Joined August 2022
- Tweets 6,042
- Following 3,954
- Followers 1,509
- Likes 42,927
521 Photos and videos
Pinned Tweet
Jan 2
Hi all, I'm interested in feedback!
You can leave anonymous comments here: admonymous.co/jsd
1
5
2,962
Jun 13
The moment is ripe for natural experiments on (R&D compute, labor) complementarities!
Jun 13
Some Anthropic researchers are probably thrilled to have tons of extra research compute for the next few days with Fable turned off 😂
1
2
51
7,910
Jun 12
I think there's people in my broader community who sacrificed a lot / dealt with enormous stress to fix some things or keep the (good parts of the) show running in the wake of the FTX catastrophe, and haven't gotten much praise for it, and I want to extend gratitude to them.
14
481
jsd retweeted
Jun 11
How big a leap is Mythos in cyber capabilities?
@timotheechauvin, @AlexBarry4, @js_denain, and @ansonwhho compiled the public evidence and found that while it’s unclear if Mythos was ahead of trend in discovering vulnerabilities, it represents a big jump in exploiting them. 🧵
8
37
350
33,578
Jun 11
This is awesome work on an important topic.
Jun 10
New paper!
Think Fast: Estimating No-CoT Task-Completion Time Horizons of Frontier AI Models
@METR_Evals showed that models' time horizons have doubled every few months. We ask: what length of tasks can models complete without any CoT?
1
15
700
Jun 10
2 virtues I care about, and want to embody more:
- wholesomeness
- seeing, and being willing to name, inconvenient/uncomfortable truths (think, Elephant in the Brain stuff)
These can often seem in tension. I'm interested in pointers for how to reconcile them.
9
2
40
2,611
jsd retweeted
I recently wrote about the effort to end the torture of pigs. On Thursday, @Dwarkesh_sp, @AvitalBalwit, and @NanRansohoff are hosting an in-person party for folks in the Bay Area to support the effort!
If you'd like to attend, sign up here! forms.gle/8EKxF6y6bnpYMfE38
13
59
685
113,955
Jun 7
I find it surprisingly hard to predict whether someone lost their faith because of "problem of evil" vs "I just don't believe in the existence of God"
I'm in the latter camp, and expected the same of most people very similar to me, but that doesn't seem to be the case!
4
6
614
Jun 2
Enjoyed this podcast a lot.
Quick clarification on Rosetta stone (≈ ECI)
① @rohinmshah says the ECI is mostly linear over time. I disagree with this, cf @AlexBarry4 and my recent analysis in "Have AI Capabilities Accelerated?" epoch.ai/publications/have-a…. (Note that this analysis is from April 2026, while the podcast was recorded in December 2025)
② I agree with something like "the ECI trend is overall very smooth compared to what you might expect from a statistical model that does not encode time information at all".
③ I think that this acceleration in ECI has come largely from increasing correlation between benchmarks and tasks that AIs are trained on directly. If we were looking at a broader set of tasks, including harder to benchmark tasks, we'd likely see less (or no) acceleration.
Jun 2
My best interview in some time.
Rohin Shah leads AGI alignment/safety at DeepMind.
And he has a lot of spicy personal takes:
We probably won’t get catastrophic misalignment (00:49)
Safety 'commitments' have severe limitations (10:38)
The intelligence explosion probably isn't imminent (1:52:44)
Why he's not working to pause AI advances (51:44)
Pre-deployment evals aren't the right focus (for catastrophic risks) (37:41)
Signalling concern for safety sometimes diverts resources from actually making AI safe (01:09:51)
Reading AI thoughts is v useful for safety – and we'll probably be able to for years to come (54:17)
Governance is somewhat more likely to be the bottleneck than alignment (43:55)
Rohin's team doesn't have a veto, and that's OK (27:36)
Central banks are a promising model for regulating AI (33:34)
Also:
Google DeepMind's actual plan for building AGI safely (1:40:29)
How external researchers can positively influence big AI companies (2:21:55)
The roles GDM most needs to hire for (2:37:03)
On the 80,000 Hours Podcast. Links below - enjoy! (@rohinmshah)
2
2
26
2,869
Jun 2
Separately, Rohin cautions against using the ECI to compare open and closed models, since open models are likely more overfit to benchmarks than closed ones.
I agree that this is a major issue with this method, but I still think it's worth doing, if only to get a lower bound on the gap.
We're updating the Limitation section here to mention this limitation. epoch.ai/data-insights/open-…
I'm also excited about analyses that compare the gap on private vs public benchmarks, for example from @htihle here: lesswrong.com/posts/rJcCrXyE….
2
2
14
949
jsd retweeted
Jun 2
My best interview in some time.
Rohin Shah leads AGI alignment/safety at DeepMind.
And he has a lot of spicy personal takes:
We probably won’t get catastrophic misalignment (00:49)
Safety 'commitments' have severe limitations (10:38)
The intelligence explosion probably isn't imminent (1:52:44)
Why he's not working to pause AI advances (51:44)
Pre-deployment evals aren't the right focus (for catastrophic risks) (37:41)
Signalling concern for safety sometimes diverts resources from actually making AI safe (01:09:51)
Reading AI thoughts is v useful for safety – and we'll probably be able to for years to come (54:17)
Governance is somewhat more likely to be the bottleneck than alignment (43:55)
Rohin's team doesn't have a veto, and that's OK (27:36)
Central banks are a promising model for regulating AI (33:34)
Also:
Google DeepMind's actual plan for building AGI safely (1:40:29)
How external researchers can positively influence big AI companies (2:21:55)
The roles GDM most needs to hire for (2:37:03)
On the 80,000 Hours Podcast. Links below - enjoy! (@rohinmshah)
24
84
847
153,125
Jun 1
2
334
Jun 1
A lot of philosophical ethics and metaethics seems very confused to me.
I like Base Camp for Mount Ethics which IMO cuts through a lot of the confusion.
nickbostrom.com/papers/mount…
1
8
739
jsd retweeted
May 29
I think most domains look like this at the moment: the returns to expenditure on agents diminish much more quickly than the returns to expenditure on human labor: (1/n)
34
99
708
182,541
May 28
Ok, but what *does* the European Commission think about UDASSA?
2
2
31
2,111
jsd retweeted
May 26
Help us produce the most useful work on AI by taking our 5-minute survey: docs.google.com/forms/d/e/1F…
(You can sign up at the end to join our compensated user research panel.)
3
7
23
9,951