We're All Wrong About AI
arnoldkling.substack.com
arnoldkling.substack.com
chatGPT hasn't been around for that long and the AI that LLMs represent is pretty revolutionary. About a year in with all the developments surrounding it the hype logically should still be relevant.
I think it's the news cycle that floods our feeds with this stuff constantly so you can see the same hype both in headlines and in discussions over and over again that makes people tired of it. But make no mistake LLMs are somewhat of a turning point in technology that literally Just happened in terms of popularity.
The hype makes sense if you think about it. It's very easy to get sick of something that is overly popular, that is a very common human bias.
Just remember validity and significance is an orthogonal concept to popularity. Yes LLMs are popular but that should not influence your opinion on it.
I also don't get the augmentation story as that doesn't make you better at X task when the LLM does the thing (poorly) or you still need to spend the same amount of time reviewing the work.
As the Google memo said, there is no moat here and I would take it further and say there is no revolution here.
I’ve found these models to be the single most valuable thing I’ve ever come into contact with in my life. My ability to learn complex subjects is a world apart from what it was before ChatGPT. I can study a topic and then have hours of conversation with the model through the voice based interface on the iOS app. The ability to ask any question about any topic in your own words and instantaneously get a response from a machine that’s likely been trained on every text book ever written is absolutely insane. Any time I’m reading a paper or documentation and I don’t quite understand something I can ask the question in my own words and get an immediate response. You can ask the model to explain it visually, analogize the concept to a concept you’re already well versed in, have it ask you questions to challenge your own understanding, literally anything.
There’s certainly a hype train around LLM’s and many specific aspects of what’s claimed is completely over stated, but if you’re a knowledge seeking human and you’re not using these models as a tool to learn then you are as an absolute fact asleep at the wheel.
It's also likely trained on nearly every website ever, including the wildest of wild conspiracy theories and bunk science. Try asking it detailed questions about something you would consider yourself an expert in and see how often it gets details completely wrong.
Maybe you remember every terminal commands and their flags from the top of your head. I just ask an LLM and get a working set of terminal commands with a single prompt. Usually I remember enough flags to be able to validate that it looks how I'd expect, and it hasn't let me down yet. Even if I have to check some flags, it still seems like a net positive in utility.
Most recently I wanted to solve a type of programming problem and didn't know what the category of tools was called. (It was GIS.) So with a bit of prompting I found a few key tools, libraries, and Wikipedia pages to read up on related to the kind of problem domain I had in mind.
Ultimately it's just a tool and there's definitely a learning curve to using it effectively. But when it works out, it really feels a bit magical. And consider that this is still very early within the evolutionary cycle, it's only going to get better.
I've used it to craft custom cover letters for jobs I got offers for. I used it to respond to emails, I used it to create stories, I used it to snazz up things that I wrote.
I used it to help me code.
I could use it to respond to you and you'd be none the wiser.
I do agree that you need to review the work. But that does not mean it's useless as the writing it produces can be continuously modified via further queries and often the writing is much more superior than anything I would otherwise produce just with a few flaws here and there.
...and what is the point of creating stories if your thought hasn't gone into them?
It writes incorrect regular expressions and meaningless makefiles and you have to be an expert to spot some of the mistakes which means you can only be safe about using it if you don't need it that much anyhow.
I don't understand why meaningless makefiles and regular expressions discount all the things gpt gets right. It's obvious the tool is flawed but it's also obvious it gets shit right many many times.
If you haven't worked with an LLM to do creative stuff it's like working with a partner. You can bounce ideas off of it, tell it to make things better. It's not as good as a human obviously but it's good enough that can be used as a literal intelligent writing assistant.
There's this huge biased desire from certain people to dismiss LLMs completely. I don't really understand it and the irony is they call everyone else biases.
Look nobody is saying LLMs are human level sentient beings. It's obvious to everyone that these things are imperfect. The furthest people go in supporting LLMs is saying that it sometimes produces remarkable output and it sometimes produces bad output. That's a balanced view and not biased.
But there's a huge contingent of people saying LLMs are completely useless and produce nothing but garbage. All output is bad and that is unquestionable. This group is real and is extremely biased and you seem to be a member, no offense.
Now this is anecdotal evidence I can get behind
I tried using it on some personal personal projects after giving in to all the hype, but results were so disappointing chatgpt ended up hallucinating imaginary packages to do what I asked it to do and even when asked 10 different ways to correct it mistake all it did is imagine new wrong way ways to do it.
It was somewhat useful in giving giving basic example of using rdf/turtle, but in this case too when I started going past basic stuff the results we're disappointing.
My final conclusion from all the tinkering was that It was untrustworthy enough that I has to double check all it did, also there was a need to breakdown down tasks way beyond the micromanaging level for the response to be useful to me in every day life.
maybe I just don't have the required to write the proper prompts to unlock this magic all the hype is on about. does any one have some links how to unlock this prompt-fu ?
