Track AGI, ASI and a conditional extinction pathway. Explore the assumptions and the evidence—not a predetermined future.
01 / SCENARIO EXPLORER
Choose an assumption set
REFERENCE CASE / MODEL v2.0
From general intelligence to a high-stakes frontier.
Fast progress continues, but reliability, infrastructure and deployment constraints still matter.
Illustrative scenarios—not calibrated forecasts. The dates are editor-selected assumptions. The final milestone assumes a catastrophic pathway occurs; humanity may survive indefinitely.
A system that reliably matches a competent human across most economically useful cognitive tasks, including unfamiliar multi-step work with limited supervision. Not one benchmark and not necessarily embodied intelligence.
02 / ASI
2035selected scenario year
Artificial superintelligence
2030–2055 assumption range
A system substantially beyond the best human experts across nearly all cognitive domains, including scientific research and strategic planning. Narrow superhuman performance does not qualify.
03 / Extinction pathway
2042selected scenario year
Conditional extinction
2032–2085 assumption range
Human extinction substantially caused by AI. Dates are conditional illustrations of an ASI-enabled loss-of-control pathway, assuming extinction occurs. Other pathways could occur before ASI; survival indefinitely is also possible.
A range of futures, not a deadline
Ranges are not statistical confidence intervals
202620582090
AGI
2031
ASI
2035
Extinction pathway
2042
AGI does not imply ASI. ASI does not imply extinction. These are separate, conditional transitions.
LAST COLLECTIONSep 21, 2026, 11:08 PM UTC
EVIDENCE COVERAGE24 reviewed / 240 sources
EXTINCTION PROBABILITYNot estimated
02 / THE EVIDENCE DESK
What’s shaping the outlook
Original sources, visible limitations. Capability gains and protective signals both count.
arXiv preprint proposing RBS-Attention, a training-free sparse-prefill method that reports large prefill-attention speedups (20x+) with small accuracy loss on long-context LLMs.
Limitation Preprint results reported on particular models/hardware; reproducibility and generality beyond the cited setups are not independently verified in the excerpt.
Why it matters & timeline effect
This is a technical preprint showing a methods improvement that materially reduces prefill cost for very long contexts and reports concrete speed/accuracy tradeoffs on specific models/hardware. That suggests an incremental capability improvement enabling more efficient long-context inference, which is relevant to overall model capabilities but is a preprint and not independently reproduced in this excerpt.
Not applied. Proposed movement: 0 months.
Reviewer confidence: medium. This describes the evidence assessment, not confidence in a forecast.
“RBS-Attention achieves 20.65$\times$ standalone prefill-attention speedup” Feed excerpt; the full article was not evaluated.
arXiv preprint introduces Attention-Aware Routing (AAR) for MoE models, reporting +3.37 pp on GSM8K and describing how routing changes propagate to reshape attention, suggesting both performance gains and insights into model circuits.
Limitation Preprint work with reported benchmarks on specific MoE setups (OLMoE); performance and circuit claims may not generalize and require independent replication.
Why it matters & timeline effect
The excerpt is a preprint that documents measurable performance improvements and an interpretability-style finding (routing-attention coupling). These are relevant to capability development and to understanding internal model dynamics, but the work is limited to the experiments described and is not independently verified here.
Not applied. Proposed movement: 0 months.
Reviewer confidence: medium. This describes the evidence assessment, not confidence in a forecast.
“AAR improves GSM8K by +3.37 pp over a routing-only SFT baseline on OLMoE.” Feed excerpt; the full article was not evaluated.
03 / NO BLACK BOX
Uncertainty is part of the model.
A useful warning system should show its assumptions, not hide them behind a precise-looking countdown.
Experimental scenario model · v2.0
Where do the dates come from?
They are explicit editorial starting assumptions, not expert consensus or an AI-derived probability forecast. The reference set starts with AGI in 2031, four years from AGI to ASI, and seven more years to the conditional extinction pathway. Fast takeoff uses 2028 + 2 + 3 years; slower progress uses 2040 + 10 + 15. Range endpoints pair earlier AGI with shorter lags. They are not percentiles. A minimum one-year gap between stages is a simplifying assumption.
What can move the clock?
Daily collection gathers allowlisted RSS feeds. An optional model reviews up to 12 excerpts per run with source quotations and limitations. Review is not independent fact-checking. By default, proposals do not change dates. Experimental automatic application requires explicit server configuration, a high-confidence assessment, a known recent publication date and a new event. Extinction-lag proposals are never automatically applied.
AGI-date and ASI-lag adjustments are capped at two months each per run and five years each cumulatively. An AGI shift propagates downstream; displayed ASI and conditional extinction can therefore move four months in a run or ten years cumulatively. Caps are engineering choices, not scientific calibration. At most one nonzero event per publisher per run is eligible; semantic event deduplication is best-effort.
Why isn’t there a risk percentage?
No validated method converts current news into a numerical probability of human extinction. The previous version’s percentage and fixed countdown were unsupported and have been removed. A faster timeline is not the same thing as a higher probability of extinction.
What’s missing?
RSS coverage is incomplete; feeds may lag or fail. Automated analysis sees excerpts, not full articles, and can misinterpret them or miss events. arXiv papers are not necessarily peer reviewed. Developer benchmarks need independent corroboration. Unforeseen breakthroughs, plateaus, policy changes and non-ASI pathways are outside this simple model. No finite range excludes “never.”
04 / OPERATIONAL TRANSPARENCY
Behind the updates
No silent fallback to “live.” Collection, analysis and storage have separate states.
Source health
Feeds are checked by the scheduled collector, not by visits to this page.