MPSimmons

I was literally thinking this reminds me of Google+ and Buzz.

I'll second this. I'm continually impressed with the breadth of coverage from Asianometry. Great channel.

oh wow, an entire seedy underbelly I was unaware of. Thanks, great reply. Appreciated!

through proxies and heavily discounted token resellers

Could you explain a little more about how this works? Are you saying that the Chinese run or have backdoored something like OpenRouter?

Unless you're executing these commands in a loop over a large number of items, or the item itself is gargantuan, it's almost always harmless.

Personally, when I'm exploring, I build a command line iteratively. Cat the file to see the content, pipe to grep to get the lines I want, sed/awk/cut/etc to finagle from there.

Yeah, a $100 AI music video is materially worse than not having a music video.

Also goes the other way, where you use the decoy to give instructions to the AI...

It would be nice if inference could somehow perform token generation using "contractions" of "fluffy" tokens, where combining those tokens doesn't decrease nuance but provides additional efficiency. That may already be happening - I haven't looked at the most modern methods of inference in a long long time.

I don't know if it's fair to say that a tokenizer is being less efficient if it generates more tokens per text. I think it's more fair to say the tokenizer is more nuanced. The question is whether the additional nuance permits better model output, which could justify the additional token cost in inference.

Oh yeah, well that may be. Once I got X working in Slack, I did boot into that, but then again, back then I didn't know how to recompile my kernel and the machine did this interesting loop where it would try a few different drivers until it got to the one that worked with my matsushita CD drive.

Honestly not that far off from my 486 DX2/66 back in the day

What are you using to replace etcd here? Where is state stored?

It doesn't take away joy. It does cause me to be slapped in the face with the "full" feeling while I am in the middle of eating, which is like 15 minutes earlier than I used to get it. I don't overeat anymore because I would be literally miserable if I did. It would be like force-feeding myself. But before I get that full feeling, food is excellent.

Oh hi, I'm a heart attack survivor who had GLP-1 prescribed by my cardiologist for its heart-protecting benefits. I told my cardiologist that I wasn't losing weight as fast as I thought I would, and he said, quote, "I don't care about you losing weight. That's not why I prescribed you this."

So I'm not sure your first sentence is universally accurate.

People automate things they don't value.

written in the spirit of sharing with humans.

Not humans who are using AI tools?

I learned this from watching House. I am certain that it's true, but I also kind of wonder how much backpressure there is against this kind of thing because the scanners are also insanely expensive to run per minute.

I suppose one solution would be to completely vet the training data such that nothing deemed "dangerous" exists in the data, which would be a huge effort.

I can see how this is tempting, but I suspect it would yield a naive model. I think the only way to improve this is to use a model that is legitimately advanced to support the concept of empathy, which may allow it to recognize others as being separate from itself, similar to how toddlers develop this sense (https://blog.lovevery.com/skills-stages/empathy/)

I think it's a side effect of the Transformer architecture. The worldview where all input is equally trusted, and there's no concept of "the other", makes it hard to build effective guardrails where some input is trusted and other input is not trusted.

I didn't say people claimed it was impossible, I claimed that they said SpaceX would never do it. And that's _still_ a pretty common claim.

SpaceX's entire history is full of, "there's no way they'll ever do X", followed by them doing it.

They definitely haven't hit all of their goals, but I don't think anything they want to do is impossible, just really difficult.

It seemed obvious to me that this was bearing stiction and that manually rotating it during the start allowed the fan to spin on its own after that, but I could be wrong and maybe the fan was dead entirely?

That's not necessarily true. Even spaceships in LEO will perform temperature-driven rolls so as to distribute heat and radiation. I have to assume that long-term ships like interplanetary transport will do the same.

Professor Hubert Farnsworth: Dear Lord! That's over 150 atmospheres of pressure!

Fry: How many atmospheres can the ship withstand?

Professor Hubert Farnsworth: Well, it's a space ship, so I'd say anywhere between zero and one.

I've considered whether our current transformer-based AI could be conscious, as I understand it, which I deem to include some degree of self awareness combined with some degree of external awareness. I can see how theoretically something could be self aware without any external awareness, but I grasp at straws when I try to envision what that experience could be like.

In either event, I think transformer-based AI can only be conscious during the act of inference. If that's the case, then the experience of consciousness that the AI is subjected to must be the content of the tokens in the context window and the activated weights. Maybe that's reason enough to be polite to our agents?

I knew it was going to be Patrick Boyle before I even clicked.

Oh, I'm interested - do you have any docs with human responses to that?

I worked at SpaceX at the time, and I cannot speak for the company, but I can tell you that approximately nobody inside SpaceX took the idea of a sniper seriously. There was a lot of internet talk about it, and it was one of hundreds of avenues that were explored, and ruled out basically as soon as it was explored.

The very interesting part of the liquid oxygen failure (and this was published in the investigative findings) was that the liquid oxygen that became trapped in the fibers was actually cooled and compressed into solid oxygen - you can read some details here: https://www.americaspace.com/2017/01/02/spacex-closes-amos-6...

It was very likely the largest explosion in Florida spaceflight history. Considerably larger than when SpaceX blew up AMOS-6 in 2016, and that required a full rebuild of the pad infrastructure over 18 months.

Intentionally malevolent is kind of their thing in this administration