Especially when it's on purpose.

You have to decode feel-good words into the concrete policy. The EAs believe that the state should prohibit entities not aligned with their philosophy to develop AIs beyond a certain power level.

All the more reason somebody else will.

Thank God for capitalism.

If you want to axiomize privacy, you can: that's a coherent philosophical position: but it's one I find curious. You're arguing that privacy breaches are harmful not because they cause harm, but because they are harm. Why is privacy, not progress, the summum bonum?

Even if we were to grant that these alleged privacy disasters causes harm, we'd have to balance them against the lost advantages of refusal to deploy the enabling technologies. It's like banning telephones on the account of everything crime anyone's ever organized over a phone call.

The pretext for these bans is always that unassailable cluster of feel-good yet vague virtues like privacy or the environment that you can make mean anything you want, but the reality on the Continent is just a rotating series of excuses for the catechism of "no, non, nein".

And it's never enough to just regulate the EU. Oh, no. The EU is the world's moral guardian, a "regulatory superpower", humanity's conscience. Obviously EU regulation should apply worldwide. The rest of humanity can't be trusted to care about privacy and the environment enough, right?

Well, I'm sick of it. How about they start saying ja to something? How about they walk about HOW we incorporate fledging technological capabilities into society instead of trying to freeze our information environment in 2008 amber?

At this point, when thinking about how we deploy new technology, I'm inclined to just leave Europe behind. Seal it off from the world of innovation with firewall rules and geofencing. The alternative is to suspend technology, the only thing that's ever in all history improved the human condition, for the sake of small-minded, small-hearted people who like mankind less than they love nein.

> This is probably due to the way larger numbers are tokenised, as big numbers can be split up into arbitrary forms. Take the integer 123456789. A BPE tokenizer (e.g., GPT-style) might split it like: ‘123’ ‘456’ ‘789’ or: ‘12’ ‘345’ ‘67’ ‘89’

One of the craziest LLM hacks that doesn't get love is https://polymathic-ai.org/blog/xval/

xVal basically says "tokenizing numbers is hard: what if instead of outputting tokens that combine to represent numbers, we just output the numbers themselves, right there in the output embedding?"

It works! Imagine you're discussing math with someone. Instead of saying "x is twenty five, which is large" in words, you'd say "x is", then switch to making a whistling noise in which the pitch of your whistle, in its position within your output frequency range, communicated the concept of 25.00 +/- epsilon. Then you'd resume speech and say "which is large".

I think the sentiment is that today's models are big and well-trained enough that receiving and delivering quantities as tokens representing numbers doesn't hurt capabilities much, but I'm still fascinated by xVal's much more elegant approach.

There's a lot of work going on in various streams towards making it possible to vary compute per-token, dynamically, e.g. universal transformers. Maybe one day it'll work well enough to beat conventional techniques.

It's possible to define your own string_view workalike that has a c_str() and binds to whatever is stringlike can has a c_str. It's a few hundred lines of code. You don't have to live with the double indirection.

It would behoove people to engage with the substance of attestation proposals. It's lazy to state that any verification scheme whatsoever is equivalent to a panopticon, dystopia as thought-terminating cliche.

We really do have the technology now to attest biographical details in such a way that whoever attests to a fact about you can't learn the use to which you put that attestation and in such a way that the person who verifies your attestation can see it's genuine without learning anything about you except that one bit of information you disclose.

And no, such a ZK scheme does not turn instantly into some megacorp extracting monopoly rents from some kind of internet participation toll booth. Why would this outcome be inevitable? We have plenty of examples of fair and open ecosystems. It's just lazy to assert right out of the gate that any attestation scheme is going to be captured.

So, please, can we stop matching every scheme whatsoever for verifying facts as actors as the East German villain in a cold war movie? We're talking about something totally different.

They'd switched away from the primary thrusters in 2004 due to this degradation. Now the backups are so degraded that the primary thrusters are better again in comparison.

Thruster clogging will kill Voyagers in about five years if nothing else gets them first. The least degraded thrusters nozzles are down to 2% of their diameter --- 0.035mm of free-flow area remaining.

The Voyagers will probably celebrate their 50th anniversary, but not much beyond that. :-(

Kind of ignominious to be done in not by the inexorable decline of radioactivity but by an everyday materials science error of the sort we make on earth all the time. In the 1970s, we knew how to make hydrazine-compatible rubber. We just didn't use it for the Voyagers.

[1] https://code.kryo.se/iodine/