iskander

Lots of practical but low impact suggestions in this thread.

I think the only real answers involve large scale decoupling from the rapidly changing social norms. Can you cobble together a social group that will go Luddite with you? If not, can you join one?

I think unpopular to mention here but John Ioannidis did a really weird turn in his career and published some atrociously non-rigorous Covid research that falls squarely in the cross-hairs of "why...research findings are false".

>we don't try to attack the actual cause

It's cars and large houses peripherally connected to amenities by car-only infrastructure.

People love this lifestyle and will fight you very energetically if you try to do anything to nudge city layouts towards the previous level of walkability.

What is Invidious? An alternative front-end to Google-hosted video content?

#1 on ASPCA's list of common toxicoses in cats: https://www.aspcapro.org/sites/default/files/zl-vetm0606_339...

Study of 286 cat/permethrin cases in London: https://journals.sagepub.com/doi/full/10.1016/j.jfms.2007.05...

Study of 750 cases in Australia: https://journals.sagepub.com/doi/full/10.1016/j.jfms.2009.12...

...we don't track this stuff nearly as well for pets as we do for humans but seems to have pretty robust evidence of toxicity.

Related: Verifiable Delay Function

https://link.springer.com/article/10.1007/s00145-020-09364-x

Only clue in the source:

<!-- this blog is proudly generated by, like, GNU make -->

I love the simple but elegant formatting of this blog.

cgadski: what did you use to make it?

Exceptions so far are Novo Nordisk and CostCo. Not sure if there are many others at scale.

Not that anyone is offering $192B but I also suspect existing doctors and the AMA might actively oppose this. Debt-free MDs would probably create downward pressure on salaries.

My dog seems to know least the names of 6 family members, 4 friends, and names of ~6-7 other dogs (in that she can go to those individuals on command). She also knows: cat, dog, cow, horse, friend, hello, "dai lapu" (Russian for give me your paw), "sidi" (Russian for sit), sneak, "bang" (for playing dead), dinner, breakfast, bath, outside, "go potties", and probably quite a few other snippets of English.

She also knows how to open windows in a car and looks for the buttons before pressing them with her paw.

Moved to the US as a kid (by way of Estonia and then Italy), went to a year of Russian language "transitional" school here. So don't have any insight on the Ukrainian school experience.

I was born in Kiev and spoke Russian at home. Can barely understand Ukrainian unless it's spoken slowly by a native Russian speaker. I can get the gist of what Zelensky is saying in an interview but can pretty much never understand native Ukrainian speakers. I think there's also a gradient of dialects and accents West to East, so I'm sure you can find some Ukrainian villager I would understand better but in general they're not mutually intelligible (to me).

I expected the core of the algorithm to be a parallel prefix scan though (isn't that the point of Mamba?):

    for i in range(l):
            x = deltaA[:, :, i] \* x + deltaB_u[:, :, i]
            y = einsum(x, C[:, i, :], 'b d_in n , b n -> b d_in')
            ys.append(y)

Keras started out great but grew clumsily, making custom use-cases cumbersome and error prone and also became inextricably linked with TensorFlow (which was a big headache compared with PyTorch).

It seems like this rewrite cleans up the meandering mess of Keras's middle period, or at least I'd be willing to give it a try.

Like, huge tangible improvements.

You can do general purpose programming in all kinds of DSLs and Rust libraries that compiles down to big-but-practical arithmetic circuits and then generate proofs that y=f(x) without revealing x for arbitrary f. That really wasn't possible until a few years ago and emerged almost exclusively within the sphere of cryptocurrency-adjacent research. You can also use the succinctness of SNARKs to batch these proofs and shrink the verifier costs to almost nothing

It's a cool model for asymmetric computing, with low capacity verifiers collecting results from high capacity provers. It'll probably find uses outside deranged gambling...

Can you still!?

It's so incredibly traceable these days and every iteration of Silk Road gets busted, no?

Merkle trees, yes! That was just an example of a technique that can be well hidden vs in the user's face.

Newer techniques like Verkle trees, SNARKs, STARKs, &c are.... well, new.

I think the initial impractical prototypes for the core techniques were in the literature for a while but there's really no comparison between eg interactive ZK protocols and the succinct non-interactive proofs used by the cryptocurrency folks. The latter are computationally general (you don't have to roll a new one for each program) and many orders of magnitude more efficient.

I think it's important to give the cryptocurrency industry credit for the few corners in which it makes real (theory/tooling) contributions even if they're motivated by nonsense

It would be a shame if this keeps happening, like: (2013) Bitcoin and its clones failed at all their stated purposes, (2017) the ICO craze left nothing of value, (2021) the DeFi/NFT bubble also left nothing but rubble. This stuff is zero sum, shuffling money without creating new things of enduring value.

I think you'll see people picking those techniques and using them in a hidden-from-the-user way (like how git and BlueSky use Merkle trees)

I gave the "industry" a pretty thorough deep dive; I learned Solidity, wrote a bunch of smart contracts, tried a lot of DeFi, and listened to ~100 episodes of the Zero Knowledge podcast.

My impression on the other end of it is that the "tech" is interesting in a purely academic sense but most of the participants are building things they know have no use and engaging in a lot of motivated reasoning about the future to justify drawing money in from the outside.

Like, it can all be a lot of fun! And some techniques, like all ZK folding schemes, are mathematical magic. But, no one wants code to be law and pretty much no one should be managing private keys tied to money.

It's a very cool but failed idea.

Scott's political ontology seems intrinsically limited to white middle class. It doesn't really make sense in poorer parts of America and really falls apart once you start trying to correlate belief systems in non-white communities.

...what happens in the next 50 years?

I'm on a laptop and it looks totally fine to me.

It's pretty readable at their large font size

They're free but not accessible, it's a universe away from compassionate use access wrt practical availability.

>This is how it already works - this is what a clinical trial is.

That's not at all how clinical trials work. If and when someone decides to advance a drug to the next stage of development, it's done because they think there's a sufficiently profitable market at the end of the road. So lots of extremely promising drugs stall out if the numbers don't make sense. Furthermore, the trial enrollment process itself is cumbersome and often your only entrypoint is email "trials@pharmacompany.com" and hope you can a response and can navigate their arbitrary screening process. Speaking of screening: trials are full of exclusion criteria. They can get copy-pasted between trial protocols and often have no real reason excluding you. What's the washout period for your last line of chemo? Ever had any kind of immunotherapy? &c &c From the point of view of whoever is trying to get a drug approved it's better to be safe than sorry wrt any kind of uncertainty in a patient's medical status or treatment history. But that means that there might be a drug that looks extremely promising from previous trials but the only current trial excludes you for a small reason.

Lots of cross-hatch patterns!

>My point of view is that anyone who wants to write multithreaded code, shouldn't be trusted to. Making it easier for people to justify this kind of footgun is a problem.

Out of curiosity, have you done any Rust programming and used Rayon?

It's hard to convey how easy and impactful multi-threading can be if properly enclosed in a safe abstraction.