@joshuagrochow
Research: TheoryCompSci, pure math, complex systems Other: climate; covid; equity, inclusion, & accessibility Assoc. Prof. @ CU Boulder Comp. Sci. & Math Views my own @joshuagrochow@mathstodon.xyz (& prev twitter) https://home.cs.colorado.edu/~jgrochow
New paper from team @aial.ie! aial.ie/research/gpa...
EU's AI Act Article 53(1)(d) is an obligation for GPAI model providers to publicly provide a 'summary' on their model’s training data. The team assessed published summaries along 6 dimensions & found that all big providers failed on all 6.
1/
Flow chart showing OpenAI, Google Meta, Microsoft, Amazon, Perplexity, ProRate.ai, IBM, Mistral, Anthropic, Nvidia, in that order with the biggest deals, with OpenAI having 21 deals, Google 17, Meta 8, Microsoft 4, Amazon 3.
Check out this interactive website by Nana Mgbechikwere, showing BigTech and "AI" companies' media capture.
No wonder the public is getting misinformed.
They say it is inspired by @surveillancewatch.bsky.social, which ❤️
nananwachukwu.github.io
email to me with a title: 2027 MSc in Artificial Intelligence Application – Research Interest in Trustworthy Generative AI & Multi-Agent Safety email body: I have been deeply inspired by your pioneering work on AI accountability, algorithmic harm governance, and ethical alignment of generative multi-modal systems. As Geoffrey Hinton has repeatedly warned the global community about the existential and structural risks of unregulated AI systems, I have long been searching for actionable, ethical frameworks to translate these high-level warnings into practical, safe AI design — and your research has been the definitive guide for me. In particular, your 2023 paper in Nature Machine Intelligence on the structural risks of large-scale generative models, as well as your AI Accountability Framework developed at the Mozilla Foundation, have fundamentally shaped my core belief: capable AI systems must be built on the premise of safety, transparency, and consistent alignment with human values, rather than pursuing functionality alone.
never published in Nature Machine Intelligence & neither do i have work on "AI Accountability Framework"
i know this is now normal but i want you all to stop & reflect on how much the future is fucked & the only way to mitigate this disaster is to ban/limit this dammed technology
i don't think i can say this any more directly:
if you want to get better at something, you have to do it a lot
there are other requirements but that one is non negotiable
This is a basic and crucial point for any discussion of LLMs’ use in teaching research or writing. Whether it can ape us, fool us, or get facts right or wrong is, in the end, irrelevant. The LLM is not the thinker we are trying to encourage; the LLM is not the writer that we are trying to improve.
i think enthusiastic LLM use is mostly a stack of cognitive biases, unacknowledged plagiarism, and unmet needs in a trenchcoat
but also my main objections aren't about them being bad at tasks so i don't care if you think they've gotten better at it
This (about the 'new' Grammarly AI toolset) is absolutely terrifying, not because the tool is good, but because it is an easy, all-in-one, way to avoid thinking any thoughts or doing any work at all *and 3000 universities say its OK to use*.
www.carleton.edu/ai/blog/gram...
If you ever want to read a paper for free and can't find it otherwise, email the lead author and politely ask for a copy. You will not be bothering the person. You will in fact make their whole entire day. I have had scientists get so excited I asked they sent me everything they ever published.
Interesting, something about the two sounds I don't think I'd ever confuse zee and cee (but I can imagine coming from other languages how one might)
¹https://www.theguardian.com/technology/2026/feb/20/amazon-cloud-outages-ai-tools-amazon-web-services-aws
"Amazon never misses a chance to point to ‘AI’ when it is useful to them .... But when a slop generator is involved in an outage, suddenly that’s just ‘coincidence’."
Pretty sure it's 1...
#MathSky
(Pretty sure SE inherited that from math, which has exactly the same problem.)
/ @imbernomics.bsky.social has an informative thread related to this (but I can't link it b/c it has quotes disabled)
(Also IIRC total dollar amount didn't decrease as much as number of grants - still very problematic)
I agree it's important to highlight the people (esp grad students) that grants support, but I'm curious why you call this indirect.
A line graph showing NSF grant awards made through 2/27/26 for fiscal year 2026 compared with grant awards for fiscal years 2021-2025.
NSF Update (Awards through 2/27/26)
Directorates to follow
1/10
When systems have to work, you need people with an accurate, nuanced mental model. Experts with a mediocre product are more capable than poorly-oriented engineers with a great product, and vibe-coding does not produce great products. Knowledge is an essential output of software engineering.
Last week, for the first time in 20 years, a man died from a treatable snake bite on my home island of Martinique. He had asked chatGPT what to do instead of going the hospital. He missed the 6 hours treatable window.
I continue to be fascinated by the phenomenon whereby an expert engages with any of the LLMs on their field of expertise and is instantly horrified by the wrong answers, and then goes on to use it for things they are not experts in as though it won’t be just as bad for those.
Ooh say more?
and also a little what a for loop is
This is so weird - have you been spying on me? I literally almost emailed you a question about this *exact* thing yesterday.
Isn't there some topologist quip like "all cohomology is H^1 of something appropriately (de)suspended"?
Even with constant coefficients, I see what you're getting at. For example, simplicial homology is about "things in the space" (synthetic) whereas simplicial cohomology is about "functions on the space" (analytic).
Changing the identity axiom of groups to "If there exists an x (any x at all), then there exists an x such that for all y, xy=y" I see.
Sure, just adjoin a new initial object to the category of groups why don't you. Have we no decorum??
I think this thread does a much better job getting at the heart of the issue I was trying to get at, about "not understanding why the LLM made the output it did": bsky.app/profile/mjcr...
Specifically, the idea of source credibility just doesn't apply straightforwardly to LLM interactions. When I get info from another person I can ask: how does the content relate to their expertise? Their lived experience? Their interests and biases?
8/
Without considering the sources of claims, efforts at epistemic vigilance become play-acting. With LLMs, we don't know the sources.
Excellent thread from @mjcrockett.bsky.social
text says "There’s no way to know which COVID-19 infection will cause long-term illness. Each new COVID-19 infection increases your risk of organ damage, Long COVID, and disability, even if you have no other health conditions that you know of, and even if your previous COVID-19 infections were mild or had no symptoms. Protect your lifelong health by protecting yourself from COVID-19, no matter your age or health history. Wear masks. get vaccinated. keep indoor air clean.” Graphics on the side illustrate people getting sick repeatedly until eventually they end up with long-term illness.
Every time you get COVID-19, your risk increases for Long COVID, organ damage, disability, & more, & there’s no way to know which infection will be the one to cause it! Protect your lifelong health by protecting yourself from COVID-19. Learn more about Long COVID: www.cdc.gov/long-covid/a...
“Almost 5 years later, #LongCovid has had one of the fastest rises in diagnoses and become one of the most economically disruptive chronic conditions in modern medicine… Another analysis… put the annual economic toll of LC at $1 trillion” — @jasongale.bsky.social www.bloomberg.com/news/feature...
