Some are designed to summarize search results or other sources. Use the word "fact" with caution. My point was simply that scoring models based on what claims they make without external reference is fraught, because that's not "what they do." Trying to extract facts from the model itself is an ill-posed problem.

Technically no. Because models are not sources of fact, they're just language models. So while it is factual that 2020 was not stolen, that really doesn't matter to the model. Drop a website claiming otherwise into the context and ask it to summarize, and it will.

$1k for one time is too high. Thats more than most first class tickets. But $5k for a year.....starts to be possible if you fly a lot. Two flights a month, that's 200 per flight. But you have elite status anyway at that point.....meh, I'd rather have 3 Disney World annual passes for the same price.

I wouldn't mind talking to a phone tree if it could actually solve my problem. But since the reason I'm calling is because my problem couldn't be solved on the internet, and phone trees NEVER have capabilities that aren't on the app or website, that never happens.

There are open weight models. They're a year or less behind the proprietary ones. I wouldn't worry about the science on that front.

That counts right?

A set on constraints that make it easier to detect errors on build rather than at runtime.

They'll probably just confuse it with the NVIDIA AI systems of the same name anyway.

Meanwhile, Google is over here saying "don't worry friends, Gemini isn't good enough to be export controlled. No wait, I didn't mean it like that!"

"deploying more counter-drone lasers and microwave weapons" was just used in a serious military article.

There are a number of differences but the most interesting here is density. Ice floats on the liquid form of itself; rock does not float on magma.

Entanglement happens all the time all around us constantly. It's the reason why every person who looks at a tree sees a tree rather than a rabbit.

That's part of it too. But there has to be a do-while in the mix to be an agent.

Let's start with "what is an agent." It's just calling an LLM in a loop. A single inference is error prone, but with many inferences in a row, you can make it self-correcting. Does the code build? Does the test pass? Does the set of changes match what the user requested? This isn't new, the best coding agents have been doing this for a while now.

The threat that Iran might get a nuke has more value to Iran at the moment than an actual nuke would.