The LLM answers your accept/reject are used as training data

That usually isn't true if you're using the agent tools aimed at companies. They are explicit about what is and isn't used for training.

They'll just ban those

They probably won't be able to.

There's a legal right to end-to-end encryption and there's nothing preventing you from making a system look like HTTPS.

It'll need to be developed extraterritorially, otherwise they'd just pick up the people developing it. They'll ban the original website, so you'll have to get it through dodgy means, which means they can insert backdoors for you (unless you're able to verify the code and compile it for yourself).

Have you heard of the hush-hush over Chat Control, a recurring theme on this site? It's getting to the point that they could shoehorn that with age verification push.

Yes, but they have to fit it around the judgements of the different European courts, and they really can't do it, so they try to say 'Oh, this is totally voluntary' and then when realise that that won't work they decide to go after companies, etc.

They have some political power, but they don't have all of it and they are quite constrained.

>there's nothing preventing you from making a system look like HTTPS.

Did you hear about the Great Firewall?

And they just have to say it's to protect "the children"/"democracy"/"to fight disinformation"/"hate speech". You can't beat politics with technology.

An mcp lets an agent call functions. These can in turn even issue queries to an LLM. E.g. an agent can issue natural language queries to a database by calling a function query("what is the answer to life, the universe and everything?") and the function will return "42" to the agent.

He also didn't argue about the rest of the math so it's likely correct that the whole model needs to be in SRAM :)

I think you get different compressed files depending on how many threads you use to compress

You're completely right, if you already have a GPU in a system adding tensor cores to it gives you better performance per area.

GPU + dedicated AI HW is virtually always the wrong approach compared to GPU+ tensor cores

Based on the graphs "up to 10 times faster" compares before/after flash attention

Accurate simulation is also an AI problem, but that should be a separate paper

How does consumption of a nutrient stop its production?

It's not stopping the production, but changing the flow.

suppose there's a flow of nutrients of type A from area one to area two. Currently, iron is the chokepoint in nutrient consumption of A, so that A is never completely consumed while going from one to two.

By adding excess iron into area one, or in the middle of the region between this flow from area one to two, you now have the possibility to consume A completely as iron no longer limits A's consumption.

So what happens to area two's consumption of A, if it became more scarce? May be nothing - or may be you now have another choke point of resources that wasn't there previously, leading to a change. If you weren't sure if this change would occur, or dont know, perhaps adding iron to area one is not a good idea, until such outcomes have been studied and acertained first.

I only use my common sense here, but it doesn't stop the production it just prevents the transport from the source to the destination through the area that was previously iron deficient and couldn't use the fully the other nutrient which passed on to another area.

No