MTS @ArtificialAnlys // RL Residency @PrimeIntellect & Tinker Grant @thinkymachines // Skipping class @UCBerkeley
San Francisco, CA
Joined July 2018
- Tweets1K
- Following2.4K
- Followers3.7K
- Likes6.1K
Scaling Era, 2019-2025 (by @dwarkesh_sp and @gleech) was a great almanac of that time period.
This week seems like it'll have many records of similar importance to that work for the 2026-202? era.
Snowballing from events kicked off by HF-OAI incident, the RSI discussions, we've seen many thought-provoking debates.
Great panel hosted by @yminsky with @Jsevillamol @EgeErdil2 @RyanGreenblatt @DKokotajlo is one for the annals, highly recommended: youtu.be/IlgvRtTwaUs?si=m0Hl…
(Also watch today's @MTSlive coverage (two great days in a row!) with Ege and @bshlgrs)
1
54
I get this sentiment. It feels this way now, but I would humbly suggest reconsidering.
The game isn’t over, yet.
@MTSlive had @ramez and @AI_Futures_ authors back to back yesterday.
Regardless of the side you are currently leaning on, it’s useful to consider refined arguments side-by-side.
One way or the other, in roughly 500 days the current “hard take-off” debate will be resolved.
Many people (and institutions) put out bold (and falsifiable) public claims, that bit I respect.
Will be fun to watch regardless.
It’s a positive development that there’s now some convergence in the AI risk community that there’s likely an intelligence explosion next year. I think it’ll be clear in a few years that this didn’t happen, these people will be largely discredited, and discussions will become more sane.
14
1,021
Adam is one of the most talented and insightful analysts evaluating AI.
We are super happy to support his efforts.
(of course he retains full independence and I’m sure we’ll continue to learn from his criticisms as well!)
5
49
1,455
@AI_Futures_ (creators of the great AI2027 project) just came out w their Q3.5 update:
• Median forecast for superhuman AI researchers set at 2027
• 35-65% odds of AI capable of automating all white-collar work before 2028
• ~1.7× AI-attributable acceleration in overall capability progress
• Authors estimate we’re ~5 months ahead of AI 2027
Very timely (given @METR_Evals report on Anthropic) and informative read
We have come full circle on em dashes. Oldest models overused them; successors were heavily RL'ed against them; now the pendulum swings back!
As a human — it feels kinda avant-garde, contrarian to use them
Fable 5.2 breaks a record that Anthropic hasn't touched since Fable 5 — most emdashes in the announcement post
3
36
1,932
Better Knowledge benefits all of us (especially amidst the slop-thunderstorm); community notes are a great example of that at scale.
If you have an idea like that, consider applying to epistemic.fund/
Thanks to @ohabryka and the team!
14
694
Certainly, there are many different ways to define the nebulous term "big model smell".
I encourage those who are curious to look at the association between Total Parameters and Accuracy (for open models) and come to their own conclusions
2
31
2,168
I like no-CoT evals; do they let us spot architectural decisions, or perhaps are breaks from the trendline signaling something else?
I ran @NeelNanda5’s no chain of thought reasoning and knowledge evals on Fable 5.2. It’s slightly above the Anthropic trend and slightly below GPT-6.1 Sol.
1
24
1,233
I think we have to collectively have a talk about what these words mean because I really think people don't mean the same thing (hopefully).
Exhibit A:
1
22
1,637
I expect this broad trend will continue. Especially wrt MLP layers, there seems to be a lot of room to prune.
You don’t need the model to memorize a random street name in Uzbekistan to write unit tests.
Life is so funny sometimes.
People spend years trying to create insanely elaborate linear algebra, deep federated learning, zk-proofs with quantum-secure elliptical curves etc…
…then a guy has a momentary lapse of judgement.
Obligatory xkcd:
1
22
847
This feels very accurate. Sample efficiency isn’t everything but if we make enough progress in it, it may solve continual learning and other challenges.
1
342
People speak past each other when debating wall vs foom. This perspective resolves that big tension. AI progress can be exponentially increasing AND have diminishing ROI.
I think this pattern of seemingly unrelated “unhobblings” is quite interesting to think about, maybe there is a higher order pattern that we’ll discover retroactively
This is like the 4th or 5th "one-time speedup" in the history of deep learning. We should expect further speedups, perhaps bc the underlying process is actually a noisy superlinear function
lesswrong.com/posts/RDG2dbg6…
1
8
614
Look at Luna (same family, smaller distill of Astra/Sol), something points to another cause than benchmaxxing perhaps?
Nice DeepSWE score! Does it survive a lowercase filename?
Introducing MaxxBench: 89 DeepSWE tasks with small changes to the instructions or starting code. Same engineering problem, slightly different requirement.
Models clustered around 74% on DeepSWE end up 23 pp apart.
Good at coding, or good at DeepSWE? 🧵
Hi @thsottiaux, I one-shotted your game via JS injection pls don't get mad at me I needed the tokens 🙏
11
148
30,901