Read documents. Not too many. Mostly PDFs. Our parents read and wrote documents. Having the world's knowledge an internet connection away doesn't change this.
Santa Rosa
Joined September 2021
- Tweets1.7K
- Following126
- Followers466
- Likes8K
Information from Documents retweeted
Local AI is quickly becoming a requirement. If you need to recover a password protected file that you created and forgot the password for.
Opus will not help you.
GPT will not help you.
Your AI, running on your hardware, will.
hashcat v8.0.0 released!
The largest update in the history of hashcat, with PCFG attacks, hash recipes that crack hashes without a hash-mode, and bcrypt on FPGA boards.
Release notes: github.com/hashcat/hashcat/b…
Download: hashcat.net/hashcat/
Information from Documents retweeted
Everyone who owns hardware with abominable software (like Sonos players) owes a debt of gratitude to Venkatraman Srinivasan and the rest of the eMudhra gang in Bengaluru, Karnataka. Without their gross negligence, it'd be a lot harder to counterfeit certificates and fix the sw🙏
Information from Documents retweeted
The reason that this tweet is so annoying to me is that OpenAI and Anthropic are playing a double game. On one hand, they tell us there is massive distillation by Chinese labs, that this proves they are far ahead, and that those labs should be restricted because it is simply not fair.
On the other hand, they refuse to show any evidence. They could. They could say: here are the benchmarks, here is how we detect it, here is our estimate of how much distillation happened. Then we could have a discussion based on facts.
Instead, they just say, " Trust us; it is very severe. And, based on that, trust us to block and restrict the competitors." That is not a fair argument.
And all of this is before we even get to the question of whether distillation is a problem at all. Companies have always copied and taken inspiration from each other. OpenAI and Anthropic built their models by distilling the entire internet. But that is a separate discussion.
Information from Documents retweeted
Folks worry about how to test if models will be trustworthy in deployment, given that models can often tell if they are being trained or evaluated, so they might behave differently to fool evaluators.
This thread shows an even subtler possibility: models can learn to notice even the presence of a *read only* probe which monitors their intermediate activations.
How can a read only probe produce any effect on downstream computation? This is what demands explanation, since the model saying “I’m being monitored” is a huge effect on the final output of computation.
One mechanism is that the probe is not entirely read only. Its presence forces the compiler to write one of the model's intermediate activations out to GPU memory where the monitor can see it. This prevents the compiler from keeping that activation inside a fused kernel. Splitting the surrounding operations, instead of fusing them, creates very small differences in the floating-point computations. Those differences are enough that the model can notice them, if it is trained to do so.
Remarkable.
Let's test some variants on Qwen3-8B to illustrate why. Let's try a simple probe implementation that stores a reference to the activation: the compiler must write this reference in GPU memory, splitting the operations at that layer into two kernels. This means there are slightly different orderings of operations (even before the rounding step you see in the figure) as the entire compiler DAG is altered, leading to tiny numerical differences that the model picks up on. Let's also try just splitting into two kernels, with no probe. Then let's try making a probe that writes to a pre-allocated buffer, without splitting the kernel. You can see that all three interventions change activations. The simple probe is easiest to learn, the kernel split alone can still be detected but takes more training, and the cleverer probe implementation could theoretically be detected sometimes, though the model never learned how.
Information from Documents retweeted
Congratulations Karl! The wonderful neurotechnologies that you’ve been developing for years now are of revolutionary importance for fundamental neuroscience as well for clinical applications. Our understanding of the brain wouldn’t be the same without you!
Information from Documents retweeted
A 59-year-old conjecture in fusion physics just fell.
Harold Grad (1967): smooth 3D plasma equilibria can't exist unless pressure is constant or there's symmetry.
This week: 3 families of counterexamples. 2 independent papers, posted a day apart. Two found with GPT-6 Astra.
Huge for stellarators.
