Did we ever think LLMs were a path to AGI...? AGI is friggin hard, I don't know why folks keep getting fooled whenever a bot writes a coherent sentence.
LLMs are the first instance of us having created some sort of general AI. I don't mean AGI, but general AI as in not specific AI. Before LLMs the problem eith AI was always that it "can only do one thing well". Now we have something on the other side: AI that can do anything but nothing specific particularly well. This is a fundamental advancement which makes AGI actually imaginable. Before LLMs there was literally no realistic plan how to build general intelligence.
Smart enough to make weapons, tame dogs, start fires and cultivate plants. Humans managed to do that even when most of their time was spent gathering food or starving.
Nobody cares about making an AI with basic human survival skills. We could probably have a certified genius level AI that still couldn't do any of that because it lacks a meaningful physical body.
If we wanted to make that the goal instead of actual meaningful contributions to human society, we could probably achieve it, and it would be a big waste of time imo.
It's mostly a thing among the youngs I feel. Anybody old enough to remember the same 'OMG its going to change the world' cycles around AI every two or three decades knows better. The field is not actually advancing. It still wrestles with the same fundamental problems they were doing in the early 60s. The only change is external, where computer power gains and data set size increases allow brute forcing problems.
I'd say the biggest change is the quantity of available CATEGORIZED data. Tagged images and what not has done a ton to help the field.
Further there are some hybrid chips which might help increase computing power specifically for the matrix math that all these systems work on.
But yeah, none of this is making what people talk about when they say AGI. Just like how some tech cult people felt that Level 5 self driving was around the corner, even with all the evidence to the contrary.
The self driving we have (or really, assisted cruise control) IS impressive, and leagues ahead of what we could do even a decade or two ago, but the gulf between that, and the goal, is similar to GPT and AGI in my eyes.
There are a lot of fundamental problems we still don't have answers to. We've just gotten a lot better at doing what we already did, and getting more conformity on how.
If by “junior dev”, you mean “a dev at a level so low they will be let go if not promoted”, then I agree.
I’ve watched my coworkers try to make use of LLMs at work, and it has convinced me the LLM’s contributions are well below the bar where their output is a net benefit to the team.
It works pretty well in my C++ code. Context: modern C++ with few footguns, inside functions with pretty-self-explanatory names.
I don't really get the "low bar for contributions" argument because GH Copilot's contributions are too small-sized for there to even be any bar. It writes the obvious and tedious loops and other boilerplate so I can focus on what the code should actually do.
Conversely, I was very skeptical of its ability to help coding something non-trivial. Then I found out that the more readable your code is - in a very human way, like descriptive identifiers, comments etc - the better this "smart autocomplete" is. It's certainly good enough to save me a lot of typing, so it is a net benefit.
> intelligence requires understanding which is not possible without consciousness
How are you defining "consciousness" and "understanding" here? Because a feedback loop into an LLM would meet the most common definition of consciousness (possessing a phonetic loop). And having an accurate internal predictive model of a system is the normal definition of understanding and a good LLM has that too.
Problem is, we have no agreed-upon operational definition of consciousness. Arguably, it's the secular equivalent of the soul. Something everything believes they have, but which is not testable, locatable or definable.
But yet (just like with the soul) we're sure we have it, and it's impossible for anything else to have it. Perhaps consciousness is simply a hallucination that makes us feel special about ourselves.
I disagree. There is a simple test for consciousness: empathy.
Empathy is the ability to emulate the contents of another consciousness.
While an agent could mimic empathetic behaviors (and words), given enough interrogation and testing you would encounter an out-of-training case that it would fail.
Uh... so is it autistic people or non-autistic people who lack consciousness? (Generally autistic people emulate other autistic people better and non-autists emulate non-autists better)
> given enough interrogation and testing you would encounter an out-of-training case that it would fail.
For one thing, this would imply that clinical psychopaths aren't conscious, which would be a very weird takeaway.
