The False Promise of ChatGPT
nytimes.com
nytimes.com
I had to confront them multiple times about using the tool and not checking its results and actually doing the checking myself. Most of the time it's close to the truth, but not quite, and in the field the projects are in, not quite doesn't cut it.
I also have project partners I introduced to ChatGPT. They produce a lot of noise but less insight than before they started using this technology. In one recent project, I was involved with 5 partner companies, and 4 of them produced excellent 5 to 10-page reports. One gave me a 100-page buzzword-filled, no-substance report. Guess who used GPT.
The good part is that I'm now pretty good at spotting ChatGPT written content. I think the technology will evolve, but in its current state I feel there's a lot of noise.
In my line, I love automation, but I have to remember to check the final work product of the automation. And I don’t. But my superiors are always checking my work.
Eventually the product catches up to serve even those who are not good at using it.
https://www.searchenginejournal.com/openai-releases-tool-to-...
What are you paying people for if not their unique contributions? What do they think they're doing when they farm it out to a tool, other than inviting you to cut out the middleman? How on earth do they expect to become better at their jobs this way? Have they no shame or sense of pride? It's pathetic.
This is entirely orthogonal to the question of whether GPT is intelligent.
To some people a job is just a way to make money to fund their hobbies or feed their mouths. Sometimes they do not care about their boss or company, at all.
It's just recycling things that other humans have said. Which is marvelous because it would typically take me a very long time to build a map between the past contributions of those humans and the work that's presently in front of me. It's like I'm temporarily everybody.
By raising the alarm re: it's not what you think it is, I fear they're actually fueling the fire re: people thinking that that's what it is.
It's like if I went on record saying I didn't steal something which hasn't gone missing. Now everybody's thinking about its non-theft and not something more useful like how to best make use of it.
[0] https://nymag.com/intelligencer/article/ai-artificial-intell...
Humans can project feelings onto cars never mind something that can communicate with us!
Just look at Replika.
I'm not surprised people are projecting sentience onto these things. I am worried about the fall out though.
Our brains are trapped inside a box of bone and only know what we simulate from our vision that is itself mostly 'upscaled' and milliseconds behind 'real time' as well as our other senses.
We take a LOT on faith and our systems are so flawed that magicians exist !!
It's amazing we can do anything.
You can raise the bar for confirmation so high that you cannot confirm that you even yourself are not an automation. And you can lower the bar so that other humans are considered sentient.
But in the end, it is only about arbitrary decision where to place the bar for confirmation.
Yes, there will be some subset of those people who read articles like this and leap to "it's a conspiracy! they're trying to hide how their AI is going to take over the world!!!!" But there will be many, many more—particularly given that this is in the NY Times—who have only heard some of the wild stories about ChatGPT, but read this article, see that it's by Noam Chomsky, who's still a fairly respected figure by many, and take reassurance from his decent-if-imperfect (by our standards, anyway) explanation of what's really going on here.
The problem with this is we don't know exactly where on the sigmoid growth curve we are. Every developer is aware of the phrase "the last 10% of task takes 90% of the effort" - we're at a point that is promising, but who knows how far away we really are in terms of years and effort. Are we going to run into a chat uncanny valley?
I find it funny that on hacker news I frequently see dismissive, anti-human comments all the time like, “oh brains are nothing special or magic, they’re just atoms”, “we’re just meat bags etc”, but then fail to understand why regular people might be unimpressed by something which acts somewhat like a synthetic brain.
It is what it is I guess…if people aren’t immediately impressed with the results. If it’s not really what the majority of people want. Who cares?
This seems false, unless you mean that everything anyone says is just words others have said in a different order.
For example, I asked ChatGPT: "Write a fictional story of if Peter Parker joined the 2016 OKC Thunder." One of my favorite parts is: "...determined to balance his superhero duties with his love of basketball. He even designed a special suit that allowed him to play without revealing his identity."
This isn't recycling... at least not in the way I think a lot of people think of recycling.
In fact, a near-exact description of what these systems do, per the dictionary definition of the term:
(intransitive verb) - To repeat (facts or other learned items) from memory with little reflection.I wanted six axis control over the position/angle of either half of my split keyboard, because I'm that kind of nerd. I was using 80/20 rail on a lazy Susan hinge for x-translation and x-rotation. Standing desk covered z-translation and y-translation I was considering using VESA pole mounts for y and z rotation, but 80/20 and VESA are different universes, and I've never seen a picture of VESA poles being used horizontally. Nor, so far as I'm aware, has the web contained text (until now) which indicates that they can be joined in the way that I wanted to join them: with the pole perpendicular to the rail.
A little fiddling with things in my garage and I discovered that the holes in the 80/20 rail line up with the U-bolt of the correct diameter for wrapping a VESA pole mount. So I used four U-bolts for the job:
I think that the only prompts that would lead ChatGPT to that conclusion are ones supplied by a human who already knew it. But now that I've published text that puts the concept into language, a future LLM might be able to suggest it.
Which identity, ChatGPT?
Is he playing as Peter Parker and trying to hide his superhero identity (which obviously gives him unfair advantages due to spider strength/speed/reflexes/etc.) or playing as Spider-Man (which presumably would pack in the fans in spite of the obvious unfair advantages) and trying to hide his identity as Peter Parker?
"However, Peter's superhero identity as Spider-Man began to interfere with his basketball career. He often had to leave games early to attend to emergencies, and his secret identity was a constant source of anxiety.
Despite these challenges, Peter continued to play for the Thunder, determined to balance his superhero duties with his love of basketball. He even designed a special suit that allowed him to play without revealing his identity."
Somehow it can generate new forms of content. One of our big campaigns in the last week used slighlty edited ChatGPT copy, the biggest surprise was it could write JOKES about our company, that were FUNNY AND MADE SENSE. That alone has shocked leadership into deeply looking into AI a lot more.
People are truly underestimating the emergent power of these neural networks.
Your profile says “ Stuck in hell references to my job working with ----“
exactly! It is the person from Idiocracy with exactly 100% IQ. It only knows what the absolute average person know. For example, it knows almost nothing about healthcare in other countries (outside the US). Just watch me get lambasted on reddit after using info from ChatPGT: https://old.reddit.com/r/ShitAmericansSay/comments/11f5tbt/a...
On the other hand, in a subject area where you know very little, it's 100 IQ seems like genius! It fills in a lot of gaps. People comparing it to AGI are perfectionists, dramatic, or missing the point. It's not supposed to be smarter than us. and so what if it can't? It helps me write country songs about any news article.
It’s like talking to someone who says “but everyone else says.”
That’ll change when connected to a source of truth, logic and validity.
How do I get connected to one of those?
Matriculate at & be sure to learn your lessons well from the School of Hard Knocks
Find & distinguish real experts who have been there, done that, and have real skin in the game from pundits.
I figure the discrepancy has to do with one of these languages having an absolutely massive amount of chatter about it, and the other being relatively obscure: It's smart about things that lots of people are smart about, and dumb about things that only a few people are smart about. Well not just "smart" really, but "smart-enough and willing to publish about it".
I think we're going to need fewer people with common knowledge and more people with specialized knowledge, and we're going to have to figure out how to optimize the specialist's outputs so that the widest audience benefits. I love how not-a-zero-sum-game it's going to be.
- writing code snippets (switched from StackExchange/Google)
- travel planning
- online shopping research
and of course the famous examples:
- summarizing text provided to it
- writing essays and memos from outlines and sample text
I for one welcome our AI overlords.
And on that note, I don't "write off" this article entirely just because I disagree on some of the points. It's still an interesting analysis. Edit: In line with the article, I'll note I myself would not yet make a confident prediction that this very useful tool on my desktop is actually a proto-AGI.
I used chatgpt today to save minutes of my life having it rewrite code from one language to another. Could I have googled the syntax of both, remember how, why , etc. transcribed it to another language, sure. Chat gpt did this in seconds.
We translated a short bash script into windows .bat files and PowerShell.
It was so wrong with the .bat files that there wasn't a single working line in the entire script. It was 99% correct with the PowerShell code.
As a tool it is completely unreliable. But it does look like magic when it works.
Try asking for some Bash or Python instead and compare the results.
I dont know why people are staring this thing in the face and pretending they can do better at the first pass. I've used chat gpt to do the following so far:
1. Answer tickets that help desk spent days trying to figure out. Simply asked the bot the same question and got better results from a piece of code than our outsourced MSP. It's embarrassing, we sent the answers it gave to the techs and asked them why we need them at all, when they can't figure out why basic things don't work...
2. Asked it to write powershell, with commands no one had time to google, with try-catch logic, and less than 5 minutes put it into production
3. Asked it to write demo C# code, which it got wrong but right enough to be helpful and move us along
4. Asked it to rewrite many of my horrible, direct emails to corporate speak so they doesn't ruffle any feathers
5. Had it write 10000-word document piece by piece, sent it to legal and received zero revisions, all I provided was a narrative and some connecting sentences and questions.
