I had a good chat with some students in my software engineering class about these issues.
It's the mythical man month! AI users are playing mid-level manager to junior engineers, and are evaluated based on how many lines of code they get those engineers to write. 🤦
Apparently there is a tension right now in industry between "tokenmaxxing" (where employees are evaluated by how many tokens they use) and non-corporate users who are paying for AI coding hitting their usage limits because problems are more complex than they expect.
www.forbes.com/sites/ronsch...
Happy birthday!
@pinkhairedcyn.bsky.social do you know any of these students? If so you should let them know they've gone viral on professor bluesky
photograph or a poster on cream colored paper. "Dear President Ambar, we are writing to you on a typewriter that is over 70 years old. This is a machine that we all know well. With it, we misspell words without the crutch of spell check or generative AI and we think intently about every phrase we pound out. As we force ourselves, for once, to slow down, we engage in a cognitive dialogue with ourselves. We do not seek perfection because we know that education is about the growing and challenging of our young minds' potential, not the chasing of institutional 'gold-star' approval. We do not believe that your so-called 'Year of AI Exploration; providing enterprise ChatGPT and Google Gemini subscriptions to every Oberlin student aligns with our college's founding principles. You claim that this year will be one of experimentation, not adoption. But even just one semester of accepted (encouraged even) chat bot use will jettison our student body down a lazy and irredeemable tunnel of intellectual destruction. We are a college grounded in learning and labor, which now risks straying from these rooted ideals. With ChatGPT at the helm, our emails, essays,and discussion posts will be generated for us, not by us. And let's not fool ourselves. This is precisely what these platforms will be used for by our busy, anxious student body. We see your vision for this year as.advancing the college's 'businessification'--an alarming trend also seen in the takeover of our beloved library cafe by a 'bookstore' with no books in stock and an app replacing customer service. In one instance, the college assumes we want efficiency at all costs through automated rather than hand pulled coffee. In the other lies the false belief that we simply desire to turn in an essay, regardless of how little we've written of it." there's more that doesn't fit in the 2000 character limit :(
OH MY HEART...the Oberlin Luddites Reject "The Year of AI Exploration"! 💚
Join @taylor-arnold.bsky.social and me for a postdoc to expand distant viewing! 2 years. Awesome liberal
arts college. Great city. Option to teach 1-2 courses a year for extra pay. Thanks to Mellon for the opportunity! Happy to answer any questions.
richmond.wd5.myworkdayjobs.com/en-US/staff_...
hearing microsoft is reorganizing its AI team under the banner of "the Copilot System." Also hearing that teams are under pressure to *reduce* AI token use, remit is that there needs to be "fiscal responsibility in AI ops" and that Claude Code usage is being reduced in favour of Copilot CLI.
This is a really unpleasant way to work, if only because it gives you "free time" but you have no control over when it begins or ends.
I know my life got much better and more productive when I started spending more time thinking and less time running lots of experiments.
It's funny to see users of AI coding agents adopt the deep learning grad student workflow, where most of the thinking is automated, but you are called into action whenever the code finishes running to review the results and set up the next experiment.
www.theregister.com/2026/03/12/i...
RIP to the man behind the funniest bit of all time: the invention of the null pointer
every quantitative measure is actually a stack of qualitative assumptions in a trenchcoat
A very good article on AI and coding from the NYT
by @clivethompson.bsky.social
Here's a gift link: www.nytimes.com/2026/03/12/m...
“A bill under consideration by the New York state legislature would prohibit artificial intelligence-powered chatbots from providing legal or medical advice, and it would allow users to sue chatbot owners who violate the ban.”
MicroGPT: a masterpiece "Minuet" by the Mozart of ML, @karpathy.bsky.social: karpathy.ai/microgpt.html
200-lines of *dependency-free* Python, yet still very readable.
If it's not, and the government is being earnest, they just handed Anthropic the best marketing they could have asked for, especially in their attempt to win over anti-AI liberals.
So the whole Anthropic thing with Hegseth and Trump is some kind of publicity stunt, right?
They've been spending a ton on advertising recently, especially through nontraditional means, and this kind of news incident gets them exactly the kind of publicity they want.
