AGI is Here. Anthropic Just Proved It.

Overview

The video argues that “practical AGI” is already here, based on internal performance data released by Anthropic in a report titled When AI builds itself. The speaker claims Anthropic’s AI—Claude—is increasingly producing the majority of the company’s code and, more importantly, is improving at the hardest kind of work: open-ended problem solving, where the correct output and even the specification are unclear.

---

Key claims and evidence presented

AGI is defined practically, not sci‑fi

The speaker argues that the key criterion is whether a model can take a previously unseen, unclear, open-ended problem and independently figure out an approach by running experiments/research—rather than only excelling at narrow, predefined tasks.

Anthropic’s internal benchmarks show rapid gains on open-ended tasks

The report is described as organizing coding work into difficulty tiers, culminating in open-ended tasks defined as having:

• No clear specification
• No known target form

Reported benchmark changes for “open-ended” problems:

• Success rate rose from 26% (6 months earlier) to 76%

Agents are doing longer work with less supervision

The video claims task completion time increased sharply:

• ~4 minutes (2 years ago)
• → 1.5 hours (1 year ago)
• → 12 hours (this year)

It also notes some newer internal models running ~16 hours continuously.

AI increasingly makes better “next-step” decisions than humans

The video claims Anthropic ran tests where, at decision points, they asked the AI what to do next and compared its choice to human researcher decisions.

Reported results:

• 51% better direction in November
• rising to 64% by April
• across 129 decision moments

Speed and autonomy in research/engineering are accelerating dramatically

Examples cited:

• A code chunk previously described as ~3x faster became 52x faster with a newer model.
• For a problem researchers were stuck on, AI agents allegedly closed 97% of the gap versus 23% by humans after about a week.

---

Scenarios for what happens next (Anthropic’s framing)

From the report, the speaker summarizes three trajectories:

1. Progress plateaus AI remains powerful, but improvements slow or stall.

2. Humans keep directing while AI compounds progress Humans set research directions and judge results, while AI does most of the work.

3. AI builds its own successor Progress becomes less constrained by human involvement and depends more on compute.

The speaker argues that scenario #2 is already effectively the present reality, leading to the conclusion that AGI (under the practical definition that matters) is already here. The remaining question is whether the system transitions into scenario #3.

---

Main concern: compounding misalignment and losing interpretability

A central warning is that errors or “misalignment” could compound as models create future models:

• Mistakes may multiply as each successor inherits flaws.
• The resulting issues could become harder to detect while increasing in frequency.
• The speaker emphasizes this is not “killer robots,” but rather that experts may lose understanding of what is being built underneath.

---

Societal impact and incentives

The video claims society splits into two groups:

• People using AI casually and dismissing it as overhyped.
• Teams using AI agents to automate work that previously required large organizations.

It cites Anthropic’s claim that:

• 100-person companies could do work comparable to 10,000–100,000-person organizations,

which the speaker argues will widen gaps between individuals and organizations.

Incentives are framed as a race dynamic:

• Everyone has incentives not to slow down.
• Stopping is unlikely because competitors won’t pause.

---

Why Anthropic might support slowdown (but verification is hard)

The speaker notes Anthropic reportedly expresses that:

• alignment is the least certain problem
• and a slowdown would help—

but only if all major labs pause and that pause can be verified.

Additional points mentioned:

• Training runs are easier to conceal than missile silos.
• The video compares verification to Cold War nuclear inspection regimes (“trust but verify”), but argues those took decades to build and “we don’t have that long.”

---

Final takeaway

The speaker ends by reframing the most valuable human advantage as shifting from:

• doing tasks to

• judgment and taste—deciding which problems are worth tackling.

The main risk is described as less about apocalyptic outcomes and more about misuse and loss of control/understanding.

---

Presenters or contributors

• Presenter/author (video speaker): The narrator/speaker (not identified by name in the subtitles)
• Organization referenced: Anthropic (Claude models, report When AI builds itself)