Why just focus on those, or any specific dates?

Yes, it's certainly shameful that slavery continued in the US until so recently, and a great deal of what is going on in general in the US is shameful.

OTOH, as far as slavery goes:

It was the Romans - Europe's ancestors - that perhaps practiced it more than anyone else.

What amounts to slavery still exists today in the middle east in countries like Saudi Arabia and the UAE.

If there were robots out on the street killing people and robbing them, would it make sense to remove them?

Removing criminals from the streets is to protect the public - it makes sense regardless of WHY people are committing the crimes,

Nobody has ever, even one time, seen a single atom, anywhere, that is not behaving per the known laws of science. So, the brain, like the rest of nature, would appear to be a machine that we can analyze and duplicate if we wish.

It doesn't leave much room for a "soul" (other than as a psychological construct), does it ?! The soul can't be controlling our neurons if science is doing it.

> the brain is not digital

The energy efficieny of our brain is not because it is analog, but rather because it is an aychronous dataflow architecture. Each neuron only updates if/when one of its inputs changes.

> we can barely model the neural network of an ant

Actually a human is estimated to have ~100T synapses, and SOTA LLMs are already over 1T params, so we're not far off in terms of compute - just need to know what to do with it.

> We have absolutely no idea if such an "algorithm" even exists

Our entire cortex - language cortex, visual cortex, auditory cortex, etc is a very regular structure. If you uncrumpled it, it's basically a tea-towel sized sheet of 6 layers of neurons, all with a very specific interconnect pattern, and feedback paths through the thalamus. It seems HIGHLY likely that it all operates the same way, and this is the "algorithm" they are looking for.

Sure, the computing substrate doesn't matter. If you can figure out how our brain learns in terms of neurotransmitters and synapses, then you can abstract that and implement it in software (and/or hardware).

I wouldn't call it a publicity stunt, but it does seem bit crazy to fund something that amounts to a wish list, apparently with no plan other than "we'll work on it" to get there.

I'm pretty sure a data center cutoff is bad on an under voltage too.

1) grid under supplies and voltage drops
2a) grid requests more power from flexible (gas) plants
2b) HUGE customers completely cut their draw
3) more power with less demand causes huge over voltage

I initially read that as:

Scientists Find Wind Blowing From Our Milky Way's Bung Hole

Calling something AI is a very low bar to meet - it just means that the system is doing something that is considered intelligent.

There are lots of things today that can reasonably be called AI, from LLMs to image generation, image recognition, etc (all completely unrelated!). It used to be that these were all called ML (machine learning) instead, and then at some point (I remember it happening) the popular press started to call it all "AI" since that sounded more sensational and headline-worthy. Initially there was some pushback from the research community, but then they got on board with the "AI" label too since VCs wanted to fund "AI" ...

So, sure, LLMs are one example of AI. A one-trick pony, but impressive nonetheless in how much capability that one trick gives you.

However anyone asking or stating whether "LLMs are AI" almost certainly doesn't mean "AI" in this correct use of the word (since the answer is obviously and trivially yes) - they are more likely talking about AGI or human-level intelligence, in which case clearly the answer is no.

I really don't understand the interest in deciding what label to slap on today's AI tech (LLMs). Clearly it's not human level, and clearly in a few years we'll have something better (not just a better LLM), even if that also isn't yet human level. I like to say that you'll know were making progress towards AGI when people stop calling LLMs AI, and start calling them LLMs again, with the (meaningless) label "AI" then referring to the new AI tech du jour.

You can call memory learning if you want to (but you're wrong - memory and learning are two different things - it's the difference between being book smart, and learning the skill of actually doing something yourself).

You can call LLMs AI is you want to (and they are).

But this this doesn't change a language model into a brain.

You seem fundamentally confused about what an LLM is - it is just a statistical predictor, generating based on the average statistics. Imagine you had a football stadium of people, each with a keyboard and a giant screen they could all see. An operator types a question, or just some text, on the screen, and now everyone in the stadium gets to vote on the next word by typing it on their keyboard. The most popular word goes on the screen, and now they vote on the next word. This is very close to what an LLM is.

If you ask an LLM if it knows something or not (maybe something it just hallucinated) then of course it will give you an answer, but the answer as always is just the statistical average of what it was trained on, not it's own answer (remember, this is not a brain, it's a language model).

