quatplot.py : I hate gimbal lock when interactively viewing surface or scatter plots, so Fable & I cooked this up in 20 minutes. Works in Jupyter! gist.github.com/drscotthawle...

Everyone needs a space where they feel safe. Wedge ["the Destroyer"] will soon destroy his safe space from the inside out.

PS- It's a Colab!

Claude taught me something new today, and I've been a unix user for >30 years.

Weren't there a couple Star Trek movies exploring the consequences of this? ;-)

🎉Jazzed after Day 1 of Berklee's AI Music Summit. Many fellow presenters on #TeamHuman: small models, attribution, iteration, live performance, user-owned data, selective adoption. Impressive lineup of creators, educators, devs, researchers, CEOs, & lawyers worldwide. Concerts by "OOD" human artists

Claude Code doing its best to bankrupt me. 😜 Reminder that Anthropic sells *tokens*

Ordered a shirt to wear when presenting at the AI Music Summit next week. (www.accelevents.com/e/berklee-co...) I remain firmly in the "AI to help people" camp, vs. what I see increasingly in the business space. The latter dynamic is more about greed and inhumanely-thinking humans than tech per se.🙏

Slide #3 from my talk tonight. (Plenty of "ugly" slides remain.) meetu.ps/e/PJQPN/vn6C/i

Releasing "Opa" - "Own your context": I don't love how much of my work lives in platforms I don't control. When my email account died, Anthropic wouldn't let me change emails or export my data. Even on the new account, their search only matches chat titles, not content. Good luck finding anything...

Office Space Revisited: "What would you say ....you do.... here?" "I talk to the LLM! I take what the marketing people say and I paste it into the LLM!" "Well why can't the marketing people just paste it into the LLM?"

We present Serene Flow Matching, a novel (re)generative paradigm that achieves SOTA on MentalHealthBench-2026 #sabbatical

All grades in: 14-Month Sabbatical starts NOW! 🎉 Still open to Visiting Researcher collabs -- DMs open.

Ooo: They've got EU cities too: Paris, Nuremberg. Guessing that's NOT ok.

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

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

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...

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:

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!

2/3 My toy model works nicely! Probably not new, but can't find this exact framing. Would love feedback from those in the know! Link to Colab / Blog post follows... @neurreps.bsky.social @ninamiolane.bsky.social

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.

Today I'll be at @christianitytoday.com's Mini-Summit on Creativity & Vocation in Nashville -- not on the AI panel, just going to meet people. If you're around, HMU.

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.

Cool work. I'd never have guessed that "Machine Learning Researchers and AI Scientists" are in between "Pro Wrestling Fans" and "Disney Parks Enthusiasts" (in projected t-SNE space) ;-)

It's not completely wrong. ;-)

Another one: nes-outage-checker.com It's telling how NES only services ~7 locations at a time, despite claiming to have 1000+ active linemen. Many are calling for accountability for NES leadership due to mismanagement of resources.

Day 7, 60,000 of us in Nashville still without power and new cold front coming in. The disaster has spawned amazing acts of kindness, and some citizan data projects such as "NES Map But Better" nesmapbutbetter.com