But also, how do you know that LMs aren't empathic? By your own admission they do "mimic empathetic behaviors", but you reject this as the real thing because you claim that with enough testing you would encounter a failure. This raises all kinds of "no true Scotsman" flags, not to mention that empathy failure is not exactly uncommon among humans. So how exactly do you actually test your hypothesis?
Great point and great question! Yes, it does imply that people who lack the capacity for empathy (as opposed to those who do not utilize their capacity for empathy) may lack conscious experience. Empathy failure here means lacking the data empathy provides rather than ignoring the data empathy provides (which as you note, is common). I’ve got a few prompts that are somewhat promising in terms of clearly showing that GPT4 is unable to correctly predict human behavior driven by human empathy. The prompts are basic thought experiments where a person has two choices: an irrational yet empathic choice, and a rational yet non-empathic choice. GPT4 does not seem able to predict that smart humans do dumb things due to empathy, unless it is prompted with such a suggestion. If it had empathy itself, it would not need to be prompted about empathy.
1) Earth has an infinite past that has always included life
2) The Earth as a planet has a finite past, but it (along with what made up the Earth) is in some sense alive, and life as we know it emerged from that life
3) The Earth has a finite past, and life has transferred to Earth from somewhere else in space
4) We are the Universe, and the Universe is alive
Or something else? I will try to tie it back to computers after this short intermission :)
LLMs have changed the world more profoundly than any technology in the past 2 decades, I'd argue.
The fact that we can communicate with computers using just natural language, and can query data, use powerful and complex tools just by describing what we want is an incredible breakthrough, and that's a very conservative use of the technology.
I'm kinda curious as to why you think that's the case. I mean, smartphones are nice, and having a browser, chat client, camera etc. in my pocket is nice, but maybe I have been terminally screen-bound all my life, but I could do almost all those things on my PC before, and I could always call folks when on the go.
I've never experienced the massively life changing effects of having a smartphone, and (thankfully) none of my friends seem to be those people who are always looking at their phones.
While many technologies provided by the smartphone were indeed not novel the cumulative effect of having a constant access to them and their subsequent normalization is nothing short of revolutionary.
For instance, I remember the time when chatting online (even with people you knew offline) was considered to be a nerdy activity. Then it gradually became more mainstream and now it's the norm to do it and a lot of people do it multiple times per day. This fundamentally changes how people interact with each other.
Another example is dating. Not that I have personal experience with modern online dating (enabled by smartphones) but what I read is disturbing and captivating at the same time e.g. apparent normalization of "ghosting"...
I don't actually see anything changing, though. There are cool demos, and LLMs can work effectively to enhance productivity for some tasks, but nothing feels fundamentally different. If LLMs were suddenly taken away I wouldn't particularly care. If the clock were turned back two decades, I'd miss wifi (only barely available in 2003) and smartphones with GPS.
Indeed. The "Clamshell" iBook G3 [0] (aka Barbie's toilet seat), introduced 1999, had WiFi capabilities (as demonstrated by Phil Schiller jumping down onto the stage while online [1]), but IIRC, you had to pay extra for the optional Wifi card.
That breakthrough would not be possible without ubiquity of personal computing at home and in your pocket, though, which seems like the bigger change in the last two decades.
Deep learning was an advance. I think the fundamental achievement is a way to use all that parallel processing power and data. Inconceivable amounts of data can give seemingly magical results. Yes, overfitting and generalizing are still problems.
I basically agree with you about the 20 year hype-cycle, and but when compute power reaches parity with human brain hardware (Kurzweil predicts by about 2029), one barrier is removed.
Human and computer hardware are not comparable, after all even with the latest chips the computer is just (many) von Neumann machine(s) operating on a very big (shared) tape. To model the human brain in such a machine would require the human brain to be discretizable, which, given its essentially biochemical nature, is not possible - certainly not by 2029.
This time around they’ve actually come up with a real productizable piece of tech, though. I don’t care what it’s called, but I enjoy better automation to automate as much of the boring shit away. And chip in in coding when it’s bloody obvious from the context what the few lines of code will be.
So not an ”AI”, but closer to ”universal adaptor” or ”smart automation”.