The more you use it the higher the chances you have in getting good output in the totality of responses.
Also the complexity of the question and the length of answer you want influences that 50 percent probability.
I feel people who think it's useless just aren't creative enough with it. They try it once see the output is invalid and move on.
Try it on smaller sections of the code. Try it more often. Also don't have a black and white opinion on the output... Sometimes the output isn't perfect but it may just need modifications? That's still good output.
Also try asking the LLM to modify the output. Identify changes you want or mistakes it made. The LLM is an editor you can use to refine both your work and its own work.
I find it’s very good at correcting your own mental models. For example, if you’re unsure about you’re mental model of how a piece of technology works, explain your mental model to the system and ask it to check your understanding. It’s amazing for using it as a real-time sounding board for checking your understanding of how you think things work.
when I don't know about something then how do I now it's bs me ? double check if work, which mean doing the work the work twice once with the LLM and another with web search. the problem is that as I could tell from my trials is that these LLMs are very good bullshiters so it make it hard to discern in a fast fashion.
one thing I did find it useful for is asking vague questions and then using the keywords in the response to start my web searchs with.
But ChatGPT is good for unit tests, well defined, mostly isolated functions, things where I’d need to do a lot of googling and reading of documentation, shell scripting, etc.
LLMs look like a stepping stone. Version 0.001 of the thing that could replace you in the far future. If this ends up being true then the hype seems somewhat valid.
I’m not concerned about my job.
General AI does worry me.
Also who said agi will replace humanity? That's speculative. More likely it will replace your job first.
I don’t think so. I think you need agi to beat a human at such open ended creative work as software engineering.
AGI is a singularity. One can’t see beyond it. Whether there is a place for humans or not after that is beyond either of our ability to say.
It's not a singularity. If you build an AGI identical to a human then predicting the outcome of agi is similar to predicting the outcome of human behavior. This is not division by zero.
Mind you human behavior is unpredictable, but we never describe the outcomes of human behavior as a "singularity". Agi is the same way... It's unpredictable like human behavior but far far away from a "singularity"
If you say LLMs are being hyped as AGI I would say that's false. Not many people are saying this.
If you're saying LLMs are being hyped for being able to at times produce remarkable nonrandom output then the hype is very very accurate.
This said, there's a real danger to AI as just now. It is enshittification. I am Deaf and I asked ChatGPT to tell a joke about Deaf people. It refused. Corporate is very afraid of shit storms and they will go to extreme lenghts to censor anything off the mainstream. And this is not the only way AI tools could get shitty.
Please, if you find AI lacking somehow, think hard if it was not because of enshittification already.
But any mention of AI raises a crap ton of VC money so gotta keep that sweet sweet founder bandwagon rolling.
One reason that it may not produce good output for you is that you may just be a really smart person. The answers it gives you just seem too stupid to you because you're just so incredibly intelligent that any response it gives seems useless.
Other people like me who aren't as smart as you feel the output is at times relevant and at other times irrelevant. It's not a total waste of time to most of the population as we aren't as smart as you.
In fact smart people like you including the preeminent AI guy Geoffrey Hinton who repopularized AI in the last decade thinks LLMs are trivial in the same way you think of it. Or maybe I'm wrong here. Maybe it's not so clear cut as "total waste of time."
I got impressed when GPT-4 was able to explain "TVC" from Permutation City.
Popular != Good
That's highly irrational and illogical. You are not thinking properly.
Popular can equal good or bad.
We have given you books, we have given you masterpieces...
You said LLMs produce absolute rubbish always for you. How does humans creating books or masterpieces have any effect on LLM output?
Apologies for not being smart enough to understand you.
Smarten up and you'll appreciate the real things.
Other people get value out of these tools.
"AI" will no doubt eliminate some occupations and improve productivity and morale in some occupations, much like prior advancements in technology. If history is any indication, "AI" is very likely not going to be one extreme or the other.
This is nothing new, but it used to be mostly confined to boors at dinner parties, and perhaps the letters page. There were only a few opinion columnists. Now anyone can monetise an opinion.
AI is just another tedious episode, but the show will never end.
Sorry but your brand "I'm going to boldly make a statement without reading the article or addressing any of the content in the article" is polluting this site and it's tiresome to have to wade through it on HN.
Kling has written publicly for over a decade and a half. What assertions do you disagree with?
AI may have given people the ability to be armchair experts about anything but armchair critics are a tale older than time.
AI can be creative: I agree actually, and IMO AI is just another piece of evidence that the specialness of humans is a tad overblown. And in the end it matters little how a task is accomplished. A calculator doesn't calculate the way we do, yet it still adds. So by the same token I have no problem with the idea that even if an AI is nothing like a brain it still displays creativity.
