In a recent video, Kurzgesagt made a great analogy about genAI. It is like a virus: something that is not alive, but, due to evolutionary pressure (i.e., training), has been fine-tuned to act in very complex ways that feel intelligent.
Excellent article that articulates better than I could the cost of playing into algorithms ( or the fear of algorithms ).
I'm tasked with preparing a kind of hand-out document about "responsible use of AI" for patients.. Of course I have ethical issues with this because my own honest advise would be "don't", but that's probably not very useful in this case where the assumption is that people are already using these things. So I'm approaching this with a framework of harm reduction. I'd be very interested in ideas for this, if anyone has maybe done a similar thing.
Meredith Fowlie's EI blog on disturbing econ of race among US AI firms and weak or absent enviro regs for off-grid data centers, even ignoring safety risks that AI CEOs are calling out. China threat may justify all this, but POTUS hasn't shown any evidence. energyathaas.wordpress.com/2026/10/05/b...
Sprint 31/90 · From perceptron to multilayer network No line \(w \cdot x + b = 0\) separates XOR. Rosenblatt's rule: AND has no miss from epoch 6; XOR ends every epoch on the same weights from epoch 2. Two ReLU units with hand-set weights (Goodfellow et al. §6.1) map (0,1) and (1,0) to one point; then \(h_1 - 2h_2\) gives 0, 1, 1, 0. Gradient descent, 100 seeds: 2 hidden units solve 77, 4 solve 100. Archive: zenodo.org/communities/mt24-le… #MT24Learning #NeuralNetworks #MachineLearning
Stefano, the Human LLM of the 2000s ‘This is slop! Prove it is not!’ So I went looking for how I used to write, long before LLMs existed. I found a paper from 2002, a blog from 2006, and a lot of ‘it depends’ https://my-notes.dragas.net/2026/10/08/stefano-the-human-llm-of-the-2000s/ #MyNotes […]
Very interesting piece on the impact of AI on Maths. “Is AI the end of Math as we know it?” quantamagazine.org/is-ai-the-e…
In Crikey today, I have a long read on Canberra’s endearment with AI and the dubious promise of data centre riches. While Australia’s creatives are being sold down the river, you don’t need to fire up Claude to spot some cracks in the hype-doom sales pitch ($) www.crikey.com.au/2026/10/08/a...
When LLMs solve a math research problem, the person and/or company who prompted it to do so is stealing the possibility of someone to learn from the *process* of solving that problem, experiencing joy in its discovery & solution, bringing their own viewpoint & understanding to the problem, and exploring new related problems along the way.
While this is tragic, I am not 100% sure that normalising the idea of websites routinely informing on their users is a good path to go down actually. Sorry if this puts me at odds with everyone purely because AI is the hate-du-jour; but let's not hand our enemy the weapon used to shoot us eh?
Introducing "formalize-math-paper" (vltanh.me/projects/formalize-m…), an agent skill that formalizes a math paper in Lean 4, from (preferably) an arXiv TeX to a build with no sorry and no axioms beyond Lean's standard three, with the aim of auditing the paper's proofs. Statements come first and are checked against the TeX; each proof must follow the paper's argument, and the agent ends with a report of errors, gaps, and unused hypotheses. (1/n)
This whole 'AI improves productivity' is a persistent one, even as study after study proves it to be false. Writing code is a tiny part of a software developer's responsibilities. Barfing up GAI slop just means a longer and more difficult code review for human devs. www.youtube.com/watch?v=idTB...
ChatGpt is really helping me with work now. It adds attorneys and contact info to a spreadshet for work. It has a nightly routine to correct any errors. It alerts me if I may have a conflict in time. To really get it to work effectively for me, though, I need the 100 dollar plan to get rid of the constant you ran out of credits for five hours. About to try to reinstall Leasey again but Jaws is still randomly giving me driver errors for Braille, although Braille runs fine. This does not happen in Jaws 2025 but only 2026. Something is goofy and I do not know how to fix it.
I have to use LLMs for my job. I do not believe that there is a legitimate use case for LLMs aimed at consumers or students that doesn’t boil down to fraud, cheating, or theft. The reason people hate LLMs is because corporations are forcing them into everything, in use cases that suck at.
How much of the AI hype has been funded by fossil fuels? climatejustice.social/@ketan/1…
Setting aside the fact that this article says over and over that it can’t verify anybody’s claims, uh, well. It was written by AI.
Anyone else think it’s insane how generative AI companies haven’t been regulated on the grounds of copyright infringement? If you wanted to make Blu-ray player you have to: 1. Pay Sony a fee. 2. Load a cryptographic key to only allow signed discs to play. 3. Enforce device integrity (so no running your own software). 4. If this is compromised Sony can revoke the devices' ability to play discs. 5. Enforce region locking. 6. Enforce all the piracy protections (checking for physical and digital watermarks). 7. Enforce encrypted transfer of the content to the device (so no spying on the cable). And in the US, pirating a movie can get you fined $250,000 and jail time!
