in retrospect, it was a mistake to teach algebra and geometry but not "statistics for everyone" in high school we are drowning in lies, partly because we never learned to swim

"amateurs" I scoff as I return to vibe-tabbing a web app with no tests

academia creates an infinite supply of methodological beef by withholding the value of "strong"

Is scale all you need? Or is there still a role for incorporating domain knowledge and inductive bias? While I was in Heidelberg, I took some time to write a short essay on this question called "The Bittersweet Lesson". theoryandpractice.org/2025/09/The%... #HLF25

(from Naming the Mind - K Danziger: www.goodreads.com/book/show/81... )

Wait, isn't review always double-blind? Why are author details a factor?

ICLR deadline vs solo parenting 2 kids and the whole family having a flu 🤬

The real existential threat isn't AI — it's billionaires. My latest, for @theatlantic.com: www.theatlantic.com/books/archiv...

computer, connect the dots for once

do hashtags work on bsky? I'm not trying to farm engagement, just to find people who might know answers 🥶

say I take N multinomial samples x1..xN from two histograms p1 and p2, and have a function X -> Bool that evaluates some sample fitness.There could be many "right" samples. Is it sensible to compare p1 with p2 by the MRR of the results of f on X? #informationretrieval #ranking #search #generative

more likely: mode-collapsing on old FIXME comments, filling PRs with Freud quotes

www.nature.com/articles/s41... kudos to the DeepSeek team, more foundation model teams should put the effort of submitting to a journal. The review team in this case was a mix of industry and academic experts. Review log is public too, so ✍️

🗣️ Can generative AI speak more than English? Cees Snoek presents NeoBabel — a multilingual, open-source text-to-image model in 🇬🇧 🇨🇳 🇳🇱 🇫🇷 🇮🇳 🇮🇷, built for inclusivity and fairness. ▶️ Watch now: youtu.be/qUadAcykUiI?...

the most useful thing LLMs can is force us to do is to reconsider the way we use linguistic dexterity as a proxy for intelligence

"Evolution produces sufficient rather than optimal adaptations." - so true and so annoying

I got nerd sniped by this one, some puzzles are definitely doable with logict in #haskell, without turning to a SMT/SAT solver , eg lobste.rs/s/8ubfdd/man...

Not only, also when the risk scenarios will not be entirely made up as they are now.

The most expensive part of training is the data, not the compute Nikhil Kandpal & Colin Raffel calculate a really low bar for how much it would cost to produce LLM training data with 3.8$\h Well, several scales more than the compute. Luckily (?), companies don't pay for the data 🤖📈🧠

As an Action Editor of TMLR, I am looking for researchers who want to serve as reviewers for TMLR (NOTE: This is not for a specific paper in my batch) If you read the guideline and want to serve as a reviewer for TMLR, send me your Google Scholar and OpenReview profiles🙋 jmlr.org/tmlr/reviewe...

one I saw recently, while landing in the fog

sqlite fails a transaction? skill issue (yours)

train stan account logging on

@nsaphra.bsky.social you confirm it's enough to ask GPT for an unbiased logic check? losing my gd mind at this GuardrailsAI example

I think mainly because they don't experience consequence, at all. They don't experience the irreversibility of time like we do.

4yo: (tries literally anything for the first time) SEE? I'm GOOD at this!

SLP 3rd edition when

tapping the sign

ICLR 1 month challenge let's go

youtu.be/ncHctrAluLM jefferson airplane in a rubadub style

my brain is a SEGA AI computer, it has a tape cassette drive and runs Prolog

7.05, family is done with breakfast 4yo is jumping on the couch wearing a tutu and skates the baby is covered in blueberry jam all shoes are covered in blueberry jam the baby is crying it is now 7.05:10

Hasukeru ga hanasemasuka?

the real question to ask is: how many Ġ's are there in blueberry?

"oh no I'm averaging attention head activations!" but then you recall fMRI has wayyyy lower resolution

Is the Language of Thought == Language? A Thread 🧵 New Preprint (link: tinyurl.com/LangLOT) with @alexanderfung.bsky.social, Paris Jaggers, Jason Chen, Josh Rule, Yael Benn, @joshtenenbaum.bsky.social, ‪@spiantado.bsky.social‬, Rosemary Varley, @evfedorenko.bsky.social 1/8

Me at work: you shouldn’t condition on posttreatment variables as it biases inferences, see Montgomery et al. (2018). Me at home: look at how sweet the kids are with each other when they’re not fighting!

I tried my hand at a sparse linalg library once*, learned a lot on numerics and data structures, but for performance and stability it's best to use the pros * github.com/ocramz/spars...

I'm quoted in this piece. In short: do not use ChatGPT or any generative AI as a “therapist.” It’s not sentient. It doesn’t care. It may even lead people toward harm. And nothing shared is protected by HIPAA or real privacy standards; it's fodder for the machine. www.dazeddigital.com/life-culture...

BREAKING: The EU Commission has released a mandatory template for AI developers to disclose training data. Unlike the Code of Practice, this is not optional. It could have global fallout, as rights holders abroad might use it to sue over copyright. digital-strategy.ec.europa.eu/en/library/e...

We should have listened when the modems screamed at us.

the title riffs on an ACL'14 keynote on semantic parsing - good overview of the field up to then: yoavartzi.com/sp14/slides/...

relevant: aclanthology.org/P18-1198/

one of many 🤷‍♂️ it's also past time to understand the world as a multipolar place

you know the "flow" mental state? has anybody studied how parenting babies is the _opposite_ of that? the constant stream of long-tail novelty, risk of harm etc

any proposals for interpretability workshops at #EurIPS ?

In my view, these result really caution against the use of LLMs for peer review. Not only are LLMs easily swayed by covert injections, they are also influenced by formulation of the "peer review" prompts. More in the write-up here: osf.io/preprints/ps...

not sure this is the most efficient implementation, but I'm quite happy it's so compact. With a tree traversal and left scan we get an annotated AST with a top-down static program analysis. #haskell #plt #compilers #staticanalysis

distinctly recall the tens of additional km you have to walk in Tokyo if you have to push a stroller. Turns out (as it's often the case in JP) the answer is in a densely printed table, in Japanese only ofc.