Ask HN: Can anyone name anything new that AI has done?
We're 2 years into this. Still looking for something that an AI did all by itself that is actually new. Everything I've seen so far still involves a human sussing out the new thing or the model is just aggregating data that is already there from disconnected sources that never got tied together.
I think it makes sense to start the discussion by identifying something that a human has done that is "actually new", and then we can use that as an anchor to discuss things that AI has accomplished. If we don't set the framing around the question before we start discussing it, I'm afraid the goalposts will be moving all over the place.
It's definitely pretty hard. Maybe there is a better question to be asked. Can training a model with specific groups of ideas redacted from the training set and see if it can recreate something it shouldn't be able to? Possibly to reach an answer for "can it actually create?"
Jazz
> start the discussion by identifying something that a human has done that is "actually new"
Yes, of course, almost everything you know and use in your life was done by other humans and, at some point in time, was "actually new". That includes LLMs and various other bots, made by humans, trained by humans and on human-generated data.
Up to this point, AI has only managed to achieve inflation, chip shortages, market bubbles and more unemployment. Of the four, only the chip shortage is actually new, so AI has achieved something too.
I was more getting at the pitfall where someone lists a movie, and someone else comes along and says "well, that's obviously derivative from this earlier film and this book series written 10 years prior..." So what I'm getting at is that most of what humans create may not be "actually new", but rather derived from something "actually new", which we know AI can already do.
Can you provide a completely made up example of something that would fit your definition of something that "is actually new" here?
I'm finding this question difficult to answer without a concrete example of what would pass the bar.
Could an AI have invented the spreadsheet in 1978 (pre-Visicalc) based only on known tasks that it would have simplified (i.e. not from a description of a spreadsheet app)? Spreadsheets were the one of the "killer apps" (https://en.wikipedia.org/wiki/Killer_application) of its era that played a large role in driving microcomputer adoption.
Can or has AI invented a new killer app yet by synthesizing known human workflow needs autonomously from a description of those needs, not a description of the resulting app? I'm not necessarily willing to say no because of the ability of AI to synthesize information across a vast array of problem domains but I'm skeptical.
For the math nerds, AI has been successful at providing proofs/disproofs of famous mathematical questions that have been unsolved for decades. What examples are there where AI has proposed new mathematical questions that the mathematical community recognizes as worthy of being famous because they would provide significant new insight into mathematics or open up new fields of study in mathematics, without being specifically prompted with the outline of the question? Can or will AI be able to come up with the next set of Millennium Prize Problems (https://en.wikipedia.org/wiki/Millennium_Prize_Problems) by only being prompted with the need for worthy problems?
Could or would? I mean if I had Claude and said I want to make a computer version of a spreadsheet I'm pretty sure it could give me sufficient help to do it. If I said I want to make a great app... I don't think it could do that but then again I don't think it could do that now.
What's funny is that if OP directly claimed that AI never did anything "actually new," most of the people here acting confused would suddenly know exactly what it means and why it's wrong, without bothering with any definitions. They'll just say "humans don't create anything new either" and consider the matter closed.
edit: or just flag the thread, I guess that works too.
It's just funny as hell. I got what I needed out of this thread and will pose this question every month when I conveniently remember to ask. There's really interesting tribal splits going on. Even being flagged is a lot of data to take away from this.
I really did just wish people would run with it and make their claims of what's new and discuss why it's new, but people don't do science anymore.
https://github.com/openai/math among many others
Maybe getting a bit philosophical here, but isn't AI just solving math problems using existing methods? It's not making up new types of math (like inventing calculus for example).. or is it?
These were unsolved problems people had tried and failed at, some for a century.
Have you looked through this repo? I'm not seeing enough here to really answer the question posed. Like some of this repo, I know came from correlation of information not connected before (like the quasi-Riemann result). Some of it is just an extension of ground work already done, not necessary new in a sense that should be interesting. We know word calculator can calculate. There's probably a lot of human credit missing here.
---
Edit: I won't be entertaining this thread further. Let's just discard the scientific method and talk out our asses here.
> Like some of this repo, I know came from correlation of information not connected before (like the quasi-Riemann result).
We give Fields medals to people who do things like that. There's no particularly good reason to bury your head in the sand and pretend AI is different. It's not surprising that the first big discoveries involve standing on the shoulders of previous researchers and their work. When work is described as "new" in some way, that's usually all we mean by it.
