RSchaeffer

We examine min-p sampling (ICLR 2025 oral) & find significant problems in all 4 lines of evidence: human eval, NLP evals, LLM-as-judge evals, community adoption claims

Best of N was shown to exhibit power (polynomial) law scaling (left), but maths suggest one should expect exponential scaling (center). We show how to resolve this "paradox", then use our insights to design methods for predicting inference-scaling capabilities that can be more sample efficient!

Thanks for taking the time to explain :)

Thanks for the advice and links! Do you know of a Render tutorial that involves getting Flask and ReactJS services to communicate with one another? Your 2nd and 3rd links demonstrate each independently. I don't know whether the same challenge will pop up in NextJS

Can you share more about the habits you built to break that cycle of scrolling for an hour in the morning?

Quitting smoking isn't a one-time event. Anyone with an addiction will tell you that it's a lifelong struggle.

Why make people search instead of quoting the relevant section?

"The human fasting mimicking diet (FMD) program is a plant-based diet program designed to attain fasting-like effects while providing micronutrient nourishment (vitamins, minerals, etc.) and minimize the burden of fasting. It comprises proprietary vegetable-based soups, energy bars, energy drinks, chip snacks, chamomile flower tea, and a vegetable supplement formula tablet (Table S4). The human FMD diet consists of a 5 day regimen: day 1 of the diet supplies ∼1,090 kcal (10% protein, 56% fat, 34% carbohydrate), days 2–5 are identical in formulation and provide 725 kcal (9% protein, 44% fat, 47% carbohydrate)."

"Subjects in the FMD cohort consumed the provided experimental diet consisting of 3 cycles of 5 continuous days of FMD followed by 25 days of normal food intake."

Then you're being far too generous with your interpretation

Does anyone know how frequently these are offered?

People frequently recommend Strang's teaching as an amazing pedagogical approach for engineers and applied mathematicians, but I find I'm frustrated every time I read his books or listen to his lectures. They don't work well for me and I've found much better alternatives

But how did their model compare against others? The article only mentions how their interpretable model compared against their own ML attempts

I just finished this fantastic class taught by Cengiz Pehlevan (https://pehlevan.seas.harvard.edu/), so I thought I might share the lectures and exercises with HN.

Why is a VAE not a generative model?

Are you hiring? This looks interesting

This is going to sound cynical, but I recently invested a week in rllib for a project before discovering that much of the under-the-hood implementation was horribly confusing, poorly documented and missing critical functionality (for instance, their IMPALA implementation only works with discrete action spaces). Does this library conceal similar problems?

I think you misread my question. I'm asking what math students will learn, not what math students should already know.

Hey! Thanks for these great courses and materials! How much additional math (beyond high school and introductory college courses) do these courses teach? For example, if I were to take both courses, would I be able to understand the papers published by Surya Ganguli (e.g., The Emergence of Spectral Universality in Deep Networks, Variational Walkback: Learning a Transition Operator as a Stochastic Recurrent Net)?

I'm curious to know how your paper differs from Learning and Querying Fast Generative Models for Reinforcement Learning. It seems relevant, but you don't mention it iirc.

IIRC, his "Bayesian interpretation of the noise" actually shows that dropout performs approximate integration over model parameters. As he says, dropout doesn't work because of the noise but despite the noise.

https://youtu.be/3ONLxYeM1Sc?t=19m21s

Go watch Yatin Gal's talk on dropout in neural networks. He shows pretty convincingly that the belief that dropout reduces network overfitting by introducing noise is wrong.

As someone thinking about pursuing a PhD to ultimately become a professor, can you link me to your talks or blog posts on the subject of what's required to be competitive?

Thanks Lubarov :)

Do you still have your replies to the speed K? I can't find them in my backfiles

yee

Peter Watts has an interesting take on this in his book Starfish.

Thanks for sharing Kerr's articles! I wasn't familiar with this issue, so I read them, and in my opinion, I think he's dead wrong (and the 3rd Circuit ruling). The argument that by disclosing his password, Doe is only admitting, "I know the password," which is a forgone conclusion, is nonsense. That statement necessarily carries with it a number of additional statements, including "Very few other people (if any) also have this password" by virtue of what a password is, and "I have read/write access to this hard drive," which when coupled with the previous statement, leads to the conclusion "I wrote the material on this hard drive to this hard drive." Kerr's argument is basically "Doe is only admitting to the premise" while ignoring that an entire chain of reasoning necessarily follows from the premise.

> For these reasons, people should consider the possibility of starting a career in tech without going to college, and employers should consider the viability of hiring such candidates.

How does this differ from the status quo?

I should have been more precise. By problem, I was referring to whatever, "Berg and Ulfberg and Amano and Maruoka have used CNF-DNF-approximators to prove exponential lower bounds for the monotone network complexity of the clique function and of Andreev's function," means, not "P ? NP."

Can someone ELI5 what this problem is, how likely the proof is to hold up to scrutiny, and whether P != NP follows?

"Alphabet Inc., through its Google Android operating system for smartphones, and Apple Inc. also have the ability to monitor how rivals' apps perform on their mobile platforms, but it isn't clear whether they use that information to shape their product road maps."

Does anyone have any other sources that can confirm or deny whether Google/Apple use their mobile OSes like Facebook uses Onavo?