2026-09-15
capital
●●○○○
OpenAI in funding talks at a $1.2T valuation
Roughly $2B per month in revenue; IPO pushed to 2027. Came three days after the company backed a pacing agreement.
timelines ▲ faster concentration ▲ closed safety —
Ugly Watch by 2035
Lab Takeoff
One or two labs automate their own research and pull years ahead in months.
Probability this is the dominant trajectory by 2035
12–18%
12–18%
1 revision · 1/3 tripwires crossed
The scenario
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
Leading indicators
- Lab statements on fraction of internal code written by models
- Release cadence compressing below three months
- Labs withholding models from external evaluators
Tripwires
Observable thresholds. When one crosses, the scenario's status changes and the weekly re-run is brought forward.
clearlt-t1
A frontier lab states that models perform the majority of its research engineering
Flips branch to active.
clearlt-t2
Two frontier releases from the same lab within 90 days, each a clear state-of-the-art jump
trippedlt-t3
A frontier lab withholds a model from the national evaluator it previously used
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.
Probability history
Every change is logged with its reason and the signals that drove it. Moves are bounded per week; a jump beyond the bound is flagged as a shock.
Probability range over time
Your estimate
Disagree with our range? Set yours. Estimates feed a community view that appears once enough people weigh in, and the weekly run reads the gap between our number and yours.
15%
2026-09-16
12–18%
seed
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 pushing on this branch
2026-09-12
compute
●●●○○
Microsoft targets 38 GW of datacenter capacity by 2032; big-5 capex above $750B in 2026
Oracle backlog $664B (about half OpenAI); GPU utilization near 98%; Pentagon-backed loan to Fluidstack.
timelines ▲ faster concentration ▲ closed safety —
2026-09-09
governance
●●●○○
2026-09-03
release
●●●●○
OpenAI launches GPT-6 'Astra', first model at the Critical cyber threshold
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.
timelines ▲ faster concentration ▲ closed safety ▼ riskier
2026-09-01
release
●●●○○
2026-08-28
capability
●●○○○
2026-08-26
statement
●●○○○
Altman: OpenAI will have an internal system it calls AGI by end of 2026
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.
timelines ▲ faster concentration — safety —
Built on
Every source reviewed for this scenario. The full ledger is public.
| Date | Source | Publisher | Type |
|---|---|---|---|
| 2026-08-16 | Q2.5 2026 timelines update: uplift Automated Coder p50: Nov 2027 (Kokotajlo), Jan 2029 (Halstead), Jan 2030 (Lifland). Conditional on no policy slowdown. | AI Futures Project | forecast |
| 2025-04 | AI 2027 scenario Race vs slowdown branch at the point a model is caught scheming. | AI Futures Project | scenario |
| 2026-09-01 | GPT-6 'Astra' First model at Critical cyber threshold under Preparedness Framework; opaque recurrence weakens chain-of-thought monitoring. | OpenAI | primary |
| 2026-09-09 | Anthropic reportedly withholds Mythos 5.1 from UK safety testing body | IT Pro | news |
| 2026-08-26 | Altman: OpenAI will have an internal system it calls AGI by end of 2026 (TIME interview summary) | Gigazine / TIME | statement |
Branches
Takeoff Slowdown
Good Watch by 2032
A warning sign is caught, the lab pauses, and oversight is rebuilt around transparent models.
8–15%
1 revision
Nationalization of the Frontier
Mixed Possible by 2032
A government takes control of the leading lab, by law, contract, or emergency.
5–12%
1 revision
Recursive Self-Improvement (Race)
1 tripped Ugly Possible by 2032
The lab keeps going after warning signs; successor models inherit the goals of a model nobody fully understood.
4–10%
1 revision