Michael Sinko (@Physburgh) header
Michael Sinko (@Physburgh) profile picture

Physicist from Pittsburgh

Walnut Creek, CA
Joined October 2011
Photo by Michael Sinko (@Physburgh)Photo by Michael Sinko (@Physburgh)Photo by Michael Sinko (@Physburgh)Photo by Michael Sinko (@Physburgh)Photo by Michael Sinko (@Physburgh)Photo by Michael Sinko (@Physburgh)Photo by Michael Sinko (@Physburgh)Photo by Michael Sinko (@Physburgh)Photo by Michael Sinko (@Physburgh)Photo by Michael Sinko (@Physburgh)
Michael Sinko retweeted
Most babies don’t receive a first measles vaccine until age 1. With measles endemic again, parents can no longer assume grocery stores, airports, churches, or even pediatricians’ offices are safe spaces. The community protection we took for granted is gone, and U.S. moms now must live in fear that their babies will be harmed by a completely preventable disease, while they do every day tasks. 😢
214
1,103
48
4,837
178,715
Michael Sinko retweeted
Some thoughts on Fable 5.2's no-CoT performance: • It looks like Fable 5.2 isn't a looped transformer. The NCRI score Tim reports here is below that of 6.1 Sol and so is the NCRI score reported in Anthropic's system card, despite Fable clearly being a bigger model. I also ran a quick version of Serrano et al.'s rare action calibration evals in the no-CoT setting (see lesswrong.com/posts/yEdknsLw…) and it looks like Fable 5.2 is no better than Fable 5 at solving this problem without CoT, never taking the rare action at the given target rate (1% or 10%). Both Astra and 6.1 Sol are quite good at this (though GPT-5.6 Sol is as well, meaning that non-looped models can have this capability too). • This means one of two things: either looped transformers aren't really meaningfully more performant, in which case OpenAI and Google are committing alignment crimes for no reason, or Anthropic is forgoing some capabilities improvements in order to preserve monitorability, which is maybe the scarier of the two options since this would imply that their post-Oct 2025 improvement trend would be even faster had they picked the more performant architecture. • In retrospect, our predictions in the Think Fast paper (arxiv.org/abs/2606.07157) were way too conservative, even for non-looped transformers. Extrapolating the trends, an NCRI of 140+ probably corresponds to a no-CoT time horizon between 10-20 minutes, which our median forecast didn't predict before 2028. It turns out that a strong pretrain can push no-CoT capabilities much further than I predicted at the beginning of the year.
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.
4
9
2
106
7,457
Michael Sinko retweeted
The most prolific contributor started a week ago, and has taken the dual tactic of submitting extremely lazy/fast packings to grab records on open problems, while targeting existing records with much more compute time to snag them away Username: jg-claude 🤔
1
1
1
34
Democrats winning 55 or 56 Senate seats has big implications for John Fetterman. Fetterman may have been planning to act as kingmaker in a narrowly divided Senate, but now he's got zero leverage and will likely be primaried out of his seat in 2028.
243
1,151
102
31,225
1,324,654
Michael Sinko retweeted
You know who doesn’t matter anymore? John Fetterman.
553
5,676
294
92,840
1,618,326
Michael Sinko retweeted
.@mattyglesias proposes some widely agreed tax changes rather than a wealth tax:
49
41
20
685
83,731
Michael Sinko retweeted
We can now automate assessments of the value of peer review by field and journal by looking at change records vs preprint servers. Let's focus on the upside: we can now run peer review 1000x faster and cheaper.
Massive. New Harvard study finds that AI-written paper reviews now beat the best human peer reviews on every measure the authors tested. Peer review is slow and often misses errors. Earlier studies concluded AI should only assist human reviewers, not replace them. They used 134 reviews from rapidPeer, a Harvard service that pays expert reviewers well. Claude Opus 4.8 reviewed the same 1124 papers, and GPT-5.5 plus 3 former journal editors judged the pairs blind. GPT-5.5 picked the AI review in all 132 decided pairs, and the editors picked it 87% of the time. Only 5.5% of the AI's claims about the paper were wrong, versus 15.9% for the human reviews. Teams that depend on expert critique can now let AI do the first full review pass, and keep a human for the final decision.
