𝐎𝐧 𝐭𝐡𝐞 𝐋𝐢𝐦𝐢𝐭𝐬 𝐨𝐟 𝐌𝐚𝐤𝐢𝐧𝐠 𝐏𝐫𝐨𝐠𝐫𝐚𝐦𝐦𝐢𝐧𝐠 𝐄𝐚𝐬𝐲 New paper with Joel Jakubovic We wrote this for my PhD supervisor Alan Mycroft's festschrift, where you can get away with having some fun when writing! With the hope it will be insightful and fun to read too :-) tomasp.net/academic/pap...

Dad telling me how great work email was when they first got it. Used to have to figure timezones for when you could call the guy in Belgium. Of course he's not in his office; leave a message; he calls back and you're not, etc. Email made it easy! You send when you work; he replies when he works.

Feeling (unfortunately) good about my decision to say no this year to SPC'ing or reviewing for any big AI conferences. Lots being written about the (generally not great) review quality, but the other side is equally bad or worse. Torrent of low-quality submissions, aggressive bad-faith authors, etc.

Man goes to doctor. Says he's depressed. Says life is nothing but postmodern pastiche of recurring references. Doctor says, "Treatment is simple. I'll refer you to great specialist Pagliacci in Antimemetics, floor 𒍌." Man bursts into tears. "But doctor..."

Expanded blog-post version of this thread, with higher-res photos and videos:

Can't believe the price of a draft pint is up to $3 now! Inflation, man.

The IEEE Conference on Games (CoG) 2026 CfP is out! Will be in Madrid, Spain, 1-4 September 2026 Full paper deadline: 17 March 2026

Saw an exhibition by Shohei Fujimoto that I liked. (At Artechouse Houston.) Split between fully digital work I'd call "generative art" and a bunch of laser installations where the lasers have enough perceived volume to look like kinetic sculpture. A few examples in the thread below.

Somehow the wording "documents" here feels like an anachronism.

At my parents' for Thanksgiving, so I'm seeing some television ads I don't normally see. Google seems to be really pushing video generation in Google Photos for regular-person consumer uses, like one-off edits that add a talking stuffed animal into a vacation photo to entertain your kids.

It's hard to build an oscillator

I like really deep dives into how a generator was developed, such as this city map generator by @tomnullpointer.bsky.social www.nullpointer.co.uk/generating-c...

Lots of interesting LLM releases last week. My fav was actually Olmo 3 (I love the Olmo series due to their full open-sourceness and transparency). If you are interested in reading through the architecture details, I coded it from scratch here: github.com/rasbt/LLMs-f...

I'd appreciate if any audio tech folks could help circulate this exciting call for a fully-funded PhD position working in multichannel audio design for galleries. It's part of an ARC funded project 'From Noise to Signal: Improving Sonic Experiences in the Gallery' www.unsw.edu.au/research/hdr...

Ominous parenthetical thrown in there at the end.

You are absolutely spot-on to mention that we are currently posting on the BlueSky social network! In fact, not only are we posting here, but we're making a *classic* and *iconic* kind of social network post, the microblog.

I get the incentives, but am getting increasingly annoyed by senior academics with no background in AI reinventing themselves as AI people. Definitely the first thing I look at now for a full professor "in AI" is: did they publish a single AI paper pre-2020?

A little puzzled why LLM RL works at all. People are solving a hard, sparse exploration problem using on-policy exploration with a number of rollouts that would barely work to solve cartpole.

The previous post was a pretty literal port of our textbook chapter. Here's a version that's more animated, repeatedly generating & playing dungeon crawler levels.

If you're looking for a language for code golf entertainment that's still LLM-resistant, I can recommend BQN.

My building still has a functional mail chute, but they're currently rebuilding the mailbox area in the lobby and I see an outgoing mail drop as part of the new layout, so I fear the chute's days are numbered.

An unrelated thing from teaching audio & image models side by side: for image models it's nice to use the time dimension for timelapses (below, 30 steps of stable diffusion generating "snowy mountain peaks, abstract painting"). But audio inconveniently already uses the time dimension for itself!

