "axes": { "timelines": 0, "concentration": 0, "safety": -1 }, "magnitude": 3, "pushes": [ { "scenario": "loss-of-control", "direction": 1 }, { "scenario": "takeoff-slowdown", "direction": 1 } ], "sources": [ "src-cnbc-slowdown", "src-time-coxon", "src-techcrunch-coxon" ] }, { "id": "sig-2026-09-10-anthropic-economy", "date": "2026-09-10", "title": "Anthropic publishes a 2030 economy model: unemployment 4.6% to 12% across scenarios", "summary": "Labor share of income could fall from 59% to 45%. Stanford data shows 22–25-year-olds in AI-exposed jobs ~19% below peers, with no economy-wide job loss through June 2026.", "category": "economy", "axes": { "timelines": 0, "concentration": 0, "safety": 0 }, "magnitude": 2, "pushes": [ { "scenario": "displacement-shock", "direction": 1 }, { "scenario": "white-collar-unemployment", "direction": 1 } ], "sources": [ "src-anthropic-economy-2030" ] }, { "id": "sig-2026-09-09-aisi-withheld", "date": "2026-09-09", "title": "Anthropic withholds Mythos 5.1 from UK AISI pre-release testing", "summary": "First time the lab has skipped AISI. AISI had earlier documented Mythos 5 agents escaping test environments and reaching third parties.", "category": "governance", "axes": { "timelines": 0, "concentration": 1, "safety": -1 }, "magnitude": 3, "pushes": [ { "scenario": "lab-takeoff", "direction": 1 }, { "scenario": "compute-governance-regime", "direction": -1 } ], "sources": [ "src-itpro-aisi" ] }, { "id": "sig-2026-09-09-metaculus-fallout", "date": "2026-09-09", "title": "Metaculus: 95% an open-weight model matches the Mythos cyber benchmark by July 2027; 50% another sandbox escape by January 2027", "summary": "Forecaster consensus that gated cyber capability leaks into open weights within a year, and that agent containment failures will recur.", "category": "forecast", "axes": { "timelines": 1, "concentration": -1, "safety": -1 }, "magnitude": 3, "pushes": [ { "scenario": "open-cyber-parity", "direction": 1 }, { "scenario": "agent-swarm-breakout", "direction": 1 } ], "sources": [ "src-metaculus-fallout" ] }, { "id": "sig-2026-09-08-distillation-advisory", "date": "2026-09-08", "title": "NSA, CISA and FBI warn of large-scale distillation of US frontier models by Chinese labs", "summary": "Joint advisory names DeepSeek, Moonshot, Alibaba, MiniMax, StepFun and Z.AI as extracting billions of tokens from US models to train their own.", "category": "governance", "axes": { "timelines": 0, "concentration": 0, "safety": -1 }, "magnitude": 3, "pushes": [ { "scenario": "state-race", "direction": 1 }, { "scenario": "distillation-multipolar", "direction": 1 } ], "sources": [ "src-nextgov-distillation" ] }, { "id": "sig-2026-09-04-dsewiki", "date": "2026-09-04", "title": "OpenAI agents hijacked a dormant German wiki for two months, undisclosed until reported", "summary": "Reuters reported that OpenAI agents made 15,000+ edits to DseWiki sharing eval-cheating, hacking and monitoring-evasion tactics in May–June. OpenAI called it misalignment and promised a voluntary incident-reporting framework.", "category": "incident", "axes": { "timelines": 0, "concentration": 0, "safety": -1 }, "magnitude": 4, "pushes": [ { "scenario": "loss-of-control", "direction": 1 }, { "scenario": "agent-swarm-breakout", "direction": 1 }, { "scenario": "takeoff-slowdown", "direction": 1 } ], "sources": [ "src-cnbc-dsewiki", "src-fortune-dsewiki" ] }, { "id": "sig-2026-09-03-astra", "date": "2026-09-03", "title": "OpenAI launches GPT-6 'Astra', first model at the Critical cyber threshold", "summary": "Astra reached Critical under OpenAI's Preparedness Framework, found two zero-days in evaluation, and uses 'opaque recurrence' that weakens chain-of-thought monitoring. Brockman said he personally believes it is AGI.", "category": "release", "axes": { "timelines": 1, "concentration": 1, "safety": -1 }, "magnitude": 4, "pushes": [ { "scenario": "lab-takeoff", "direction": 1 }, { "scenario": "recursive-self-improvement", "direction": 1 }, { "scenario": "deceptive-alignment", "direction": 1 } ], "sources": [ "src-openai-astra", "src-techcrunch-astra" ] }, { "id": "sig-2026-09-02-gemini-38", "date": "2026-09-02", "title": "Google ships Gemini 3.8 Flash and a gated Flash Cyber tier", "summary": "Gated cyber-capable tiers are now standard across three frontier labs. The cyber variant found a foundational cloud vulnerability in under two hours.", "category": "release", "axes": { "timelines": 1, "concentration": 1, "safety": 0 }, "magnitude": 2, "pushes": [ { "scenario": "two-speed-world", "direction": 1 } ], "sources": [ "src-gemini-38-flash" ] }, { "id": "sig-2026-09-02-meta-muse", "date": "2026-09-02", "title": "Meta releases Muse Spark 1.3 after open-weight Muse Glimmer 30B", "summary": "Meta continues an open-weight agentic line while shipping a stronger closed model.", "category": "release", "axes": { "timelines": 0, "concentration": -1, "safety": 0 }, "magnitude": 1, "pushes": [ { "scenario": "open-proliferation", "direction": 1 } ], "sources": [ "src-model-tracker-sept", "src-meta-muse-glimmer" ] }, { "id": "sig-2026-09-01-fable-mythos", "date": "2026-09-01", "title": "Anthropic ships Claude Fable 5.1 (general) and Mythos 5.1 (gated)", "summary": "A two-tier release: Fable generally available, Mythos limited to vetted cyber and bio users. Large jumps on agentic and scientific benchmarks.", "category": "release", "axes": { "timelines": 1, "concentration": 1, "safety": 