I'm on a LifeFitness treadmill in Nashville TN and it's throwing up a video walk/run of Buenos Aires-- *actual* video: tourist' faces unobscured (a la Google Street View). Doubt they got release forms from everybody (as with movies/TV). ...Is this ok? @404media.co @hypervisible.blacksky.app

How do we make LLMs faster and lighter? Don’t force the GPU to adapt to sparsity. Reshape the sparsity to fit the GPU! Our latest work with NVIDIA introduces new CUDA kernels & data formats for faster inference and training of sparse transformer language models: Blog: pub.sakana.ai/sparser-fast...

Very well written, engaging and accessible. Recommended! (I read a pre-release version)

"Hacking the Canvas API (Extension, Customization & Automation)" is the title of a talk I volunteered to give to my college in January — crickets, except for two Data Science profs. Still think the open API is Canvas's greatest strength. The title, however, may need revisiting!

The timing of this Canvas hack made me laugh, since this is my class’s very next assigned reading (which they now can’t access lol) eric.ed.gov?q=lms+AND+k1...

Ooo neato! I must investigate this.

Google tools for customizing searches | Discussion

Good luck with AI without humanities, UK.

Another "required reading" post from @sedielem.bsky.social . Sander's posts are always a treasure troves. 👏👏👏

Russell Moore: "I use AI at the *end* of the creative work, to push back on what I said and to challenge my position. 'Give me the strongest argument against what I've written.'"

Hello there! On this special occasion, we want to introduce a new kind of Force users in the galaxy: micro‑Newton thrusters.

AI slop and ripping people off are bad. But having watched both films, I don't see it. Very low correlation strength IMHO. I suspect confounding factors. www.cartoonbrew.com/student/nicc...

This confirms my preexisting biases. ;-) www.psypost.org/general-inte...

University-sponsored Microsoft Copilot training is going well./s Bonus: Teachable Moment about how Internet scraping + Maximum likelihood = Stereotypes! ;-)

I still read. I don't ask an LLM to summarize. But I use @howard.fm 's SolveIt for "close reading" to chat as I read. Here's Alberto Pepe's 228-page PhD "Machine Learning with Geometric Algebra: Multivectors for Modelling, Understanding & Computing" in SolveIt form share.solve.it.com/d/8d9bd19c3a...

Hey remember MOOCs? Let’s do that again but make it much much worse for everyone involved by remixing random course content

"Compacting..." is Claude's way of saying, "Forgetting everything we've been talking about..." (Even its own responses shortly before compacting.)

"Write your evals first" is a good principle for MANY forms of scientific work -- not just for LLM-powered software engineering. ;-) We would call it building instruments and diagnostic subsystems.

Unable to fly to Rio for #ICLR @iclr-conf.bsky.social , yet here's my poster for "Flow Where You Want"! BlogPost & Colab: iclr-blogposts.github.io/2026/blog/20... Video: iclr.cc/virtual/2026...

The Living Wage for Musicians Act would create a new, additional royalty that is ONLY available to human artists. No AI music is eligible for the royalty. Support the LWMA here: weareumaw.org/make-streaming-pay

Oooo! What happens when a consortium of numerical analysis experts turn their "attention" to Attention, specifically fast approximation techniques at inference time? 42 pages of yeeeaa. arxiv.org/abs/2604.01757

Excited that tomorrow's keynote speaker for @Belmont University's SPARK Symposium is Destin Sandlin of the Smarter Everyday YouTube channel. Mostly he does fun physics-related videos, but sometimes it gets serious... Check it out

I still have yet to try xLSTM's but I find it bizarre that only 400 people on here follow @hochreitersepp.bsky.social and Bsky needs more technical activity, so reposting!

Why do the feeds of so many [social] media platforms eventually become dominated by "Suggested Posts"? Not just ads, yet they're still noise that devalue the platform (ie make me want to *stop* scrolling & not return). Facebook, Reddit, Instagram,... Glad BlueSky hasn't done that -- yet.

More technical content posts like this my feed, please! Let's diminish the need to check the other site to stay up to date in ML

Real example, not "/s": bsky.app/profile/asa-...