Information from Documents retweeted
home.uchicago.edu/dtson/pape…
Really awesome result from Son+Claude. I could not find the error during my PhD. Turns out they lost a factor of 1/2 🫠
Always good advice, that I didn't always take. With AI's current rapid advances, I am exposed to new information faster than usual. So do I focus on learning a about it in order to become more effective later? Or apply the visible new information for short-term gain?
Information from Documents retweeted
Best community note in a while.
We're used to thinking of open-source models as an unadulterated good. But in the case of AI, they can actually pose additional dangers, as @ReidHoffman and I got into at #CGI2026. I appreciated this nuanced discussion.
Information from Documents retweeted
The worst of it is that south sweden gets to pay a higher electricity rate *because* of the renewables of the danes and the germans. I'd be furious.
Information from Documents retweeted
learned @yaroslavvb's protocol: if it's generated by your agent, i'll have my agent read it, if it's human generated i'll read it
This explanation of LLM reasoning traces reminds me of LLM users' lack of learning I see. I've learned a lot by having my incorrect answers rejected. Some of this learning was mechanical process that machines do better (e.g. solving PDEs). But some of it was useful principles.
there’s a lot of signals you can use during RL, eg you could:
penalize correct-but-inefficient solutions
monitor cots and reject rollouts with correct outcome but bad reasoning
have rubrics during train that target partial success
hence my quote in “just” RL 😅 but you can get this behavior w/o having to directly supervise on tokens.
also i think it’s far from perfect, otherwise we would have solved interpretability. I find reasoning traces to extremely difficult to check for accuracy, I can’t review them for correctness w/o an agent
Information from Documents retweeted
More writing output never did mean more knowledge gained. We knew that. And yet we continued to act like it did. This isn't AIs fault. It's science's.
We have to stop using "paper count" as a useful scientific metric. More writing output no longer means more knowledge gained
washingtonpost.com/education…
Information from Documents retweeted
Code reviews today are mostly about gut checking if the complexity is worth the benefits.
Information from Documents retweeted
most important release this week: caring about the data (obviously a lot missing from official arxiv release) still hard but undefeated.
pleased to share the entire arXiv site as a dataset on HuggingFace
huggingface.co/datasets/sece…
3,148,796 papers, every version, in LaTeX, PDFs, PostScript, HTML, 16 TB in total
This is the nicest introduction to Bell's experiment I have seen. facebook.com/reel/1008138521…
Information from Documents retweeted
here's the whole 1996 conversation with @mckewmax, @rodneyabrooks, doug lenat, plus hugo degaris at the start. the part on RSI (not original with me; i.j. good had the idea in 1965) is around the 28 minute mark.
youtube.com/watch?v=WFeeVJR5…
Information from Documents retweeted
Navier-Stokes Millennium saga continues. Here are my slides and a summary from a recent talk. Disclaimer: this is a physicists view of the problem & solution! 🧵
Information from Documents retweeted
ok i think i figured it out:
- "independent third-party evaluators" --> spies that will share info on biggest competitors
- "pacing" --> stop spending and start monetizing
Information from Documents retweeted
editting students' ICLR/workshop submissions this weekend. a lot my job is now to teaching them to post-math after they give me a draft that they + gpt/claude have pre-mathed and written.
the drafts usually start out basically unreadable. but younger students can't articulate why they are unreadable. they are unreadable because they're written in an unnatural Claudish way that throws results at you without explaining what they're supposed to mean. they're not organized in a way for another human to progressively digest it. i.e. lack of post-math alignment
i propose separating “math” into pre-mathing and post-mathing:
pre-mathing - getting an answer/proof asap.
post-mathing - aligning that answer to human preference for deeper insight and understanding.
pre-mathing is what oai did. the lack of post-mathing is why the 25 fields medalists are protesting.
it used to be all just “mathing” because you did both at once, but with ai you can now do them separately and people haven’t adapted to thinking about them like that yet.