Pretty nice in any case. And if true AI is possible, the automations enabled by this will probably be part of the narrative how we reach it (just like mundane things like standardized screws were part of the narrative of Apollo mission).
> Hype and announcements, sure, but this is the first time there's actually a product.
No, its not. Its just once the hype cycle dies down, we tend to stop calling the products of the last AI hype cycle "AI", we call them after the name of the more specific implementation technology (rules engines/expert systems being one of the older ones, for instance.)
And if this cycle hits a wall, maybe in 20 years we'll have LLMs and diffusion models, etc., embedded lots of places, but no one will call them alone "AI", and then the next hype cycle will have some new technology and we'll call that "AI" while the cycle is active...
As an outsider, I can talk to AI and get more coherent responses than from humans (flawed, but it's getting better). That's tangible, that's an improvement. I for one don't even consider the Internet to be as revolutionary as the steam engine or freight trains. But AI is actually modifying my own life already - and that's far from the end.
P.S. I've just created this account here on Hacker News because Altman is one of the talking heads I've been listening to. Not too sure what to make of this. I'm an accelerationist, so my biggest fear is America stifling its research the same way it buried space exploration and human gene editing in the past. All hope is for China - but then again, the CCP might be even more fearful of non-human entities than the West. Stormy times indeed.
Mainly because LLMs have so far basically passed every formal test of ‘AGI’ including totally smashing the Turing test.
Now we are just reliant on ‘I’ll know it when I see it’.
LLMs as AGI isn’t about looking at the mechanics and trying to see if we think that could cause AGI - it’s looking at the tremendous results and success.
“You're in a desert, walking along in the sand when all of a sudden you look down and see a tortoise. You reach down and flip the tortoise over on its back. The tortoise lays on its back, its belly baking in the hot sun, beating its legs trying to turn itself over. But it can't. Not with out your help. But you're not helping. Why is that?”
got-3.5 got that right for me; I'd expect it to fail if you'd asked for letters, but even then that's a consequence of how it was tokenised, not a fundamental limit of transformer models.
This sort of test has been my go-to trip up for LLMs, and 3.5 fails quite often. 4 has been as bad as 3.5 in the past but recently has been doing better.
if this is the test you're going to then you literally do not understand how LLMs work. it's like asking your keyboard to tell you what colour the nth pixel on the top row of your computer monitor is.
An LLM could easily answer that question if it was trained to do it. Nothing in its architecture makes it hard to answer, the attention part could easily look up the previous parts of its answer and refer to the fourth word but it doesn't do that.
So it is a good example that the LLM doesn't generalize understanding, it can answer the question in theory but not in practice since it isn't smart enough. A human can easily answer it even though the human never saw such a question before.
> the model doesn't have a functionality to retrospectively analyse its own output; it doesn't track or count words as it generates text. it's always in the mode of 'what comes next?' rather than 'what have i written?'
Humans doesn't do that either. The reason humans can solve this problem is that humans can generate such strategies on the fly and thus solve general problems, that is the bar for AGI, as long as you say it is unfair to give such problems to the model we know that we aren't talking about an AGI.
Making a new AI that is specialized in solving this specific problem by changing the input representation still isn't an AGI, it will have many similar tasks that it will fail at.
> also, again, tired of explaining this to people: gpt models are token-based. they operate at the level of tokens - which can be whole words or parts of words - and not individual characters. this token-based approach means the model's primary concern is predicting the most probable next token, not keeping track of the position of each token in the sequence, and the smallest resolution available to it is not a character. this is why it can't tell you what the nth letter of a word is either.
And humans are a pixel based model, we operate on pixels and physical outputs. Yet we humans do generate all the necessary context, and adapts it to the task at hand to solve arbitrary problem. Such context and inputs manipulations are expected of an AGI. Maybe not the entire way from pixels and 3d mechanical movement, but there are many steps in between there that humans can easily adapt in. For example humans didn't evolve to read and write text, yet we do that easily even though we operate on a pixel level.
If you ask me to count letters my mind focuses on the letter representation I created in my head. If you talk about words I focus on the word representation. If you talk about holes I focus on the pixel representation and start to identify color parts. If you talk about sounds I focus on the vocal representation of the words since I can transform to that as well.
