Jacob Buckman

Best part of the article is the footnote, where the author characterizes his own papers: > This paper is bullshit, this paper (a NeurIPS oral) is bullshit, this paper is complete bullshit, this paper is mostly good science but also has a sprinkling of bullshit. Apologies to my co-authors. https://jacobbuckman.com/2021-05-29-please-commit-more-blatant-academi…

https://jacobbuckman.com/2020-01-17-a-sober-look-at-bayesian-neural-ne… "But in practice, BNNs do generalize to test points, and do seem to output reasonable uncertainty estimates. (Although it’s worth noting that simpler approaches, like ensembles, consistently outperform BNNs.)"

> my claim was that we need some mechanism to decide who gets a position and who doesn't I don't disagree. My claim is that the current mechanisms are nearly indistinguishable from noise but we pretend that they are very strong. I have also made the claim that Goodhart's Law is at play and made claims for the mechanisms and evidence of this. I'm not saying "trust me bro," I'm saying "here's my claim, here's the mechanism which I believe explains the claim, and here's my evidence." I'll admit a bit messy because a HN conversation, but this is here. What I haven't stated explicitly is the alternative, so I will. The alternative is to do what we already do but just without journals and conferences. Peers still make decisions. You still look at citations, h-indices, and such, even though these are still noisy and problematic. I really just have two claims here: 1) journals and conferences are high entropy signals. 2) If a signal has high entropy, it should not be the basis for important decisions. If you want to see evidence for #1, then considering you publish in ML, I highly recommend reading several of the MANY papers and blog posts on the NeurIPS consistency experiments (there are 2!). The tldr is "reviewers are good at identifying bad papers, but not good at identifying good papers." In other words: high false negative rate or acceptance is noisy. This is because it is highly subjective and reviewers default to reject. I'd add that they have a incentive to too. If you…

This is also a huge problem in offline Rl (learning a policy using only a dataset). If done naively, the learned policy will keep accumulating errors due to enter areas that are not well covered. So the trick is to avoid these areas. In offline rl they do this by measuring epistemic uncertainty and using this as a regularization term in the loss function such that the model learns to avoid these areas. This a good blog post that explains it way better https://jacobbuckman.com/2020-11-30-conceptual-fundamentals-of-offline…

I fully agree with you. While I can't speak for Keras, after reading [1] it seems Tensorflow would very much benefit from strong types. You'd easily be able to catch most of the errors presented in the article with higher-kinded and linear types, and maybe you don't even need that much power. [1]: https://jacobbuckman.com/post/tensorflow-the-confusing-parts-1/

> if you're saying i'm a racist I wasn't, sorry if that's what you interpreted. > The status quo however should be changed by people,not people with machines I'm not sure I understand what this means. People/companies own machines (and ML models) and use them. So shouldn't we make sure that the machines' decisions align with what people/companies _want_ them to do? (i.e. that the people's ethics align with the ML model's ethical consequences; I'm 100% sure that people who deploy "racist models" don't do it on purpose or out of malice) > You can't call something ethical just because you think it should be; it must be argued out On one hand this sounds like a strawman. No one thinks that something is ethical because someone randomly declared it so. On the other hand... ethics are a human construct, and will continue to evolve as our culture evolves over decades and centuries. Shouldn't we construct ML models which are flexible in that they can align themselves with the ethics we collectively decide? We don't know how to do that yet! > Using AI to shut out some more of that argument will only create a universal standard, not necessarily the correct one. You seem to be under the impression that the fields of AI ethics is dedicated to brainwashing people into some particular unpopular moral philosophy. This is simply untrue. Within the field of AI ethics there is a lot of diversity of thought and disagreement on how human morals should be "encoded" so that AI can "align"…

> That's a sloppy statement. You haven't defined what standard. Actually, I think I established a basis of discussion then later illustrated this basis with the way HN has grown and tends to enforce its standard. > A "rhetor" is a teacher of rhetoric. This is not the correct word in this case. A rhetor is a person practicing rhetoric, or a person delivering persuasive or effective communication. Teachers are indeed a subset of that, but also public speakers, negotiators, and e.g. authors who write to influence their audiences' understanding or perception. I think you may also misunderstand the function of ethos in rhetoric. > A knowledgeable audience can decide whether to trust something based on the actual content. Perhaps we're in agreement? If the content reflects reflects clear understanding it can improve the efficacy of a rhetorical artifact, while sloppiness can reduce its persuasiveness. Errors can diminish ethos. Correcting errors can amplify it: https://jacobbuckman.com/2020-01-17-a-sober-look-at-bayesian-neural-ne…