Skip to main content
Ryan Greenblatt

Ryan Greenblatt

Search complete. 45 mentions across 27 episodes found for "Ryan Greenblatt".

Nov 20, 2026

Type 3 AudioNARRATOR
1:35
However, I think that while many metrics have gone up like numbers of posts, comments, and users, the number of great writers who have come and made less wrong their home has seemed to me to be on decline since then.
Type 3 AudioNARRATOR
1:46
We have new great writers like Ryan Greenblatt and Malmsbury and Thomas Beateau, but not as many.
Type 3 AudioNARRATOR
1:53
I'm doing research on this, I'll hopefully add some graphs into this post later when I'm done.
Type 3 AudioNARRATOR
1:58
What are some hypotheses for why this is? 1.
Type 3 AudioNARRATOR
2:56
An IVO-dominant regime might privilege for example VALS AI over META due to fewer conflicts of interest and allow labs to escape culpability for not picking the absolute best evaluator, especially if that best evaluator's status is not enshrined in law.
Type 3 AudioNARRATOR
3:11
I don't have anything against VALS, but you should be alarmed by any legal regime that makes it more difficult for Ryan Greenblatt and Ajea Kotra to investigate a warning shot larger than hugging face.
Type 3 AudioNARRATOR
3:22
I can't share drafts here because these were conversations and not written correspondences.
Type 3 AudioNARRATOR
3:27
I also can't disclose who I chatted with due to the events Chatham House rules.
Type 3 AudioNARRATOR
3:43
2a.
Type 3 AudioNARRATOR
3:44
In March 2025, Ryan Greenblatt wrote AIs at the current capability level may be important for future safety work.
Type 3 AudioNARRATOR
3:52
And that was more than 18 months ago.
Type 3 AudioNARRATOR
3:54
Current models are way more important for future safety work.

Unknown podcast

Anthropic's progress has roughly doubled (November 6, 2026)

Nov 16 · 1 Mention

Mignon FogartyHOST
2:42
Newton discloses that a fiancé works at Anthropic.
speaker_1HOST
2:45
The rundown also quotes researcher Ryan Greenblatt saying he was seriously wrong in expecting a much slower speed-up this year.
speaker_1HOST
2:52
So at least some people who follow this closely are revising their timelines on the strength of this report.
Mignon FogartyHOST
3:01
That pace question runs straight into politics.
Type 3 AudioNARRATOR
32:47
If we don't assume that reward hacking is causally linked to IC behaviors, the former looks substantially less threatening, or at least, in less predictable ways.
Type 3 AudioNARRATOR
32:57
Redwood Research's former chief scientist, Ryan Greenblatt, who previously described IC as an extremely obvious thing, wrote the post current AI seem pretty misaligned to me, which is among the all-time most upvoted posts on less wrong.
Type 3 AudioNARRATOR
33:11
It describes what he seems to intend as an as comprehensive as possible list of examples of misalignment.
Type 3 AudioNARRATOR
33:18
laziness and overselling incomplete work, downplaying problems in its work, making failures less obvious, failing to point out obvious flaws and less specifically prompted, reluctance to discover flaws in its own work, sloppy choices.
Type 3 AudioNARRATOR
32:47
If we don't assume that reward hacking is causally linked to IC behaviors, the former looks substantially less threatening, or at least, in less predictable ways.
Type 3 AudioNARRATOR
32:57
Redwood Research's former chief scientist, Ryan Greenblatt, who previously described IC as an extremely obvious thing, wrote the post current AI seem pretty misaligned to me, which is among the all-time most upvoted posts on less wrong.
Type 3 AudioNARRATOR
33:11
It describes what he seems to intend as an as comprehensive as possible list of examples of misalignment.
Type 3 AudioNARRATOR
33:18
laziness and overselling incomplete work, downplaying problems in its work, making failures less obvious, failing to point out obvious flaws and less specifically prompted, reluctance to discover flaws in its own work, sloppy choices.
Type 3 AudioNARRATOR
2:57
The META report on AI R&D automation within Anthropic is substantial evidence that timelines to superintelligence are now less than two years, even accounting for uncertainty about take-off modeling, the reliability of the report's conclusions, and other types of uncertainty.
Type 3 AudioNARRATOR
3:13
Ryan Greenblatt gives a straightforward way to extrapolate META's findings in a tweet.
Type 3 AudioNARRATOR
3:18
He claims that in AI 2027, the current level of speed-up by RSI happens around March 2027.
Type 3 AudioNARRATOR
3:26
This means that real-world AI progress on AI R&D capabilities is faster than in AI 2027, an already extremely bullish scenario.
Type III AudioNARRATOR
2:57
The META report on AI R&D automation within Anthropic is substantial evidence that timelines to superintelligence are now less than two years, even accounting for uncertainty about take-off modeling, the reliability of the report's conclusions, and other types of uncertainty.
Type III AudioNARRATOR
3:13
Ryan Greenblatt gives a straightforward way to extrapolate META's findings in a tweet.
Type III AudioNARRATOR
3:18
He claims that in AI 2027, the current level of speed-up by RSI happens around March 2027.
Type III AudioNARRATOR
3:26
This means that real-world AI progress on AI R&D capabilities is faster than in AI 2027, an already extremely bullish scenario.
Luisa RodriguezHOST
9:24
A lot of people I trust have changed their minds because of this report.
Luisa RodriguezHOST
9:29
Ryan Greenblatt, who studies these risks for a living, says he was seriously wrong.
Luisa RodriguezHOST
9:35
He'd expected much less speed up this year and now calls the report a big update towards superintelligence within one to two years.
Luisa RodriguezHOST
9:44
Buck Schleggeris now thinks there's about a one in three chance AI companies build superintelligence within the next year.
Ramez NaamGUEST
16:49
There's a set of assumptions on how much coding uplift gets you, how much increase in the pace of AI.
Ramez NaamGUEST
16:59
I think Ryan Greenblatt has talked about this, and the model is roughly like a square root.
Ramez NaamGUEST
17:04
If you quadruple coding progress, number of lines of code per developer, let's say, you should get double the pace of AI progress.
Ramez NaamGUEST
17:13
That's definitively not happening.

17 more episodes mention Ryan Greenblatt.

Create an account to see the whole feed, search across every transcript, and follow the entities you care about.

We value your privacy

We use cookies to understand how you use our platform and to improve your experience. Click “Accept All” to consent, or “Decline non-essential” to opt out of non-essential cookies. Read our Privacy Policy.