carbocation

> In recent years, transformer-based models have gained prominence in multivariate long-term time series forecasting

Prominence, yes. But are they generally better than non-deep learning models? My understanding was that this is not the case, but I don't follow this field closely.

Meta-comment: the prior discussions basically happened a little too early for the west coast folks.

As a physician I have the same expectations as you for those two words. Especially given that the link is to an anchor on the middle of the page, I was thinking in units of "OOM". Essentially I was thinking "OK an OOM is a unit for the doubling of RAM" (wrong).

Hypertension is not caused by the heart pumping too well, so this would not work for the purpose of curing hypertension. But if you could invent a durable heart replacement therapy, you'd save a lot of lives.

Nominative determinism strikes again.

Almost every noncommunicable disease of adulthood becomes more common with age. (The diagnoses in this paper were things like "cardiomegaly" which are mostly just X-ray findings and, while they have ICD codes, are not a meaningful diagnosis that a practicing physician would care about.)

> Or you might prefer a model that looks at the image only

These models are only looking at the images. They are inferring demographics.

I believe that imbuing what the model does to make a prediction with the idea of "shortcuts" is oversimplifying a more complex issue. I don't think it's helpful to describe a model that can distinguish demographics as taking "shortcuts". I think that adds a layer of jargon that we then need to cut through to understand what is going on.

There are plenty of tools that have been developed to enforce that models perform in a manner that is unbiased across some dimension (e.g., sex, or hospital, etc). (For example unsupervised domain adaptation.) I think that splitting the field with jargon makes it more difficult to follow the breadth of the field.

There is no reason to believe that an X-ray model's correct estimation of age is due to "spurious correlation". Rather, it seems to be "undesirable correlation".

It's not too hard to train yourself to identify sex or approximate age. (I claim this because of time spent reviewing model output for work that I have done to build models to estimate age.) The reason radiologists don't do this is that there is no clinical reason to do it, so it's not a skill that they develop. (Deep neural networks can also beat radiologists at bird classification, another task that is not relevant for their job.)

The authors refer to a literature describing "shortcuts" as "correlations that are present in the data but have no real clinical basis, for instance deep models using the hospital as a shortcut for disease prediction". It feels like a parallel language is developing. Most of us would call such a phenomenon "overfitting" or describe some specific issue with generalization. That example is not a shortcut in any normal sense of the word unless you are providing the hospital via some extra path.

They call demographics like age and sex "shortcuts" but I find this to be a frustrating term since it seems to obscure what's happening under the hood. (They cite many papers using the same word, so I'm not blaming them for this usage.) Men are typically larger; old bones do not look like young bones. There is plenty of biology involved in what they refer to as demographic shortcuts.

I think you could take the same results and say "Models are able to distinguish men from women. For our purposes, it's important that they cannot do this. Therefore, we did XYZ on these weakly labeled public databases." But perhaps that sounds less exciting.

For starters, I'd love to just see a rock-solid neural network replacement for screened poisson surface reconstruction. (I have seen MeshAnything and I don't think that's the end-game.)

The behavior should be the start and the end of the regulation. No need to specify AI, algorithms, etc, at all.

I'm focusing on a tiny technical detail here but from the description, it sounds like the ISOs weren't set up with an 83(b) election which is another bummer.

Very cool. I looked at the Bay Bridge and it gets the towers but probably not the bridge itself (this is a trivial point except for folks right by the bridge, but fun to look for edge cases).

Thank you.

Does this university press release name-check Cell but not link to the article? A bit frustrating.

News article about it: https://www.cbsnews.com/sanfrancisco/news/supreme-court-ruli...

Also on GCP, the vCPUs are usually hyperthreads (except for t2d- instance types, and perhaps a few others). So that machine you've described has 1 CPU core.

The substance of my comment to Sen Wiener hinges on a similar point to what you raise here.

It's incredibly difficult to imagine doing a good job of regulating model training, especially in a few years when the available flops are high enough that this limit is being hit often.

It's much more straightforward to regulate actions: constructing WMD is illegal, synthesizing drugs is illegal, etc.

If the state wants to tighten up its laws about various activities, go for it. That's the right place to act. Injecting itself into the model training process seems very unlikely to yield any substantive benefits and very likely to hinder progress.

I think this advice is incomplete. For those of us who live in California, shouldn't we be contacting our representatives?

None of the petty crimes he was wanted for seemed worth the effort he put into maintaining the false identity. It makes me wonder whether there is more to uncover.

I would like to see the new VoiceCraft model on that list eventually (weights released yesterday, discussion at [1]).

1 = https://news.ycombinator.com/item?id=39865340

I agree with you that EpiPen is far more widely known. But (in my opinion) the ambiguity in the title is not worth the name recognition.

The phrase "auto injector" would have been a much more helpful substitute for "EpiPen" here.

> Selatogrel is a P2Y12 receptor antagonist

So this is basically like chewing on ticag, but faster.[1]

1 = https://jamanetwork.com/journals/jamacardiology/fullarticle/...

(You replied to a meta-joke.)

Great point. Maybe you get clever with it and make it an amenity for your building and charge a lot for others who want to use it (etc).

The pro-land value tax response is that, if someone really believed people highly valued washing their cars in the city, one could open up a car wash on expensive land. If people really don't want to drive 30 minutes but also really want to wash their car, they can pay more to use this expensive in-town car wash. It is of course possible that the economics wouldn't make sense to support such a use of urban land in the densest/highest value areas.

My impression is that you don't need eminent domain takings for this, you just need higher level government (e.g., states) to set more rules around how restrictive local residential zoning is allowed to be.