No, they have to stay one step ahead. It'll be x64.

Just curious...

Training adjusts the weights. It doesn't adjust the network structure, except to the extent that some weights go to zero.

Great Scott!

A while back someone figured out that neural nets needed history, and RNNs (recursive neural networks) were born. The current "time" in the network referenced previous "times." The problem was, RNNs kind of sucked.

The attention mechanism is a fancier version of the same idea. Instead of referring to a few previous times, you refer to ALL previous times with context-dependent weights.

Transformers were the end state of replacing recursive nets with attention in a much of different places.

You need to understand that neural networks have structure. A network to recognize images is fundamentally different from a network to recognize voice. There are "LEGO blocks" of structure they use to architect these things. Transformers, long short-term memory (LSTM), etc.

The real advances come from identifying new LEGO blocks to use.

A big company will hire a CEO named Dat Ash. I look forward to the headlines.

If you ask Google Maps for a route, it will find one pretty efficiently. But if the mountain pass that is the fastest way is closed, and the next fastest way is 3 hours longer, it's going to do a LOT of searching across the graph before it finds that. These hacks are no different. The agent can't do the obvious thing so it starts trying everything to find a way through. It's still just state space search in the end.

Everyone has figured out by now how to manipulate Trump via flattery.

Iran doesn't agree to terms they don't like. The fact that they agreed to the JCPOA and were again willing to put the removal of their enriched material on the table in February shows pretty clearly that nukes are not their primary objective. No....they want the monetary concessions that the THREAT they might get nukes gets them. Money is the source of Iran's power and influence.

And if the MOU had held up, they would've gotten it. There was a lot of money in there. A LOT. They even had Rubio start spinning it as "unfreezing isn't the same as paying," remember? But then that fell apart.

Let's take that as an AI.

"You're a rabbit, Mal!" -Firefly, according to the captions.

Try asking it to make a plan before it starts coding. You can iterate on things like "use this annotation" there where it's easy to fix. If it still does something unexpected, ask it why. It will either point out something you didn't think of, or decide on its own it's being stupid, or at least point out what consideration is driving that so you can tell it not to worry about that thing.

Every now and then an opportunity arises to apply a better algorithm, and I genuinely value those moments. But most of software engineering is not that. It's understanding the existing system well enough to know where the next part is most easily bolted on. It always was; AI just makes the bolting easier.

We need to define what "thinking" is before saying they can't. What they do closely resembles thinking at a surface level. There are obvious differences but nailing them down isn't so easy. It seems to me what they do may be an approximation of some aspects of what we do.