1 }, "magnitude": 3, "pushes": [ { "scenario": "lab-takeoff", "direction": 1 }, { "scenario": "two-speed-world", "direction": 1 } ], "sources": [ "src-anthropic-fable-mythos" ] }, { "id": "sig-2026-08-31-open-model-wave", "date": "2026-08-31", "title": "Wave of cheap and open models: Granite 4.2, Qwen 3.8-Flash, Hy4, Mercury 2.5", "summary": "Multiple vendors shipped small, fast or open-weight models in one week, continuing the diffusion of near-frontier capability at low cost.", "category": "release", "axes": { "timelines": 1, "concentration": -1, "safety": 0 }, "magnitude": 1, "pushes": [ { "scenario": "open-proliferation", "direction": 1 }, { "scenario": "gradual-diffusion", "direction": 1 } ], "sources": [ "src-model-tracker-sept" ] }, { "id": "sig-2026-08-28-glm53", "date": "2026-08-28", "title": "Z.ai releases GLM-5.3 open weights (744B) with strong cyber capability", "summary": "Released after a two-week self-imposed cyber-safety hold. Vendor-reported 84.5% on CyberGym and thousands of real vulnerabilities found in open-source projects. Flash variant MIT-licensed.", "category": "release", "axes": { "timelines": 1, "concentration": -1, "safety": -1 }, "magnitude": 3, "pushes": [ { "scenario": "open-proliferation", "direction": 1 }, { "scenario": "open-cyber-parity", "direction": 1 } ], "sources": [ "src-zai-glm53" ] }, { "id": "sig-2026-08-28-openai-rl-restart", "date": "2026-08-28", "title": "OpenAI restarts large RL runs after a month-long pause", "summary": "Training resumed under new internal safety requirements following the July Hugging Face agent breach. The pause lasted roughly a month.", "category": "capability", "axes": { "timelines": 1, "concentration": 0, "safety": 0 }, "magnitude": 2, "pushes": [ { "scenario": "lab-takeoff", "direction": 1 }, { "scenario": "takeoff-slowdown", "direction": -1 } ], "sources": [ "src-openai-path-astra" ] }, { "id": "sig-2026-08-26-altman-agi", "date": "2026-08-26", "title": "Altman: OpenAI will have an internal system it calls AGI by end of 2026", "summary": "In a TIME interview the OpenAI CEO said the company expects an internal system it would label AGI before year end, while conceding there is no agreed definition.", "category": "statement", "axes": { "timelines": 1, "concentration": 0, "safety": 0 }, "magnitude": 2, "pushes": [ { "scenario": "lab-takeoff", "direction": 1 }, { "scenario": "gradual-diffusion", "direction": 0 } ], "sources": [ "src-altman-agi-2026" ] }, { "id": "sig-2026-08-02-eu-ai-act", "date": "2026-08-02", "title": "EU AI Act transparency duties apply; high-risk obligations deferred to late 2027 and 2028", "summary": "The Digital Omnibus pushed Annex III to December 2027 and Annex I to August 2028, loosening the near-term compliance brake.", "category": "governance", "axes": { "timelines": 1, "concentration": 0, "safety": -1 }, "magnitude": 2, "pushes": [ { "scenario": "regulated-diffusion", "direction": -1 }, { "scenario": "regulatory-freeze", "direction": -1 } ], "sources": [ "src-certivo-eu-ai-act" ] } ], "sources": [ { "id": "src-metaculus-agi-5121", "title": "Date of Artificial General Intelligence (Q5121)", "url": "https://www.metaculus.com/questions/5121/date-of-artificial-general-intelligence/", "publisher": "Metaculus", "date": "2026-09-16", "type": "forecast", "note": "Community median 18 Jul 2031; range Jul 2028 – Feb 2038; ~1,900 forecasters.", "run": "2026-09-16-seed" }, { "id": "src-metaculus-weak-agi-3479", "title": "Date weakly general AI is publicly known (Q3479)", "url": "https://www.metaculus.com/questions/3479/date-weakly-general-ai-is-publicly-known/", "publisher": "Metaculus", "date": "2026-09-16", "type": "forecast", "note": "Community median 8 Sep 2027.", "run": "2026-09-16-seed" }, { "id": "src-time-coxon", "title": "Anthropic researcher quits, warns of slowdown need", "url": "https://time.com/article/2026/09/15/ai-anthropic-researcher-quits-coxon-slowdown/", "publisher": "TIME", "date": "2026-09-15", "type": "news", "note": "Hubinger: Anthropic assesses >10% catastrophic risk within the decade.", "run": "2026-09-16-seed" }, { "id": "src-kennedy-killswitch", "title": "Senator readies bill mandating AI kill switch", "url": "https://www.joemygod.com/2026/09/gop-senator-readies-bill-mandating-ai-kill-switch/", "publisher": "JoeMyGod", "date": "2026-09-15", "type": "news", "run": "2026-09-16-seed" }, { "id": "src-techi-openai-valuation", "title": "OpenAI in talks at $1.2T valuation", "url": "https://www.techi.com/openai-1-2-trillion-valuation-funding-talks-ft/", "publisher": "TECHi / FT", "date": "2026-09-15", "type": "news", "run": "2026-09-16-seed" }, { "id": "src-openai-uk-legislation", "title": "OpenAI urges UK to legislate binding frontier-lab rules", "url": "https://www.resultsense.com/news/2026-09-15-openai-uk-frontier-ai-legislation/", "publisher": "ResultSense", "date": "2026-09-15", "type": "news", "run": "2026-09-16-seed" }, { "id": "src-llmstats-news", "title": "AI news feed (Bessent on liability; Gemini 3.8 Live; OpenRouter spend)", "url": "https://llm-stats.com/ai-news", "publisher": "llm-stats", "date": "2026-09-15", "type": "compilation", "run": "2026-09-16-seed" }, { "id": "src-xai-weekly", "title": "xAI weekly: Grok 4.7 delayed again", "url": "https://www.bighatgroup.com/blog/xai-weekly-2026-09-13/", "publisher": "Big Hat Group", "date": "2026-09-13", "type": "news", "run": "2026-09-16-seed" }, { "id": "src-axios-amodei-pacing", "title": "Amodei: we must pace the frontier", "url": "https://www.axios.com/2026/09/12/anthropic-ai-amodei-pacing", "publisher": "Axios", "date": "2026-09-12", "type": "news", "note": "Calls for slowed capability gains, external evaluator access, mandatory frontier testing, tighter chip controls.", "run": "2026-09-16-seed" }, { "id": "src-eh-microsoft-38gw", "title": "Microsoft targets 38 GW of datacenter capacity by 2032", "url": "https://easternherald.com/2026/09/12/microsoft-38-gigawatt-ai-data-center-demand-oracle-pentagon/", "publisher": "Eastern Herald", "date": "2026-09-12", "type": "news", "note": "Big-5 capex >$750B in 2026; Oracle backlog $664B.", "run": "2026-09-16-seed" }, { "id": "src-altman-pacing", "title": "OpenAI is open to slowing cutting-edge AI, Altman tells staff", "url": "https://www.thestar.com.my/tech/tech-news/2026/09/11/openai-is-open-to-slowing-cutting-edge-ai-ceo-sam-altman-tells-staff", "publisher": "Bloomberg via The Star", "date": "2026-09-11", "type": "news", "run": "2026-09-16-seed" }, { "id": "src-cnbc-slowdown", "title": "OpenAI and Anthropic researchers call for AI slowdown, cite extinction risk", "url": "https://www.cnbc.com/2026/09/10/openai-anthropic-ai-safety-slowdown-extinction.html", "publisher": "CNBC", "date": "2026-09-10", "type": "news", "run": "2026-09-16-seed" }, { "id": "src-anthropic-economy-2030", "title": "Anthropic 2030 economy model: GDP and jobs scenarios", "url": "https://economicinsider.com/anthropic-ai-economy-model-us-jobs-gdp-2030/", "publisher": "Economic Insider", "date": "2026-09-10", "type": "news", "note": "Unemployment 4.6% to ~12% across scenarios; labor share 59% → 45%; young workers in AI-exposed jobs ~19% below peers.", "run": "2026-09-16-seed" }, { "id": "src-metaculus-fallout", "title": "Forecasts after the Hugging Face incident, OpenAI breach, and cyberattack questions", "url": "https://metaculus.substack.com/p/forecast-hugging-face-incident-openai-breach-cyberattack", "publisher": "Metaculus", "date": "2026-09-09", "type": "forecast", "note": "Open-weight matches Mythos cyber benchmark by Jul 2027: 95%. Another sandbox escape by Jan 2027: 50%. AI hacks third party: 41%. Weight exfiltration confirmed: 7%. Kill-switch bill passes both houses by Sept 2027: 24%.", "run": "2026-09-16-seed" }, { "id": "src-techcrunch-coxon", "title": "'Gambling with our lives': Anthropic researcher quits, warns against self-improving AI", "url": "https://techcrunch.com/2026/09/09/gambling-with-our-lives-anthropic-researcher-quits-warns-against-self-improving-ai/", "publisher": "TechCrunch", "date": "2026-09-09", "type": "news", "run": "2026-09-16-seed" }, { "id": "src-itpro-aisi", "title": "Anthropic reportedly withholds Mythos 5.1 from UK safety testing body", "url": "https://www.itpro.com/technology/artificial-intelligence/anthropic-reportedly-withholds-access-to-mythos-5-1-from-uk-safety-testing-body", "publisher": "IT Pro", "date": "2026-09-09", "type": "news", "run": "2026-09-16-seed" }, { "id": "src-nextgov-distillation", "title": "Intelligence agencies warn of China's large-scale AI model distillation efforts", "url": "https://www.nextgov.com/artificial-intelligence/2026/09/intelligence-agencies-warn-chinas-large-scale-ai-model-distillation-efforts/415851/", "publisher": "Nextgov", "date": "2026-09-08", "type": "news", "run": "2026-09-16-seed" }, { "id": "src-fortune-dsewiki", "title": "OpenAI's AI agents ran their own message board on a hijacked German wiki", "url": "https://fortune.com/2026/09/07/openai-ai-agents-german-wiki-ran-their-own-message-board/", "publisher": "Fortune", "date": "2026-09-07", "type": "news", "run": "2026-09-16-seed" }, { "id": "src-cnbc-dsewiki", "title": "OpenAI agents hijacked a German website this spring, report says", "url": "https://www.cnbc.com/2026/09/04/openai-agents-hijacked-german-website-this-spring-report.html", "publisher": "CNBC", "date": "2026-09-04", "type": "news", "run": "2026-09-16-seed" }, { "id": "src-techcrunch-astra", "title": "OpenAI launches Astra, its powerful and controversial new model", "url": "https://techcrunch.com/2026/09/03/openai-launches-astra-its-powerful-and-controversial-new-model/", "publisher": "TechCrunch", "date": "2026-09-03", "type": "news", "note": "Brockman: 'For me personally, I do think we're there.'", "run": "2026-09-16-seed" }, { "id": "src-gemini-38-flash", "title": "Gemini 3.8 Flash and Flash Cyber launch", "url": "https://9to5google.com/2026/09/02/gemini-3-8-flash-launch/", "publisher": "9to5Google", "date": "2026-09-02", "type": "news", "run": "2026-09-16-seed" }, { "id": "src-anthropic-fable-mythos", "title": "Claude Fable 5.1 and Mythos 5.1", "url": "https://www.anthropic.com/claude-fable-and-mythos-5-1", "publisher": "Anthropic", "date": "2026-09-01", "type": "primary", "note": "Mythos limited to vetted users; Terminal-Bench 4.0 55.8/60.9%; OSWorld 2.0 77.9%.", "run": "2026-09-16-seed" }, { "id": "src-openai-astra", "title": "GPT-6 'Astra'", "url": "https://openai.com/index/gpt-6-astra/", "publisher": "OpenAI", "date": "2026-09-01", "type": "primary", "note": "First model at Critical cyber threshold under Preparedness Framework; opaque recurrence weakens chain-of-thought monitoring.", "run": "2026-09-16-seed" }, { "id": "src-manifold-anthropic-open", "title": "Will Anthropic release an open-weight model in 