I think LLMs are pretty well described by what I said. How aren’t they?

What you’re linking to is the opposite of a logical fallacy.

I do not endorse this slippery slope argument and I think it’s dumb

Not even close

There’s simply no reason a trillion parameter language model should be involved in choosing my Starbucks order

I don’t think so and I do think we’re overshooting it a tad

The fact that this is what people sound like when they are sleep talking makes me think the brain has a language model

Perhaps this is in the umbrella of (a) but there is a third meaning where an agent is a rational actor who strategizes and pursues goals and whatnot which is also quite complicating nowadays

I don't really see how

So it’s the fact that it takes input and produces output that precludes it being intelligent?

Lol ok man. I surrender for real this time. But I do recommend that you take a closer look into this if you want to understand these systems because you are not correct about this.

If you wanted to you could use it in a different way that it was not designed to support in order to generate text deterministically, but that would be your own non-standard way to use it. What it is designed to do, and the way people use it, is to sample text randomly from the learned distribution.

The training objective is to find a probability distribution which could have generated the training data. The reason to make that the training objective is that you’re going to then sample from that probability distribution to generate new text. All the big LLM providers generate text this way.

It really could not be more inherent. Look. This is from the original GPT paper. The big P stands for Probability. The whole point of this thing is model text generation as a particular random process, and then to generate text by carrying out that process. cdn.openai.com/research-cov...

“It reminds me of antivaxxers” no it’s not really the same. Antivaxxers are bad because their misconceptions are broadly harmful to others. This is just a person with an opinion that is kind of silly.

To the vibe coder reading this: you don’t have to get too worked up when you see this. You can just let it roll off your back. It’s not a big deal.

Every time they force Grok to recite another race science line it puts another sql injection vulnerability in there

Imagine how many zero days X has because it’s vibe coded entirely using emergent-misaligned Grok instead of noble Claude

So you think this is just a lie?

I do think you are confusing vector search and text generation though. That also explains your earlier comment about databases. You would probably benefit by getting that straightened out.

Yes of course if they were different then they would be different. I'm talking about how they are actually implemented today.

Ok but you’ll note that in context, we were talking about how LLM generate text, not about neural networks more broadly.

Idk what to tell you my man, that’s how LLMs have worked since GPT1

Alright man I surrender

Yes, as I said, you could invent your own deterministic way to generate text using an LLM, but the standard approach is to randomly sample tokens according to the output probabilities. That’s why they’re called probabilities.

Yeah I’m extremely familiar with how they work lol

Alright if you say so

No. The transformer network outputs a vector of probabilities over the next token. The calculation of this vector is indeed not random. But to generate text, a token is randomly sampled from the distribution that is defined by that vector.

You could generate text deterministically from an LLM if you wanted to but that would be a very non-standard way to use an LLM. All of the LLMs that people use regularly generate text randomly.