Natalie, PhD

Can confirm from personal experience that glucose spikes are totally multifactorial.

A large part of the reason I debated with obgyn on gestational diabetes diagnosis and care was that drivers for me were stress and sleep deprivation.

They were like "you should take thiz 2-hour nutrition class."

Please can you do a similar study in gestational diabetes?

They still make us come in for the 1 hour glucose test, and if failed 3 hour glucose test. Specificity of the test is awful.

It's also a PITA and doesn't really reflect real world -- AT ALL.

Again, do I think this indicates models are 'reasoning'? No.

But I do wonder if they're 'pattern-matching' to human systems of thought and approaching problems with the same attempts to minimize attention and reduce cognitive load. Which would be interesting, but could also generate some problems.

Thinking, Fast and Slow - Wikipedia

But to me this looks like an attempt to shift from System II thinking to System I thinking, a la what Kahneman lays out in "Thinking, Fast and Slow."

en.wikipedia.org/wiki/Thinkin...

The author points out that these models are not 'trained' to write out every step of these problems manually. I agree here - users are often looking for quick answers, and the creators are looking to minimize energy costs. Asking for full solutions like the above is counter to this.

But the second post *DOES* bring up something really interesting - the author tried to get DeepSeek to solve a 10-disk Tower of Hanoi problem and found it resistant to doing so because it would be such a long chain of thought.

Instead it looked for shortcuts.

I don't think the second link (critiquing this paper) addresses these two issues very well.

I also disagree with them on the premise - I do think the logic puzzles used here (e.g., Tower of Hanoi, river crossing) do add a new dimension to test reasoning and do nicely scale in complexity.

The punchline for me is 'no', or at least 'not in the way we would consider humans to reason'.

This is because:
- LRMs often find the right answer early on but continue to iterate
- users providing the solution to the problem in the prompt did not improve success rate or time to success

Slowly getting around to massive tabs backlog and I know it's old news but read these two things on whether Large Reasoning Models - sort of extended versions of LLMs - can actually reason.

ml-site.cdn-apple.com/papers/the-i...
www.seangoedecke.com/illusion-of-...

#LLM #AI #benchmarking #logic

(Un)fun and horrifying fact: you can sunburn so badly that your skin literally blisters.

This summer, please don't be me and remember to sunblock ANY exposed part of your body when you're going to be out in direct sun for more than a few minutes.

Using our OncoSim toolkit—combining ML + mechanistic models across 1,000+ cancer cell lines—we’ve predicted 100+ novel synthetic lethal pairs. Some are validated and may guide therapy & resistance prediction.
Poster by Oliver Purcell @ GRC👇

www.deeporigin.com/academic-pos...

Japan has a whole industry prepared for this

For anyone wondering about the oranges, it's probably due to citrus greening disease.

Have made the bizarre discovery that Kinder Buenos don't cause a glucose spike for me, but bran flakes do.

Is this free license to eat as many kinder Bueno bars as I can while pregnant?

So if you're asking for yourself, look at the measures to avoid reinfection and decide wherher they are worth the risk reduction. I don't have a right to decide for you.

In the same way, one person does not have the right to decide that question for all of society. It's collective.

The exchange above is a good example. One person likely thinks closing physical schools and shifting to online learning, for that duration of time, was a mistake. The other likely thinks it's a mistake we have in person schooling now.

Who is 'correct'? This isn't easily answerable, if at all.

The hardest part about adjusting to COVID (IMO) is that everyone's risk tolerance is different, as is their willingness to take different measures to avoid reinfection. But we have to come to some sort of social agreement on what risk we will collectively tolerate and what measures we should take.

Yes, but I think the struggle for most people is recognizing there's a cost to doing that, and that cost varies.

Some measures are low cost: wash your hands, wear a mask in enclosed areas with a lot of human trafficking

Some are relatively high cost: closing schools, not traveling anywhere

It's possible there is data out there supporting this, I've only spent a couple of hours reading papers and haven't kept up with the COIVD literature since late 2021.

To be clear, I'd still recommend not getting COVID. But if say, I had to pick between getting COVID and measles, I'd pick COVID.

Ultimately science/medicine aims to build a diagnostic test that should predict, with some (ideally high) confidence, who has PASC.

for long COVID in kids (and probably adults!) we are...not near that.

These news outlets are using numbers prematurely IMO, without context.

Does that make the data invalid? No, it just means it's noisy and we need more research to get to an accurate number (and definition of PASC). Not satisfying but what comes now are prospective studies, ID of molecular mechanisms, etc.

Like, holy smokes, this is a very broad definition.

Particularly in a population that is:
1. Undergoing a lot of growth and bodily changes
2. Frequently getting exposed to new environments and illnesses
3. Cannot always self-report symptoms well (not never, but it is harder to get data)

This is particularly problematic in the context of the NIH definition of long COVID in kids:

"PASC refers to ongoing, relapsing, or new symptoms, or other health effects occurring after the acute phase of SARS-CoV-2 infection that is present 4 or more weeks after the acute infection."

There are a lot of reasons using this data is a stretch. One is that it's very fragmented in the U.S.

second, you often don't get much context in EHRs. It could look something like "COVID PCR positive" then 4 months later "respiratory distress". Could be long COVID. Could be they got RSV.

This led me down a rabbit hole of reading a few other papers from RECOVER...they may have multiple datasets but the main one I see for kids/pediatric is a massive electronic health record (EHR) dataset.

Which...is basically them trying to figure it out based on doctors records and diagnosis codes.

Why does this happen? Usually because the writer is having to multiply two numbers from different papers, or their own work plus another paper, to get a rough estimate of real world impact.

So this 5.8M number would be fine as inference if the 10-20% stat was referenced. But it's not.

(that random '6' halfway through the sentence is the reference)

The actual sentence: "Given that ∼20% of COVID cases in the United States are in children,6 and that current pediatric postacute sequelae of SARS CoV-2 (PASC) prevalence estimates are 10% to 20%, PASC is estimated to affect up to 5.8 million children, representing a significant community impact."

...but don't seem to add a citation for the second part of that sentence, where they state incidence of PASC (colloquially, 'long COVID in kids') is 10-20%.

First is this paper, which is cited in the above Forbes article: publications.aap.org/pediatrics/a...

The 5.8M number almost certainly originates here, but is only half referenced. The authors reference a paper to note that 20% of covid cases are in children...