Something that came to mind after teaching diffusion models this week: Why do "decent" audio diffusion models seem possible to train with much less data than image models? For example, Stable Diffusion was trained on 5 *billion* images, but Stable Audio Open used only 500k audio samples.

It was a huge honor to work with Evan and turn one of my blog posts into an IEEE spectrum article! I feel like almost a real writer now. spectrum.ieee.org/robots-foldi...

It’s unfortunate to watch neuroscientists debating about whether spike rates or oscillatory dynamics are more important for cognition when Kant already showed that it’s done by the transcendental ego and so won’t appear in any time-series data.

Two years later and this is still true!

YOU ARE LIKELY TO BE LICENSED BY A GRUE arstechnica.com/gaming/2025/...

Finally! You can now pay with contactless credit cards on all WMATA buses and trains instead of having to buy a DC-specific transit card. Pretty much the #1 nice thing as a visitor to a city is if I can pay for transit normally.

we released Olmo 3! lot of exciting stuff but wanna focus on: 🐟Olmo 3 32B Base, the best fully-open base model to-date, near Qwen 2.5 & Gemma 3 on diverse evals 🐠Olmo 3 32B Think, first fully-open reasoning model approaching Qwen 3 levels 🐡12 training datasets corresp to different staged training

Tbh I'm surprised there are still scientists who didn't give up on online surveys 10+ years ago. Kids in my high school were using survey bots to make money, and the bot detectors were always a step behind!

The timing of Thanksgiving & Christmas this year makes it so the semester feels almost over. There's a half week next week with online classes, then we return for just one week of classes, then finals week. Students certainly perceive it as basically over apart from final projects/exams.

another big earnings day for King Jensen

The two Schools here most excited about using generative AI for education are the School of Business (which I expected) and the School of Education (which I didn't expect). Pedagogy people, or at least some of them, seem very into AI for personalized education, rethinking the classroom via AI, etc.

"Post Doctoral Research Associate position in human-centered machine listening" at NJIT/NYU, working with Mark Cartwright

The next lab from my tiny systems course is up. Learn how to write your own small implementation of the logic programming language Prolog in #fsharp! www.youtube.com/watch?v=KC7A...

it's Engineering Students Constructing Illegal Radio Station Day, for those who celebrate

GDM WeatherNext 2 8x faster than v1, it can compute extreme situations and game out scenarios in one minute flat on a single TPU (as opposed to hours of supercomputer time for traditional algorithms) will be available in all of Google’s weather apps blog.google/technology/g...

Long faces in the Math department this Monday as they received the official news that they've been dropped from the M in STEM, to be replaced by Machinelearning.

Last week, Harvard released a report on curtailing "grade inflation," and (some) Harvard students kind of freaked out in an embarrassing way. It's easy to call the students coddled, but why grade inflation became a problem is a knotty, decades-long nightmare of a problem, which I wrote about:

The posting: www.linkedin.com/jobs/view/42...

So this got presented Friday, by our undergrad first author who had never been to an academic conference before (Jonathan Morse), and he did a great job! (I'm pretty sure my first conference presentation, by contrast, was not well delivered at all.)

Outdoor boarding! Nice to get a bit of fresh air.

At the airport flying back from #aiide25. Lots of interesting research that I may post about later. I think I've now been to each of my "home" conferences once post-pandemic (AIIDE 2025, FDG 2024, CoG 2023), which were all somewhat surreal experiences for a variety of reasons.

Surprised by a mention in this keynote that Claude Shannon attended the 1989 World Computer Chess Championship at the University of Alberta. I was alive then!

So good to have Jonathan Schaeffer back for the final keynote on AI in games #AIIDE25

A weird thing about switching back to glasses (with pretty high myopia) after 20 years or so of contacts is 3d color effects, like reds on a screen in particular "pop out". Makes the AIIDE logo pretty trippy.

"Solving a Million-Step LLM Task with Zero Errors", from the inimitable Elliot Meyerson (et al.)

👀

Trying out a style of conference attendance where I go to bed early enough to sleep eight hours and actually wake up for the morning keynote. Has its pros and cons.