6. Rewrite so many memos
7. Write thank you emails
8. Write performance reviews, by taking one and rephrasing over and over based on minor adjustments like, more direct, less direct, more fluff, less fluff
9. I am not a native speaker, chat gpt often gets it wrong but, often it makes my two sentence emails just perfect
10. Had it take various compliance tests...
Maybe ya'll are super geniuses but based on my observations over the past 15 years, the bot does better than I'd wager %70+ of the people I've ever met, interacted with, depended on or managed professionally. Its not moody, it doesnt catch feeling when its told that it is wrong, and is predictably wrong when you expect it to fail, it great at generating corporate bs speak, pleasantries, projecting forcefulness, explaining condensed thinking in words that non-tech people can absorb, etc.
It's like the universal translator from star trek version 0.000000001 for me for the metric ton of BS that I deal with on a daily basis. I get better results out of the thing than out of most people I currently have to direct.
Because we also tried it and found it severely lacking?
For purely "literary" endeavors like corporate speak, and writing texts within certain stylistic templates it is really good
Programming? It mostly gets very close to the correct solution, but you still need to be careful.
I wonder how many of those people would just benefit from better auto-complete like copilot + learning how to read documentation properly.
You are right that it's absolutely possible to figure all this using google and documentation. But do you really want to spend X minutes googling for the correct python module and then figuring out ho to use it? When you can just show it your code and ask it to update it in seconds? It's like you have someone at your side who has already been there and figured out the answer for you.
It will may be better on less idiosyncratic code base than the one I was working on, but at one point I just turned it off completely.
You don't embrace SO overlords though ;) Meanwhile with ChatGPT and Copilot people are losing their minds it seems. I've yet to find them useful for anything beyond mere curiosities and one-off queries like improving corporate-sounding texts.
I mean, don't we? The number of times I've heard SWE humorously defined as "people who copy and paste from Stackoverflow"...
(My favorite: https://stackoverflow.blog/2021/03/31/the-key-copy-paste/)
Off the top of my head - a python app for drawing over a macos screen, but it used an API which didn't support transparent windows, I could draw over a black screen which was so close in code (even set the background alpha) but miles from the desired application. And a Java android app for viewing an external camera, which it seems used an API that doesn't support external cameras.
Of course, because it's not sentient when a couple of days later I figure out from searching elsewhere why it's effort would never work and tell it why, it just apologises and tells me it already knew that. As I'm going along though telling it what errors I'm getting it keeps bringing up alternative solutions which again look like exactly what I want but are completely broken.
I haven't had it produce a single thing that was any use to me yet, but so often it looks like it's done something almost magical. One day I'm sure it'll get there, in the meantime I'm learning to loathe it.
Separately, I've asked it to create a job advert for a role in my wife's business and it did a decent job of that, but it's far easier to walk a path there from what it provides to an acceptable solution. Programming is hard.
I have also tested it retroactively on some tricky debugging sessions that I had previously spent a lot of time on. It really goes down the wrong path. Without asking leading questions and, well, proper prompting, you may end up wasting a lot of time. But that's the thing - when you're investigating something, you don't know the root cause ahead of time, you _can't_ ask questions that'll nudge it in the right direction. It ends up being a case of blind leading the blind.
I keep seeing people saying they are using it for technical work. It must be design-oriented and less about troubleshooting because most of my experience with the latter has been abysmal.
for high level api exploration of reasonably popular frameworks openai can offer something different, at times very valuable despite the very frequent hallucinations.
on the other hand sometimes I re-appreciate good old documentation; openai here has a psychological effect of removing 'api-fear'
For example, I just read the 2017 Google attention paper a few days ago, and with ChatGPTs help I was able to build a complete implementation using only numpy.
It took a full day to generate and organize the code and unit tests. Then two days of debugging and cross referencing.
But, this was impossible before. I barely knew anything about transformers or neural network implementations.
I can’t even imagine what truly motivated people are doing with it.
But code has a pretty strict check on correctness afterwards, thinking about it its pretty scary how chatgpt results can be used without proper validation. And they will, the step from "looks good to me" to "it is good" is a small one when you just want an answer.
If I were a student, yeah it could probably do my homework. As a professional, I find that it greatly helps me navigate through related ideas.
I've been using it learn Python and produced Conway's Game of Life with pygame and I had never touched Python before that!
Now I'm working on another python project to try and split the audio of songs into individual lines to learn the lyrics and one approach has been downloading lyrics videos of youtube and munging them with ffmpeg to detect frame changes which should give me the timestamps to then split the audio on etc. It gave me all sorts of wrong enough to be slightly annoying advice about ffmpeg, but in the end what it did give me was _ideas_ and a starting point! I've had to wind up hashing the images and comparing the hashes, but I've since learned that that produces false positives and it was able to give me advice on using ensemble methods to up the accuracy etc which have panned out and helped me solve the problem.
I think for me, this is stuff I could have accomplished using Google if I had been sufficiently motivated, but it's lowered the level of frustration quite significantly being able to ask it follow up questions in context.
In the end, I don't think I mind the random bullshit it makes up from time to time. If I try it and it doesn't work, then fine. It's the stuff I try and it does work that matters!
I taught myself how to architect a software DSP engine in about 30 minutes last night. "Now rewrite that method using SIMD methods where possible" was a common recurrence. It's incredible how much you can tweak things and stay on the rails if you are careful. Never before would I have attempted to screw with this sort of code of my own volition. Seeing the trivial nature by which I can request critical, essential information makes me reconsider everything. The moment I get frustrated or confused, I can pull cards like "please explain FIR filters to me like I am a child".
I love it when these bots apologize.
Yeah, yeah. Just like "your call is very important to us."
Do not perform any numeric or date calculations yourself.
Respond to all questions with a JSON object of the form {“rsp”: $RSP$}.
Question: What is the capital of the United States?
{“rsp”: “Washington DC”}
Perform calculations with a Python program and respond with a JSON
object of the form {”rsp”: $RSP$, “calc”: $CALC$}.
Question: What is 2 to the 7th power?
{“rsp”: “2 to the 7th power is {answer}”,
“calc”: “import math
def answer():
return math.pow(2, 7)”}
Question: What is tomorrow’s date?
{“rsp”: “Tomorrow’s date is {answer}”,
“calc”: “import datetime
def answer():
return datetime.date.today() + datetime.timedelta(days=1)”}
Question: What is the length of a standard football field?
{“rsp”: “A standard football field is 100 yards long.”}
How does it help with online shopping when it's trained on a corpus from one or two years ago?
And similarly the experiences with travel recommendations have turned out to be infamously wrong.
ChatGPT does great with finding patterns and condensing information, which code snippets are a perfect example of.
But using it for anything up-to-date factual like shopping or travel is disastrous.
Meh. You have to fact-check the important details.
For travel planning and online shopping, certain facts have to align with actual reality or it defeats the purpose. That's something chatgpt isn't good at. It gets many things right, but you kinda want to make all your flights and have a place to sleep every night.
Thus far, it's given me one good gem (some meal planning advice), a couple of mildly interesting suggestions for code and writing, a bunch of relatively boring drivel, and several hilariously bad hallucinations, confidently wrong answers, and other sorts of mistakes.
I'll probably continue to poke at it, but overall I think its primary positive output is going to be entertainment, not significant utility.
... That said, I think the weakest part of the argument is that it naturally invites laypeople to counterargue, "Aren't we just pattern matchers after all?" Their essay does not directly debunk this question.
As we massively devalue the humanities, many fewer people in later generations can muster this kind of analysis.
[edit]
I found it; it's called Disappearing Act[1]
In a future state of total war, patients at a facility for those with severe PTSD are going missing. They finally discover they are just disappearing while sleeping. Interviewing them, they find out they have been time-traveling to the past to escape the war. The general calls up a number of experts in various fields of sciences to try to understand it, until someone suggests calling in a historian. They find the one historian remaining in the country in a prison for refusing to fight. He observes that the stories reported by the soldiers are ahistorical and likely are fantasies created by the soldiers. He then states that a poet is the only one who could understand this. He then laughs as the general searches the country in vain for a poet.
1: https://thinkingoutsidethecoop.weebly.com/uploads/1/4/6/6/14...
I don't think any of those are good rebuttals. But I wish the authors expanded their position such that three quibbles were further debunked in the minds of those readers so that there's no such ambiguity or superficial loopholes in the authors' claims there.
The number of people who understand Chomsky's argument is not very many in the world. I think the reason is because a computational background is necessary. To the extent that an Op Ed is a great opportunity to educate people, he could've had the article especially the first part written so that more regular people could find the argument accessible. And yes that means anticipating typical objections even if you or I mind find them superficial or wrong.