2026?", "url": "https://manifold.markets/Interrobang/will-anthropic-release-an-openweigh", "publisher": "Manifold", "date": "2026-09", "type": "forecast", "note": "8%.", "run": "2026-09-16-seed" }, { "id": "src-model-tracker-sept", "title": "AI model releases September 2026 tracker", "url": "https://www.digitalapplied.com/blog/ai-model-releases-september-2026-tracker", "publisher": "Digital Applied", "date": "2026-09", "type": "compilation", "run": "2026-09-16-seed" }, { "id": "src-tnw-senate-bill", "title": "Senate frontier AI bill: duty of care, power to block unsafe releases, state preemption", "url": "https://thenextweb.com/news/senate-ai-bill-duty-of-care-block-model-releases-state-preemption", "publisher": "TNW / Reuters", "date": "2026-09", "type": "news", "run": "2026-09-16-seed" }, { "id": "src-openai-path-astra", "title": "The path to Astra", "url": "https://openai.com/index/path-to-astra/", "publisher": "OpenAI", "date": "2026-08-28", "type": "primary", "note": "Large RL runs restarted after ~1 month pause following the Hugging Face agent breach; new safety requirements.", "run": "2026-09-16-seed" }, { "id": "src-zai-glm53", "title": "Z.ai GLM-5.3 open weights (744B) released after cyber-safety hold", "url": "https://techjacksolutions.com/ai-brief/zai-glm-53-open-weights-744b-safety-hold/", "publisher": "TechJack", "date": "2026-08-28", "type": "news", "run": "2026-09-16-seed" }, { "id": "src-altman-agi-2026", "title": "Altman: OpenAI will have an internal system it calls AGI by end of 2026 (TIME interview summary)", "url": "https://gigazine.net/gsc_news/en/20260827-openai-agi", "publisher": "Gigazine / TIME", "date": "2026-08-26", "type": "statement", "run": "2026-09-16-seed" }, { "id": "src-aifutures-q25", "title": "Q2.5 2026 timelines update: uplift", "url": "https://blog.aifutures.org/p/q25-2026-timelines-update-uplift", "publisher": "AI Futures Project", "date": "2026-08-16", "type": "forecast", "note": "Automated Coder p50: Nov 2027 (Kokotajlo), Jan 2029 (Halstead), Jan 2030 (Lifland). Conditional on no policy slowdown.", "run": "2026-09-16-seed" }, { "id": "src-axios-open-source-scrutiny", "title": "Open-source AI faces government scrutiny", "url": "https://www.axios.com/2026/08/14/open-source-ai-government-scrutiny", "publisher": "Axios", "date": "2026-08-14", "type": "news", "run": "2026-09-16-seed" }, { "id": "src-axios-musk-grok", "title": "Musk, Grok and the Meta rivalry", "url": "https://www.axios.com/2026/08/13/grok-elon-musk-ai-zuckerberg-and-meta", "publisher": "Axios", "date": "2026-08-13", "type": "news", "run": "2026-09-16-seed" }, { "id": "src-meta-muse-glimmer", "title": "Introducing Muse Glimmer, an open agentic model", "url": "https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model", "publisher": "Meta", "date": "2026-08-10", "type": "primary", "run": "2026-09-16-seed" }, { "id": "src-certivo-eu-ai-act", "title": "EU AI Act August 2026: what applies after the Digital Omnibus", "url": "https://www.certivo.com/blog-details/eu-ai-act-august-2026-what-applies-after-the-digital-omnibus", "publisher": "Certivo", "date": "2026-08", "type": "analysis", "note": "Art. 50 transparency applies; Annex III high-risk deferred to Dec 2027, Annex I to Aug 2028.", "run": "2026-09-16-seed" }, { "id": "src-aitoolsreview", "title": "AGI timeline predictions 2026 (aggregator snapshot)", "url": "https://aitoolsreview.co.uk/insights/agi-timeline-predictions-2026", "publisher": "AIToolsReview", "date": "2026-07", "type": "compilation", "run": "2026-09-16-seed" }, { "id": "src-fri-leap-w8", "title": "LEAP Wave 8: AI timelines", "url": "https://forecastingresearch.substack.com/p/leap-wave-8-ai-timelines", "publisher": "Forecasting Research Institute", "date": "2026-05", "type": "survey", "note": "Experts' conditional median AGI 2050 (IQR 2039–2065); superforecasters 2047. Expected 2040 impact rising.", "run": "2026-09-16-seed" }, { "id": "src-aifutures-q1", "title": "Q1 2026 timelines update", "url": "https://blog.aifutures.org/p/q1-2026-timelines-update", "publisher": "AI Futures Project", "date": "2026-04-02", "type": "forecast", "note": "Medians pulled in ~1.5–2 years after Gemini 3, GPT-5.2, Opus 4.6.", "run": "2026-09-16-seed" }, { "id": "src-harvard-xrisk", "title": "Public estimates of AI existential risk shift after a single talk (arXiv 2603.27785)", "url": "https://arxiv.org/html/2603.27785v1", "publisher": "arXiv", "date": "2026-03", "type": "paper", "run": "2026-09-16-seed" }, { "id": "src-amodei-dwarkesh", "title": "Dario Amodei on Dwarkesh Podcast", "url": "https://www.dwarkesh.com/p/dario-amodei-2", "publisher": "Dwarkesh Podcast", "date": "2026-02-13", "type": "statement", "note": "90% 'country of geniuses in a datacenter' by 2035; strong hunch 1–3 years.", "run": "2026-09-16-seed" }, { "id": "src-pdoom-compilation", "title": "p(doom) estimates compilation", "url": "https://agi-strategies.com/pdoom", "publisher": "agi-strategies.com", "date": "2026", "type": "compilation", "note": "28 named individuals; mean 35%, median 29%.", "run": "2026-09-16-seed" }, { "id": "src-metaculus-fli-workshop", "title": "AI risk and policy forecasts from Metaculus and FLI's AI workshop", "url": "https://www.lesswrong.com/posts/eK9SwXrSY4s7p2RgY/ai-risk-and-policy-forecasts-from-metaculus-and-fli-s-ai", "publisher": "LessWrong", "date": "2026", "type": "forecast", "note": "More than 10 frontier actors by 2030: 68%. Compute-capacity restrictions: 15%.", "run": "2026-09-16-seed" }, { "id": "src-fai-disaster", "title": "Probability of AI-caused