carbocation

I'm perplexed by this university press release. Do they really not link to the underlying research article that they're discussing?

The front page of Reuters right now is a story about a major presidential proposal[1]. I think that is certainly headline news in the traditional sense. Still, it would be nice if there were additionally a news wire that didn't cover statements, only events.

1 = (Not describing the content because that's not the point.)

(Update: it’s visible for me now.)

I don't even see it in the US right now.

For segmentation? Sure. For their role as a clinical consultant? Not yet.

A few responses:

- Since when have the radiologists ever been worse than AI at reading images?

- AI is doing mechanical tasks for the radiologists, so it stands to reason that this makes radiologist more efficient.

- The radiologist is a liability sponge, so if at all possible it of course makes sense to augment the radiologist with AI rather than to try to do away with them. (This roughly gets at your point about the system being unchangeable.)

Offer a Turing award to the bot-trainer?

I remember when Hinton said that that we should "stop training radiologists now" in 2016[1]. Meanwhile, radiologists are in high demand and are getting paid better than ever. I believe the same will be true for programmers in the future. Sure, some of the boilerplate will be handled for you, just like segmentation is for radiologists. That's great for everyone.

1 = https://www.youtube.com/watch?v=2HMPRXstSvQ&t=30s

My kingdom for renaming this paper to something like "Tensor Product Attention is a Memory-Efficient Approach for Long-Sequence Language Modeling"

I turned these summaries off, but then turned them back on because I find them humorous.

A bit notable: the AI summarization of spam texts makes them seem much more credible.

> For non-commercial use

What the hell

The gemini result is great. I modified your prompt to encourage more detail ("Generate an SVG of a pelican riding a bicycle. The degree of detail should be surprisingly high and should spark delight for the viewer.")

This is what o1-pro yielded: https://gist.github.com/carbocation/8d780ad4c3312693ca9a43c6...

I think you misunderstand me. I’m not telling them to give up or expressing hope that they don’t succeed. I’m identifying what I see as a major barrier to adoption. I’d be interested in a response that addresses those points.

FZROX gives me 0% fees, can be bought in my retirement accounts, and is attached to a company with something like $1 trillion AUM. The latter gives me faith that it will still be around next year. I appreciate that the 0% fee options are limited, but personally I’d rather deal with 0.03% fees than entrust my money to a small shop. Especially when the reason to do so is not some trading edge, but saving a small amount on fees. I think this is going to be the main barrier to getting people to sign up.

There is no assertion that those individuals were randomized, simply "treated".

According to the final paragraph of the article, Nicholas Nimeskern remains employed in law enforcement.

An uncited statement in a review article!

If I'm rich in time and money enough to routinely take the time to wander down to my friendly beekeeper and buy their small-batch artisanal honey, for so many reasons I will live longer than someone who isn't.

Someone should randomize people to honey vs fake-honey-sugar-mix and settle the issue.

Seems like something that would be worth pinging OpenAI about because it's a pretty important claim that they are making on their pricing page! Unless it's a matter of counting tokens differently.

According to the pricing page, 32K context is for Plus users and 128K context is for Pro users. Not disagreeing with you, just adding context for readers that while you are explaining that the 4o API has 128K window, the 4o ChatGPT agent appears to have varying context depending on account type.

Pro does have longer context windows, specifically 128k. Take a look at the pricing page for this information: https://openai.com/chatgpt/pricing/

If this is also available via API, then I could easily see myself keeping the $20/mo pro and supplementing with API-based calls to the $200/mo Pro model as needed.

Is this just an announcement of a grant renewal?

One thing that I noticed from watching the first ~minute of the video: rather than simply pressing on the skin, it looks like the finger is mostly pulling the skin down over the edge. My intuition is that this is because there isn't all that much deformation from simply pressing, and so this pulling action triggers a stronger response. But I might be overinterpreting from a few seconds of video.

I feel like this comment was made in good faith so I wanted to explain my downvote: I think it's just too far off-topic for this particular announcement. It's better as a separate discussion.

All of Us is great for disease labels and for genetic validation (and discovery). I like it a lot. I just have... ideas for how they could make it more valuable. Ideas that I stole from Rory Collins. https://www.ukbiobank.ac.uk/media/gnkeyh2q/study-rationale.p...

Are you thinking of the Million Veteran Program? All of Us is readily accessible to most US institutions and is rolling out a plan to expand access.

The imaging reports are being imported, but not the images.

The bigger issue is that clinical imaging is done due to medical indications. Medical indications dramatically confound the results. A key element of the UK Biobank's value is that imaging and ECG are being done without any clinical indication.

Sadly the All of Us program (of which I am a research subject [and researcher]) hasn't done any sort of imaging or ECG. They did draw blood, so that allowed for genome sequencing. That may also, in principle, allow for assaying new biomarkers (I don't believe anything like that has been funded though).

Huge fan of the SAM loss function. Thanks for making this.

The only thing I'm 'shocked' by is the idea that anyone needs to pay to access my academic writing for model training. I would hope that using my academic writing to train models would be considered fair use.