It’s Official: We’ve Entered AGI

It’s Official: We’ve Entered AGI

More

Summary

Julia McCoy argues that AI has crossed into the AGI era, anchoring her case in results from the ARC-AGI-3 benchmark, which drops AI agents into unfamiliar game worlds with no instructions and measures how efficiently they learn.

The video recounts how frontier models scored under 1% at launch, how OpenAI’s GPT-5.6 reached 7.78%, and how GPT-6 Astra later hit 99.9% using OpenAI’s own setup while using fewer moves than the median human on 96% of levels. McCoy also notes the ARC Prize team’s caveats: the games are closed and rule-based, and on a neutral, barebones setup Astra scored 62.7%. She uses that gap to argue that the environment around a model matters as much as the model itself.

The second half looks at agentic coding with Claude Code and Codex, citing METR’s time-horizon measurements, from about four minutes of human-equivalent work for GPT-4 in 2023 to at least 16 hours for Claude Mythos Preview. Examples include Claude Fable 5 and 5.1 builds, a DataCamp comparison with Claude Opus 5, and Codex runs lasting many hours. The video closes with practical prompting advice: define a clear finish line and let the model test its own work.


📺 Source: Julia McCoy · Published November 15, 2026
🏷️ Format: Opinion Editorial

1 Item

Channels