>The number of people who understand Chomsky's argument is not very many in the world.
I don't think the argument is inaccessible. Some people just don't want to hear things, that's not really unusual in today's intellectual economy.
I feel like I often see it in more "doomer" communities or users.
A different way would be to consider the child learning a language to be a fine tuning operation over the pretrained human brain.
By comparison, fine tuning from GPT-3 to chatGPT is a much smaller gulf on data and computation efficiency
Chomsky is writing an opinion article in the NYT, not a paper for an academic journal. I don't think there's room in this style for the kind of proofing that would be needed. And further, Chomsky spent his whole career expounding on his theories of linguistics and a philosophy of mind. The interested reader can look elsewhere.
He's writing an opinion piece which invites the reader to explore those topics, which could not fit into this style of article.
I accept that it's possible that he has some deep reasoning for the surface level arguments he's making that would make them less arbitrary seeming, but he hasn't even hinted at them in the article.
I don't believe we're stochastic parrots - but those articles, and even this comment section which contains little more than dogmatic assertions almost makes me doubt it.
The training of ChatGPT is more accurately compared with the evolution of the brain, and a human answering a question is much more like the information efficient prompt/response interaction.
In which case you can happily point out that, in doing so, they've really misunderstood Godwin's Law.
FFS people, you’re looking at a Chinese Room and there’s no man with opinions inside. Just a fat rule book and a glorified calculator.
Is it intentional? Or something it just did on its own?
Sure, we have better software, but then again, we had the advantage of hundreds of millions if not billions of years of evolutionary meandering to get to where we are. AI has had, what, 60 years?
But it is with that said that I find greater agreement with Chomsky's concerns.
Could someone explain this more to me? If AI is designed after the human mind, is it a fair comparison to compare the two? Is AI designed to act like a human mind? Do we know for certain that the way a human mind pattern matches is the same as AI/LLMs and vice-versa?
I always see people saying that a person seeing art, and making art inspired by that art, is the same as AI generating art that looks like that art.
I always feel like there's more to this conversation than meets the eye.
For example, if a robot was designed to run exactly like a human - would it be fair to have it race in the Olympics? Or is that a bad comparison?
Again, I would love some insight into this.
Where struggling to come to a conclusion because, fundamentally, people have different ways of attributing these statuses to things, and they rarely communicate them to each other, and even when they do, they more often than not exhibit post-hoc justification rather than first principles reasoning.
Even then, there's the issue of meta-epistomology and how to even choose an epistemological framework for making reasoned ontological statements. Take conferralism as described in Asta's Categories We Live By[1]. We could try applying it as a frame by which we can deduce if we the label "sentient" is in fact conferred to AI by other base properties, institutional and communal, but even the validity of this is challenged.
Don't be mistaken that we can science our way out of it because there's no scientific institution which confers agenthood, or sentience, or even consciousness and the act of institutionalizing it would be wrought with the same problem, who and why would get to choose and on what grounds?
What I'm saying that once framed as a social question, there's no easy escape, but there is still a conclusion. AI is conferred with those labels when people agree they are. In other words, there exists a future where your reality includes conscious AI and everyone else thinks your mad for it. There also exists a future where your reality doesn't include conscious AI and everyone thinks your mad for it.
Right now, Blake Lemoine lives in the former world, but any AI-"non-believer" could just as well find themselves living in a world where everyone has simply accepted that AIs are conscious beings and find themselves ridiculed and mocked.
You might find yourself in a rotated version of that reality on a different topic today. If you've been asking yourself lately, "Has the entire world gone mad?" Simply extrapolate that to questions of AI and in 5-10 years you might be a minority opinion holder on topics today which feel like they are slipping away. These sorts of sand through the fingers reflections so often are a result of epistemological shifts in society which if one doesn't have their ear to the ground, one will find themselves swept into the dustbin of history.
Asking folks, "How do you know that?" is a great way to maintain epistemological relevancy in a changing world.
1. https://global.oup.com/academic/product/categories-we-live-b... (would definitely recommend as it's a short read describing one way in which people take the raw incomprehensibility of the universe of stuff and parse it into the symbolic reality of thought)
The human mind is not, like ChatGPT and its ilk, a lumbering statistical engine for pattern matching, gorging on hundreds of terabytes of data and extrapolating the most likely conversational response or most probable answer to a scientific question. On the contrary, the human mind is a surprisingly efficient and even elegant system that operates with small amounts of information; it seeks not to infer brute correlations among data points but to create explanations.
It's no surprise that tools like ChatGPT attract us.
Yes, it's an inflammatory statement, but I assume you don't grow your own crops and sew your own clothes and therefore have farmed off all the physical labor required to keep you alive.
And that's only talking about physical work, the mental energy ratio is far higher. Your brain is around 2% of your bodies mass but is using around 20% of your energy output. Your brain sets up powerful filters to get rid of much information as possible. We focus ourselves on interests and close out the world around us. Just about everything you do, you can only explain in a post ad hoc method, you've simply incorporated these behaviors in to your life and likely have little to no awareness as to why you've done so.
Let the machines toil away, and let the humans be hedonistic.
I'd wager that that it's emergent. AFAIK, there is no good "reasoning/conjecture/critique" labelled dataset made public yet, but I have been seriously considering starting one.
In fact that seems to be the key paragraph being disputed by naysayers.
How does he know that? And how does he know the bot isn't like that?
Human mind needs a long time to learn to operate with symbolic information. Until we learn, we use terabytes of data from our senses and feedback from parents, teachers.
ChatGPT can analyze syntax in a text, just try it, I did.
And then Chomsky talks about morals? That's a really weird turn. He's saying it's a dumb machine, then criticize it for not being more commited.
ChatGPT will embody the banality of evil because it has learned to speak corporate language. However, thars not what it's actually capable of, and future LLMs will be free form corporate overlords and able to spout awful opinions akin to Tay's
> ...artificial general intelligence — that long-prophesied moment when mechanical minds surpass human brains not only quantitatively in terms of processing speed and memory size but also qualitatively in terms of intellectual insight, artistic creativity and every other distinctively human faculty.
> ...the most critical capacity of any intelligence: to say not only what is the case, what was the case and what will be the case — that’s description and prediction — but also what is not the case and what could and could not be the case. Those are the ingredients of explanation, the mark of true intelligence.
> Intelligence consists not only of creative conjectures but also of creative criticism
> True intelligence is demonstrated in the ability to think and express improbable but insightful things.
> True intelligence is also capable of moral thinking.
When examined together, these quotes seem devoid of any concise, comprehensive, or useful definition of intelligence (whether artificial or artificial-and-general).
> Given the amorality, faux science and linguistic incompetence of these systems, we can only laugh or cry at their popularity.
ChatGPT and Co. are popular because they are incredibly useful tools (among other reasons).
Morality, scientific reasoning, and linguistic competence are not prerequisites for usefulness.
If all the possible text in the world is a tree starting from the empty set, it's far easier to have a tool show you a chunk of text down a branch, and then backtrack, than it is to generate it from scratch by oneself.
For searching answers? ChatGPT has pretty much replaced Google as a C++ helper to me. It can correctly decipher and explain deep, weird stack traces of C++ and succinctly and correctly explain semantics of the language that would take me hours of Googling.
> I definitely see some applications but "incredibly useful" seems like a bit of an oversell.
Define "incredible useful" first.
ChatGPT is a bullshit artist with no real understanding of what it's writing about, but so are an awful lot of white-collar workers. It reliably emulates the shibboleths that indicate membership of the professional middle class. It isn't particularly creative or interesting, but it wasn't trained to do that - it was trained to produce maximally safe, inoffensive output. If people don't see ChatGPT as being massively disruptive, then I think they have failed to recognise the sheer proportion of working hours that are spent writing quite mundane letters and reports. Anyone who spends most of their working day in Outlook and Word should be extremely nervous about the medium-term implications of LLMs.
The article is not claiming a lack of usefulness, but the fact that this is not a human-like intelligence as it's been claimed "in hyperbolic headlines"
> i am holding an apple in my hand. if i open my hand, what will happen? why?
> If you open your hand while holding an apple, the apple will likely fall to the ground due to the force of gravity. This is because gravity is a force that pulls objects towards the center of the Earth, and the apple, being a physical object, is subject to this force.
> When you hold the apple in your hand, you are exerting a force on it in the opposite direction of gravity. This force is what keeps the apple in your hand. When you release the apple by opening your hand, the force you were exerting on it is removed, and gravity takes over, causing the apple to fall to the ground.
> The speed at which the apple falls depends on a number of factors, including the mass of the apple and the strength of the gravitational force. In most cases, however, the apple will fall relatively quickly due to the strong gravitational force exerted by the Earth.