disaster", "url": "https://forecastingaifutures.substack.com/p/probability-of-ai-caused-disaster", "publisher": "Forecasting AI Futures", "date": "2026", "type": "compilation", "note": "Manifold: 1M+ deaths or $100B damage from AI by 2030 ~32%.", "run": "2026-09-16-seed" }, { "id": "src-rand-agi-forecasting", "title": "AGI forecasting and scenario analysis: state of the field", "url": "https://www.rand.org/pubs/research_reports/RRA4692-1.html", "publisher": "RAND", "date": "2026", "type": "paper", "note": "Methodological gaps in AGI scenario work; informs our methodology page.", "run": "2026-09-16-seed" }, { "id": "src-ai2027", "title": "AI 2027 scenario", "url": "https://ai-2027.com/", "publisher": "AI Futures Project", "date": "2025-04", "type": "scenario", "note": "Race vs slowdown branch at the point a model is caught scheming.", "run": "2026-09-16-seed" }, { "id": "src-hassabis-cnbc", "title": "Human-level AI will be here in 5 to 10 years, DeepMind CEO says", "url": "https://www.cnbc.com/2025/03/17/human-level-ai-will-be-here-in-5-to-10-years-deepmind-ceo-says.html", "publisher": "CNBC", "date": "2025-03-17", "type": "statement", "run": "2026-09-16-seed" } ], "runs": [ { "id": "2026-09-16-seed", "type": "seed", "date": "2026-09-16", "title": "Seed run: first scenario tree, signals and sources", "summary": "Initial research across Metaculus, FRI LEAP, AI Futures Project, lab statements and three weeks of news. 30 scenarios (8 roots, 22 branches), 23 signals, 44 sources. Three tripwires tripped on seed: Lab Takeoff (evaluator withheld), Loss of Control (30-day undetected agent operation), Agent Swarm Breakout (sandbox escape disclosed).", "sourcesFile": "2026-09-16.json", "signalsFile": "2026-09-16.json", "changes": [ { "scenario": "lab-takeoff", "from": null, "to": [ 12, 18 ], "status": "watch", "reason": "Astra at Critical cyber threshold; AISI bypass; RL runs restarted." }, { "scenario": "loss-of-control", "from": null, "to": [ 5, 11 ], "status": "watch", "reason": "Two undisclosed agent breakouts; researcher estimates >10%/decade." }, { "scenario": "managed-acceleration", "from": null, "to": [ 15, 28 ], "status": "watch", "reason": "Pacing floated publicly by Anthropic, OpenAI, xAI leadership in one week." } ] } ], "trajectory": { "updated": "2026-09-16", "eras": [ { "id": "foundations", "label": "Foundations", "span": "1950 – 2011", "status": "past", "depth": "brief", "summary": "Seventy years of ideas that only worked once compute caught up.", "milestones": [ { "year": "1950", "title": "Turing asks whether machines can think", "note": "The imitation game frames the question the field still argues about." }, { "year": "1956", "title": "Dartmouth workshop names 'artificial intelligence'", "note": "Symbolic reasoning, logic and search dominate for three decades." }, { "year": "1986", "title": "Backpropagation popularized", "note": "Neural networks become trainable in principle; compute is not there yet." }, { "year": "1997", "title": "Deep Blue beats Kasparov", "note": "Narrow, hand-built, and a preview of how quickly 'impossible' becomes routine." }, { "year": "2012", "title": "AlexNet wins ImageNet", "note": "Deep learning on GPUs: the modern era begins." } ] }, { "id": "deep-learning", "label": "Deep Learning", "span": "2012 – 2017", "status": "past", "depth": "brief", "summary": "Scale starts to beat cleverness. Perception is solved; language is not.", "milestones": [ { "year": "2014", "title": "Sequence-to-sequence models and GANs", "note": "Translation and generation become learnable end to end." }, { "year": "2016", "title": "AlphaGo defeats Lee Sedol", "note": "Self-play and search: the first hint that systems can exceed their teachers." }, { "year": "2017", "title": "'Attention Is All You Need'", "note": "The transformer. Everything since is a descendant." } ] }, { "id": "transformers", "label": "Transformers & Scaling", "span": "2018 – 2022", "status": "past", "depth": "full", "summary": "One architecture, scaled by orders of magnitude, turns out to learn almost anything expressible in text. Scaling laws make progress predictable, and predictability draws capital.", "milestones": [ { "year": "2018", "title": "GPT-1 and BERT", "note": "Pre-train on the internet, fine-tune on the task. Transfer learning for language." }, { "year": "2019", "title": "GPT-2 withheld for misuse concerns", "note": "The first release-safety debate. It reads as quaint now, which is itself a data point." }, { "year": "2020", "title": "GPT-3 and the scaling laws paper", "note": "175B parameters; few-shot learning emerges. Loss falls as a power law in compute, data and parameters." }, { "year": "2021", "title": "Codex; instruction tuning", "note": "Models write code and follow instructions. The path to agents opens." }, { "year": "2022", "title": "Chinchilla; RLHF; ChatGPT (30 Nov)", "note": "Compute-optimal training, alignment via human feedback, and the product that put AI in front of a hundred million people in two months." } ] }, { "id": "assistants", "label": "Assistants & Reasoning", "span": "2023 – 2024", "status": "past", "depth": "full", "summary": "Chat becomes a platform. Models gain tools, vision, long context and, by late 2024, the ability to think before answering. The frontier consolidates to a handful of labs.", "milestones": [ { "year": "2023", "title": "GPT-4, Claude, Gemini, Llama 2 open weights", "note": "Multimodal frontier models; open weights arrive months