The AI safety research organization METR has published a new report on how Anthropic is using its own AI systems to speed up the development of future models and tools.
AI Safety · METR on Anthropic
Is AI Speeding Up the Building of AI?
METR's new report examines how Anthropic uses its own AI systems in engineering and research, and whether that assistance is shortening the time needed to build future models and tools.
The feedback loop at the heart of the debate
Why "AI accelerating AI" is a question worth measuring
1
AI assists the work
Coding, debugging, experiment design and data analysis
→
2
Development speeds up
Less time needed to build models and tools
→
3
Next models arrive sooner
Competitive pace and oversight needs shift
↺
4
Stronger AI feeds back
More capable systems can assist the next cycle
The case for acceleration
Frontier labs already use their own models across coding and research tasks.
Some researchers argue the feedback loop could compound progress over time.
Independent measurement from an evaluator like METR adds credibility to the claim.
The critical view
Productivity gains are hard to measure and may be uneven across tasks.
Risks grow if human oversight does not keep pace with AI-assisted work.
Team changes and shifting priorities can be mistaken for AI's contribution.
What makes the claim verifiable
Acceleration is only meaningful if outsiders can see how it was defined and measured.
Definition
How the report defines acceleration
Baseline
What the speed-up is compared against
Task type
Routine engineering versus genuinely new research
Attribution
AI's share separated from other factors
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article...