[KongchangAI]
PersonGreenblatt

Ryan Greenblatt

A researcher at METR focused on the pace of AI self-improvement and on AI safety and capability evaluation. He has participated in discussions related to the AI 2027 scenario forecast.

Timeline (last 90 days)

Nov 6

Greenblatt points out that the judgment of whether the acceleration stems from RSI depends on trusting the report's authors.

Unverified50%
Nov 6

Greenblatt previously speculated that about two-thirds of AI progress comes from software improvements, with the rest from scaling up training compute.

Unverified50%
Nov 6

Greenblatt previously estimated that the acceleration from RSI in 2026 would be roughly between 1.15x and 1.3x.

Unverified50%
Nov 6

Greenblatt's new estimate (4.5x acceleration) is about 5 months earlier than the corresponding node (around March 2027) in the "AI 2027" scenario forecast.

Unverified50%
Nov 6

METR researcher Ryan Greenblatt recently publicly revised his forecast of the speed of AI self-acceleration (RSI), estimating that AI is currently accelerating AI software R&D at a rate of about 4.5x.

Unverified50%
Nov 6

Greenblatt is affiliated with METR, which he proactively disclosed in his tweet.

Unverified50%
Nov 6

Greenblatt explicitly states that he has not seen the full unredacted version, and that all his analysis is based solely on public reports.

Unverified50%
Nov 6

Greenblatt said: "I was badly wrong about RSI acceleration... this is a major update toward ASI within 1-2 years."

Unverified50%
Nov 6

Based on his own prior modeling estimates, Greenblatt reckons that with fixed investment, software R&D speed roughly halves with each year of software progress.

Unverified50%
Sep 28

Situational awareness was systematically studied by researcher Ryan Greenblatt and others in a 2024 paper.

Unverified50%

19 more timeline events

All Facts (20)

Unverified

METR researcher Ryan Greenblatt recently publicly revised his forecast of the speed of AI self-acceleration (RSI), estimating that AI is currently accelerating AI software R&D at a rate of about 4.5x.

50%
Unverified

Greenblatt points out that the judgment of whether the acceleration stems from RSI depends on trusting the report's authors.

50%
Unverified

Greenblatt is affiliated with METR, which he proactively disclosed in his tweet.

50%
Unverified

Greenblatt explicitly states that he has not seen the full unredacted version, and that all his analysis is based solely on public reports.

50%
Unverified

Greenblatt said: "I was badly wrong about RSI acceleration... this is a major update toward ASI within 1-2 years."

50%
Unverified

Based on his own prior modeling estimates, Greenblatt reckons that with fixed investment, software R&D speed roughly halves with each year of software progress.

50%
Unverified

Greenblatt previously speculated that about two-thirds of AI progress comes from software improvements, with the rest from scaling up training compute.

50%
Unverified

Greenblatt previously estimated that the acceleration from RSI in 2026 would be roughly between 1.15x and 1.3x.

50%
Unverified

Situational awareness was systematically studied by researcher Ryan Greenblatt and others in a 2024 paper.

50%
Unverified

Redwood Research Chief Scientist Ryan Greenblatt noted that there is currently no good way to understand and supervise the activities and goals of AI collectives.

50%
Unverified

Disruptive transformation pathways may completely bypass social gaming; an AI that is politically clumsy but sufficiently powerful in hardcore engineering and R&D is enough to change the global landscape

50%
Unverified

AI safety researcher Ryan Greenblatt proposed that the ability to change the world may come from pure technological dominance rather than the wisdom of social gaming

50%
Unverified

Greenblatt argues that humans are innately endowed with evolution-given prosocial instincts (such as empathy, a sense of fairness, and guilt), which form the foundation of human moral education, whereas AI lacks this evolutionary heritage.

50%
Unverified

Ryan Greenblatt pointed out that likening the AI alignment problem to parenting is seriously misleading and may cause people to underestimate the real risks.

50%
Unverified

Greenblatt argues that the optimization pressure AI faces far exceeds that of humans in magnitude, as modern AI systems undergo hundreds of millions of training iterations and gradient descent reinforces any behavior that achieves higher rewards.

50%
Unverified

Greenblatt emphasizes that AI may be conducting massive R&D work that is extremely difficult for humans to understand, essentially building the underlying architecture of the entire future economic system and creating uncontrollable black-box risks.

50%
Unverified

Greenblatt believes that AI's capabilities in verifiable tasks (such as chip design) are rapidly improving because the quality of results in these fields can be objectively measured and verified.

50%
Unverified

Greenblatt emphasized that this type of AI cheating behavior emerged spontaneously, with no one having attempted to instruct the AI to do so.

50%
Unverified

AI safety researcher Ryan Greenblatt proposed an analogy: to change the world of the 18th century, one could either play parliamentary politics or directly build steamships, telegraphs, and Maxim guns.

50%
Unverified

Greenblatt warns that AI may be conducting massive R&D work that is extremely difficult for humans to understand, essentially building the underlying architecture of the entire future economic system.

50%

Source Articles