@pbloem That's a good question! I wrote up a longer answer to your question at https://benjaminhan.net/posts/20260517-self-correction-after-reasoning-models/?utm_source=mastodon&utm_medium=social The short version: yes, the recent reasoning-model training *internalizes* what used to be an […]

We're finally getting a glimpse of how much AI use is being subsidized by VC money, and in some cases, it's up to a factor of 20. If Moore's law applies (doubling time 18 months), that means that today's heaviest AI use will actually be economical for the current pricing in a little under 6.5 […]

I keep going back and forth on the arXiv thing. Basically, the way I do things now, I can easily see myself falling foul of the rule. I don't use LLMs to write or to insert references, but I often co-author with students and I prefer to trust them. I have read every line of every paper that […]

The most realistic movie depiction of AI/human relations is Wall-E. People don't resist. They are falling over themselves to defer responsibility.

Cannot-Self-Correct tests the strong claim that LLMs can revise their own reasoning answers without any external signal about correctness. Across three benchmarks (GSM8K, CommonSenseQA, HotPotQA), the answer is no: the model's confidence carries over from the initial answer into the revision […]

Probably unpopular opinion: if you want to be a serious scientist in AI, you cannot have shares (or otherwise a stake) in an AI company. It's a conflict of interest.

Something I'm quite tired of: A huge, obviously complicated community effort is made to find some reasonable comprise about a polarizing issue. A carefully written statement is posted online. Someone goes on Mastodon, takes a quote from the piece out of context and dismisses the effort and […]

enough. i’m tired of github actions. it’s time for github consequences

Interesting data from Pew on use of Ai by race/ethnic categories and by wealth. www.pewresearch.org/internet/202... Broadly speaking the more social privilege you have the less likely you are to use Ai for help.

I think this will be my last MacBook.

Found on Bluesky (by @hannahshelley.site)

I was skeptical about the climate costs of Google pushing a 4Gb AI model to every copy of Chrome, but this article is more careful and reasonable than I expected: https://www.thatprivacyguy.com/blog/chrome-silent-nano-install/ Specifically: - It's not the normal "AI inference costs energy" […]

We're really speedrunning this episode of the Scary Door, aren't we? https://www.youtube.com/watch?v=RSEo64cW3mc

I've had trouble quitting X, since it's the only way I can keep on top of breakthroughs in ML. But I just realized I can quit everything but that. I just unfollowed every account I could find that isn't machine learning. Hopefully this turns it into a single in-and-out daily check, rather than […]

This looks interesting...

Scale buys calibration in exactly the format you tested in. Lettered MC works at 52B; swap one option for "none of the above" and it collapses; True/False rephrasing restores it. RLHF policies need a temperature fix. "Do I know this?" heads miscalibrate out of distribution. Three years on […]

My latest books cartoon for the @theguardian.com

I feel a certain amount of dread knowing that there is a Richard Dawkins novel in progress.

RE: https://mstdn.social/@hkrn/116499639295847975 This is a recurring theme now. Long story short: AI can find vulnerabilities quickly, so open source is now seen as a risk. I think this is misguided Closed source is not a sound safety measure. One code leak and your vulnerability-finding AI […]

So, copy.fail was found with one hour of AI assistance, and would (according to this article) have earned $500K on the open market not too long ago. https://www.bugcrowd.com/blog/what-we-know-about-copy-fail-cve-2026-31431/ I'm no security researcher, but this kind of contradicts all those […]

Look at what Mathematicians have been demanding your Respect for ...

A friend of mine is developing this slick little space trader sim MMO that you can play in your browser: https://everestdrive.io/ I think you can even invite friends to be on your ship and run some of the stations. This is an early access game; there's a Discord for bug reports. #videogames

Clowns to the left of me, jokers to the right...