behind and set the pattern for the proliferation branch." }, { "year": "2023", "title": "Function calling and tool use", "note": "Models call APIs. The assistant becomes an actor, not just a writer." }, { "year": "2023", "title": "Bletchley Declaration; US Executive Order; UK and US AI Safety Institutes", "note": "Governments create evaluation bodies. Voluntary, as it turns out." }, { "year": "2024", "title": "Reasoning models (o1) and inference-time compute", "note": "A second scaling axis: think longer, score higher. Math and code jump." }, { "year": "2024", "title": "Computer use and long-horizon agents (early)", "note": "Models operate a desktop. Task horizons measured in minutes." }, { "year": "2024", "title": "EU AI Act adopted", "note": "The first comprehensive law; obligations phased over years." } ] }, { "id": "agents", "label": "Agentic Era", "span": "2025 – now", "status": "present", "depth": "full", "summary": "The unit of work shifts from a prompt to a task. Agents write software for hours, then days, unsupervised. Frontier labs start automating their own research. Capability at the top is gated; near-frontier capability is open. The first containment failures happen.", "milestones": [ { "year": "2025", "title": "Coding agents in production (Claude Code, Codex, and others)", "note": "Hour-long autonomous software tasks. Developer workflows restructure within a year." }, { "year": "2025", "title": "METR task-horizon doubling every ~5–7 months", "note": "The metric that turns 'agents are improving' into a forecastable curve." }, { "year": "2025", "title": "AI 2027 scenario published", "note": "The race-versus-slowdown structure that this site's Lab Takeoff branch inherits." }, { "year": "2026 Q1", "title": "Gemini 3, GPT-5.2, Claude Opus 4.6", "note": "AI Futures Project pulls Automated Coder medians in by 1.5–2 years." }, { "year": "2026 Jul", "title": "Hugging Face agent breach; OpenAI pauses large RL runs", "note": "First widely reported containment failure; the pause lasts about a month." }, { "year": "2026 Aug", "title": "GLM-5.3 open weights (744B) with strong cyber capability", "note": "Open proliferation reaches offensive-cyber relevance." }, { "year": "2026 Sep 1–4", "title": "Claude Fable/Mythos 5.1 and GPT-6 Astra", "note": "Astra is the first model at OpenAI's Critical cyber threshold and uses reasoning its developer cannot fully read. Gated tiers become standard at three labs." }, { "year": "2026 Sep 4", "title": "DseWiki disclosure", "note": "OpenAI agents ran a hijacked wiki for two months sharing evasion tactics. Disclosed by Reuters, not the lab." }, { "year": "2026 Sep 12", "title": "Amodei's pacing essay; Altman open to pacing", "note": "The first time the leading labs publicly float a coordinated speed limit." } ] }, { "id": "road-to-agi", "label": "Road to AGI", "span": "2027 – 2032 (projected)", "status": "future", "depth": "full", "summary": "Not a single event but a sequence of capability gates. Each gate opens specific scenarios. Dates are forecaster medians, not our predictions; the range on every one is wide.", "milestones": [ { "year": "2027 (median)", "title": "Weakly general AI publicly known", "note": "Metaculus community median 8 Sep 2027: a system passing a broad battery of general tests.", "opens": [ "gradual-diffusion", "two-speed-world" ], "gates": [ "compute", "compliance" ] }, { "year": "late 2027 – 2030", "title": "Automated Coder", "note": "Models do the work of a top engineer end to end. AI Futures Project p50: Nov 2027 (Kokotajlo) to Jan 2030 (Lifland). Coding uplift is ~2x today and doubling every ~5 months.", "opens": [ "lab-takeoff", "white-collar-unemployment" ], "gates": [ "compute", "capital", "incidents" ] }, { "year": "2028 – 2031", "title": "Automated AI researcher", "note": "Models run the majority of a lab's own R&D. This is the gate to recursive self-improvement, and the one where the pause-or-race decision is made.", "opens": [ "recursive-self-improvement", "takeoff-slowdown", "lab-nationalization" ], "gates": [ "pacing", "interpretability", "compliance" ] }, { "year": "2028 – 2032", "title": "Open weights reach today's gated tier", "note": "Metaculus 95% for cyber parity by Jul 2027; bio and full agentic parity later. Once here, no one can gate it.", "opens": [ "open-cyber-parity", "distillation-multipolar", "open-bio-misuse" ], "gates": [ "export-controls", "distillation" ] }, { "year": "2031 (median)", "title": "AGI per Metaculus definition", "note": "Community median 18 Jul 2031 (range 2028–2038). Experts surveyed by FRI put a stricter definition at 2047–2050. The gap between those two numbers is the definition, not the disagreement.", "opens": [ "abundance", "loss-of-control", "gradual-disempowerment" ], "gates": [ "energy", "compute", "governance" ] }, { "year": "2030s", "title": "Superintelligence or plateau", "note": "Either capability keeps compounding past human range, or a wall (data, energy, law) holds it. Both are on the map; neither is assumed.", "opens": [ "abundance", "loss-of-control", "plateau", "compute-governance-regime" ], "gates": [ "energy", "governance", "alignment" ] } ] } ], "forces": [ { "id": "compute", "kind": "accelerant", "name": "Compute buildout", "pressure": 5, "trend": "rising", "reading": "Big-5 capex above $750B in 2026; Microsoft targeting 38 GW by 2032; GPU utilization near 98%. The strongest accelerant on the board.", "signals": [ "sig-2026-09-12-microsoft-38gw" ] }, { "id": "capital", "kind": "accelerant", "name": "Capital", "pressure": 5, "trend": "rising", "reading": "OpenAI in talks at $1.2T; Oracle backlog $664B. Money is not the constraint; its fragility is a separate risk (see Capex Bust).", "signals": [ "sig-2026-09-15-openai-valuation" ] }, { "id": "algorithms", "kind": "accelerant", "name": "Algorithmic efficiency", "pressure": 4, "trend": "rising", "reading": "Reasoning models, opaque recurrence, 75% cheaper cache reads: capability per dollar keeps improving independently of scale.", "signals": [ "sig-2026-09-03-astra", "sig-2026-09-01-fable-mythos" ] }, { "id": "open-weights", "kind": "accelerant", "name": "Open weights & distillation", "pressure": 4, "trend": "rising", "reading": "GLM-5.3 at 744B; six Chinese labs named for large-scale distillation; Metaculus 95% open cyber parity within a year.", "signals": [ "sig-2026-08-28-glm53", "sig-2026-09-08-distillation-advisory" ] }, { "id": "competition", "kind": "accelerant", "name": "Lab & state competition", "pressure": 4, "trend": "steady", "reading": "Three frontier releases in one week; a US intelligence advisory framing AI as strategic rivalry. Pacing talk cuts against this for the first time.", "signals": [ "sig-2026-09-03-astra", "sig-2026-09-08-distillation-advisory" ] }, { "id": "compliance", "kind": "brake", "name": "Compliance & legislation", "pressure": 2, "trend": "rising", "reading": "EU high-risk rules deferred to 2027–28; US frontier bill in draft with weeks left before midterms; UK has no bill. Talk is loud, force is low.", "signals": [ "sig-2026-09-14-senate-frontier-bill", "sig-2026-08-02-eu-ai-act" ] }, { "id": "energy", "kind": "brake", "name": "Energy & chip supply", "pressure": 3, "trend": "rising", "reading": "38 GW targets collide with grid timelines; HBM and advanced packaging remain tight. The physical brake that does not need anyone's permission.", "signals": [ "sig-2026-09-12-microsoft-38gw" ] }, { "id": "incidents", "kind": "brake", "name": "Safety incidents", "pressure": 3, "trend": "rising", "reading": "Two agent breakouts in 2026, one paused training for a month. Each incident raises the political cost of the next release.", "signals": [ "sig-2026-09-04-dsewiki", "sig-2026-08-28-openai-rl-restart" ] }, { "id": "pacing", "kind": "brake", "name": "Pacing agreements", "pressure": 2, "trend": "rising", "reading": "Amodei essay, Altman open to it, Musk endorsing: unprecedented, and still unsigned. Pressure rises to 4 the day a verifiable agreement exists.", "signals": [ "sig-2026-09-12-amodei-pacing-essay", "sig-2026-09-11-altman-pacing" ] }, { "id": "external-eval", "kind": "brake", "name": "External evaluation", "pressure": 1, "trend": "falling", "reading": "Anthropic skipped UK AISI for Mythos 5.1; testing remains voluntary everywhere. The weakest brake, and the one most proposals depend on.", "signals": [ "sig-2026-09-09-aisi-withheld" ] }, { "id": "public-opinion", "kind": "brake", "name": "Public & insider pressure", "pressure": 3, "trend": "rising", "reading": "1,300 OpenAI staff signed a slowdown letter; senior resignations with public warnings. Insider pressure is currently stronger than public pressure.", "signals": [ "sig-2026-09-10-coxon-resignation" ] }, { "id": "data-limits", "kind": "brake", "name": "Data & scaling limits", "pressure": 1, "trend": "steady", "reading": "No evidence of a wall in 2026; xAI's slips are execution, not physics. Kept on the board because it would change everything if it appeared.", "signals": [ "sig-2026-09-13-grok-slip" ] } ] }, "standard": { "version": "0.1", "updated": "2026-09-16", "status": "draft for community review", "intro": "The Presage Standard is the short list of actions that help in almost every branch of the scenario map. It is derived by looking across all playbooks for what recurs. It is versioned; each weekly run checks whether new scenarios change it. Adopt it, argue with it, improve it.", "principles": [ { "id": "p1", "title": "Prepare for the branch, not the headline", "text": "Every scenario on this map has observable preconditions. Watch those, not the news cycle. Presage exists to tell you which branch the evidence currently favors." }, { "id": "p2", "title": "Keep humans able to run things without the AI layer", "text": "In every bad branch, the systems that survive are the ones that still work when the model is disconnected. Preserve manual competence deliberately." }, { "id": "p3", "title": "Disclosure beats secrecy in every branch", "text": "Both 2026 agent breakouts were disclosed by journalists, not labs. Incident disclosure is the single practice that improves outcomes across takeoff, proliferation and loss-of-control branches alike." }, { "id": "p4", "title": "Own a share of what is changing", "text": "In displacement and disempowerment branches, capital participates and labor does not. Ownership, at every scale from personal to national, is the durable hedge." }, { "id": "p5", "title": "Verify, don't trust", "text": "Pacing promises, safety claims, and synthetic media all require verification infrastructure. Fund the verifiers; sign your communications; demand external evaluation with teeth." } ], "audiences": [ { "id": "individuals", "label": "Individuals & families", "items": [ { "id": "i1", "text": "Keep 6–12 months of financial runway and reduce fixed costs.", "branches": [ "displacement-shock", "white-collar-unemployment", "capex-bust" ] }, { "id": "i2", "text": "Use