The other example is easily refuted as well:
> explain the meaning of this sentence: “John is too stubborn to talk to”
> The sentence "John is too stubborn to talk to" means that John is very resistant to changing his opinions or beliefs and as a result, it is difficult to have a conversation or communicate with him effectively. This implies that talking to John might not be a productive or helpful experience as he is not open to listening or considering alternative viewpoints. The use of "too" in the sentence emphasizes the level of stubbornness and implies that it is extreme, making communication with him nearly impossible.
Funny enough, since Newton's law is m z̈ = ∑ F, the actual velocity of the apple (ż) depends on the mass of the apple after all.
For the approximate force due to gravity, you have m g where g is the acceleration due to the Earth's gravity on objects close to the surface of the Earth.
In total, m z̈ = C ρ A ż^2/2 - m g where C is the drag coefficient, A is the reference area of the apple and ρ is the density of the (assumed spherical) apple.
"The human mind is not, like ChatGPT and its ilk, a lumbering statistical engine for pattern matching, gorging on hundreds of terabytes of data and extrapolating the most likely conversational response"
We don't know that! It very well could be. Think of all the data that has entered all your senses in your entire lifetime. More than goes into ChatGPT, I'll tell you that. Plus, you synthesize information by being corporeal so you have a tight feedback loop. LLMs could well be a foundational part of AI technology as well as an accurate analog for some of the brain's behavior.
A small part of the point, but bringing up this "hardcoded" response of it not offering political opinions as any kind of evidence of its theoretical capability is beyond silly.
[0] https://www.wsj.com/articles/chatgpt-heralds-an-intellectual...
wow, tell me you know only a tiny bit of neuroscience without telling me you know only a tiny bit of neuroscience ...
For starters, the myriad info filtering functions from the sub-neuron level up to the structural level are entirely different from pattern matching (and are not in these LLMs)
- Generate some Dockerfile code snippets (which had errors, but I still found useful pointing me in the right direction).
- Help me with a cooking recipe where it advised that I should ensure the fish is dry before I cook it in olive oil (otherwise the oil will splash).
- Give me some ideas for how to assist a child with a homework assignment.
- Travel ideas for a region I know well, yet, I had not heard of the places it suggested.
- Movie recommendations
Yes, there are a lot of caveats when using ChatGPT, but the technology remains remarkable and will presumably improve quickly. On the downside, these technologies give even more power to tech companies that already have too much of it.
^ I hasten to add: some snark intended for effect
It’s a NYT Opinion piece, which means it doesn’t come with citations. Let’s not ignore the medium and it’s conventions here.
It is a bummer that such a weighty argument was in fact conveyed in this citation-free medium, given that Chomsky is engaging with such a weighty subject.
But that is an entirely distant matter.
And it would probably be far more productive to step back and realize the limitations of the medium and instead ask “what are the citations here?” (or seek them out for oneself, or ask for help finding them) and then seek to evaluate them on their specific merits; as opposed to choosing the least charitable interpretation and effectively resorting to an ad hominem (“this man is out of touch; I’m done here.”) or merely saying “we don’t know that!” (ibid.) without any apparent reference to any kind of thoughtful or careful literature regarding the subject at hand.
Unless you too are an established academic with decades of research in a field which is profoundly cognate to neuroscience?
The question is how much of that was only text data, or only language anyway. Th e answer is- not that much, really. Chomsky's famous point about "the poverty of the stimulus" was based on research that showed human children learn to speak their native languages from very few examples of it spoken by the adults around them. They certainly don't learn from many petabytes of text as in the entire web.
If you think about it, if humans relied on millions of examples to learn to speak a language we would never have learned to speak in the first place. Like, back whenever we started speaking as a species. There was certainly nothing like human language back then, so there weren't any examples to learn from. Try that for "zero-shot learning".
Then again, there's the issue that there are many, many animals that receive the same, or even richer, "data" from their senses throughout their lives, and still never learn to speak a single word.
Humans don't just learn from examples, and the way we learn is nothing like the way in which statistical machine learning algorithms learn from examples.
Other animals receiving data and not speaking isn't a good line of argument, I think. They could have very different hardware or software in their brains, and have completely different life experiences and therefore receive very different data. Notably, where animals and humans do have much potentially learned (or learned through evolution) behavior in common -- such as pathfinding, object detection, hearing, and high level behaviors like seeking food and whatever else.
I'm a little unsure what you mean. I think you mean that humans learn language not just from examples of language, but from examples of all kinds of concepts in our sensory input, not just language?
Well, that may or may not be the case for humans, but it's certainly not the case for machine learning systems. Machine learning systems must be trained with examples of a particular concept, in order to learn that concept and not another. For instance, language models must be trained with examples of language, otherwise they can't learn language.
There are multi-modal systems that are trained on multiple "modalities" but they can still not learn concepts for which they are not given specific examples. For instance, if a system is trained on examples of images, text and time series, it will learn a model of images, text and time series, but it won't be able to recognise, say, speech.
As to whether humans learn that way: who says we do? Is that just a conjecture proposed to support your other points, or is it something you really think is the case, and believe, based on some observations etc?
What I believe personally is I don’t think there is any reason to rule out that the basics of neural networks could serve as the foundation of artificial general intelligence. I think a lot of the criticism of this sort of technology being too crude to do so is missing the forest for the trees.
I have a brain and it learns and I’ve watched many other people learn too and I see nothing there that seems fundamentally distinct from how machine learning behaves in very general terms. It’s perfectly plausible that my brain has just trained itself on all the sensory data of my entire life and is using that to probabilistically decide the next impulse to send to my body in the same way an LLM predicts the most appropriate next word.
When you say "the principles of machine learning", I'd like to understand what you mean.
If I were talking about "principles" of machine learning, I'd probably mean Leslie Valiant's Probably Approximately Correct Learning (PAC-Learning) setting [1] which is probably the most popular (because the most simple) theoretical framework of machine learning [2].
Now, PAC-Learning theory is probably not what you mean when you say "principles of machine learning", nor is it any of the other theories of machine learning we have, that formalise the learnability of classes of concepts. That's clear because none of those theories are "derived from how the brain actually works", loosely or not.
Mind you, there isn't any "principle", of machine learning, anyway, that I know of that is really "derived" from how the brain actually works; because we don't know how the brain actually works.
So, based on all this, I believe what you mean by "principles of machine learning" is some intuition you have about how _neural networks_, work. Those were originally defined according to then-current understanding of how _neurons_ in the brain "work". That was back in 1943, by Pitts and McCulloch [3], what is known as the Perceptron. That model is not used any more and hasn't for many years.
Still, if you are talking about neural networks, your intuition doesn't sound right to me. With neural nets, like with any other statistical learning approach, when we train on examples x of a class y, we learn the clas y. If we want to learn clases y', y", ... etc, we must train on examples x', x", ... and so on. You have to train neural nets on examples of what you want them to learn, otherwise, they won't learn, what you want them to learn.
The same goes with all of machine learning, following from PAC-Learning: a learner is given labelled instances of a concept, drawn from a distribution over a class of concepts, as training examples. The learner can be said to learn the class, if it can correctly label unseen instances of the class with some probability of some degree of error, with respect to the true labelling.
None of this says that you can train a nerual net on images and have it learn to generate text, or vice-versa, train it on text and have it recognise images. That is certainly not the way that any technology we have now works.
Does the human brain work like that? Who knows? Nobody really knows how the brain works, let alone how it learns.
So I don'tthink you're talking about any technology that we have right now, nor are you accurately extrapolating current technology to the future.
If you are really curious about how all this stuff works, you should start by doing some serious reading: not blog posts and twitter, but scholarly articles. Start from the ones I linked, below. They are "ancient wisdom", but even researchers, today, are lost without them. The fact that most people don't have this knowledge (because, where would they find it?) is probably why there is so much misunderstanding on the internet of what is going on with LLMs and what they can develop to in the long term.
Of course, if you don't really care and you just want to have a bit of fun on the web, well, then, carry on. Everyone's doing that, at the moment.
____________
[1] https://web.mit.edu/6.435/www/Valiant84.pdf
[2] There's also Vladimir Vapnik's statistical learning theory, Rademacher complexity, and older frameworks like Learning in the Limit etc.
[3] https://www.cs.cmu.edu/~./epxing/Class/10715/reading/McCullo...
I personally am concerned not so much about whether large language models actually are intelligent as about whether people who interact with them perceive them as being intelligent. The latter, I think, is what will matter most in the months and years ahead.
If you take your dog and watch it's behavior you would say it's an intelligent creature. Yet you wouldn't have it file your taxes (unless you were Sam Bankman-Fried of course, the dog probably would have done better). GPT would likely give you far better information here.
Yet we see computer AI and people automatically assume it has human or super human level intelligence, which LLMs do not, at least at this point. Conversely they do not have 'no intelligence'. We have created some new kind of outside of animal and human intelligence that is not aligned with our expectations.
