Dashiell
banner
dashiells.bsky.social
Dashiell
@dashiells.bsky.social
Machine learning haruspex
This is not my kindest post, admittedly
November 10, 2025 at 4:45 AM
If the things being described are small and dumb enough that your equations don't fight back, then it can be accurately described by math
November 10, 2025 at 2:33 AM
I like the understated pun
October 23, 2025 at 2:36 AM
The (n-1, 1) representation ?
October 18, 2025 at 11:56 PM
Building an 11th century castle with period accurate tools and techniques in the Bay Area
October 15, 2025 at 11:01 PM
I am rigorously and provably sane as long as I'm integrated against any Schwarz test distribution
October 13, 2025 at 11:01 PM
I look forward to it! I know representation theory, but don't know any physics. I've always been intrigued by the connection
October 7, 2025 at 5:16 PM
I'd second group representations in particle theory
October 7, 2025 at 4:49 PM
Uses as much water as 640 acres of corn
September 27, 2025 at 6:57 PM
The points about the necessity of real touch data are super interesting though
September 27, 2025 at 2:57 AM
The inductive bias of ConvNets is probably why it was the "imagenet moment" and not the "alpha fold moment". But we've largely seen a consolidation of architectures, you only need the domain expertise for that last bit of oomph
September 27, 2025 at 2:57 AM
I disagree with how he characterizes the successes of end to end learning. Vision transformers do extremely well with basically no inductive bias, you can get a pretty good LM trained end to end on bytes, etc...
September 27, 2025 at 2:55 AM
No Free Lunch Theorem for Virtue
September 22, 2025 at 5:35 AM
Honestly being too good at math just makes it more likely you're going to get lost down some elegant-but-useless rabbithole
September 20, 2025 at 1:17 AM
I wonder if it would work better on TPUs? The Jax compiler is a lot more flexible than torch's
September 3, 2025 at 4:39 PM
I humbly recommend Lem's Summa Technologia
August 26, 2025 at 7:33 PM
Triumph of the Will and Birth of a Nation are both straightforwardly propaganda and are also two of the most influential films ever made. You have "almost never" doing a lot of work there
August 18, 2025 at 5:26 AM
*interiority
August 13, 2025 at 3:56 PM
It doesn't scan, but it's way more clever than I could come up with. Especially on the spot
August 13, 2025 at 1:28 AM
The person who breaks down what LMs do in terms of human understandable mechanisms like "multiplexing" will win a Turing award and Anthropic will give them a gazillion dollars
August 10, 2025 at 8:28 PM