For all Copilot's failings, it is EXCELLENT at Windows technical support! I am amazed. Seems this was a priority for Microsoft -- and/or what they had the most training data for. ;-)

Tiny quality-of-life utility: do_every(). Instead of mod-based integer interval checkers (e.g. 'if step % ckpt_every == 0:', this can let me easily do wall-clock-based interval checking, e.g. 'if do_every("15min"):' Each do_every() gets unique ID for indep. tracking gist.github.com/drscotthawle...

Update: Yes this method works. No, I haven't found any practical use for it yet ;-) (at least not compared to other existing methods). There may be some practical use, I've just been working on other things (that don't work, lol).

Amazing tutorial/explainer on LLM model quantization by Sam Rose @samwho.dev : Interactive visualizations,...Just wow! ngrok.com/blog/quantiz...

Fun music visualization. Crazy rhythm :-) www.youtube.com/watch?v=ZDyk...

Thinking of starting to replace "AI" with "quantum computing" in future edtech-hype discourses, e.g. "We need to teach all students [quantum computing] literacy", "We Are a [Quantum Computing] Campus". /s

This is so, so well-articulated.

Former student Billy Mitchell just defended his PhD at U. North Texas! Now seelking postdoc positions. Strongly Endorsed! Billy can do it all. He first-authored a JAES paper w/ me on neural audio effects as an undergrad, been working on photons & positronium since.. www.linkedin.com/in/williammi...

"Guy from Memento" is such an apt metaphor: they have no memory, no persistent state, only a running chat log prepended to each query. I find myself asking an LLM "Why did you just do that?" but the decision is long gone. I'm asking the guy from Memento to read a chat log and make up an explanation.

TIL that despite 'sandboxing' Claude Code can still *read* from anywhere on your machine. Solution: Put it in a dev container. Pretty painless, actually. :-) (Use case: trying karpathy's autoresearch on my Mac via CC).

The video for @freya.bsky.social's great intro to Geometric Algebra at #GAME26 is now posted on YouTube. Very accessible, highly recommended. www.youtube.com/watch?v=QFzo...

Happy Pi Day, y'all youtu.be/AlB9pXJ-T4k

The post of the month...

Saw early screening of "Project Hail Mary" movie. It was good. Note that they removed (almost) all the science that was in the book. Don't expect hard sci-fi, but it's a fine as a "space fantasy adventure" e.g the film that started Ethan Hawke & River Phoenix's careers www.imdb.com/title/tt0089...

Congrats to authors who pushed back on Grammarly's ill-conceived "Expert Review". As noted on ./: Grammarly could have offered ROYALTIES & LICENSING instead of the typical zero-sum game of "rip everyone off" or "shut it down". There's a middle path: Opt-in, Give a cut. slashdot.org/story/26/03/...

"need a human to take responsibility" applies to coding too. arstechnica.com/ai/2026/03/a...

Update: On real data, combining the "factorization loss" with an "attraction loss", SIGReg loss, and MAE decoder loss... is messy -- esp. the latter. I'll replace the MAE decoder with a masked embedding decoder (cf iJEPA). Also: The earlier diagram had a mistake showing pitch ^ time. Should be this:

Commentary: Anyone Else Have Those Weird Dreams Where Sobbing Future Generations Beg You To Change Course?

While hilarious at first, bad idea in retrospect: It subtly encouraged me toward anthropomorphic thinking and regarding the LLM as some kind of friend -- even in me! And I've read enough @j2bryson.bsky.social to know better. Let's reaffirm: "The model is not my friend"

GPU-poor tip for #LeJEPA: chunk your SIGReg slices. 256 slices / chunks of 32 = 8x less VRAM, and it even completes *faster* on my 4090(MaxQ). Giant complex exponential tensors were my bottleneck. One extra loop. Now more room for batch_size!

Live stream of Geometric Algebra Mini Event (GAME) '26! www.youtube.com/watch?v=Bdhz...

Been musing on the geometry of disentangled representations & factorization: Doing it by construction seems like a limiting inductive bias. What if you just add a soft geometric constraint to the loss and let the latent space sort itself out? Easy and tunable. 1/3

qwen3-coder:30b (18 GB) as a "junior dev" in Claude Code on my (64 GB) Macbook Pro means the laptop gets warm, and every few hours CC pops up asking to do something I don't want. Then it disappears for another few hours before asking if it may run "ls" 😜. Not quite the revolution I was promised.