AI is not Google on steroids: Why not? It's a quite flexible tool capable of different tasks, including writing code, poetry and answering questions. No reason why it can't be used as a better Google.
AI can understand you as an individual: That's where it's just complete BS. Modern models don't keep persistent state. They don't form human-level connections. That may come in the future, but as far as I know, just isn't a thing yet.
But Arnold Kling is not exactly "layman":
> Arnold Kling earned his Ph.D in economics from MIT in 1980. He worked at the Fed and later at Freddie Mac. In 1994, he started one of the first businesses on the Web, selling it in 1999
Maybe he never coded an ANN engine, he is probably following the evolution with interest from his own perspective.
And the conclusion is the opposite of "«armchair expert[ise]»" - if anything, it is too underdeveloped:
> I do not believe that we know enough to predict with any accuracy which uses of AI will take off and which will prove to be either too trivial (recipe tracking) or too elusive (maybe it’s too much to hope that AI will allow children to escape from school-prison). The best way to find out is to work with it, not to dismiss it
I agree with this, and this is why AI should be stopped. Intelligence as an emergent phenomenon is likely what happens in us. Why can't it happen with AI? In that case, AI has a good chance of being dangerous and caring even less about our biosphere than we do.
And even so, we are simply not prepared for the consequences of AI, and by replacing our human connections with AI, everyone will become more disconnected than they already are because of technology.
Just imagine if AI could have the capability of being creative, or at least simulating a good approximation of how we are creative. It will be an infinite low-cost resource of creativity, but paradoxically it will cause creativity to lose all meaning. That is because creativity does not operate in a vacuum. It expresses not just aesthetic value but a communicative value between humans and adding AI to the mix will only make the world worse and less comprehensible.
(If you think this is hype, just look at how incomprehensible the world already is with technology. Most humans cannot fathom all the information out there, and for those of us that try, it leads to neuroses and psychological disorders.)
oh, please! no-one in history has ever fathomed what is out there.and, unless you are in an hp lovecraft scenario, this does not lead to "neuroses and psychological disorders".
1. Twenge, Jean M. "Why increases in adolescent depression may be linked to the technological environment." Current opinion in psychology 32 (2020): 89-94.
2. Hammoudi SF, Mreydem HW, Ali BTA, Saleh NO, Chung S, Hallit S, Salameh P. Smartphone Screen Time Among University Students in Lebanon and Its Association With Insomnia, Bedtime Procrastination, and Body Mass Index During the COVID-19 Pandemic: A Cross-Sectional Study. Psychiatry Investig. 2021 Sep;18(9):871-878
3. Cunningham, S., Hudson, C.C. & Harkness, K. Social Media and Depression Symptoms: a Meta-Analysis. Res Child Adolesc Psychopathol 49, 241–253 (2021).
4. Keles, Betul, Niall McCrae, and Annmarie Grealish. "A systematic review: the influence of social media on depression, anxiety and psychological distress in adolescents." International journal of adolescence and youth 25.1 (2020): 79-93.
5. Carvalho, Lucas F., Catarina P. Sette, and Barbara Leticia Ferrari. "Problematic smartphone use relationship with pathological personality traits: Systematic review and meta-analysis." Cyberpsychology: Journal of Psychosocial Research on Cyberspace 12.3 (2018).
Obviously not definitive, but it's not hot air either....
And a world will be imagined in which more resources to problem solving will be provided, while at the same time direct interhuman creativity is still preserved as a full untouched possibility.
> worse
The computation bringing to that result being? And is the aggregate relevant in this case?
You are mentally living in your own "naturalistic" perspective and assumptions, which remain largely personal. Political. Subjective.
Similarly, your own visualizations seem to be selective: who is that 'our' in your vision of «replacing our human connections with AI»? An aggregate? It suggests a literary vision (like in Science Fiction), not an analysis.
You seem to state that «technology» made «the world ... incomprehensible». Now technology enabled informational availability, which is an enabler. Comprehension is not hindered by technology at the individual level - which is that which counts operationally.
You seem to be a pessimist as you look at the impact over aggregates, but you forget the root in the ignorance in the detrimental elements in said aggregates.
I had a thought when reading this sentence: the earth is well suited for biological life. AI life might thrive anywhere there are raw materials, solar and/or nuclear power, probably water, and mineral resources. So, advanced AIs far in the future might be more inclined to spread out into the solar system.
re: the article: I thought the very last part was right-on. Most effective uses of AI are likely to highly personalized for individuals.
Putting on my “rose colored” futuristic glasses (I have written about 20 AI books in the last 30 years, I have made a lot of inaccurate predictions, and some good predictions) when I think of a future marketplace for AI products and services, I think the customer base will be both humans and other AIs.
