LLMs are mostly composed of regular old fully connected ANNs, and the remainder, the transformers, are also ANNs. ANNs certainly can learn the rules of chess, and you can train one to play chess at a level that is generally regarded as superhuman. There's also a proof that any 2+ layer ANN of sufficient size can learn any IO function.

So there's nothing about the structure of an LLM that would make it unable to learn and follow the rules of chess. The fact that they don't, or don't do so very well, means that the way they are trained is an inefficient way to learn chess. Which shouldn't really be surprising to anybody who's read a book about something and then gone off a tried to do it with zero experience.

Have a citation for such an advertisement?

https://ancillary-proxy.atarimworker.io?url=https%3A%2F%2Fblog.samaltman.com%2Fthe...
"Humanity is close to building digital superintelligence"
"we have recently built systems that are smarter than people in many ways"
"In some big sense, ChatGPT is already more powerful than any human who has ever lived."
etc.

Amazon is not some random dude on the Internet. Big companies generally support their gift card programs and certainly would if they were a major payment method, because not doing so would compromise their actual business.

Adversarial noise isn't "noise" like static or random junk. It's specially crafted to make the model see things that aren't visible to humans, to alter their behavior.

Benn Jordan created a pretty good video about audio-specific implementation. Examples include perfectly normal sounding audio clips tricking digital assistants into thinking they're getting voice commands and having music completely misidentified. The practical application means an artist can apply adversarial noise to their work and have it sound perfectly normal to a human audience, but any generative model that tries to train on it will end up producing inappropriate and useless output.

There are also methods to do similar with images. Text may be a bit harder but it's still possible with websites through embedding or invisible text and similar tactics.
=Smidge=

They clearly did, if this is that much faster than dmg then they fucked up with dmg.

Nothing I wrote makes it even seem like I'm confused about that, except to people who are confused by mice with more than one button.

Yes, I read the fine summary. Some of us do that.

What was it about the implementation of dmg that made it so much slower, and why did apple think it was okay?

There are really two parts of a vaguely modern vehicle that are attacked by the results of using ethanol fuel, fuel pumps and the injection equipment whether that's a carburetor or a fuel injector. Those parts have steel bits, including jets or nozzles, and the ethanol draws water in from the atmosphere and then it evaporates. That leads to corrosion of these parts.

Ethanol is a potential problem for hoses and seals, but this is only usually an issue for much older vehicles and the fix is pretty easy, except where carburetors are involved. Then they need to be re-sealed, and if there's not a kit available, that requires making new seals on a laser cutter. And those are still moderately expensive.

The fuel isn't the problem. Basing ethanol production on topsoil is the problem.

Sandia NREL proved in the 1980s you can grow algae economically in open raceway ponds, and you don't even need to add algae. The air will do it for you, and the most efficient algae to produce at your latitude and in your local conditions will outcompete other strains so you will automatically get the most beneficial species for production in your location.

The focus at the time was for lipids for biodiesel production. But you could as easily produce ethanol. Or more intelligently, you would make butanol using the ABE process, which also produces some acetone and some ethanol. Octane can be adjusted by mixing the butanol and the acetone. The ethanol can be used as an industrial solvent, far away from fuel systems, where it draws in water.

This is closely related to what I said when we discussed this last time: When I went to college, I was required to apply for financial aid, which means doing the FAFSA. This pushes ID verification off onto the federal government.

They're just getting ready for the war between the USA'n'SRs vs Europe.

Some of us have always wanted smart glasses. But we also want them to not be crap. I have pretty limited requirements for the graphics capabilities, but it does include overlay. But they also need to be in basically the same form factor as ordinary glasses, and they have to not be under the control of someone who's going to piss me off all the time showing me a lot of sponsored fuckery, and any processing has to be done on a device on my person and not someone else's computer. And I really don't want to be around other people who are streaming video to teh cloud 24/7, either.

What we're going to get will be very different from that description for the foreseeable future.

It doesn't matter where the cluster is located, if you don't own it then you have the same issues with your data on someone else's systems. If you're going to load it into someone else's rolling DC, you might as well load it into someone's remote DC. Then nobody has to drive the DC around.

"Don't put China and North Korea into the same basket. Both claim to be communist, neither one is."

Don't say they are the same, that's my job!

This is my pet peeve. AI has been turned into a marketing term for things that are not the traditional definition of AI.

How so? What is the traditional definition of AI? Are you sure you're using the correct one?

ChatGPT is advertised as AI, approaching human level. AI is building machines that exhibit human behaviour and capabilities.

So they made the thing play a computer chess algorithm and it made excuses and demanded a rematch. Sounds like what most humans with no chess experience would do. It didn't flip the board and stomp off though.

I don't disagree. But 20% is a very, very low estimate.