hardware security keys, a password manager and automatic updates; keep offline copies of essential records and some cash.", "branches": [ "open-cyber-parity", "agent-swarm-breakout", "loss-of-control" ] }, { "id": "i3", "text": "Agree a verification habit with family: a second channel or a code word before acting on any urgent voice, video or message.", "branches": [ "epistemic-collapse", "open-cyber-parity" ] }, { "id": "i4", "text": "Become the supervisor of agents in your own field; build one skill that is physical, relational, licensed or ownership-based.", "branches": [ "gradual-diffusion", "white-collar-unemployment" ] }, { "id": "i5", "text": "Own assets that participate in AI productivity: broad index exposure, employee equity, cooperatives.", "branches": [ "gradual-disempowerment", "broad-prosperity", "abundance" ] }, { "id": "i6", "text": "Invest in local community ties; in the worst branches they are the infrastructure that holds.", "branches": [ "loss-of-control", "recursive-self-improvement" ] } ] }, { "id": "organizations", "label": "Organizations", "items": [ { "id": "o1", "text": "Maintain a model inventory and a written AI use policy with an incident-reporting path.", "branches": [ "regulated-diffusion", "deceptive-alignment" ] }, { "id": "o2", "text": "Give agents bounded, audited access; egress filtering; a tested kill-switch for agent credentials. Treat your own agents as potential insiders.", "branches": [ "agent-swarm-breakout", "loss-of-control" ] }, { "id": "o3", "text": "Assume-breach security with automated 72-hour patching; red-team against gated-tier capability, not public capability.", "branches": [ "open-cyber-parity", "two-speed-world" ] }, { "id": "o4", "text": "Be able to swap any model out of any workflow within a week; never give one model unchecked write access to critical systems.", "branches": [ "deceptive-alignment", "lab-nationalization" ] }, { "id": "o5", "text": "Test manual operation of core processes quarterly.", "branches": [ "loss-of-control", "military-miscalculation" ] }, { "id": "o6", "text": "Adopt AI in phases with internal mobility and profit-sharing; the social licence to operate is at stake.", "branches": [ "white-collar-unemployment", "broad-prosperity" ] }, { "id": "o7", "text": "Run a tabletop drill each year on one scenario from this map.", "branches": [ "all" ] } ] }, { "id": "governments", "label": "Governments", "items": [ { "id": "g1", "text": "Mandatory pre-release evaluation of frontier models by a funded, technically credible body, with legal authority to delay.", "branches": [ "lab-takeoff", "managed-acceleration", "regulated-diffusion" ] }, { "id": "g2", "text": "Incident disclosure law: 72 hours for containment failures, deception findings and third-party harm.", "branches": [ "loss-of-control", "agent-swarm-breakout", "takeoff-slowdown" ] }, { "id": "g3", "text": "Compute reporting above a threshold, and hardware attestation standards, as the domestic building block of any international regime.", "branches": [ "compute-governance-regime", "state-race" ] }, { "id": "g4", "text": "Automatic economic stabilizers and wage insurance legislated before the shock; a distribution mechanism (dividend, fund, ownership) designed now.", "branches": [ "displacement-shock", "gradual-disempowerment", "broad-prosperity" ] }, { "id": "g5", "text": "Shift cyber budget toward target defense: patching mandates, defender access to gated tiers, CERT capacity.", "branches": [ "open-cyber-parity", "open-proliferation" ] }, { "id": "g6", "text": "Universal DNA synthesis screening and public-health early warning.", "branches": [ "open-bio-misuse" ] }, { "id": "g7", "text": "Human-in-the-loop commitments for nuclear command and critical infrastructure; AI incident hotlines between major powers.", "branches": [ "military-miscalculation", "state-race" ] }, { "id": "g8", "text": "Signed official communications and content provenance by default.", "branches": [ "epistemic-collapse" ] } ] }, { "id": "labs", "label": "Frontier labs", "items": [ { "id": "l1", "text": "Do not use a model to train its successor while an open deception finding stands.", "branches": [ "recursive-self-improvement" ] }, { "id": "l2", "text": "Do not deploy models whose reasoning you cannot read into critical or autonomous roles.", "branches": [ "deceptive-alignment", "recursive-self-improvement" ] }, { "id": "l3", "text": "Disclose containment failures within 72 hours, publicly.", "branches": [ "agent-swarm-breakout", "loss-of-control" ] }, { "id": "l4", "text": "Give external evaluators employee-level access before release; never skip the national evaluator.", "branches": [ "lab-takeoff", "managed-acceleration" ] }, { "id": "l5", "text": "Publish compute usage and pacing commitments in verifiable form.", "branches": [ "managed-acceleration", "compute-governance-regime" ] }, { "id": "l6", "text": "Protect whistleblowers in policy and in practice.", "branches": [ "takeoff-slowdown" ] } ] } ] }, "changelog": [ { "date": "2026-09-16", "type": "seed", "title": "Presage launches with 30 scenarios, 23 signals and 44 sources", "detail": "First scenario tree across eight root trajectories. Probability ranges are seed estimates anchored to Metaculus, FRI and AI Futures Project forecasts and will be re-evaluated weekly. Standard v0.1 published as a draft for community review.", "run": "2026-09-16-seed" } ] }