Echoes of the AI Winter
13 points by kchanqvq
13 points by kchanqvq
No reasoning is given for the prediction. My take is different. LLMs are a genuine contribution and advance to computing power. They affect both information retreival and generation. They will make some categories of jobs redundant and create new ones. In short they will push us to more strategic and decision making roles while tasks like synthesizing information, generating code and producing reports (e.g. summaries and diagrams) will become more and more automated.
The frontier that has been slow to be breached is mechanical manipulation of the world, but that is coming, and at present it seems China will lead, though that is surprising to me because I always assumed countries/areas like Japan, the EU and the US with severe skilled labor shortages would be pushing the frontiers of robotics.
How are you supposed to make good decisions if you've given up on trying to understand things?
You have uttered my favorite word: “understand”. I don’t know about you, but the whole of my life is making decisions in a world I can not understand. The obscurity of large language models is but a tiny part of what I can’t comprehend.
I don’t know about you, but the whole of my life is making decisions in a world I can not understand. The obscurity of large language models is but a tiny part of what I can’t comprehend.
Sure. But for a lot of people – from carpenters to physicists to medical professionals to car mechanics – there is at least one professional corner of the world that is your very job to understand and to know. I watch in trepedation as people seem to let more and more of that be left to LLMs, too.
A lot of life involves the ability to make reasonably good decisions without a good understanding, I agree. But also a lot of life involves the ability to actually understand. I'm afraid the human condition is such that both can be replaced by something if that something is "easy".
Am I worried that people replace "click whatever the top Google result is" with "do what the LLM says" when it comes to figuring out the best recipe for birthday cake? No. Am I worried that people do that for things they're actually supposed to be experts at? HELL YES!
I think competence is a social trait. Technology has never changed it. There have always been people who take charge of their work and those who dodge their way through it. I don’t think technology has ever changed that.
The entire field of knowledge work is about selling understanding of problems.
I thought it was the delivery of software that enabled a business need.
Can you explain the parts of that job that aren't entirely about understanding?
The goal of the business is as you say, but the people doing the work accomplish it almost entirely by understanding problems. The remaining work is typing, and if you remove the need to extract understanding from brains, the typing also becomes unnecessary.
You can make an argument that you also want someone you can fire when things go wrong, to insulate the folks in charge from blame.
Old enough to remember people saying exactly the same thing in regards to using C instead of assembly or using GC instead of managing memory by hand.
Yes, I also remember high schools having luddite clubs spring up over C compilers, and artists get enraged about the wholesale strip-mining of their copyrighted work. I remember how when people were polled about C compilers, something like half of them believed we were building a worse future, and there were more people in the world who believed in telekenisis than both knew what a C compiler was and believed C compilers were good for humanity.
Wait. No, I don't.
I remember how when people were polled about C compilers, something like half of them believed we were building a worse future
That attempt at sarcasm right there is how I definitively know you weren't there because that's precisely how people reacted.
Good job on ignoring nearly all of the post. And as for the part you responded to: No, people from the general public at the time did not hate C compilers. Most of them didn't know what C was. This is not true for AI.
I don't think you're going to be able to snow people on this one.
It appears that you are still trapped in a state of terminal literalism being unable to grasp basic abstraction. Historical parallels do not require perfectly symmetrical events down to the last detail. Demanding that an analogy perfectly mirror the exact cultural footprint of a past era demonstrates your inability to generalize concepts.
Pointing out that artists did not protest a systems programming language half a century ago is hardly the devastating rhetorical checkmate you seem to think it to be. C was a specialized tool built for and by a niche demographic which was every bit as passionate about technological change as people are today. The difference with generative AI is that it operates in universally accessible domains, hence there's a broader public debate about it. Yet, the point I was originally making, and you studiously ignored, is that the underlying human psychology remains completely unchanged in both cases.
It's the exact same type of moral panic where people who base their whole identity on a specific set of skills start to lament about erosion of authentic skill and paint a dystopian future caused by their specific type of labor being automation.
Look, if you're going to attempt a shallow, snarky response (upthread, "old enough to remember..."), best not get upset when others respond in kind.
This analogy is tired and flawed. A compiler translates a well-specified language into another by well-specified rules. An LLM not so much.
This analogy is simply inconvenient for you because it exposes how your own argument is flawed. The code LLM produces is every bit as well-specified as the code you write by hand. Nothing stops you from reading and understanding this code, nor does it magically become non deterministic because it wasn't written artisanally by hand.
Of course the code is well-specified. That wasn't my point. You conveniently ignored the last part of my sentence, where the point resides. I wrote
A compiler translates a well-specified language into another by well-specified rules.
Emphasis added this time. The way an LLM turns a prosaic description of a program into (well-specified) code is decidedly not well-specified. The comparison with compilers is terrible.