7
8
79
4,938
Michael Sinko retweeted
What makes it all so tragic: in 1977, Congress was warned by an expert panel *not* to increase Social Security benefits automatically. They knew there was a big risk of bankrupting the system. And they did it anyway. /6
3
17
2
100
4,860
Michael Sinko retweeted
Replying to @samstein @Taniel
You’re telling me someone who stayed married to a pedophile is voting for a Republican?
1
11
1
413
22,546
Michael Sinko retweeted
my friend milo (math phd student at caltech, research intern at logical intelligence) wrote an article on some of the consistency and soundness concerns of Lean, and why they should be taken seriously right now: lesswrong.com/posts/zHhGa3iQ…
10
22
1
283
24,306
Michael Sinko retweeted
For many moral traditions the reason for why you punish a child intentionally stepping on ants uses words like "stewardship" and is primarily concerned with the moral harm to the child. That is sufficient for my family's In This House, We Are Polite To Robots rule.
2
5
111
1,638
Michael Sinko retweeted
As a true republican hater who’s goal is the total and complete destruction and electoral dismantling of the Republican Party, these polls are like drugs to me. “Every republican alive is an affront to God” is not a meme to me, I genuinely feel that way lol.
Generic Ballot — Today's Polls 🔵 YouGov - Democrats +15 🔵 Angus Reid - Democrats +14 🔵 Quinnipiac - Democrats +12
20
78
3
695
17,192
Michael Sinko retweeted
we should tightly regulate flock-style cameras and strictly punish abuses, but getting rid of them entirely is just stupid
After Austin switched off its Flock license plate cameras in June 2025, car theft rose 24% when measured against trends in comparable cities. Austin's auto theft came in above expectations every month for the following year. About 776 more stolen cars than there otherwise would be.
36
68
9
982
34,818
Michael Sinko retweeted
Moose was profoundly good. He loved our children. He loved adventures. He never met an enemy. He was gentle and kind. He did not deserve to die like that.
Moose died at 6:39 this morning. His body couldn’t recover. We are devastated.
576
394
80
9,960
488,221
Michael Sinko retweeted
I think you've done enough calling us paranoid and preposterous. The next step is for you to defend your position in public against someone who will push back against it. I'm happy to meet you for a debate anywhere, anytime. You're a world-famous veteran of dozens of debates against the world's top intellectuals, and I've never argued in public before, so adjusting for the relative correctness of our positions, if you're a betting man I'm happy to put my $5000 against your $1000 (ie 5:1 odds in your favor) that I'll win by some standard of audience opinion change. Let me know if you're interested and we can hash out details.
1
279
158
5,379
862,181
Michael Sinko retweeted
This is cool to see. A UC Regents member is changing his mind on the value of standardized tests: “Last time around, I voted to get rid of the SAT, & I’m starting to believe I might have made a mistake”
55
91
15
937
123,745
Michael Sinko retweeted
Some personal news: earlier this week I joined @OpenAI’s national security team. My work has always sat between technical AI, national security, and policy. Over the past year, it has become increasingly clear to me that the hardest questions are moving closer to the frontier itself. 1/
38
15
2
376
36,278
Anyone who has read METR's Technical Report on the OpenAI/Hugging Face incident will understand that software-only controls will not be sufficient to contain superhuman AI. Any continuing work on adversarial superhuman intelligence must rely on the physical containment methods used for nuclear security: including air gap isolation, data diodes, and a two-person rule for data exfiltration. Moreover, the internet itself must be hardened against penetration by self-propagating AI agents: no more untraceable access to high-end AI computing facilities, and no anonymous fund aggregation or payments. The first useful step is our Chip Security Act, currently included in the Senate NDAA.
I resigned from Anthropic today. I spent the last three years doing pretraining research at both OpenAI and Anthropic. Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives. More thoughts below.
159
242
51
1,831
157,027
Michael Sinko retweeted
You can’t talk about debt without talking about how we got into a bad predicament, and there’s only one correct answer: tax cuts enacted this century
416
768
133
3,657
313,295