- Yann LeCun @ JMM'25. ‼️
Found this talk to be extremely interesting - and also a hint at how much more we could do with AI if we had... even MORE compute (and rely less on 'dumb' forms of learning). Related, but at the opposite end of the spectrum - I was listening to this talk by Hinton and found it generally underwhelming, lots of things just 'extrapolated' out of thin air. Was still interesting to listen to it, but hard to agree with most of the ideas.
- History of Lisp machines.
Very nice talk on the history of lisp machines and of the first "AI winter". I wouldn't try to use this too much to divine how the current AI trend will pan out, but it is a nice talk to listen to. Even if you think you already know about lisp machines - I imagine that not most are really familiar with their history.
- Genesis GPU physics.
Not sure if this is really 'AI' - but a bunch of of people from different universities and companies created a new, very fast and parallel multi-modal physics simulator, aimed at generating data to train AIs. The results seem impressive (and the various applications to training are amazing), even if at the time of writing, the corresponding paper is not out yet. But what's surprising is all of this seems to be a collection of (mostly differentiable, but not always) "conventional" solvers. Not a ML prediction model itself, which is surprising as ML can be good at accelerating complex physics. I guess they wanted "ground truth" data here.
Retrocomputing.
- Voodoo 2 on windows 10.
Insane. Drivers.
- Add an ISA slot to Modern Motherboards!
In the spirt of the other link... about adding a Voodoo 2 to Win10.
- Vintage 3d.
Well I guess as I added a link to how to use a Voodoo2 and how to add ISA to a modern PC - now you'll want to know which GPU is best... see also Virge deep dive. I also like to look at emulators (PCem, 86box) to see how these early "mainstream" 3d GPUs used to work exactly. Back then, when I was a kid playing with computers, of course I did not know any of this...
Miscellanea.
- #rtr Collection of attempts at realtime GI.
Very nice! Albeit to this day, I'm not persuaded that ReSTIR is such a great idea, compared to other ways of temporally accumulating data for importance sampling... But For that I'd have to put my money where my mouth is and start coding something.
- #rtr Lygia shader library.
Seems well maintained. Could be useful - I used to have something similar myself (but much less extensive, and much less cross-platform(.
- #gaming Mario Kart 64 Decomp. ‼️
Amazing work, decompilation is the new emulation? See also decomp.me
- #gaming Battlefield 6.
Trailer is out. While you wait you might enjoy blockfront or battlebit (albeit I hear the latter is dead-ish?).
- #cgi Path-traced lens flares.
J.J.Abrams approved. One day we'll do the same for (realtime) DoF? It has always been something in the back of my mind to try to compute a table/function to get a better "blur function" (CoC) depending on the location relative to the focal plane. Should be quite easy, but I've never seen anyone doing it.
- #math Pendulum fractal.
Cool idea - exploring and visualizing the conditions that make double pendulums fractals. Science youtube is fantastic.
- #demoscene 4k image behind the scenes.
Neatly explained.
- The Physics of Dissonance. ‼️
I've always been curious about the mathematics of music, but never fully grok them - this video does a fantastic job illustrating one of the most accredited theories explaining where harmony comes from.
- The big OOPs - BSC 2025. ‼️
I'm not a fan of Casey's online persona, nor of the maximalist take that every single cycle counts at all times, and modern software sucks (as we didn't have... basic... or VMs in the 80ies and so on). So my praise for this talk should count even more? It's a fantastic, very well researched, very reasonable talk. A must-see! The conference in general seems quite interesting Better Software Conference and well organized! I'll probably watch a few more videos from it, when I find time.
- Fortran Hints.
TIL - Fortran had a "frequency" statement as a hint of how... frequently a statement was going to be executed. Wish we had it today! C++ compilers do often times have hints you can set on branches (likely/unlikely) but nothing more - the rest I guess is the realm of profile-guided optimization.
- GradIEEEnt half decent: The hidden power of imprecise lines. ‼️
Tom7 - enough said. You should watch everything video he posts.