mjewkes

> However, a proof of the Riemann hypothesis, say, may need new ideas that are strictly outside of the convex hull of current mathematical ideas

[...]

> After I finished writing this blog post, and had already sent it to Terry, OpenAI released a huge treasure trove of results in mathematics [including] the resolution of the so-called quasi Riemann Hypothesis

We ought to all be careful about underestimating the speed and magnitude of the change that is coming.

> if you are a student who is passionate to learn what is new and what is left to do, then a PhD is definitely the right path for you

This is an awful lot of confidence to put behind career advice in a wildly changing world. Markets are real and tradeoffs bite. We're not in gay communist space utopia yet.

>an awful lot of modern professional life consists of producing artifacts whose primary audience is other people producing artifacts. Slide decks for slide decks, strategy documents about strategy documents

This is because thinking, communication, and collaboration are extremely valuable.

Sure, they can build it here—but protectionism doesn’t deliver long-term success.

In the short term, it drives up prices and makes consumers poorer. In the long run, tariffs shield companies from global competition and limit their ability to scale. Without that pressure or opportunity, businesses grow weaker—not stronger.

That weakens U.S. industry, not strengthens it.

How do you get 55%? Wikipedia says 49.8%.

https://en.wikipedia.org/wiki/2024_United_States_presidentia...

CAR-T may have a ~50% cure rate for blood cancers: https://ashpublications.org/blood/article/141/19/2307/494672...

I agree completely with the first part. But SWA-1380 was a commercial operating fatality in 2018. Not a crash into terrain, but the engine definitely crashed into the fuselage.

Centrifugal forces on the spinning object will tear it apart once it goes “really fast”

The FDA approves new modalities that do things “in their own way” all of the time!

Yescarta, Kymriah, the mRNA vaccines, the approved VHH, scVF and “mimetics” in the article. Here’s a whole page of cell/gene therapies, many of these were produced “in their own way” or in ways that didn’t exist at all 10 years ago. https://www.fda.gov/vaccines-blood-biologics/cellular-gene-t...

Warp speed did accelerate review of COVID vaccines, but it didn’t change trial design nor manufacturing requirements - I’m not sure that supports your argument. I have seen the FDA be quite receptive to innovative trial design - baysean protocols, conditional endpoints, genetic diagnostics in inclusion criteria, etc

If you think there is a whole wave of innovation being held back by the FDA - can you point to examples of drugs it rejected or innovations happening in any other country that should be happening here?

Lots of stem cell “medicine” is happening in Korea where the requirements on cellular medicine are more lax… none of that is translating to effective medicine here afaik.

As you say, biology is hard and messy! And it is the job of the FDA to prevent five 20 year olds in a garage from injecting desperate cancer patients with some hacked together MVP. We’re agreed on that.

But I just don’t see any evidence of a giant pool of untapped innovation at the preclinical level that the FDA is blocking.

>I can say this with pretty strong confidence: had the FDA not being so strict about its trials, lots of these experimental type of antibodies would have already received massived research and commercialization

If the FDA was not so strict in requiring evidence, medicine would look like the supplement aisle, and the incentive to actually do the stupidly hard work of developing effective drugs & proving their efficacy would disappear.

Why even bother investing in developing a complex new antibody if you can sell some mushroom powder from a beach and claim whatever you want about it, ala Tim Ferris & "BrainQUICKEN".

There are plently of valid "incremental" critiques about the FDA. IMO, their devices and diagnostics divisions are weaker than CDER and CBER. Even those two (drug focused) divisions can be too slow moving and occasionally capricious about when they want more evidence. But their final decisions are generally pretty good, with a few exceptions. And most of those exceptions are drugs that they approved that they shouldn't have (e.g. Aduhelm) - not vice versa.

Can you point at a promising drug candidate with statistically solid efficacy data that CDER or CBER has rejected?

Also - note that there are approved scFVs and VHHs and mimetics - as stated in the article. Each is useful in particular situations, but none is some sort of wild medical breakthrough. There is just no evidence that any of these modalities would be attracting extraordinary investment "if only it weren't for the FDA".

If you are truly netflix-to-be, you should be able to affordably incentivize your 12 key employees with equity.

An exit for Employee < 50 at a netflix-to be will be in the hundreds of millions.

Apples are better than bananas. Cherries are worse than apples. Are cherries better than bananas?

--

GPT-4 - wrong

gpt2-chatbot - wrong

Claude 3 Opus - correct

Notably, if you ask it to transform the statements to formal logic, you get a correct response! This stuff is truly magic.

https://chat.openai.com/share/81e45fef-a72b-4258-98d6-5c8190...

For sure. It's not a fair prompt at all. I'm super bullish on LLMs and am using GPT-4 in production right now. This stuff is magic.

It's actually hard to find short, simple, "plain english" failure cases like the above.

The "chain of reasoning" that the modern models deploy before the fail is funny too. This is GPT-4:

---

To determine the relationship between cherries and bananas based on your statements, let's break it down:

  1. Apples are better than bananas.
  2. Cherries are worse than apples.

From statement 1, we know apples rank higher than bananas. Statement 2 tells us cherries rank lower than apples. By this logic, since cherries are lower than apples, which are higher than bananas, it follows that cherries are also lower than bananas.

Therefore, based on these comparisons, cherries are not better than bananas.