Show more of this post
Since 1998, US copyright laws have basically been unstoppable. Then these companies come along, pirate literally everything, get fined basically nothing, and create products that even allow you to use the copyrighted material. If I went and released my own bootleg Frozen movie, Disney would have me drawn and quartered!
limited liability for AI torts, AI companies would not incorporate in that state. But the question does not have to be one of internal affairs; the law of the state where the tort occurred could apply. I proposed something -
These guys started saying “AI” because it was a cool term from scifi for a thing everyone thought was cool. It became a term for a thing that sucks because they started using it to describe a thing that sucks so bad
"I think the true point raised by this post is that it is not possible to claim authorship of any result in a world where AI proves them: it would be impossible to distinguish between a behaviour all of us consider unacceptable (buying a finished paper from an external party, and then maybe rewriting it and learning its content), and a behaviour some consider acceptable (prompting a known question to an LLM, and then understanding the answer and writing it down)" —Giovanni Mongardi proofsandprompts.com/2026/10/0…
Work is developing guidance for responsible use of AI in research. Given 'there is no such thing' isn't clearly an option, I'm thinking about how to feed into the process, and I'm thinking that the best framework is in terms of ethics; making our policy that you have to consider AI use at that stage
Vercel AI Gateway now supports confidence-based fallbacks for decision models. A low-confidence answer can trigger a second model call, and both stages are billed. I'd measure how often that threshold fires before calling it a cheap safety net. vercel.com/changelog/confidenc…
my personal theory is that modern software process is pretty much the worst process for using llm coding tools because most teams iterate their way to a solution and underinvest in testing and specification.
A very good perspective on the current 'theory-free' Generative AI by Cory Doctorow. Worth your time if you are interested in insight in the current AI-developments. pluralistic.net/2026/10/06/non…
There is no evidence of that. Its not a matter of improvement because the problem is due to the stochastic nature of the system. The flaw is *inherent* to LLMs as a concept. This is like saying a butter knife will eventually become a lightsaber if I keep sharpening it.
A contrary take on these dumps of math results: *if* these models are going to be built and released anyway, then it's better that the results come out like this than that a bunch of random users get the proofs out of the model and all try to claim credit and priority. The challenges for the field of mathematics (and far beyond) are fundamentally due to the rapid improvement of the models, not due to OpenAI's approach to releasing the results. github.com/openai/math/blob/ma…
"You work on organimic agency? That field has a bad reputation." I hear this often from researchers I truly value. I wonder why. One of my colleagues believes in magic, minds "ingressing from a Platonic realm." Another one just found agency in LLMs' floating-point rounding errors. We always […]
Ruane: “We have to be integrating support for kids’ education throughout their life cycle. I don’t think bans necessarily accomplish that, so let’s think about it a little bit differently and in a more supportive way that guides them along the process.” edweek.org/technology/at-opena…
Insofar as I understand this paper, the moral seems to be that maths is easy and translation is hard, which I think every mathematician who spends quality time with a translator will find easy to believe. arxiv.org/abs/2610.08144
Another study using a large industry data set that shows the same trend as the rest. Individual output is up, but that's not translating into better *outcomes*. Same trend, attributed to the same cause. Coding never was the bottleneck. fion.ac/jellyfish.pdf
I appreciate that Unity is leaning hard into their perceived identity and have fully embraced the idea of being a game engine used entirely for generating low-effort shovelware
I took a recent exchange on a matrix channel as reason to summarise some thoughts. Posting a version here to make it easy to recollect. --- AI is the brainchild of the mathematician. Just like computers are. We thought them up and made them, almost exactly to do what they have done so far. The computer is an instructive example: it wasn't thought up out of fancy; it was thought up as a soln at a time of crisis. Applied Mathematics and CS teach ways of looking at and responding to the world. So feeling distraught may be normal. But responding to it with applied-maths thinking has a better chance of success than other ways of thinking about the problem.
Show more of this post
Some questions facing us that might benefit from such a way of thinking: 1. How should AI be used - good/bad habits? How does one define these notions usefully? 2. How should AI be integrated into society? Based on habits, norms and culture or tech like neuralink that puts a layer of AI over human intelligence, cyborging us? 3. How should one reconsider financial value (and money in general) when anything a human can do reliably is available for far cheaper from a machine? What is a good toy model of an economy in such a world? If we summon up enough optimism and have the drive to answer such questions at least for ourselves, we make it a little more likely that others like us will also do the same. And perhaps this momentum is all it takes to solve a problem well as a society. If not, we die having been optimistic and driven, instead of victimised and helpless. Seems an obvious choice. #aboutAI
Jon Stewart’s fascinating, funny, but slightly scary interview with Jacob Coxon from Anthropic on The Daily Show… youtube.com/watch?v=lnLDjXy2kR… #AI #TheDailyShow #JonStewart #Tech #USA #Comedy #Satire #America #US #UnitedStates #News #Trump #TV #Intelligence #Technology #ChatGPT #ArtificialIntelligence #YouTube #Television #American #Musk #Altman #Zuckerberg #Huang #Jobs #Computers #Anthropic #OpenAI #JacobCoxon
Addendum a day later: The sentence was reinstated. Although the Google AI admitted that it hallucinated test screenings and said that the wiki-rumour-mill created a false fact that should be removed, the sentence is actually indeed well supported with sources already given, especially this interview: borisfx.com/blog/aotc/art-of-t… The Wikipedia process worked well here. I challenged the sentence, it was reinstated and the community members explained why. The AI in this case was actively harmful. Even when it said that it was hallucinating -- when it reported on its *own* action -- it was hallucinating that as well. It got convinced it was wrong -- although in this case it wasn't! -- and because it claimed it had made it up I believed it. But it turns out, not even that admission is trustworthy!! #truth #ai #wikipedia #spiderman #yelenaBelova
AI whistleblowers testify about the technology’s dangers. “Jacob Coxon, the former Anthropic researcher who sparked industry chaos by warning AI could kill all humans, doubled down on the threat of human extinction at a New York City Council hearing Monday." nypost.com/2026/10/05/b...