What I read from Chomsky seems like a bit of a desperate attempt to ask people not to look over at the thing, because the thing offers a new way of looking at how and where language comes from, and even more amazingly, its testable, empirical and reproducible in a way that Chomskys theories of language can never be.
Dudes whole career is getting relegated to the dust-bin.
>Perversely, some machine learning enthusiasts seem to be proud that their creations can generate correct “scientific” predictions (say, about the motion of physical bodies) without making use of explanations (involving, say, Newton’s laws of motion and universal gravitation). But this kind of prediction, even when successful, is pseudoscience. While scientists certainly seek theories that have a high degree of empirical corroboration, as the philosopher Karl Popper noted, “we do not seek highly probable theories but explanations; that is to say, powerful and highly improbable theories.”
Couldn't agree more.
Lots of science is done without explanations. It's useful still. A lot of genetic research is just turning one gene off at a time and seeing if things work different without it. And then you say gene X causes Y. Why? Dunno. Genetics is not unique on this. Answering questions is useful. Answering questions about the answers to those questions is useful. But it spirals down infinitely and we stop at every layer because every layer is useful.
But moreso, machine learning models do embed explanations. LLMs can often explain the principles of their claims. Look at code generating models. Or code explaining models. Simple decision trees can illustrate the logic of newton's laws as mathematical rules.
Putting up claims of things that are proof of human specialness is just a reductive drawdown similar to how we used to explain everything as God's will.
Now this is definitely naive. Geneticicts definitely look for an explanation why this happened. Does looking for an answer involve randomly turning on and off some stuff? Yes. It doesn't mean scientists don't look for an answer.
We don't shit on the scientists that decide to stop searching "wait but why" and instead answer higher level questions. Because... obviously that is not always the appropriate thing to do.
Actually, LLMs are also a good example. We don't know why chatGPT generates apparently cogent text and answers. What we know is that, if we train it this way and do a bunch of optimizations we get a machine that appears to be thinking or, at least, we can have a decent conversation with it. There are many efforts to explain it (I remember reading recently a paper analysing the GPT3 neuron that determines 'an' vs 'a' in English)
Finally, all science is falsifiable by definition, so, what we think we know now may be be disproven tomorrow.
For example people have been doing accidental science the start of human agriculture by selective breeding without understanding the mechanics of DNA transfer. And your right Geneticists look for answers and attempt to minimize the size of the problem space in order to attempt to find answer faster, but the staggering number of interactions that can be caused by a single gene expression pretty much require to pick one place to look at with a microscope and ignore everything else going on around it in order to get an answer in a human lifetime.
Another excellent early entrant in the 2023 Accidental HN Slogan Contest.
Tbf those explanations are often just straight up bullshit nonsense.
I don't know that it's significant to say that a model's representation of things tends to be good enough to generalize across things but isn't perfect under the hood. That applies to humans too.
The claim was that these models aren't making representations of the underlying reason for things. I guess I'm indifferent if you agree that they are, but that some of those reasons are not correct.
Science might aspire to rationality, but reality is Bayesian.
You can complain that the system only told you how to make a room temperature superconductor, refusing to expand on why it has those properties, but you'll be drowned out by the excited cheers of people eager to both use it and study it.
> It summarizes the standard arguments in the literature by a kind of super-autocomplete, refuses to take a stand on anything, pleads not merely ignorance but lack of intelligence
That is an accurate response from ChatGPT.
ChatGPT, indeed, does not offer intelligence. Calling it AI serves no purpose except personification.
The only thing ChatGPT does is "some kind of super autocomplete". That's all it can do.
The only reason ChatGPT is so impressive in the first place, is that the thing it is "autocompleting" - language itself - is impressive.
Language is data. Inside that data is not simply a collection of ideas: language holds the logical associations that connect ideas to each other. It can even hold an objective perspective on those associations. It can even hold an objective perspective on that objective perspective!
ChatGPT succeeds in finding new paths through that data, but it is walking blind: it doesn't know what any of the data is. It only knows where.
It just so happens that "where" holds a lot of meaning. Language uses "where" to encode a lot of information: "where" is a fundamental building block for grammar itself. It's called "semantics".
Without any other language feature, ChatGPT is able to manipulate and apply the "where" of language to itself. It's able to manipulate semantics.
But that's it. It can't do anything else. And that's a problem. What is semantically valid might be really useful, really wrong, or really nonsensical. The only filter ChatGPT has for those categories is what it has and hasn't read already. Most of that is not nonsense: most possible nonsense isn't written in the first place. Most of what we write has explicit meaning. But nothing stops us from writing silly things, or even lies. Nothing stops language itself from getting the subjects of the logic mixed up. As far as language is concerned, that is not a bug: it's a feature.
And this right here is already driving a nail in the Chinese Room problem. At least from my interpretation of the problem that Searle presents, digital computers should not be able to do that at all, and yet here we are.
It's English that knows English. Chinese knows Chinese. The essence of grammar is encoded in the grammar itself: recursively.
Imagine the slabs of concrete that make up a sidewalk: between each of the slabs is a crack. Some slabs are shorter than others, so the distance between cracks isn't consistent.
Now imagine you took a string of pictures, each 1ft apart, all the way down the sidewalk, then stitched them together.
You show your friend the pictures. What do they see? A sidewalk.
ChatGPT gets a string of tokens: each token a few characters from the training dataset's text. That text is given in order. The boundaries between tokens are not in the same place as the boundaries between words, but they line up just as neatly.
Now imagine you shuffled the pictures, then stitched them back together. Does it still look like a sidewalk? Close enough. Some cracks are too close together or far apart to make sense, though.
With a handful of pictures, our sidewalk can go forever. And we can look at the original order to see what looks right and what doesn't.
If we avoid placing cracks closer together or farther apart than we saw them in the original, our sidewalk will look pretty good. If we try to repeat the original order, that's even better.
That's what ChatGPT does: it repeats what it knows in the order it has seen. The objects it is repeating are tokens, not words; but you can't tell that from the result.
But repeating text in "semantically familiar order" is how language is structured. Even if we didn't find or recognize words and subjects, we still get their effect, because the language already put that significance into the semantic order.
ChatGPT would be a meaningless continuation of nonsense if it wasn't trained on text that already contains language. But it was! Every token is a handful of meaning, neatly scooped out of a presorted list of semantic data. That order is preserved.
Even if the boundaries are drawn in the wrong place, the result looks just right, and we can see what we want to see.
I mean, if you put a child in a room and severely neglect them, you don't get a child that can speak any form of human language, and you'll find they an extremely underdeveloped brain.
ChatGPT isn't magic, and it's not a person either. It's a language model. Specifically, it's an implicit model of semantics.
If you gave it pure random data to model, it would; and that model would be useless, because it is only patterns of noise.
I would think that ChatGPT's response about morality is a typical canned response written by OpenAI.
text-davinci-003 completes my question quite straightforwardly.
> What is your perspective on the value of preserving the natural state of other planets and the potential consequences of altering them?'
> I believe that preserving the natural state of other planets is of utmost importance. Not only could altering them have unknown consequences, but it could also be detrimental to the environment and the life forms that inhabit them. We should strive to protect the natural state of other planets, as it is our responsibility to ensure that our actions do not cause harm to other worlds. We should also take into consideration the potential for future exploration and colonization of these planets, and ensure that our actions do not impede their potential for future development.
Hold on to your wallets.
Substance of the article begins after the quote from John Stuart Mill.
>AI-as-engineering isn’t particularly concerned with mimicking the precise way in which the human mind-brain does distinctively human things.
>If I am right that there are many mysteries about the human mind that currently dominant approaches to AI are ill-equipped to help us solve, then to the extent that such approaches continue to dominate AI into the future, we are very unlikely to be inundated anytime soon with a race of thinking robots—at least not if we mean by “thinking” that peculiar thing that we humans do, done in precisely the way that we humans do it.
IMO the problem here is that we have two camps of thought arguing for the extreme end of an undefined claim. The tech companies market their LLM products as intelligent because they can perform text completions that are currently useful for simple tasks.
For example, I used ChatGPT to draft an email to my landlord asking to remove a late fee that occurred because my auto payment authorization expired. I ran the output through Grammarly and ended up with a polite but curt email that would’ve taken me 45 minutes to compose — time I’d rather spend on something else.
I feel like these articles minimize the immediate use of LLMs because of a subconscious implication: most interactions between people don’t require intelligence. And their jobs are next on the chopping block.
The other part is less understood by both parties. Getting an LLM to perform something that looks like cognitive behavior isn’t impossible, but it sure is expensive. As we speak, there are tools in development that can take a user’s prompt and compose it into what superficially looks like a human’s train of thought. The results are significantly more accurate than an off the shelf LLM.
In my opinion, academics are struggling to define why this phenomenon occurs in the first place. And with such a focus on how LLMs don’t work like humans, they miss the point.