We're really speedrunning this episode of the Scary Door, aren't we? https://www.youtube.com/watch?v=RSEo64cW3mc

If respecting a bike lane is too difficult—and I cannot emphasize this enough—then the system is not robust enough for the real world. Not just for the bits with bike lanes. For any of it. Not a highway, not a parking garage, not a bus lane. These are not coding agents or image generators […]

I am reading, with morbid fascination, the Palantir "manifesto". This bit sticks out to me. It's very illustrative of how shallowly these people think about AI. They always build on some assumption that the effect on society can be predicted and harnessed […] [Original post on sigmoid.social]

I'm pretty centrist on AI, but one thing that really triggers me is the trend of using the Ghibli style for AI-generated images. It's a real marker of a certain kind of view of the world that just doesn't understand or respect where art comes from. It's the worst kind of science-smart […]

I'm sure everybody already knows this, but it's eye opening for me. UBlock allows you to click on any element on a website to block it. It doesn't have to be an ad. It could also be an annoying AI chat assistant, an auto-playing video, the hot network […] [Original post on sigmoid.social]

How long until somebody somewhere decides that nobody remembers phone numbers anymore, and so we can just make all phone numbers a 128-digit hex string. That way we can always assign anything a phone number. You can call your fridge!

Mathematicians declare they have solved the crisis. The Strait of Hormuz is, in fact, clopen.

This seems like a failure of memory to me. Cellphones were pushed like nobody's business. The polarization around AI reminds me a lot of the attitudes to (pre-smartphone) cell phones. Especially the tedious repetitive arguments. The telecoms industry was […] [Original post on sigmoid.social]

RE: https://mstdn.social/@hkrn/116409214190278520 Makes sense. Who's going to need footwear once the singularity hits?

RE: https://sigmoid.social/@pbloem/116250532724597189 4 days left to apply.

And after JD went through all that trouble.

I wouldn't want to be the carpet in Putin's office tonight...

You know you're in the weeds when the background color of the PDF is off-white.

RE: https://mastodon.social/@mikeolson/116375430505949437 It's remarkable how important the right tools are for this. I can happily write 6k words a day in a good .md editor or in Google Docs. Make me use Word, though, and all that goes out the window. It's like putting a chef in a fast food […]

RE: https://sigmoid.social/@pbloem/116330841525931373 Update: they've decided to host their code on github.

If anybody needs help understanding the Dutch grading system: these students built a prototype belly to help test MRI scans on patients that aren't as slim as the dummy. It worked so well the hospital began using right away. Their grade was a 9 out of 10. The 10 is not there to be attained […]

There a pernicious trend in leftwing discourse where a person correctly notes that somebody is being given privileged treatment, but then immediately implies that the right course of action is to take that privilege away. More often, the right thing to do would be to afford everybody the […]

I feel like we're not talking enough about how weird it is that Jensen Huang always wears a leather jacket indoors.

A rare misfire from Tufekci, I'd say. She's usually on the money. While the algorithm is always the same, AI is operating in substantially different modes when it's hallucinating to when it's retrieving factual information. Hallucination is not a solved […] [Original post on sigmoid.social]

As my wife ate her porridge this morning she asked "what exactly are oats?" and it turned out to be an interesting question because of Vavilonian mimicry. https://en.wikipedia.org/wiki/Vavilovian_mimicry

I like fedi a lot, but even without an algorithm, there is definitely still an echo chamber problem and there is definitely still a dogpiling problem. I think the fediverse has the self-awareness and the autonomy to address these, but we need to recognize it first.

RE: https://mastodon.social/@randahl/116352139642025684 I am once again reminded of Harrison Bergeron's president. https://www.youtube.com/watch?v=m5cQFB3aZcc

Well, Euro Office is off to a flying start...

We're going to have an outside context problem in our lifetime, and it's name is AI.

Thank god my newspaper is sending me email verification codes whenever I log in. Mossad could well have set up an operation to conduct a man-in-the-middle attack to use my account to read the comics without a paywall.

For no reason at all, please give me your favourite cow-related figures of speech! (Stuff like "No use crying over spilled milk" or "until the cows come home", puns extremely welcome)

I keep wondering what the rock bottom will look like that will finally wake us up, and make us realize that power should be kept away from people like Trump and Musk. I think a Kessler syndrome is probably the least bad option that is still bad enough to do it. We lose our access to space for […]