My current favorite “LLM breaker” below. GPT4, Claude, and this all fail.

—-

Apples are better than bananas. Cherries are worse than apples. Are cherries better than bananas?

Editorial write-up: https://archive.ph/0T2XI

Worth checking out the talk page. Looks like this whole article is out of date and that much of the problem is resolved.

This seems credible: https://ascopubs.org/doi/full/10.1200/EDBK_288015

As fabiancook mentions, credible sources all end up on pubmed within a year or two of publication, so it's an excellent resource.

That said, Pubmed is an aggregator, not a curator. So you'll find a fair amount of crap there too. Also, there are ~600 results per year related to ovarian cancer, which is likely too much to meaningfully sift through.

The things that come to mind for me are:

- Try to get second, third, fourth opinions from well regarded oncologists at well regarded hospitals, ideally who specialize in ovarian cancer. There is a pretty big gap, especially in cancer treatment, between those who keep up with the latest recommendations, and those who use "still ok" recommendations from 5-10-20-30 years ago

- ASCO is the major scientific conference for oncologists. In terms of understanding what is cutting edge, looking through their material is useful. I would suggest reading what has been presented at ASCO and then taking that to your doctors / second opinions.

- If there are specific biomarkers associated with her cancer, understand those, and make sure you understand if there are latest recommendations specific to those biomarkers. These are the types of specific case information that a generalist oncologist might miss (e.g. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7998656/)

- Another place you might be able to complement / compensate your oncologist's care is understanding options for "cancer support" (chemotherapy support as fabian says). Basically, the faster you can recover between treatment sessions, the more drug you can be given faster, and the more cancer gets killed. Oncologists are often (but not always) more up to date with the latest chemo/immuno/surgery/radiological interventions than with the options for chemo support.

https://www.rand.org/pubs/research_briefs/RB9412.html and The Great American Drug Deal, by Peter Kolchinsky. Specific claims:

- The US pays more for new drugs than the rest of the world

- The US contribution is disproportionate, and is major driver of the drug development industry.

-If the US paid less for new drugs, the entire industry would slow, and we would have fewer more effective treatments.

No assumptions. The rand article addresses that directly.

Hey biorach - I would suggest: https://www.rand.org/pubs/research_briefs/RB9412.html and The Great American Drug Deal, by Peter Kolchinsky.

Specific claims:

- The US pays more for new drugs than the rest of the world

- The US contribution is disproportionate, and is major driver of the drug development industry.

-If the US paid less for new drugs, the entire industry would slow, and we would have fewer more effective treatments.

...

Routine medical procedures being crazy expensive is unrelated and pathological.

The claim is largely true for drug development, but not for routine procedures. Those are expensive for unrelated, and largely pathological, reasons.

I would suggest: https://www.rand.org/pubs/research_briefs/RB9412.html or The Great American Drug Deal, by Peter Kolchinsky.

The claim is largely true for drug development, but not for routine procedures. Those are expensive for unrelated, and largely pathological, reasons.

I work in the drug development industry, its a shame this has been downvoted, it's largely true - at least for new drug innovation. The fact that routine procedures are expensive in the US is unrelated.

The US is ~35% of the global pharmaceutical industry, which is the topline driver of new drug development.

When the US pays top dollar for new drugs, it does effectively subsidize European access to the new research.

>Do you have any data to backup your claim?

I would suggest https://www.rand.org/pubs/research_briefs/RB9412.html

or the Great American Drug Deal, by Peter Kolchinsky.

>Do you have any data to backup your claim?

I would suggest The Great American Drug Deal, by Peter Kolchinsky.

Also - https://www.rand.org/pubs/research_briefs/RB9412.html

>For example Cuban healthcare/treatment/outcome is excellent based on its own research

Cuba does not contribute substantially to global pharmaceutical research.

>Roche(Swiss) nor Merck ( German, not related to US Merck) don't benefit much (if at all) from US subsidies.

Roche makes 55% of its revenue from US drug sales (https://assets.cwp.roche.com/f/126832/x/db9d31e8a7/fb20e.pdf p22)

EMD Serono (formerly Merck KgAA) is similar.

These revenues are the topline driver of the drug development industry.

Does anyone here recommend a modern-ish guide to GRF/settings/Gamescripts?

It seems like there are so many features like CargoDist/FIRS/City growth - many of which are "essential", but don't play well together, or have questionable balance.

Exact phrase matching is what google (sometimes? used to?) do for you if you put your search terms in double "full quotes".

It returns only results that contain the exact multi word sequence in exactly the same order.

ILIKE itself is a linear scan. The only way to index them are trigram indicies, which are very inefficient (and sometimes not usable) if you're searching document-length content.

Whether or not it works in your specific situation depends on your use case.

Try https://simonwillison.net/search/?q=%22your+own+benchmarks%2...

Looks like we get 37 results, of which 2 are true positives.

Looks like "your" and "own" are both contained in the english.stop stopwords list. So you could fix this by removing stopwords from your dictionary.

While disabling the stemmer is relatively easy (use the 'simple' language setting for your ts_query), altering the stopword dictionaries is more involved, and not easy to maintain or pass between developers/environments, and not at all easy to share between queries.

And so the most common suggestion is to use ILIKE.

Lucene has no problems with any of this.

Exact phrase matching. This generally requires falling back to ILIKE, which is not performant.