I didn't ignore it, I addressed it directly. Emphasis added this time.
Nothing stops you from reading and understanding this code, nor does it magically become non deterministic because it wasn't written artisanally by hand.
You're making a straw man implying that once LLM generates code then nobody reads it or understands it. Perhaps that's how you do things, but then it's a problem with your process as opposed to a problem with LLMs.
The code that's produced is obviously well specified. It can be read, understood, and tested just like any other code. You're making a deeply disingenuous argument, and I'm frankly amazed you chose to double down on it.
You've shifted the goal-posts. In your first response, you drew an analogy between catastrophizing about C compilers being used to produce machine code and catastrophizing about LLMs being used to produce (say) C code. In this analogy, the LLM is analogous to the C compiler. @gspr is responding to this by saying that a C compiler is deterministic and well-specified, while an LLM is not. In your most recent responses, you're implicitly conceding this (while misdirecting, acting as if @gspr's claim was that the output of the LLM would be nondeterministic) and instead arguing that it doesn't matter, because you can always review the analogical equivalent of the machine code by hand to see whether it is sensible or correct.
I didn't shift any goal-posts. In my original response I was pointing out that people had similar type of moral panic over previous advances in programming tools. @gspr then went to make a straw man about how LLM is used as a tool. Nowhere in my original comment was I talking about whether LLM produces deterministic output. That has nothing to do with the point I was making. Let me know if this is still unclear for you.
Now you've side-tracked to a general statement about LLM-generated code. That wasn't what we were discussing. We were discussing whether LLMs are analogous to compilers or not. Can we finish that discussion first?
You actually make my point very well here when you write
It can be read, understood, and tested just like any other code.
("It" being LLM-generated code).
Of course! But the very fact that this is necessary sets the LLM distinctly apart from a compiler! Most people use compilers without reading, understanding or verifying their output. And that's perfectly fine. It's not a fine approach with LLMs. So your original analogy is bad. That's all I'm trying to say here.
You're the one who's been doing all the side tracking. What I originally said was that the reaction to new technology was the same, not that the technology works in the same way. That's just the straw man you built to argue against.
I then proceeded to address your straw man by explaining that a stochastic tool is not used in the same way as a deterministic one. Nowhere did I say these tools were equivalent in the way they work. In fact, my point has nothing to do with how these tools work, merely the reaction people have to them. My original analogy has to deal with people.
I apologize. I thought you meant to say that the reaction to compilers was the same as to LLMs because they represent the same kind of change.
If you're really just saying that you believe the reactions are the same, and are making no comparison between what the respective changes entail, then it's all very superficial. By the same token, I could string together two entirely unrelated things that elicited similar reactions to each other. Hardly a useful rhetorical tool.
Claiming these are unrelated things is frankly absurd, and there's nothing superficial here. All these advances were transformative in how we work with code, and completely changed the industry going forward. It doesn't mean these advances are all of the same nature however. LLMs are just as transformative as compilers or introduction of GC was, but they work in a different way, and we currently don't know what programming practice will look like as a result. However, the moral panic people are having each time such a big change happens is the invariant.
My favorite take is that is large frontier models may not necessarily succeed because right now its a race to be a loss leader. However smaller models in domain specific deployments may become more ubiquitous - especially ones that can be self hosted
I suspect that will be the post hype peak steady state, yes. Natural language interfaces for most things, perhaps personal oracles to do our bookings and organization for us.
Perhaps we might even see a fallback where external services revert to bare API endpoints and everyone creates their own personal interfaces to them.
So, for example, shopping on an online portal is completely personalized with our personal oracle interacting with the shop’s oracle.
What isn't mentioned is that advances made before the AI winter began were valuable. The results were just not nearly as earth-shattering as they were marketed to be. Utility was gained but it didn't change everything like they believed.
I can't deny that LLMs and diffusion models are a technological advancement. But I don't buy the hype that they are intelligent, that they will produce a "permanent underclass" or that they will replace software engineering entirely. I think they still may dwindle in popularity for serious coding projects--especially as they get more expensive. I personally believe it's a trap to use them, and I also don't want to increase my electricity consumption by using them.
Some other technology that uses LLMs as a component may certainly change the game. As is, they aren't as incredible as they are touted to be.
The “winter” is the collapse of an investment bubble combined with the anti-hype ironically leading to a reduction in non-business funding.
This happens not because each “ai” wave was useless but because it’s very difficult to privatize the benefits of ai in a way that leads to money being funneled into investors pockets from “ai” tech companies rather than users of ai.
The collapse in government spending is ironic because technologies that convey broad benefits that are felt widely but not privatizable is the exact area that liberal and neoliberal economic theory would suggest should be provided by government.