uh_uh

Was it hallucinating here, or are the commenters hallucinating? What OP is saying is just not true. A CT scan and normal daily commute in Grand Central station are NOT comparable in terms of radiation received. Somehow this is controversial because an AI said it?

Weird you don't have this requirement for the OP spewing his urban myths above.

In your opinion how many hours spent in Grand Central station equal the radiation received from a CT scan?

> Remember that you'll get comparable levels of radiation even if you commute through the grand central station every day.

Gemini says this:

> A single typical CT scan delivers a dose that is roughly 1,000 to over 5,000 times higher than the dose you'd get from spending a few hours in Grand Central Terminal.

Where did you get that from?

1. "It is not saying that humans confidently give wrong answers if they do not know correct ones." And I didn't say that they do either, so you might have hallucinated that.

2. What are you arguing about? I didn't say they're always correct obviously.

> We do know how LLMs work, correct?

NO! We have working training algorithms. We still don't have a complete understanding of why deep learning works in practice, and especially not why it works at the current level of scale. If you disagree, please cite me the papers because I'd love to read them.

To put in another way: Just because you can breed dogs, it doesn't necessary mean that you have a working theory of genes or even that you know they exist. Which was actually the human condition for most of history.

Examples of emergence:

1. Multi-step reasoning with backtracking when DeepSeek R1 was trained via GRPO.

2. Translation of languages they haven't even seen via in-context learning.

3. Arithmetic: heavily correlated with model size, but it does appear.

I could go on.

Albeit it's not an LLM, but a deep learning model trained via RL, would you say that AlphaZero's move 37 also doesn't count as emergence and the model has no understanding of Go?

Sorry but this is nonsense. Do you have a theory about when certain LLM capabilities emerge? AFAIK we don't have a good theory about when and why they do emerge.

But even if knew how something works (which in present case we don't), shouldn't diminish our opinion of it. Will you have a lesser opinion of human intelligence, once we figure out how it works?

The LLM is a glorified autocomplete in as much as you are a glorified replicator. Yes, it was trained on autocomplete but that doesn't say much about what capabilities might emerge.

1. Many humans don't have an idea of the limits of their competence. It's called the Dunning–Kruger effect.

2. LLMs regularly tell me if what I'm asking for is possible or not. I'm not saying they're always correct, but they seem to have at least some sense of what's in the realm of possibility.

We don't really have a clue what they are and aren't capable of. Prior to the LLM-boom, many people – and I include myself in this – thought it'd be impossible to get to the level of capability we have now purely from statistical methods and here we are. If you have a strong theory that proves some bounds on LLM-capability, then please put it forward. In the absence of that, your sceptical attitude is just as sus as the article's.

This top to bottom drawing – does this tell us anything about the underlying model architecture? AFAIK diffusion models do not work like that. They denoise the full frame over many steps. In the past there used to be attempts to slowly synthetize a picture by predicting the next pixel, but I wasn't aware whether there has been a shift to that kind of architecture within OpenAI.

I hope you're joking. Sometimes they don't even know which company developed them. E.g. DeepSeek was claiming it was developed by OpenAI.

Could be. Also, as you imply, they'd have to loosen the regularization penalty on θ, and maybe it's difficult to loosen it such that it won't become too prone to overfitting.

Maybe their current setup of keeping θ "dumb" encourages the neural network to take on the role of the "algorithm" as opposed to the higher-variance input encoded by z (the puzzle), though this separation seems fuzzy to me.

Given that all parameters are trained jointly at inference time and a single sample of z is supposed to encode ALL inputs and outputs for a given puzzle (I think), I don't quite understand the role of the latent z here. Feels like μ and Σ could be absorbed into θ (and an additional variance parameter).

Although if they partition z such that each section corresponds to one input and run f_θ through those sections iteratively, then I guess it makes sense.

This is actually a pretty cool accidental mirror test.

Pundits were saying that deep learning has hit a plateau even before the LLM boom.

This. Articles like this are examples of motivated reasoning and seem to be coming from a place of insecurity by programmers who feel their careers threatened.

So why are they SOTA translators? Would you consider old translation software bullshit generators? Because LLMs can do their job, and more.

The choice unfortunately seems to correlate with the person's age. Younger generations will have no trouble treating LLMs as actually intelligent. Yet another example of "Science progresses one funeral at a time.”

> There is no abstracted concept of a "colour" in there. There's just a lot of imagery tagged with each colour name, and if you select a different colour you get a vector in a space pointing to different images.

It has been observed in LLMs that the distance between embeddings for colors follows the same similarity patterns that humans experience - colors that appear similar to humans, like red and orange, are closer together in the embedding space than colors that appear very different, like red and blue.

While some argue these models 'just extract statistics,' if the end result matches how we use concepts, what's the difference?

Is there anything AI (or more like AGI) could do for you that would change your opinion? Eternal life, interstellar travel, curing diseases, educating people with superhuman patience and competence would not be worth it?

Aren't you at least a little bit excited for the potential benefits?

So you don't have a concrete suggestion to solve the scamming problem?

Are you suggesting government action against putting up code like this to GitHub? It’s ok if you are, but I want to put into more concrete terms what we’re talking about.

Cooperation works if the potential damage caused by a rouge actor is sufficiently low. Otherwise, it's too easy to sabotage things. This is why we don't want random rouge states to have nukes. AI will give so much leverage to rouge actors that it will significantly shift the game theory in favour of not cooperating.

Demanding responsible behaviour from everybody is not going to work. Some people don't care about negative externalities that much and it's enough if only a few of them decide not to play ball. So either grandma needs to adapt which will upset some people or distributing the tech should be regulated/prosecuted which will upset another group of people.

This tech is going to be ubiquitous, it's just too easy to distribute it. Grandma better starts adapting now.

Interesting write-up. Does this mean that all of OP's competitors are facing the same legal issues? I wonder what (if anything) they do about this.

Why is "without prompting" a requirement? If a hypothetical LLM comes up with the best neologisms that one can imagine, but only after prompting, then they're disqualified?