@srush_nlp +1, I wonder what this means for the relevance of autograd in general and NN libraries like PyTorch and Jax. What is the value of the flexibility they offer if we're all using the same model architecture?
English
Jonathan Malmaud
133 posts
One ML trend I've been grappling with is that we are post-abstraction. Since everyone is building roughly the same "one big model" there isn't really much need generality. We're roughly converging to giant, straight line, assembly-coded systems that get recoded each year.
English
@tejasdkulkarni This is exactly the right direction for AI asset generation. Congrats on the launch!
English
Text -> Image -> 3D -> AI-assisted UVs -> AI Retexturing.
This is now live for all our paid users. I am incredibly excited about hybrid Artist + AI workflows.
Common Sense Machines@CSM_ai
AI-generated assets have often faced criticism for not being production-ready due to their non-editable internal representations. Now, Cube supports AI-assisted UV unwrapping to fix this. Simply go to one of your refined meshes and click on 'UV Remapping'. This feature is available now to all paid users: 3d.csm.ai/purchase-plan. AI retexturing can be used on top to achieve assets with unprecedented quality. For more information, visit our blog: csm.ai/blog/ai-genera…
English
@roydanroy Man this is making me sad that I work out of a suburban office (albeit in a nice building).
English
Jonathan Malmaud أُعيد تغريده
Jonathan Malmaud أُعيد تغريده
Excited to be attending #ICML2023! Will be presenting our paper on Wednesday in hall 1. Feel free to dm to chat about program synthesis, AI4Code, or any other topics!
Gabe Orlanski@GOrlanski
📢Measuring The Impact Of Programming Language Distribution We present the BabelCode framework for multi-lingual code evaluation and an investigation into the impact of PL distributions in training data. Paper: arxiv.org/abs/2302.01973 Code: github.com/google-researc… 🧵
English
Jonathan Malmaud أُعيد تغريده
Today I was talking to a coworker and we had a moment where we agreed that a senior exec's most critical leverage is the numerous job opportunities they have. If they don't, maybe they are not qualified. Without that leverage, I don't know how it's possible to have a backbone.
English
@SimonPJByrne Ya, good points. Definitely agree that there are lots of times when people aren't doing scientific computing and would just prefer the intuitiveness of easy fixed-point.
English
@malmaud Languages provide other numeric types which aren't native (e.g. 128-bit integers)?
English
Why did programming languages stop offering native fixed-point decimal types? Especially since they've stuck around in databases (e.g. DECIMAL types)
Dr Owain Kenway@owainkenway
This is very cool. People will remember I've repeatedly tweeted this article (medium.com/the-technical-…) by @bellmar about the challenges of numerical representation. Well there's another way to store numbers, which Common Lisp does by default which is to store them as fractions.
English
@MilesCranmer How you say this compares with Flax's `nn.compact` syntax flax.readthedocs.io/en/latest/guid…?
English
I cannot wait for this proposed deep learning layer API to be merged upstream for Flux.jl. It's so nice!
github.com/FluxML/Fluxper…
It exploits Julia's expressiveness to let you create custom deep learning layers, *in-line*, with the absolute minimum amount of code:
English
@sirbayes Isn't this just a generic (but accurate) argument against null hypothesis testing? It's always true that with enough data, every statistical test will reject every point hypothesis since almost no such hypothesis is ever 100% true IRL.
English
“Choosing a distribution is not a statistical question; it is a modeling decision. No statistical test can tell you whether a particular distribution is a good model for your data” - yes!
Allen Downey@AllenDowney
In the last week, three people on reddit/r/statistics have asked about testing whether a sample came from a Gaussian distribution. The answer is that you should never test for normality. The result is a non-answer to the wrong question. allendowney.com/blog/2023/01/2…
English
@roydanroy Oh man, seeing this tweet at 5am PT was my first time learning about the layoffs. Luckily my wife and I are ok.
English
@TomerUllman Chatgpt prompt: "One bitten, twice shy. Twice bitten, * shy. What is *?"
Twice bitten, three times careful
Twice bitten, much more shy
Twice bitten, even more cautious
Twice bitten, even more wary
English
@johnmyleswhite I've just come to accept that the value of pundit predictions is to provide entertainment and conversation topics. Its accuracy is irrelevant.
English
It really is remarkable how little people value predictive accuracy in their pundits
M. Nolan Gray 🥑@mnolangray
You have to admire Ehrlich's staying power, he's been on this grift for literally a half century, making spectacularly wrong predictions about overpopulation, mass extinctions, famines, environmental disasters, etc. and yet the media incredulously reprints his every prediction.
English
@johnmyleswhite I've been thinking of doing the same thing, but in the Bay Area - any tips from your experience?
English
Constantly amazed by the intensity of the contrast between the rhetoric of a greener NYC created by Local Law 97 et al and my own personal experience struggling to renovate a NYC apartment to use a heat pump and induction cooking.
English
I'm guessing you've had this experience: you're listening to someone give a talk, a pitch, a lesson.
It's fine. But they seem...too polished. It's like they've hit play on an internal recorder. They're on automatic, they're "not there".
And you find yourself drifting off.
English
I am atingle about this new pre-print, “Negative Evaluations of Rote Teachers".
Work by @LonnieMBass, @cristian_e, @E_Bonawitz (and me):
psyarxiv.com/pmun5/
English
@sirbayes Agreed, although:
1) Neurosymbolic methods that combine LLM code generation with reasoning (eg, automated unit-testing of potential samples) might help.
2) How does the base rate of humans making subtle errors compare to LLMs? Should we hold LLMs to a higher standard?
English
I worry with all this LLM frenzy, we are building on a bed of sand. Subtle auto-generated errors will creep into our code, which will be published on github, memorized by the next round of "intelligent" agents, and then our autonomous cars crash etc.
Alexander D'Amour (@alexdamour.bsky.social)@alexdamour
Awesome CLI tool, but that answer? 🤔 Algorithm description (correct) doesn't match the code. The code (wrong) samples a shifted scaled Rayleigh distribution (which is non-negative when "mean"=0). But the answer sounds so helpful!
English
@roydanroy @vmansinghka Wasn't Lyric Labs in Cambridge working on something like that? Not sure whatever happened with those efforts.
English
Imagine if we invested as much $$$ building hardware to accelerate MCMC as we did to accelerate Ax=b.
@vmansinghka
Max Dax@maximilian_dax
Can you trust deep learning for scientific inference? And what can you do when results are inaccurate? We address these questions for ML-based inference of complex gravitational wave models and get highly accurate and reliable results. arxiv.org/abs/2210.05686 1/12
English