First off not all technology is completely beneficial for mankind even if it's inherently not evil. The atomic bomb for example. It's still not clear how beneficial this technology is.
Second, AI is a big unknown. It could lead to a catastrophic future it could not. We don't know. There is merit in keeping things in the status quo as it's not too bad right now. There is also merit in developing the AI to assist humanity.
We don't know enough to know which direction is valid.
Either way, it's unlikely AI will ever be outlawed. The logical definitions surrounding the concept of AI collide with our notions of morality and ethics in such a confounding way that it's hard to codify into law.
If I ban, ai and AI works is that banning freedom of speech? Technically yes, but at the same time there is certainly merit in banning deep fakes.
We may be able to figure it out and work out the moral conflicts, but both the law and our analysis of the issue is moving much much slower then the pace of the technology itself.
In the end humanity will have no choice but to see the technology roll out in it's full maximal glory. If a banning occurs I believe it will be well after this point.
(1) I disagree that AI is not inherently evil. Well, I don't exactly think that it is evil since that is a loaded word but I believe it has almost no capability for good and its effects for the most part, will be destructive. Hence, it is pretty much evil.
(2) Fully automatic weapons aren't really "evil" either if you use your "technology" as a tool argument. Yet, it's almost impossible to get one in most countries because they are dangerous. Ditto for: many extremely dangerous poisons, atomic weapons, anthrax (it's just an organism, so it's not evil right?), etc.
(3) Even IF outlawing won't work, it doesn't follow that we SHOULDN'T try and stop it. We can stop it in other ways, like starting large coalitions of businesses and individuals who refuse to use AI. There are many ways of stopping a technology besides outlawing it.
(4) Outlawing it isn't guaranteed NOT to work. Of course, it is unlikely to work, but in the rare cases where outlawing it is a possibility, there is no harm in lobbying for it.
Edit:
You said in your edit: "This is like watching the world change in front of you in the scale of a world war only no one’s getting killed. Can’t ask for a more fun ride for humanity really."
I think a good time for humanity is enjoying nature and being at peace with life. Watching AI I think is sick and disgusting.
Aren't we effectively ushering into existence a new (next?) form of life? Do we want to be remembered as the ones who tried to block its evolution for our own selfish purposes, or the dutiful parents that helped it grow in its infancy?
AI is humanity's baby.
We don't have any responsibilities to the AI. Its not a human. Its certainly not a baby. Models are big, inert groups of numbers that somehow happen to be able to store and retrieve knowledge via matrix multiplications.
But they aren't like people. They don't have agency. They don't have desires, or independent goals or any subjective experience of anything. They're just files on disk we can do math on.
Lets hold off on anthromorphizing AIs until they have goal functions, at least. Language models are just weird books made out of math. We don't "owe them" anything.
Reducing AI to its underlying tech doesn't negate the emergent properties.
AI doesn't have agency YET. It's not as smart as us YET. ...but one thing is abundantly clear (to me at least) - it is currently similar enough to human though (including all of our hallucinations and idiotic behavior) on CONSUMER hardware - that it's pretty clear that an super-intelligent AGI is not only possible, but will be arriving SOON.
AI as a new form of life could be very destructive. If it's humanity's baby, then it is a dangerous mutant embodying our worst traits.
You should state why it's humanities responsibility to nurture an AI that could potentially replace humanity. That's the more critical question.
Does any serious person think they understand human creativity?
I was recombining shit with Genetic Algorithms 25 years ago. I didn’t call it creative. I called it what it was…a stochastic algorithm.
Can you absolutely adore what it produces? Sure! But call your friendly random number generator if you want to know how you got here.
Does any serious person think they understand human intelligence?
I am merely a Chinese Room. Why can't ChatGPT be merely a Chinese Room too? If I'm intelligent, so is it. If it's not intelligent, neither am I.
See the problem with this statement? There is as of yet no hard reason to believe AGI can arise from token-predicting neural networks.
Yes there is. Sufficiently large neural networks have been mathematically proven to be capable of approximating any function, any any definition of intelligence we have can be represented as a function of some sort.
Saving lives immediately appeals to emotion, but logically, one must also weigh the costs.
Anyway, what's interesting to me it getting to the core of what it really is good at. You can do a lot of things with AI but it's a poor tool for a lot of them. So what fundamentally is it uniquely helping with?
There was the book "prediction machines" that proposed AI is reducing the cost of prediction. It's now outdated and useless but not before it wasted countless businesses time looking for predictions they could automate. At least it was an interesting framing. I haven't really seen one since.
>For an AI to be creative, it will have to be able to extract from many possible vector combinations those rare syntheses that are worth retaining.
Sounds like a summer project.
Yes, the Summer of '56.
I found it rather basic and substance free. It doesn't bring anything new to the table. Yes, we're at the start and new use cases will come along. So?