If you give it 99 training samples that say "I don't know what the capital of France is", and 1 training sample that says "the capital of France is Paris", then ask it "do you know what the capital of France is", then it will say "no" because that is the average answer. They don't deal in facts - they deal in statistics.

> You keep insisting that to qualify as AI, an LLM must replicate all aspects of human intelligence. I don't think that's a fair standard.

No, that is not what I am saying. There is nothing wrong with calling an LLM AI. What I am pointing out is exactly what you note - that an LLM is far from being able to do what a human can, or even a dog for that matter. It's a one-trick pony - an auto-regressive language generator.

The label "AI" just refers to a computer doing something (maybe just one thing) that previously only a human could do. If you want to refer to a broader set of capabilities, then the usual term to use if "AGI" (G=general), but even that doesn't mean "capable of everything intelligence-related that a human can do".

Of the big companies working on AI, only Google DeepMind defines AGI as broadly human level (and note that it'll require a number of major breakthroughs to get there).

So, yes, per the above definitions LLMs are one form of AI, but they are not AGI, and certainly not human-level AGI.

Ask 100 people what "intelligence" is, and you will probably get 100 different answers, but most would intuitively include ability to learn as a key part of that. Even a bird or a mouse can learn, but an LLM can not - it is more like a fixed database or expert system of knowledge.

Ask an LLM today should you walk or drive to the nearby car wash and it'll say walk since it is nearby. Now explain to it that you can't wash the car if the car is not there and it'll probably admit you are right. Now come back tomorrow and ask the same question and it'll give the same wrong answer. Is that human level intelligence to you?

Creativity is another major thing missing. An LLM is only capable of generating output composed out of the language and phrases it was trained on. It can generate a lot of variety for sure, combining the stuff it was trained on in various ways, but it is never going to be able to come up with something truly creative for a variety of reasons.

1) LLMs are not designed to do that. The name says it all - they are language models. They are designed and trained to predict (i.e. copy) language. They are copying machines, not creativity machines.

2) An LLM has no mechanisms of curiosity or boredom, or knowledge of what it "knows" and does not know (this is why they hallucinate - they are just a bunch of statistics). They are just not built to be able to explore and learn based on the realization that they don't know something.

3) LLMs can't learn, so of course they can't learn anything new - be creative as opposed to just rehashing what they are trained on.

I could go on and on ... but if you are even a tiny bit familiar with how our brain is built and works you would have realized the above. Our cortex is at least 50% feedback paths that are there to detect prediction failure (reality != prediction) and to be able to learn from that.

You can call LLMs whatever you like, but the reality is that they are just copying/prediction machines. That is what they are built to do, and it is therefore what they do.

I just gave you two examples of what is lacking, but why not think for yourself and look at a Transformer and brain and consider for yourself everything that is lacking!

As noted though "AI" doesn't actually mean anything (nor does "AGI"), so your list of what is lacking is going to depend on what you are comparing it to.

If you are comparing to a human brain, then a LOT is missing.

Dyson was lucky he came to this realization in the 90s -- he only had 1 building and 1 relic to blow up. How do you blow up all the data centers of all the frontier AI companies unless you are the US Air Force or NSA and prepared for enormous civilian (including US citizen) casualtites?

Realistically, even USAF can't do it since they don't want to trigger global thermonuclear war, which is why you're left with needing an international agreement between the US and China (at the very least) to materially manage AI risks.

I wouldn't call it odd or surprising - this just reflects one of the limitations of the Transformer - that it consists of a smallish (~100) fixed number of layers of transformations (cf thought steps). If you need to try to get it to do something that requires more than 100 steps of "thought" then the old way was "think step by step" prompting, which these "thinking" models now do automatically.

Every word/token the LLM generates gets fed back in as an input, giving it another 100 layers of computation for the next token, etc, etc.

Yes, but I'm sure Anthropic is well aware that China will never slow down, and even if they could convince Trump to ban Chinese models that would just put the US at a disadvantage.

Therefore, rather than this being a plea to not ruin a good thing by racing towards a plateau, I think the main goal here is Anthropic's now typical MO of fear-mongering as marketing, trying to juice the demand for their upcoming IPO by "warning" about how powerful AI is becoming - and will continue to become - unlimited investor upside !

FWIW I think that smaller models are going to eat Anthropic's lunch, as well as that of OpenAI, with the market for increasingly capable smaller, and much cheaper, models being much larger than that for these huge SOTA models whose differentiated capabilities are going to become less and less relevant for most use cases (especially the high volume ones of coding and business automation).