Gennady Gorin

Maybe it's a Meta issue, maybe it's about incentives, maybe it's not representative of the tech. But after about a hundred garbled Far Sides the excuses become tedious and unconvincing. Especially in such a simple and visible use case (cf. Google Search snippets)

My information sources are that my Facebook feed was, for a period of three months, saturated with emetic Far Side upscales. Sometimes it's about direct experience rather than theory.

(Preserved at www.reddit.com/r/FarSideCom...)

I know I posted this on Twitter ages ago, but I came across it again last night and thought some may not have seen it yet.
(From the Not the Nine O'Clock News spin-off book, 1980.)

Anyway, this has little to do with testing the training set, which is to say the conclusion that performance is far worse on datasets/packages that were not in the original velocity papers. This means the methods don't generalize.

You could say that. I would put it a different way— that the embeddings and neighborhoods are remarkably unstable and arbitrary (see Figs. 6, 7, 10 in our paper). By the way, did you happen to check calibration of the p-value distribution?

I have found the candles

Programming skill is just a set of rules you've inferred via personal suffering

To achieve good velocity embeddings, test on training data

iCloud has a fun Easter egg where it deletes all your books from the local device at random intervals so you can't use Apple Books when you get on a plane

Sure, I got that, and I completely agree. There are plenty of sources that make this exact point. This should be an encouragement to understand, characterize, and exploit the noise, rather than to elide it as ultimately deterministic, especially since it is in part controlled by the genome itself.

I would prefer to strongly push back on the assumption, because I do not see how it follows. And the end game could be either; statistical modeling helps both, and insisting on a deterministic formulation hurts both.

Those are the conceptual and practical issues. But there's no need to go that far. I have read enough papers that torture their data to fit deterministic models to realize this approach is harmful.

Not really! The goal is to analyze, and having a model likelihood, limited as it may be, helps more than having an intractable hairball of unknown and unknowable interactions. I also don't see why we ought to assume those processes are deterministic.

I would say it is not only parsimonious, it is practically mandatory as a starting point to identify signatures of unmodeled factors. You have to have some kind of null hypothesis or working model.

Because a seed is an initial condition for a specific deterministic process, and I don't really see how it helps define stochastic processes, even by analogy.

hard no.

Yes, because as soon as variability is treated as random, we can use the usual suite of statistical and probabilistic tools to quantify and evaluate that assumption.

In a manner of speaking. Because it makes using probabilistic methods and defining ergodic limits needlessly hard or impossible.

We could, but there are good reasons people do not.

I appreciate the aleatoric/epistemic uncertainty distinction as much as anyone. But to throw out biological noise is to throw out a century of really good tools for no benefit at all. (I also don't see why the quantum effects would disappear at the level of molecules.)

In my experience, treating universal and pervasive biochemical noise as somehow fake hurts more than it helps.

Intuitively yes, ranks are unstable, and one should hope that providing different data should yield different results! And yet I am not even sure about that (e.g. if the methods are ultimately relying on "uninteresting" robust signals, and throw out all other information)

By "optimistic" I mean there may be a sensible way to identify, resolve, explain, and account for biases from different technologies, but this may be both impossible (because of data access) and futile (because every analysis is bespoke) for downstream pipelines

A question I have often wondered about but have not yet managed to formulate in a quantitative way. By what measure?
I think the pipeline question is far too optimistic because source studies use a truly staggering variety of pipelines, references, and hyperparameters

Congratulations! Any speculations on whether activation in males uses similar mechanisms?

How do the people at Google and various other AI users feel about the page-specific search snippets being poorly AI-paraphrased, and what implications do they draw, or avoid drawing, from this fact

Hungarian writer László Krasznahorkai wins the Nobel Prize in literature Hungarian writer László Krasznahorkai has won the Nobel Prize in literature for what the Nobel committee called his compelling and visionary work.

Hungarian writer László Krasznahorkai, known for philosophical and bleakly funny novels, has won the Nobel Prize in literature.

Amazing! None of the answers are even in the same ballpark as the physics usage (neutral net+symbolic regression) or my usage (neutral net parametrizing some physical model)

"We're going to create superintelligence" How about making outlook search work first. How about that

The optimistic point of view, then 🙃