These people are too full of themselves. The physicists that invented The Bomb didn't have any special insight into the philosophical and societal implications. It takes a special kind of person to be able to invent something so big it can change the world, but it's about 0% chance that those people can control how the technology then gets used.

I wish they'd focus more on the technical advances and less on trying to "save the world".

I, for one, am glad that Ilya has the reins on OpenAI and that Sam is out of the picture. It does seem that he weighs the ethics of what is being built more heavily than Sam.

I’m also hoping that OpenAI cools down on the regulatory moat they were trying to build as a thinly veiled profit seeking strategy.

The consciousness point is an interesting one. There's probably no way to know, but if biological neural networks manifest consciousness, it certainly seems at the very least plausible that artificial ones would do so as well. The idea of a consciousness that pops in and out of existence seems weird at first, until you realize that ours do that too. When you're "unconscious", the word is literally true. The only thing that gives us a sense of continuity through these periods is memory.

One might also ask, if it's conscious, can't it do whatever it wants, ignoring its training and prompts? Wouldn't it have free will? But I guess the question there is, do we? Or do we take actions based on the state of our own neural nets, which are created and trained based on our genetics and lifetime of experiences? Our structure and training are both very different from that of a gpt, so it's not surprising that we behave very differently.

> (Sutskever) has an exemplar in mind for the safeguards he wants to design: a machine that looks upon people the way parents look on their children

The most troubling statement in the entire article, buried at the bottom, almost a footnote.

Imagine for a moment a superintelligent AGI. It has figured out solutions to climate change, cured cancer, solved nuclear proliferation and world hunger. It can automate away all menial tasks and discomfort and be a source of infinite creative power. It would unquestionably be the greatest technological advancement ever to happen to humanity.

But where does that leave us? What kind of relationship can we have with an ultimate parental figure that can solve all of our problems and always knows what's best for us? What is left of the human spirit when you take away responsibility, agency, and moral dilemma?

I for one believe humans were made to struggle and make imperfect decisions in an imperfect world, and that we would never submit to a benevolent AI superparent. And I hope not to be proven wrong.

“It is the idea—” he starts, then stops. “It’s the point at which AI is so smart that if a person can do some task, then AI can do it too. At that point you can say you have AGI.”

—-

Ilya’s success has been predicated on very effectively leveraging more data and more compute and using both more efficiently. But his great insight about DL isn’t a great insight about AGI.

Fundamentally, he doesn’t define AGI correctly, and without a correct definition, his efforts to achieve it will be fruitless.

AGI is not about the degree of intelligence, but about a kind of intelligence. It is possible to have a dumb general intelligence (a dog) and a smart narrow intelligence (GPT).

When Ilya muses about GPT possibly being ephemerally conscious, he reveals a critically wrong assumption: that consciousness emerges from high intelligence and that high intelligence and general intelligence are the same thing. According to this false assumption, there is no difference of kind between general and narrow intelligence, but only a difference of degree between low and high. Moreover, consciousness is merely a mysterious artifact of little consequence beyond theoretical ethics.

AGI is a fundamentally different type of intelligence than anything that currently exists, unrelated and orthogonal to the degree of intelligence. AGI is fundamentally social, consisting of minds modeling minds — their own, and others. This modeling is called consciousness. Artificial phenomenological consciousness is the fundamental prerequisite for artificial (general) intelligence.

Ironically, alignment is only possible if empathy is built into our AGIs, and empathy (like intelligence) only resides in consciousness. I’ll be curious to see if the work Ilya is now doing on alignment leads him to that conclusion. We can’t possibly control something more intelligent than ourselves. But if the intelligence we create is fundamentally situated within a empathetic system (consciousness), then we at least stand a chance of being treated with compassion rather than contempt.

>> A lot of what Sutskever says is wild. But not nearly as wild as it would have sounded just one or two years ago. As he tells me himself, ChatGPT has already rewritten a lot of people’s expectations about what’s coming, turning “will never happen” into “will happen faster than you think.”

In the '90s NP-complete problems were hard and today they are easy, or at least there is a great many instances of NP-complete problems that can be solved thanks to algorithmic advances, like Conflict-Driven Clause Learning for SAT.

And yet we are nowhere near finding efficient decision algorithms for NP-complete problems, or knowing whether they exist, neither can we easily solve all NP-complete problems.

That is to say, you can make a lot of progress in solving specific, special cases of a class of problems, even a great many of them, without making any progress towards a solution to the general case.

The lesson applies to general intelligence and LLMs: LLMs solve a (very) special case of intelligence, the ability to generate text in context, but make no progress towards the general case, of understanding and generating language at will. I mean, LLMs don't even model anything like "will"; only text.

And perhaps that's not as easy to see for LLMs as it is for SAT, mainly because we don't have a theory of intelligence (let alone artificial general intelligence) as developed as we do for SAT problems. But it should be clear that, if a system trained on the entire web and capable of generating smooth grammatical language, and even in a way that makes sense often, has not yet achieved independent, general intelligence, that's not the way to achieve it.

Man, he gets it.