I'm a cognitive scientist with an interest in epistemic vigilance, and this essay that's been going around gave me pause.
I don't think it's straightforward to apply the concept of epistemic vigilance to interactions with LLMs, as this essay does.
🧵/
sbgeoaiphd.github.io/rotating_the...
This whole project started from an side project I did during lockdown making weird gifs with CycleGAN: samgoree.github.io/assets/cycle...
Thanks to my labmate Vibhas Vats, it is now a real paper with an enormous amount of inquiry beyond my initial experiment.
T-SNE plots of three scenarios on MNIST, illustrating the latent manifold collapse across many generations of retraining or re-generating.
A table showing eight kinds of collapse according to three metrics: sigma-intra, M_LB and PR_G. Each is characterized by its local and global collapse or expansion. Coherent Expansion exhibits local and global expansion, with both local metrics increasing Wrinkled Expansion exhibits local expansion and global contraction, with both local metrics increasing Anisotropic Expansion exhibits local expansion and global contraction, with M_LB decreasing Oblate Expansion exhibits local expansion and global contraction with M_LB decreasing Coherent Contraction exhibits local contraction and global contraction with all metrics decreasing Anisotropic Contraction exhibits local contraction and global expansion with both local metrics decreasing Wrinkled Contraction exhibits local contraction and global contraction with M_LB increasing Oblate Contraction exhibits local contraction and global expansion with M_LB increasing
Other patterns of collapse follow different patterns, which we categorize into an eight-category framework. The T-SNE plots show the MNIST latent feedback scenario, which exhibits oblate explansion, and two retrained scenarios which both exhibit coherent contraction into wrinkled contraction.
Spark plots showing the curves for each of five experiments. LGR-MNIST exhibits CC and WC LF-MNIST exhibits OE LGR-ImageNet5 exhibits CC LF-ImageNet5 exhibits WE CycleGAN exhibits CE Acronyms are explained in the next post
We use three metrics for dimensionality, and found, interestingly, that these feedback patterns result in very different kinds of collapse. But the feedback loops that exhibit directional contraction and obey the Markov property all exhibit the same kind of collapse, which we call neural resonance.
diagram of four experimental scenarios -- Lucier's feedback loop, CycleGAN, Latent-feedback models, retrained models, with diagrams for each.
We explored this phenomenon in four modes -- the Lucier setting, CycleGAN, in a diffusion model with latent feedback (where we train a model to recover the latent input and loop those), and in a fully retrained diffusion model. Then, we study each loop's dynamics in latent space.
This is, at a high level, the same thing that is happening when you ask ChatGPT to make you a copy of an image over and over again, which went viral last year. x.com/genekogan/st...
It is also what is happening when we retrain models on their own output.
Alvin Lucier's sonic art piece, I am sitting in a room, is a classic example of minimalist music composition. www.youtube.com/watch?v=fAxH...
It starts with a spoken message from Lucier, which is recorded, played back and rerecorded into the same room over and over again until the speech is gone.
A preprint for a paper I've been contributing to on-and-off for over five years is now out! We look at model collapse, the phenomenon where AI models trained on their own output stop working, using the theory of Markov chains, and the music of Alvin Lucier arxiv.org/pdf/2602.19033
If you ever wonder if your research idea is too simple or obvious to get published, submit it anyway. Worst case they reject it. Best case, everyone who walks by your poster instantly understands what you were doing and why.
The really fun part of this poster for me, at least, is that I could explain it to anyone in a sentence. Every conference talk/poster I've done in the past has been complex and niche, so only a dozen people are even interested, and it takes a few minutes to explain even to those people.
I had the opportunity this week to present a fun little poster at SIGCSE.
We took data structures students on a field trip to the library to help them better understand how the library functions as a data structure, and the algorithms that librarians use to maintain it. The activity worked great!
Today I was proud to share my 2026 ACM SIGCSE Technical Symposium opening keynote, titled Love, Learning, and Computing Education. Here's a rehearsal, in case you weren't there. Thank you to the organizers for inviting me to speak and to our community for listening and learning. youtu.be/iaT_zvNN_zc