I think if you are strong headed enough, you can ignore that the sky is blue and day is light
Some of us lack reading comprehension and show up defending a a repo as proof without ever having read what's in it. We are not the same.
Doesn't read it, doesn't parse it, accuses others of "talk out our asses".
So you already know that it works you just don't admit it?
The school system has failed you. Can you please go review how the scientific method works and report back with a citation as to what you think proves that it has already?
On my end, I'm literally open to being impressed by a verifiable instance of "LLM did something that only humans can do cognitively" and will even settle for a discovery that we literally could have never figured out even with all our data to date.
But people like you show up and want to argue that I have no clue what I'm talking about.
I don't expect anything from you to be honest, because your low effort scoff at a simple question doesn't leave me think you have any integrity to follow up with any words of substance.
> the model is just aggregating data that is already there from disconnected sources that never got tied together.
Ignoring if it can do anything "new" by itself, is it not worth it as a tool if it can turn hours of work that a human would normally have to do synchronously into an async session done in minutes?
There's plenty of valid criticisms for the uses of LLMs, but you pointed out an unbelievably useful case for the tool that you're trying to question.
LLMs are cool and useful, but that's not the question being posed. Next time you're in a group of people you trust to talk about AI with, ask them what they think it can do, then try to figure out if they know the limits of the tech. Right now I know many people that love to argue that it can create new things, but never once have I had anyone prove that it can. I also don't think it can because math, but I'm open to seeing something genuinely new that hasn't been done before, done by the LLM all by itself.
"Actually new" seems worth defining. What is an example of something (even human created) that is "actually new" and does not have any origin in "data" (in the most general sense, like one's memories or knowledge) being aggregated in some way?
Development of programming languages? Building the transistor? Like, pretty much every human invention ever.
Yeah I've read so much about how it's going to solve all these problems. So far it's only created and exhasterbated existing problems
protein folding problem? https://deepmind.google/science/alphafold/
AlphaFold made no contribution to the protein folding problem. It merely made a dramatic improvement to the process of static structure prediction, which is somethign that was already being done using probabilistic models.
(that's not to denigrate what they did- it was a great achievement- but not really anything "new").
There’s a difference between protein folding and protein structure determination. The former is an energetic problem that is very difficult to model. The latter is something they’ve been doing via similarity for quite some time (protein homology). Improving the latter, while useful, isn’t “impressive”.
It is the modern world. Everything has been thought of and there is nothing new.
or maybe
What would something new even look like? Would we recognize it?
There's like different levels of capability I suppose and most of humanity by in large is just remixing. At some point in the past though, the general purpose computer was probably a new idea. Someone had the idea first and someone eventually implemented it, so there's more than one place that can be measured for the thing being "new".
It is a bit far fetched, but if LLMs showed up with a viable means of time travel or making a couple of astronomical units traversable for humans somehow, that would inherently be provable and very unlikely be something we already have enough data to solve at all.
To your point though, maybe an LLM out there solved ant language and knows how to communicate with the ancient species, but the reviewer seeing it didn't understand the significance of the output and it found it's way into the recycling bin.
Really good question.
Vulnerability research and someone is definitely abusing the unreported.
It's given customer service departments at companies the ability to say "I don't know, that's what the AI says", then you get the blank stare and realize they won't help you.
Yep, it is all over. Nihil novi and all that. It it is totally pointless. No sane person uses AI. Can you step a little closer over there? Don't worry. It is totally not a trap.
A new IT hype that probably sucks up money. Seriously, I see problems with housing affordability and climate change, and nothing can be solved by new data centers or frontier LLMs. I won't say that these may just be my problems, given that I have an iPhone, but I haven't opened the App Store in years. I also use programs that were written in the 80s and 90s and can't imagine what all this solopreneurs will invent with AI.
We don't really need LLMs or AIs to figure out how to solve housing, food insecurity, climate change, etc. We already know exactly how to do that, we just choose not to (in ways that often vary by nation).
Not really what the post is about, but I would say the part you're bothered by is a people problem. Greedy people in control of the companies pushing it and desperate people consuming it hoping to get out of the cycle of constantly working their life away.
I pose this specific question because people still talk about it being this creative amazing thing and it just isn't. Though I am curious, have we seen it happen yet? Which it would be interesting to spot in the wild.