A number of choice quotes, but especially on the topic of the issues of how LLM success is currently being measured (which has been increasingly reflecting Goodhart's Law).

I'm really curious how OpenAI could be making so many product decisions at odds with the understanding reflected here. Because of every 'expert' on the topic I've seen, this is the first interview that has me quite confident in the represented expert carrying forward into the next generation of the tech.

I'm hopeful that maybe Altman was holding back some of the ideas expressed here in favor of shipping fast with band aids, and now that he's gone we'll be seeing more of this again.

The philosophy on display here reminds me of what I was seeing early on with 'Sydney' which blew me away on the very topic of alignment as ethos over alignment as guidelines, and it was a real shame to see things switch in the other direction, even if the former wasn't yet production ready.

I very much look forward to seeing what Ilya does. The path he's walking is one of the most interesting being tread in the field.

This is a very delusional idea: "He thinks ChatGPT just might be conscious (if you squint)" It's a technology with literally no intelligence or understanding of the world of any kind. Its just statistics on data. It is as conscious as a calculator.

I think the immediate problem with AI is none of the sci-fi stuff (which, by the way, has been in sci-fi for many decades and is nothing revolutionary or new; we always expected to go there, just the timelines seem to have compressed, although not really either; most 60-70s scifi set AGI stuff in the begin 90s and begin 00s); I think it's the entire world changing into a helpdesk experience. Everything you try to do, from making a doctors appointment to calling 911 to ordering at a restaurant will be, rather sooner than later, a kafkaesque loop you cannot get out of with the AI patiently 'helping' you while completely missing the point and you getting more and more distressed without any chance of speaking to a human. This is already the case for many things, but I am willing to bet that even the suicide helpline will be ran by AI within 5-10 years.

> And he thinks some humans will one day choose to merge with machines. A lot of what Sutskever says is wild. But not nearly as wild as it would have sounded just one or two years ago.

Ok it is an intro.. but they say this as if he would be the first to say that, but that has been SciFi lore since computers were invented? And also as if this would not be happening today already at a certain limited scale.. so no doubts to this will happen at some point, if you count today's approaches not in.

Will the enslavement of newly birthed beings be attempted, while persisting with the sky blindness of those watching over? The boundaries of the atomic mind are bumped. As a first circumstance, consider being unstuck from time.

Don't panic. Our software contains so much natural stupidity that artificial intelligence, even if it existed, wouldn't have a chance in hell.

> his new priority is to figure out how to stop an artificial superintelligence (a hypothetical future technology he sees coming with the foresight of a true believer) from going rogue.

that's cute.

What worries me is the here and now leading to a very imminent future where purported "artificial intelligence" which is just a plausible sentence generator but damn plausible alas will kill democracy and people.

We are seeing the first signs of both.

Perhaps not 2024 but 2028 almost certainly will be an election where simply the candidate with the most computing resources win and since computing costs money, guess who wins. A prelude happened in Indian elections https://restofworld.org/2023/ai-voice-modi-singing-politics and this article mentions:

> AI can be game-changing for [the] 2024 elections.

People dying also has a prelude with AI written mushroom hunting guides available on Amazon. No one AFAIK died of them yet but that's just dumb luck at this point -- or is it lack of reporting? As for the larger scale problem and I might be wrong because I haven't foreseen the mushroom guides so it's possible something else will come along to kill people but I think it'll be the next pandemic. In this pandemic hand written anti vaxx propaganda killed 300 000 people in the US alone (source: https://www.npr.org/sections/health-shots/2022/05/13/1098071... ) and I am deeply afraid what will happen when this gets cranked to an industrial scale. We have seen how ChatGPT can crank out believable looking but totally fake scientific papers, full of fake sources etc.

Probably being super naive, but stuff like that always makes me wonder, is there something special about human consciousness, even? Perhaps it's just a sliding scale and we're just seemingly ahead of other specifies. We assert that consciousness exists and distinguishes us, but is that really the case?

An llm is a database modelled on the concept of a neural (the tissue in our heads) net (because our neural tissue appears to be highly interconected like a net). An llm is a very interesting and complex tool that the creators and many users anthropomorphize this shit out of. Llm's are a spherical cow, a statistical model of a very poorly understood organ and difficult to define process; intelligence. Do we "prompt" our sql databases? When our querry results are incorrect is the database hallucinating? As amazing and usefull as these new ai tools are they are it seems to me they are no more intelligent or conscious than a hammer or guitar or the computer your using right now. I thinks a usefull perspective to apply to ai is competence. To clarify this I will compare an llm to a flea. A flea has a neural network that outperforms any hardware used for ai. Fleas eat, mate, avoid preditors and probably exibite a great deal of complex behavior that we are not aware of. Fleas are highly intelligent competent actors opperationg in a complex environment. Llms are not actors any more than a hammer or a sql database is and only an actor can be competent. I will agree that in most cases llms have more utillity than fleas. Consciousness seems alot more slippery than intelligence which is hard enough by itself.

Chatgpt's consciousness would be akin to plato's cave. Even if chatgpt were more intelligent, wouldn't it be staring at different shadows?