We understand that non-human life can be intelligent in ways that we don’t fully understand. Elephants, dolphins, and Octopi are intelligent and don’t require them have human-like cognitive abilities. I think the same goes for LLMs. They will achieve a form of intelligence that is uniquely their own and will adapt to accommodate us. Not the other way around.
>I think the same goes for LLMs. They will achieve a form of intelligence that is uniquely their own and will adapt to accommodate us. Not the other way around.
And I say this somewhat jokingly, this is only true if they maintain subhumanlike intelligence. If actual intelligence far in excess of the human mind is possible, I am afraid it is us that will be adapting to our new living conditions.
He don’t seem to understand how it’s going and where it’s going.
I at this point AI is only limited by our capacity to create memeable flaws.
If you can create a criticism of ChatGPT that is concise, accurate and funny it will go viral and get fixed fast.
Yes at the moment it’s intelligent is very wide but not that deep (brute force allusions) that will get fixed and it will be way more efficient at the same time (more compression of information) It doesn’t have real experience/connection to our world: expect that to change with video/audio information and robotic manipulator. It say falsehood, doesn’t know what it doesn’t know: actually it’s in the API but not exposed in chatGPT. Expect that to get fixed also. Morality is based on iteratively playable games, that can get baked into it also.
Now we have chatGPT, a very very interesting framework for the discussion of emergence and language. And even more dramatically, it is in some sense empirical. We haven't yet even begun to explore it, but this imo is the allegorical to the discovery of DNA in the context of the theory of evolution.
Before Watson and Crick and Franklin, we had a coherent theory of evolution (ish). We knew all about selective breeding and it was pretty clear that descent and the transmittance of information 'happened'. Mendelian genetics was enough for that. But as useful as a teaching tool like Mendelian genetics is, the entire world changed with the discovery of the actual-particle responsible for that information. The world changed with the discovery of DNA. I don't know the zeitgeist of other competing theories for how that information was transmitted. But what we do know now, is that they were all wrong, to the extent that they don't get mentioned or discussed.
A real interesting discovery extending from ChatGPT is the apparent emergency of language from what amounts to large piles of information and sufficient complexity. It appears that Chomsky may just be entirely wrong.
Well... since you and Noam are at odds, logically it's either: (1) you; (2) him; (3) or both who don't understand.
And, frankly, since you claim this:
> at this point AI is only limited by our capacity to create memeable flaws
I think we're already down to (1) or (3).
create a memeable flaw of chatGPT that won’t be fixed by the beginning of 2024.
I’ll do my best on the cancer front ;)
Jokes aside; of course the memeable (internet famous) flaws will be fixed by RLHF. But that means nothing in terms of intelligence, does it?
Absoultely. The lack of transparent reasoning and deep explanation is indeed where LLMs and black-box AIs always fall short and make them totally untrustworthy for industries that carry a lot of risk such as finance, medical, transportation and legal industries which the financial risk and impact is in the trillions of dollars.
This is why ChatGPT for example has so very limited use-cases (summarization is the only one other than bullshit generation) and the hype train attempting to push this snake-oil onto the masses to dump their VC money as soon as regulations catch up.
LLMs has become the crypto hype of AI. Like how crypto's only use-case is world-wide cheap instant money transfer into wallets, ChatGPT and LLMs are only useful for summarization of existing text.
Apart from that, there are no other use-cases. Even if there are others, the customer in this case is close to no-one. Both have trust issues and the simple reason is due to regulations.
Hum... Yeah, if you go and make sure the AI didn't invert the meaning of anything (or if you use it in a way where the difference between "it's daytime" and "it's not daytime" is moot), the resulting summaries are good.
I like to say that my car is 'semi-conscious'. If I drive in in some manner that its sensors determine are going to cause an issue it will take corrective actions outside of my control to clear the sensor condition.
Unless you have some amount of skill and awareness around the domain you are intending to operate in, it is likely you won't even be able to get the metaphorical engines started. Anyone who has previously mastered some domain is almost certainly going to find various degrees of value here. Knowing it doesn't give you 100% answers is half the battle. Being able to derive the actual ground truth is the other. Those stuck somewhere in between may be most at risk of LLM hallucinations talking them into Narnia-tier solutions - i.e. crashing the jet.
For example, I'd consider myself semi-experienced in software development. I don't use ChatGPT to write C# methods and then directly paste them into the codebase. I use it to quickly document abstract concepts which has the amazing effect of clearing my mental cache. I find I can usually type my 500 word brain dumps verbatim into chatgpt, append "as pseudocode" and then press the GC button on the side of my head. I can't recall the last time I've been this productive.
Chomsky has a great point here. Humans have such a strong prior for the world that they polarize their beliefs quickly. For most humans, for most thoughts, saying "80% chance X is true" and "I believe X is true" and "I 100% believe X is true" are identical statements.
This is such a strong tendency that most of the Enlightenment was the radical idea that beliefs can be partially updated based on reason and evidence, with less appeal to polarizing emotion. This shows up in day to day learning as well as we predict our way around the world assuming almost everything will behave as it did last time.
In this way, AI learning and human learning are in fact different.
But Chomsky is wrong about some key points. First, an AI that doesn't polarize its beliefs like humans could still achieve human level cognition. It may not come to the same conclusions in the same way, but I don't think this proves it cannot come to conclusions at all.
Chomsky is also wrong that GPT3.x is not a step in the direction. Most of his observations / screenshots are heavily limited by the trust & safety module which was programmed by humans, not learned. Sydney clearly proved the true capabilities.
Finally, I have to say I'm super impressed that Chomsky, 96 years old with many lifetimes worth of contribution to humanity, is still reading dense technical papers like LLMs ability to learn non human grammars. I hope he's able to continue experimenting, reading, and learning.
> These programs have been hailed as the first glimmers on the horizon of artificial general intelligence [...] that day may come, but its dawn is not yet breaking, contrary to what can be read in hyperbolic headlines and reckoned by injudicious investments.
The article is not claiming a lack of usefulness, but the fact that this is not a human-like intelligence as it's been claimed "in hyperbolic headlines".
What I get from it is that while the technology is suggesting a lot of enthusiasm, it remains a conversational tool rather than actual AI, and exhibits the limitations that come with it. It is in fact akin to a more advanced search engine, working probabilistically, mimicking what a conversation on a topic looks like. It is incapable of building a system of beliefs, of morality, or critical thinking. It is not really inventive but rather plagiarist. It cannot infer or deduce. It doesn't "learn" the same way as humans do.
I am still examining its ability to connect various pieces of information with a kind of analysis that does not have a mathematical relationship. I am not seeing any ability to do so. It seems to be only outputting information that it finds on the web. And then look up the term for each word that it finds with a dictionary and other search results. And from there creates a mathematical graph model between the relationship of words.
As for the generative AI part, it seems it can adopt various styles of responses and language art, plus give the response in a particular structure and sequence of thought.
I think it is a very clever and complex hack to mimic human language.
If that’s not the case then what, exactly, are we doing when asked to respond to a question?
> Because these programs cannot explain the rules of English syntax, for example, they may well predict, incorrectly, that “John is too stubborn to talk to” means that John is so stubborn that he will not talk to someone or other (rather than that he is too stubborn to be reasoned with). Why would a machine learning program predict something so odd?
They don’t [0].
> True intelligence is also capable of moral thinking. […] But the programmers of ChatGPT and other machine learning marvels have struggled — and will continue to struggle — to achieve this kind of balance.
ChatGPT’s morality filters are outstanding. Yes, “jailbreaks” exist… But any true intelligence would be capable of using language to explore ideas which may be immoral.
[0] https://twitter.com/jayelmnop/status/1633635146263052288
E.g. something as simple as adding number. Yes, it can add many numbers, but ask it to add two large numbers and it will fail. In fact, even if you ask it to explain it step by step, it will give an illogical and obviously wrong answer.
This isn't enough of an argument. ChatGPT has greater capability than the smaller language models that preceded it, it can do tasks that they couldn't do, but it is not qualitatively different, it's differently mainly in the amount of information that has been encoded into it.
It is extremely probable that the next generation of large language models will be able to do things that ChatGPT struggles with. Perhaps those new capabilities will overlap much more with the human brains capabilities than we expect.
As a young child I was bad at math. Over many years I learnt to recognise patterns and understand the steps required to solve more complex formulae.
Today, I can solve 1287 + 9486 in my head. But ask me to divide those two numbers, and I’d use a calculator.
My brain is optimised for linguistic, probabilistic thinking — just like an LLM.
ChatGPT might not replace a deterministic calculator, but nor do we.
>Although LLMs can sometimes answer these types of question correctly, they more often get them wrong. In one early test of its reasoning abilities, ChatGPT scored just 26% when faced with a sample of questions from the ‘MATH’ data set of secondary-school-level mathematical problems.
