Are we living in the singularity now?
Retrieval got lighter, benchmarks shuffled, and yardsticks multiplied. The singularity still has forms to file.
The strict definition.
Not “AI did something strange.” A singularity requires broadly superhuman systems improving the systems that improve them, fast enough to break ordinary forecasting.
Threshold audit
Four locks, no shortcuts- Superhuman generalityRetrieve-for-Train concerns specialized retrieval, Epoch's index still shows uneven math and software-engineering results, and Anthropic reports process metrics; none establishes reliable best-human performance across most consequential domains.Not met
- Closed-loop AI improvementThe sources describe an offline training method, benchmark measurements, and oversight metrics for AI-assisted R&D; none verifies a system autonomously improving its own capabilities in a closed loop.Not met
- Self-sustaining accelerationThe reported retrieval technique, benchmark update, and proposed reporting measures may aid engineering and oversight, but none shows AI-generated improvements sustaining faster or more capable subsequent rounds without continued human direction.Not met
- Forecasting regime brokenA targeted systems method, updated benchmark scores, and proposed oversight measurements remain legible within ordinary technical and institutional forecasts; they do not show a broken forecasting regime.Not met
Three exhibits from this week
Evidence, not vibes- 01Watch the workshop
Anthropic proposed measurements for monitoring the pace of frontier AI development
Anthropic proposed reporting coverage, review latency, and escalation rates for agents working on AI research and engineering. It also described a system for overseeing and intervening in agent actions. These are transparency and oversight measures, not evidence that an AI system autonomously builds its successor.
Anthropic · 17 Sept 2026 ↗ - 02The scorecard moved
Epoch AI reported GPT-6 Astra leading its composite capability index
Epoch AI reported GPT-6 Astra at 166 on its Epoch Capabilities Index, ahead of Claude Fable 5.1 at 164 and GPT-5.6 Sol at 162. Astra set the reported Math-ECI record, while Fable led the reported software-engineering index; the mixed benchmark result does not establish general superhuman ability.
Epoch AI · 16 Sept 2026 ↗ - 03Search, packed for travel
Google Research described Retrieve-for-Train for lower-overhead complex retrieval
Google Research described Retrieve-for-Train, which uses reinforcement learning offline to turn reward-driven exploration into training data, then distills the behavior into a lightweight diffusion prior for retrieval. The work targets inference latency and compute overhead in specialized retrieval, rather than autonomous improvement of a general-purpose system.
Google Research · 15 Sept 2026 ↗
The verdict follows the strict intelligence-explosion tradition: greater-than-human general capability, closed-loop AI improvement, sustained acceleration, then a genuine break in predictability.