They spent a lot of time and money making sure CUDA worked right. For a while AMD's compute API wasn't backwards *or* forwards compatible. You had to do some rewriting and a recompile every time a new API was released.

Intel has gone through three completely different, and mostly incompatible, hardware stacks. Remember Phi? Altera? Now it's AVX for some compute tasks, and Xe for other tasks.

My very first PC build was a 486 built out of pulled parts from the computer lab where I worked. My first real DIY build was a K62-300 with a Matrox G200 video card. Ran every bit as good as my friend's P2 at half the price.

It's all statistical connections between words. It's not a conceptual model. There is no understanding of morality, ethics, or basic reason or logic. The only way to fix it is to bolt stuff on after training.

Also, you need an enormous dataset to get enough useful weighting for the model to work. For example, they didn't use chat logs because they wanted to, but because they needed the training data to get the models to function. They are still looking for more. You could prune back sources, but the models will perform worse.

Came here to say this. There is a *lot* of garbage in podcast land. The handful I listen to have been around for over ten years. I haven't heard a new one interesting enough to subscribe to in, probably, five or six years.

Futures generations will look back at these times and how "simple" and "awesome" they were. You're part of something bigger and you have a front row seat.

Coding is not development. AI can code. It cannot develop. For more information go back to the developers using AI to write a C compiler in Rust. It took a significant amount of developer effort. They didn't do a lot of coding beyond unit tests, but there was a whole lot of manual development work going on.

This is an incredible hack. And I just resurrected my father-in-law's Compaq Armada P3 laptop and it's looking for something to do...

Lack of hardware virtualization shouldn't be a problem as Windows 9x lets you do almost anything. Heck you can change addressing modes, which is verboten in WinNT+.

Compare the code, if it's similar then Claude is relying on stuff it's been trained on. If not, it's generating novel code that does the same thing.

They were government-required audit logs, tied to user accounts (21 CFR Part 11) The underlying feature is un-ban-able. We asked the companies for them so we could figure out the adoption rates of new features. Most companies didn't care. We didn't do anything else with the data.

The company I used to work for data-mined telemetry to find out what product features people actually use. Would that be banned?

Also, a data-mining ban from the government will never come, as they want that data, too.

I've seen videos of Amazon drivers chucking packages from the sidewalk onto the porch. The arc could have very well been 10'.

The auctions are in the Futures and Options trading. Beans, gasoline, everything is in there.

My wife loves Allbirds shoes. She has knee issues and they are the most comfortable walking shoes she has found. AEG is the holding company where brands go to die. Time to stock up before they start churning out expensive garbage.

You're overthinking things. You don't care about the riff-raff who hang out in the Lil' Ale'inn. You make it seem like it's easy to get a janitorial job at the facility and monitor the heck out of those who apply. Also any government workers whom seem over-eager to get their hands on documentation coming out of the site, which is all, also, disinformation, so who cares if any other country gets it.

"UFO" balloons aren't to monitor aircraft. They are to test air defense radar. How good an enemy's radar is at picking up specific signatures, say roughly cruise-missile-like, is much more useful information than aircraft capability.

I'm confusing what you want and don't want to be upwind/downwind of :)