For the housing, climate, and bad neighbor issues, start keeping track of your local, state, and federal politicians. Something these models are really good at is data aggregation. Automate a report card with what you care about, do checkins to make sure your reps are doing what they should be doing, and share your data at publicly and loudly as you can when things start trending in the wrong direction.
I asked the same question a month ago. I think there will be more and more people who ponder it.
There's the "is this human or AI-generated?" question (suspicion) that now hovers over every piece of writing, images, etc.
Definitely wasn't there before. Definitely AI that did that.
Some pretty interesting financial schemes
I think people did that, I can't hold the almighty AI god accountable for the actions of greedy humans.
What has any human done all by itself? Everything you do comes from a synthesis of your memories of the past with your senses of the present.
> "Everything I've seen so far still involves a human sussing out the new thing or the model is just aggregating data that is already there from disconnected sources that never got tied together."
Yes, that describes people as well. You think ideas appear from...what...handed to you by God through the aether from the spirit realm? We're literally meat computers.
Your question seems to boil down to whether current AI has any kind of intrinsic motivation to do anything. It doesn't.
For a long time we've heard people saying things like "it's not really intelligent it's just forming a prediction based on data it's seen before" like that's not the fundamental basis of all learning, or "it can make mistakes that are easily proven incorrect" like people's own judgment and memory isn't fallible - ultimately won't it just boil down to some subjective choice of factors like
- "where do you want to coarse-grain the concept of agency"
- "what qualifies as a novel synthesis of information"
- "how consistently correct must it be to count as intelligent"
?
> Your question seems to boil down to whether current AI has any kind of intrinsic motivating force. It doesn't.
Isn't an AI intrinsic motivating force its initial system prompt?
I wonder how many people's initial system prompt just looks something like "You are a sentient being attempting to maintain a sense of constancy in an ever-changing world, for which you inherently find this fact dis-satisfactory" ?
Would we say they don't have motivation?
> "Isn't an AI intrinsic motivating force its initial system prompt?"
It probably could be, but I'm talking about something abstractly analogous to "explore and create, compete and thrive in the world" (i.e. "good luck, have fun, don't die"). And I think it probably only counts if accompanied by the right scaffolding that allows the system to operate continuously. To my knowledge that hasn't been done. Doing it creates a very hairy ethical dilemma.
Hm; I was under the impression this was some of what was happening at the frontier labs? Like feedback loops of agents orchestrating tests against other mutations of their models, selecting for best results, then repeating that process with new tests?
[FWIW I would personally count that entire system as 'conscious', for some definition of Integrated Information Theory's phi value, though maybe not 'sentient'; I don't think a single LLM (in the sense you can run a model) counts as a conscious system, but it makes sense to me that even one in a feedback loop where it can alter its weights probably counts]
I’ve heard that some rando invented the Web. And a couple of randos made C and Unix. As a huge outlier, Einstein’s 1905 publications were damn impressive.
Which part of "everything that you've ever done was done through the synthesis of your memories and senses" doesn't make sense to you? What exactly do you think was happening inside Einstein's brain at the time? What do you think thinking is?
That's AI for ya. What you are seemingly looking for is AGI and that we do not yet have.
Maybe someone shows up, makes a solid case for something new, and surprises us. There's still people out there claiming it can, so I'm going to keep asking this every month until someone wows me or the topic dies at a societal level.
It created a novel form of psychosis.
We already had that. Did everyone forget about QAnon?
It’s a little different, most people with LLM psychosis think they “ensouled” their chatbot. I’m sure it’s not completely new but it’s pretty novel at this scale. I at least had never heard of someone believing they had unlocked the soul and consciousness of a computer program, making them a kind of Chosen One.
The “spiral” neo religious movement (if you want to call it that) from the GPT 4o era is pretty novel even if it has antecedents in Nick Land and other CCRU people
There's like different sects of users here. That ensouled thing is just one of them. There's also people using them as AI girlfriends or to just fill the void of human interaction because it's scary.
I don't know, but suspect the same features that made QAnon happen are the same mechanisms that cause people to think the LLMs are alive.
Either way, it's just depressing, but I can't say the LLMs are the responsible parties here. The LLMs didn't invent the psychosis, but did what they actually do and fed out exactly what you would get from echo chambers on the internet.