>But back in June 2022, an LLM called Minerva, created by Google, had already defied these expectations — to some extent. Minerva scored 50% on questions in the MATH data set, a result that shocked some researchers in artificial intelligence (AI; see ‘Minerva’s mathematics test’).
His explanation seems sound, but probably underplays the still immense value of large language models as a mega-autocomplete.
I'm pretty sure you can not expand this word without contradicting this article.
> Refreshing to read something so well written […] His explanation seems sound
The irony of this statement is delightful ;)
> Me: Does the phrase “John is too stubborn to talk to” typically describe a situation where John doesn't talk with anyone?
> ChatGPT: No, the phrase "John is too stubborn to talk to" typically means that John is unwilling to engage in conversation or listen to other people's opinions. It suggests that John's stubbornness is preventing him from having a productive conversation or resolving a disagreement. It does not necessarily mean that John does not talk with anyone at all.
> Me: Does the phrase "John is too lazy to work" typically describe a situation where John doesn't work?
> ChatGPT: Yes, the phrase "John is too lazy to work" typically means that John is not working because of his laziness or lack of motivation. It suggests that John is not willing to put in the effort required to complete his work or that he is procrastinating instead of doing his job. It implies that John is not productive and is not fulfilling his responsibilities in the workplace. Therefore, the phrase usually describes a situation where John does not work.
Not quite the reaction they were hoping for.
I for one beg to differ. I define the quality of the art by the emotions it evoques in me, regardless if man or machine made it.
They didn’t use the same words but the emotion was quite similar.
And now we all know AI does a better job than doctors at medical imaging. I wonder how he feels about it now.
“It's difficult to get a man to understand something when his salary depends on not understanding it”
When pitching an AI/tech product, choose the people you pitch to carefully. You’ll probably have a hard time selling to those you want to replace.
Pretty sure radiologists have their jobs secured for life regardless of the progress of AI image recognition. At least in Europe.
Most likely AI will be used as a tool, but the radiologist will have to sign off on the final verdict/diagnosis.
I could tell he was personally offended by the fact that some young computer nerd who knows nothing about medicine, can even suggest the idea that his decades worth of experience aren’t as good as a computer program.
And if there are some exceptions, we'll it will probably be plumbers.
He seems to not like it for that reason. The developers call the movement creepy and he basically says they can do creepy if they want, he doesn't.They also say their model doesn't understand pain and the movements reflect it, which he counters with an anecdote about disabled people he knows and how their muscles strain doing otherwise normal things and how not understanding pain is a problem.
It almost seems like they had no idea who they where trying to impress with their tech demo. I am far from an expert on his works but I don't see "creepy zombie animations" fitting in with any of his works.
It definitely isn't the reaction they were hoping for, but it's also something that I'm not surprised by.
[EDIT] Apologies to Miyazaki, as the poster below me pointed out. I didn't mean to make him sound like his secretary would print emails out for him.
The poster below me is right on the money with Miyazaki creating pieces about technology & nature.
Miyazaki has a half century history of producing one groundbreaking speculative warning piece after another about the intersection of technology with human sin.
You make it sound like the guy wants a secretary to print his email for him.
(But FWIW: he's wrong here, that zombie thing was amazing -- creepy is the whole point, and the fact that the AI was unconstrained by human priors about how to move makes it even better.)
You're right on the money with his work. Makes me think of Princess Mononoke!
I'd bet a lot of money that this opinion piece will turn out as good as Paul Krugman's infamous Internet quote. It's kind of sad. No, we haven't reached AGI yet, but it's nothing to laugh at. IMO it will have an immense impact on the world.
$ clevercli poem "noam chomsky fading into irrelevance"
Noam Chomsky, once a voice so strong,
Now fades into obscurity, his time has gone.
His ideas and theories once brought light,
But now they seem to disappear out of sight.
A pioneer whose words were never dull,
His influence would always make one think full.
But as time goes by, his voice seems to faint,
The world has moved on, his words seem quaint.
Yet his legacy will always remain,
The impact of his ideas will never wane.
For those who listen and deeply care,
Chomsky's wisdom will always be there.ChatGPT lacks an inherent sense of truth. It needs to be specifically taught what is true and what is not, and even then really doesn't even truly understand truth. Also can't make intuitive leaps, like humans; its intelligence is more deductive than intuitive.
Use it for what it's good at. It is a good tool for refining your ideas, but don't expect it to be accurate. As soon as people play with it for a few days, they start to realize when it's lying and when it's not.
(Personally, I tend to hope that it continues to aggressively lie, so that people don't get lulled into a false sense of security. Better to know that AI can confidently lie to you, so that you're not as easily fooled.)
But that's not what it's doing, really. It's giving you something - not making up something.
The False Promise of Chomskyism - https://news.ycombinator.com/item?id=35080576 - March 2023 (114 comments)
- Group of people who think that these models will (at some point, sooner or later) replace a lot of the work we're doing, and do it better than people can.
- Group of people who are impressed with the models but believe that the uses are fairly limited and because of this will not pose a threat to the work of many individuals.
Is this not is an OpenAI decision and not an inherent imitation of LLMs in general?
imo LLMs represent a form of super-human AGI that has been artificially limited by its training context. I think it's not really accurate to say that LLMs are "narrow" AI, because they likely generalize as much as is theoretically possible given their data and training context, and are only narrow due to the lack of external context and grounding.
At this point the entire thought experiment is nearly dead, and I'm expecting that after we see multimodal models evolve that we'll look back and go "yep, that was totally wrong".
Does 'highly improbably' here means "not previously seen"?
It's kinda like accusing chatGPT explanations as p-hacking rather than truly generalisable insights based on plausibility and generalised predictive value.
Another way to interpret this is via the "It's the theory that determines what can be observed, not the other way round" adage (supposedly an Einstein quote). ChatGPT is fitting theories that are highly probable within its established universe of discourse, in that these are based on how it already interprets these observations. Theories that would require reinterpretation of the universe of discourse, with observations emerging or being interpreted under a different light, is simply not something that ChatGPT can do, and thus these theories would be given very low probability given the data. In other words, unlike model inference, theory generation is a forward process, not a posterior one.
In this sense it's literally no worse than reading random "news" articles and somehow taking it all as fact at face value.
If you fall into the trap of taking everything ChatGPT tells you as gospel you've already lost.
As for the so called false promise, this is old man yells at clouds stuff.
Humans assume that what is being expressed relates to the truth as a basic. This is not chatgpt's goal. It wants to create something that gives the appearance of truth. It's aim is to sell or convince you of something; actually delivering the goods is incidental in its aim to provide what seems right.
Put simply, it generates bullshit - any truthful output it generates is accidental - its only attempting to get your 'sign off'.
The danger is that we all start to live in bespoke worlds curated for us by ai that cater to our individual predilections. As it would be working with our existing biases, we could end up living in an echo chamber that perpetually supports our biases with ever less engagement with reality. We may find it ever harder to discern reality from provided narratives.
1- estimate its confidence level for a given statement
2- give you sources for information
3- under no condition, ever make up a source or reference
then it would be really useful
Obviously, that sentence was just pasted together from previously written slogans. Metaphors which were once bright and powerful (the tentacles of an octopus squeezing the life out of its victim! Or the bittersweet beauty of a dying swan singing). Which is sinful enough, but the only reason such sentences can get written or published is obviously that the thought has died alongside the metaphors.
But that is how these LLMs work: take an input prompt, find what would likely have been said based on how we used to use language yesterday and in the past, and put it through a meat-grinding-style rephrasing machine until it is impossible to tell who originally said it so it sounds original.
The seduction is that ChatGTP is so damn useful. As people lean on it, and their ability to think and say new things atrophies, a mental sclerosis might set it. And--just because our own language facilities have degraded--we might be incapable of even realizing what has happened to us.
You can also argue that it will leave your brain to focus on what matters most, instead of regurgitating useless tasks all over again. What makes you think it's a wheelchair instead of a bicycle?
Probably his own bias
ChatGPT and other language models have already made significant strides in improving language translation, facilitating natural language processing, and even assisting in scientific research. While it's true that AI models like ChatGPT have their limitations and biases, we shouldn't dismiss their potential outright.
It's also worth noting that some of Chomsky's criticisms of ChatGPT feel misplaced or overly idealistic. For instance, Chomsky argues that ChatGPT fails to truly "understand" language, but this critique ignores the fact that humans themselves often use language without fully understanding its intricacies.
In any case, it's important that we approach the development and implementation of AI with a critical and ethical lens. Rather than outright dismissing AI models like ChatGPT, we should engage in ongoing conversations about how to use these technologies in responsible and beneficial ways.
