AGI is Here. Anthropic Just Proved It.

The video opens with the striking statistic that over 80% of the code Anthropic ships is now written by its own AI, Claude, sourced from an internal report titled 'When AI Builds Itself.' The creator uses this as a launching point to argue that AGI — defined practically as a general model that can take an open-ended, unsolved problem and figure out a solution autonomously — has already arrived.

The creator rejects sci-fi definitions of AGI and instead focuses on a functional definition: can you hand a model a problem with no clear answer, and have it independently research, experiment, and return a working solution? Using this lens, he walks through Anthropic's internal data. Anthropic categorizes coding tasks into four tiers — trivial, routine, substantial, and open-ended — and Claude's success rate on the hardest 'open-ended' tier jumped from 26% to 76% in just six months. The duration of tasks AI can handle autonomously has also grown exponentially: from 4-minute tasks two years ago, to 1.5-hour tasks a year ago, to 12-hour tasks now, with projections suggesting week-long tasks could be within range by 2027.

The report also reveals that when AI was tested against human researchers at decision points in real projects, it chose the better next step 64% of the time by April (up from 51% in November). On a specific optimization task, the AI made code 52 times faster compared to 3x a year ago, and on a research problem that stumped humans (who recovered 23% of lost ground in a week), AI agents working continuously recovered 97%.

Anthropix outlines three future scenarios: a plateau, continued compounding gains with humans still directing research, or full AI self-improvement where progress is limited only by compute. The creator argues that scenario two is already the present reality, and that scenario three — the dangerous one — is an open question even Anthropic cannot answer. He highlights the alignment risk: flaws in today's models could compound invisibly as AI builds its successors, becoming more frequent and less understood over time.

The creator zooms out to social implications, noting a widening gap between people who use AI superficially and those leveraging it to run operations that previously required large teams. Anthropic itself claims 100-person companies could do the work of 100,000-person organizations. He acknowledges Anthropic's stated desire to slow down, but notes this is contingent on all major labs pausing simultaneously and verifiably — something nearly impossible since AI training runs, unlike missile silos, are invisible to outside observers. The video closes by emphasizing that the most valuable human skill going forward is judgment: knowing which problem to point the AI at, seeing the bigger picture, and thinking beyond the immediate task.