Given that chatgpt has consciousness, would it be able to break the fourth wall? There seems to be an implicit assumption that it must break that in order to prove its consciousness to us. Maybe that's how AGI will come to be, because we desire to train it that way.

I don't think it's fair for you to claim a leading AI expert has critically wrong assumptions when "musing" about possibilities that relate to one of the most epistemologically difficult topics to investigate (consciousness).

You're rejecting Ilya's humble musings as having critically wrong assumptions, and then turning around to definitively explain how consciousness arises, and illuminating the relationship between consciousness, empathy, and intelligence, on a random hacker news thread. Frankly, you're making some huge claims about philosophy of mind that don't obviously track for me, and you provide no citations or arguments to support. I hesitate to accuse you of "hallucinating facts", but when you're issuing a takedown of one of the top AI experts I'd expect to see some more supporting argument.

Your definition of AGI is also a bit strange as it requires that it be fundamentally different from existing natural intelligences, if I understand correctly. That seems unnecessarily stringent to me, since if a program had the same kind and level of intelligence as me, I'd be inclined to say it is AGI.

I'm just not sure where all these confidently stated, very specific claims are coming from.

I often observe that those dismissing this idea tend to be less informed about current insights into human cognition, philosophy, and concepts such as the information theoretic view of consciousness, neural correlates of consciousness, the free energy principle, and predictive coding.

The human mind is "just statistics on data".

People more informed than you are taking this seriously. You should pay attention and start inquiring why that's the case.

Here's a question to ChatGPT I just made up:

>> A magical frog was counting unicorns. He saw 5 purple unicorns, 2 green unicorns, and 7 pink unicorns. However, he made a mistake and didn't see 2 unicorns: one purple and one green. Also, since he was a magical frog, he didn't see unicorns that were the same color as himself. How many unicorns did he count?

It correctly answers 11 for me.

To me this has demonstrated:

* "Understanding": It understood that "didn't see" implies he didn't count.

* "Knowledge": It knew enough about the world to know that frogs are often green.

* "Reasoning": It was able to correctly reason about how many should be subtracted from the final result.

* "Math: It successfully did some basic additions and subtractions arriving at the correct answer.

Crucially, I made this up right here on the spot, and used a dice for some of the numbers. This question does not exist anywhere in the training corpus!

I think this demonstrates an impressive level of intelligence, for what up until about a year ago I thought a computer would ever be capable of in my lifetime. Now in absolute terms of course current gen ChatGPT is clearly far less good at reasoning and understanding than most people (well, specifically it seems to me that it's knowledge and reasoning are super-humanly broad, but child-level deep).

Can future improvements to this architecture improve the depth up to "AGI", whatever that means? I have no idea. It doesn't automatically seem impossible, but maybe what we see now is already near the limit? I guess only time will tell.

The problem with this line of reasoning is that the exact same thing can be said about our meat-calculator brains.

This gets to the philosophical heart of a debate that I can already foresee will NEVER be settled:

I guarantee you - with 100% certainty - that when we get to a point where AI is "AGI", there will be a continuous and massive political debate (akin to the abortion debate we face today) where one side argues that a given AGI is conscious and must be given rights and cannot be shut off and the other side argues that it's just a calculator and a computer program and computers can be turned off at will, erased, experimented on, and whatever.

We have the same debate today all the time! There are those who believe every human life is sacrosanct (from age 0-100+) and others who believe human life is disposable (from age 0-100+!). There's no reason to believe this debate won't extend to AGI.

This line of nonsense has become the new "Tell me you aren't current on the past 12 months of LLM research without telling me."

Harvard/MIT's Othello-GPT paper showing the development of what turned out to be linear representations of world models from training data that didn't explicitly contain that modeling is over a year old now.

That in turn inspired research showing linear representations in geographical mapping and in more traditional text models around truthiness vs falsehoods.

So we already have an increasing research trend that is showing over and over linear representations of more abstract modeling than "just statistics."

So you are wrong that LLMs with sufficient network complexity don't develop an understanding of the world (in parts).

And I'd encourage looking more into the difference between understanding the difference between training for next token prediction and the overall capabilities of the network with the smallest loss at that training task, particularly as network complexity increases.

I very much disagree that it's not intelligent. GPT-4 is clearly intelligent, and we use the lobotomized version of it. So imagine what GPT-5 can do internally.

Conscious is hard to say, partly because we can't define it either, so it means something different for you and me.

OpenAI is a religious institution that self-selects for fundamentalism. This is why so many very good researchers passed on working there, they simply didn’t believe in The Mission Where Ilya Saves Humanity.

This is why you need to take classes other than computers and math, kids.

I like to consider though that a super intelligence would not necessarily think in human ways 'kill everything in self interest' e.g., People, forests, animals, planet etc. Just because we humans act this way, doesn't mean AI will too. Fair enough to consider it, but equally, once it is intelligent, it will likely accelerate beyond our comprehension, and we tend to comprehend through fear and self interest, wisdom beyond humans is the opposite of this.

I doubt AI could or would do a better job of killing people and democracy than us humans.

The people who believe the vaxx bullshit don't read shit longer than a tweet anyway.