@1rreverant@aksharikuu@secemp9 I saw it. That's why I said you can approximate gradients through nondifferentiable functions, but they're approximations. You can read about Gumbel softmax or straight through gradient estimators.
Its not pedantic. You cannot differentiate through a conditional. A relu is fully differentiable, it is not a branching statement. It is a piecewise linear function. You can approximate these gradients in many ways like you reference, but it is approximation. You cannot approximate a conditional, there's no world where an if statement is differentiable. What would that even mean? Where is the slope?
@1rreverant@secemp9 The demo is resizing a window. That is a differentiable operation. An 'if' statement is not differentiable, no matter if you express it in js or tensor operations. This guy seems to be implying that it is, which is untrue. You cannot backpropagate through a conditional.
@secemp9 This is the problem with coding agents... It'll do whatever dumb shit you ask for without question. Then you look like a fool when you claim a breakthrough.
@loveofdoing No. It's possible to represent code as matrices, yes. But code is not automatically differentiable. LLMs are not symbolic. If you represented an LLM as code, it would be low resolution.
@jpschroeder None of this is new besides it being in javascript. Read about RASP, Tracr, ALTA. It has serious practical limitations, eg loops and branching.
@SGRodriques Cuts both ways, though. You can't have a model tell you about defense without also telling you how the defense is beaten. Same issue in info sec.
@Samuel_Gregson@ilbue63@Planarwalk You two are characterizing her as a populist conspiracy theorist crackpot, which she is not. That's why it looks like she struck a nerve. Her criticisms of physics research are pretty valid imo, she does not believe physicists are 'nefarious'. Just misguided.
@ilbue63@Planarwalk@Samuel_Gregson I get it but I think you're proving her point. If we won't know if it's applicable for decades, maybe it's not appropriate for the prize. There's other discoveries in basic science that are more directly beneficial, eg quantum error correction or materials.
@ilbue63@Planarwalk@Samuel_Gregson Maybe you could explain what benefit it has? Its no doubt a great discovery but let's not pretend its practically useful. I get you all dislike Sabine because she dogs on your field but maybe it hurts because she's right.
@Planarwalk@Samuel_Gregson If you watch the 15 second clip she actually says the discovery doesn't benefit humanity, which is the verbatim criteria of the Nobel.
3K Followers 2K FollowingAssistant Prof. @ChicagoBooth | OM & Applied AI | Most Cited Paper: Neural Collapse in AI Training | BSE @Princeton, PhD @Cornell, MS/PostDoc @Stanford
129 Followers 3K FollowingThe Lord bless you and keep you; the Lord make his face shine on you and be gracious to you; the Lord turn his face toward you and give you ...✝️
12K Followers 11K FollowingProfessor @UAM_Madrid
Writing on what matters in AI for Science: https://t.co/sNvjOIcqgU
"IA para estudiantes de ciencias": https://t.co/Zs5o2gszjj
264 Followers 473 FollowingCoordinating research and community around AI-driven algorithm discovery.
AIDDA 2026 Conference - June 9 and 10, 2026 - https://t.co/ForgcRZoF8
33 Followers 397 Followingpeptides , sarms and gears available 💪
available . click link join all my social media and don’t forget to join my discord 👇👇
WhatsApp :
612K Followers 60K FollowingThe automated life | San Francisco/Silicon Valley AI, robotics, BCIs | Ex-Microsoft, Rackspace, Fast Company | Wrote eight books about the future.
8K Followers 7K FollowingCenter for Language and Speech Processing at @JohnsHopkins #NLProc #MachineLearning #AI https://t.co/6IXR5OSQtw
@[email protected]
61K Followers 0 FollowingExploring theories of everything, consciousness, Ai, and God. Early access episodes: https://t.co/9HHWJeJbKw
For business inquiries: [email protected]
682K Followers 20 FollowingThe pricing error account they tried to shut down… We find the glitches. You make the profit. Turn 🔔 on to not miss deals! (We may earn affiliate commission)
1K Followers 989 Following🦾 Head of AI @TheHumanoidAI
💻 Ex @nebiusai, @Yandex, @MSFTResearch, @CHARMTherapeutx
🧠 Interested in (M|D|R)L, AGI, rev. Bayes
🤤 Opinions stupid but my own
18K Followers 12 FollowingPrivacy-first AI assistant by @ProtonPrivacy, built in Europe. Lumo does not track or record your conversations.
Ask me anything, it's actually confidential.
3K Followers 2K FollowingAssistant Prof. @ChicagoBooth | OM & Applied AI | Most Cited Paper: Neural Collapse in AI Training | BSE @Princeton, PhD @Cornell, MS/PostDoc @Stanford
177 Followers 228 FollowingCEO at @pingram_io
Ship Emails, SMS, Voice to production in 2 minutes.
I tweet about startups, software, cloud, and AI.
Prev. HeyOrca (exit 2019)
80K Followers 714 FollowingSharing insights on Probability, Statistics, ML, DL and AI research. Subscribe for recent research paper discussions at $2/month. DM to collaborate.
90K Followers 1K FollowingCo-founder and CTO at @huggingface 🤗. ML/AI for everyone, building products to propel communities fwd. @Stanford + @Polytechnique
86K Followers 1K FollowingNYT bestselling author of EMPIRE OF AI: https://t.co/wh5mCWc2mo. ai reporter. american national magazine award winner. in @MorePerfectUS @theatlantic @wsj @techreview