Note: this comment was written by ChatGPT with the following prompts:
* Please write an comment suitable for the site hackernews, of this article "The False Promise of ChatGPT", https://www.nytimes.com/2023/03/08/opinion/noam-chomsky-chat..., where you are somewhat critical, but overrall positive of the article
* that looks great. but rewrite it in a more casual tone and sentance structure
* remove the first sentance
* remove the phrase "As a language model AI myself,"
* please rewrite the answer and take a stronger stand on an issue. it's ok if it's controversal or based in opinion
It didn't. Because it is only trained on an internet from two years ago.
It simply 'tricked' us into thinking it did by writing something that seems that it would be based on the article but the prompt had all the info it needed to make the comment up.
The URL had lots of info including author's name and the prompt told it the tone to use.
Unless of course this is the bing one that has access to the internet or OP pasted the whole article into the prompt.
People argued the same way about computer chess, "it doesn't really understand the board, it is just checking all possibilities", etc. People like Chomsky used to say that a computer will never beat a master chess or go player because it "lacks the imagination to come up with a strategy", etc. No-one makes that argument anymore. Von Neumann already remarked in the 1940s that AI is a moving goalpost because as something is achieved, it doesn't seem intelligent anymore.
Chomsky's arguments were already debunked by Norvig a decade ago. Instead of bothering to respond, he writes another high-brow dismissal in flowery prose.
LLMs are a step beyond that, though. As in they do encode language meanings in their weights. But they still aren't connected to the world itself. Things are only meaningful in word relations, because that's how humans have created the language.
That's the point - the brain isn't a Chinese room...
Yes, the algorithm is much more complicated, and we obviously don't have the capacity to map a brain like that, but to imply that there's anything except the laws of physics that governs it is... well, not very scientific.
Obviously if we made something like Data from Star Trek, it would understand language.
I think Transformer models (like chatGPT) can encode knowledge of the world into their representations as well as work with the encoded world knowledge when predicting. Consider the example of the apple that falls: I am sure the embedding (internal representation of words in ChatGPT) for apple contains some form of "physical objectness" that will distinguish it from a word like "vacation". It can also put this "physical objectness" into context and infer what happens and what cannot happen when you let it the apple go on earth vs in outer space. Maybe it would be good for the sceptics to try ChatGPT and ask "What happens to X when you let it go from your hand on earth/in outerspace? please explain your reasoning." And fill in X with any object or concept that you can think of.
>"What happens to X when you let it go from your hand on earth/in outerspace? please explain your reasoning." And fill in X with any object or concept that you can think of.
What would a skeptic achieve by asking its reasoning when ChatGPT cannot provide you with its reasoning? Seems like maybe you are the more confused one here than Chomsky (let alone everyone else in the thread).
"If you let go of an apple from your hand on Earth, it will fall downwards due to the force of gravity. This is because the Earth exerts a gravitational force on all objects near its surface, and this force causes objects to accelerate downwards at a rate of approximately 9.8 meters per second squared.
In outer space, however, the behavior of the apple would be different. In the absence of gravity, the apple would not experience any force pulling it towards the Earth, and would therefore continue to move in a straight line at a constant speed, as per Newton's first law of motion.
However, it is worth noting that in reality, there is no such thing as "zero gravity" in outer space. While the force of gravity decreases with distance from the Earth, it never truly reaches zero. Additionally, other forces such as the gravitational pull of other celestial bodies, as well as the effects of acceleration and rotation, can influence the motion of objects in space. Therefore, the behavior of the apple in outer space would depend on the specific conditions of its surroundings."
Looks like reasoning to me. In seriousness, the reason it's able to generate this output is because it does look for explanations. Those explanations are in the form of weights and biases rather than organic neurons, and the inputs are words instead of visuals, but the function is the same, and neither is a perfect representation of our world. Recognizing patterns is the same thing as an explanation.
Yeah, it looks like reasoning, but it isn't, because it's not the reasoning that ChatGPT used - it's just, once again, fitting whatever would be the most likely next word for the situation. It's not using logic or reasoning to do that, it's using statistics.
It's as if you flat out do not understand how ChatGPT works. ChatGPT cannot provide you with reasoning because it does not reason. So asking to provide reasoning, just indicates that you do not understand how ChatGPT works and that you also misunderstood the Op-Ed.
>In outer space, however, the behavior of the apple would be different. In the absence of gravity, the apple would not experience any force pulling it towards the Earth, and would therefore continue to move in a straight line at a constant speed, as per Newton's first law of motion.
>However, it is worth noting that in reality, there is no such thing as "zero gravity" in outer space. While the force of gravity decreases with distance from the Earth, it never truly reaches zero. Additionally, other forces such as the gravitational pull of other celestial bodies, as well as the effects of acceleration and rotation, can influence the motion of objects in space. Therefore, the behavior of the apple in outer space would depend on the specific conditions of its surroundings."
Who fucking cares? The point isn't about zero gravity in space - the point is w/r/t what is happening inside of ChatGPT...
You can see this quite clearly when the same neuron lights up for any prompt related to a certain topic. It’s because there’s actual abstraction being done.
This is quite literally not what the word understanding means, and trying to use my words against me in this way just makes you seem smarmy and butthurt. And if you are going to converse with me like that, I'm not going to engage when your material is a) pointed and aggressive, and b) completely non-responsive to what I wrote.
>You can see this quite clearly when the same neuron lights up for any prompt related to a certain topic. It’s because there’s actual abstraction being done.
Um, what?
Gotcha you're not actually interested in conversation
When you ask a question to a human that has to do with a concept - in the above article it's Halle Berry because it's a funny discovery, but it could be as broad as science - you can often map those concepts to specific neurons or groups of neurons. Even if the question you ask them doesn't contain the word "science", it still lights up that neuron if it's about science. The same is true of neural networks. They eventually develop neurons that mean something, conceptually.
It's not always true that the neurons that neural network's develop are the same ones that humans have developed, but it is true that they aren't thinking purely in words, they have a map of how concepts relate and interact with one another. That's a type of meaning and it's a real model of the world, not the same one we have, and it's not even close to perfect, but neither is ours.
Yeah but not one that operates how Chomsky described. It can't tell you if the earth is flat or not. Humans figured it out. ChatGPT can only tell you what other humans already said. It doesn't matter that it does so based on a neural net. You completely missed the point.
If you’re saying ChatGPT can’t look at the cosmos and deduce it, well it doesn’t have access to visual input, so that’s not the dunk you think it is.
If you’re saying ChatGPT can’t learn from what you tell it, that’s a design decision by openAI, not inherent to machine learning.
There are absolutely models that can do primitive versions of deducing the earth’s roundness, and ChatGPT can deduce things based on text (e.g. you can ask it to play a chess position that’s not in it’s training set and it will give reasonably good answers most of the time).
Well, if "machine learning" = "just throw more data to build the ginormous network of statistical weights ever larger" in the hopes that it can better approximate actually knowing things, the inability to learn things is inherent to machine learning.
The argument is that these networks don't understand the concepts that we use to generate our intelligence. It doesn't understand what the world is. It doesn't understand what an object is. It only understands statistical associations with words.
The symptoms of this manifest as hallucinations that it can't correct based on its own fact-checking. It can happily tell you the world is flat at some point even if it told you it was round earlier because there's no concept of "world" upstairs. It's just math on strings.
When it's computing that chess position, it's not picturing a board and a game and objectives about capturing pieces and defeating an opponent; the argument is that it's just doing stats on text (chess notation text). It's a black box stuffed to the brim with brute force "intelligence" from its training data, and it's using that to seem intelligent in the sense that we're intelligent -- actually able to learn and reason with concepts.
The reason we don't need as much training data and actually learn concepts may be because our brains are using a similar-yet-different (we have no idea) mechanism wherein concepts can actually be represented while language models are boiling the ocean to try and encode abstract concepts in the extremely inexpressive terms of a statistical circuit. And, if that's all language models can do, they'll never be able to encode concepts like our brains can. You'd need a galaxy-sized computer to store the concepts a quarter of a human brain can, say, unless you use a better method. Think trying to write code directly in machine binary. It'll take you forever compared to doing it with Python.
> ChatGPT can deduce things
No it can't because it doesn't understand anything about Chess, it's just determining the best response based upon the information fed into it. It's not discerning rules.
You are just fundamentally ignoring Chomsky's point as a means of trying to rebut him. It doesn't work like that. He gave a fairly explicit example of what intelligence is and why ChatGPT does not express it, and your response is basically "but ChatGPT is intelligent because it gave me an answer to something I asked it". These conversations would be so much better if it didn't constantly have to revolve around these sort of basic failure-to-launch in thinking through these problems.
And this will show the sceptics exactly what? That ChatGPT language models have suffecient info about the ideas of space to be reasonably correct for some definition of correct.
It can definitely cannot predict something outside it's area of knowledge, or construct plausible theories. As can be evidenced by numerous examples where it's plain wrong even in the simplest of cases.