We would expect an AGI to make similar translations when needed, from the token space you talk about to the letter space or word space etc. That ChatGPT and similar can't do this just means they aren't even close to AGI currently.
I reckon an LLM with a second pass correction loop would manage it. (By that I mean that after every response it is instructed to, given the its previous response, produce a second better response, roughly analogous to a human that thinks before it speaks)
LLMs are not AIs, but they could be a core component for one.
Every token is already being generated with all previously generated tokens as inputs. There's nothing about the architecture that makes this hard. It just hasn't been trained on this kind of task.
The following are a part of my "custom instructions" to chatGPT -
"Please include a timestamp with current date and time at the end of each response.
After generating each answer, check it for internal consistency and accuracy. Revise your answer if it is inconsistent or inaccurate, and do this repeatedly till you have an accurate and consistent answer."
It manages to follow them very inconsistently, but it has gone into something approaching an infinite loop (for infinity ~= 10) on a few occasions - rechecking the last timestamp against current time, finding a mismatch, generating a new timestamp, and so on until (I think) it finally exits the loop by failing to follow instructions.
I don’t think the test of AGI should necessarily be an inability to trip it up with specifically crafted sentences, because we can definitely trip humans up with specifically crafted sentences.
I disagree about the claim that any LLM has beaten the Turing test. Do you have a source for this? Has there been an actual Turing test according to the standard interpretation of Turings paper? Making ChatGPT 4 respond in a non human way right now is trivial: "Write 'A', then wait one minute and then write 'B'".
Your test fails because the scaffolding around the LM in ChatGPT specifically does not implement this kind of thing. But you absolutely can run the LM in a continuous loop and e.g. feed it strings like "1 minute passed" or even just the current time in an internal monologue (that the user doesn't see). And then it would be able to do exactly what you describe. Or you could use all those API integrations that it has to let it schedule a timer to activate itself.
By completely smashes, my assertion would be that it has invalidated the Turing test, because GPT-4s answers are not indistinguishable from a human because they are, on the whole, noticeably better answers than an average human would be able to provide for the majority of questions.
I don’t think the original test probably accounted for the fact that you could distinguish the machine because it’s answers were better than an average human.
LLMs can't develop concepts in the way we think of them (i.e., you can't feed LLMs the scientific corpus and ask them to independently to tell you which papers are good or bad and for what reasons, and to build on these papers to develop novel ideas). True AGI—like any decent grad student—could do this.
that being said, it is highly intelligent, capable of reasoning as well as a human, and passes IQ tests like GMAT and GRE at levels like the 97th percentile.
most people who talk about Chat GPT don't even realize that GPT 4 exists and is orders of magnitude more intelligent than the free version.
That’s just showing the tests are measuring specific things that LLMs can game particularly well.
Computers have been able to smash high school algebra tests since the 1970’s, but that doesn’t make them as smart as a 16 year old (or even a three year old).
Answers in Progress had a great video[0] where one of their presenters tested against an LLM in five different types of intelligence. tl;dr, AI was worlds ahead on two of the five, and worlds behind on the other three. Interesting stuff -- and clear that we're not as close to AGI as some of us might have thought earlier this year, but probably closer than a lot of the naysayers think.
This isn't the same thing. This is a commanded recital of a lack of capability, not that its confidence in it's answer is low. For a type of question the GPT _could_ answer, most of the time it _will_ answer, regardless of accuracy
I just noticed that when I ask really difficult technical questions, but for which there is an exact answer, It often tries to answer plausibly, but incorrectly instead of answering "I don't know". But over time, It becomes smarter and there are fewer and fewer such questions...
I read an article where they did a proper Turing test and it seems people recognize it was a machine answering because it made no writing errors and wrote perfectly
I've not read that, but I do remember hearing that the first human to fail the Turing test did so because they seemed to know far too much minutiae about Star Trek.
Are you kidding? Have you seen the reactions since ChatGPT was released, including in this very website? You'd think The Singularity is just around the corner!