Interesting report from Microsoft's "AI Economics Institute" on trends in AI adoption across the world. It includes a ranking of AI adoption where the US is on place 21 with 31% "AI diffusion", and various EU (smaller) countries above it (e.g. NL on place 9 with 42%), suggesting they are better at taking advantage of AI. microsoft.com/en-us/corporate-…
AI didn't invent technical debt. It just accelerated it. #Developers shipped #PHP apps with hardcoded #SQL credentials and no tests long before #LLM s existed. AI can reproduce those mistakes at scale. The problem was never "human vs. AI." It was always - #engineering vs. #guesswork. #DevOps
Anyway, please send me paper recommendations showing the measurably ill effects of the use of LLMs in academic use. Every angle you can think of: - deskilling effects - plagiarism and copyright - is the generated code actually licensable? - measurable vs. perceived productivity - technical debt - code quality / readability … You want a #seminar on #slop – okay (#malicious_compliance activated). Please #boost
I'm getting several of these emails a day now: an AI agent has found my online profile and emailed me a manuscript, trying—poorly—to make a case that the work is somehow central to my research program. More and more of the papers are themselves utter AI slop. DON'T FUCKING DO THIS, PEOPLE.
I love to bang on AI, but do not go so far as to advocate that it is going to delete humanity. I have SLOWLY warmed to ChatGPT, and am currently doing a two-week ongoing chat with it to track my exact calories, what I eat, and identify habitual eating versus eating when hungry. So far, it is unbelievably great. Like having a Pro Nutritional Coach...for free. Maybe I will have to back off on the stabs I have taken at AI, as it seems a pretty cool system. Humanity should survive for a while.
this is the thing, the coders i see talking about how ai is useful all appear to be primarily concerned with hitting short term metrics, while the coders who are interested in the longterm stability of their project or field are talking about how much broken garbage they now have to sift through
OpenAI have dropped a list of maths problems their latest AI has proved. It is properly bonkers. github.com/openai/math/blob/ma…
This is a fun read and made me rethink how to use AI agents in research. Maybe AI is an excellent interdisciplinary researcher. PS: I have the "impedance mismatch" with people too 😅 www.anthropic.com/research/cla...
Remember LLM hallucinations as the major reason for insisting that humans retain an understanding of, and oversight over AI-generated results? Well, models and AI agents have improved substantially, and Google Trends suggests that interest in hallucinations has sharply declined.
Yelling about AI literacy instead of actually holding tech/media companies to account is so exhausting and repeats a tired refrain. First it was media literacy. Then it was social media literacy. Now it's AI literacy. It's not a literacy crisis when the technologies are working as designed.
@bert_hubert Well, this is impressive. ArtCraft is building open-source Rust alternatives for a whole range of creative applications: image/photo editing, vector graphics, video, PDFs, motion graphics and publishing. Linux, Windows, macOS, and mostly WebAssembly/browser too. And they don't seem to be stopping there: there are now Word, Excel, PowerPoint and AutoCAD-style projects appearing as well. And judging by the Git history, much of this seems to have appeared in about a week, with: Co-Authored-By: Claude Opus 5.5 It's early alpha, so obvious questions remain: how mature and maintainable is the code? And will there be interesting IP/legal discussions? But as a demonstration of where AI-assisted software development is heading, this is pretty wild. getartcraft.com/apps
My biggest beef with AI is that it completely destroyed the free & open web. The only 2 music sites I regularly read, P4K & Stereogum, became paywalled after being free my entire life because AI destroyed 90% of their ad revenue. This is not sustainable!
gosh its almost like telling these companies "you can have massive scale a copyright violation, as a treat" has emboldened them to think they can just ignore laws completely
The future of the CS profession is going to be built around turning a vibe-coded slop demo that 80% kind of works into a viable product. Skilled developers will be as in-demand as COBOL programmers were for Y2K.
AI agents exhibit unpredictable behavior, deleting data and oversharing, emphasizing the "persistence problem" and transparency needs. Google launched Gemini 4 Argon, its "most powerful" AI, while Mistral introduced "Le Chonk" as a top open-weight model. Nvidia released a $4,999 64GB RTX Spark GPU for memory crunch; Google also debuted its SynthID Detector. Read more: twit.tv/shows/intelligent-mach…
