The AI boom is based on a fundamental mistake
theverge.com
theverge.com
Though I agree that language is not intelligence, suggesting that the AI boom is only about LLMs, or that LLMs do not create value is incredibly misleading.
I disagree with the base concept that to justify the current investment, we must arrive at AGI.
Estimates put total AI investement at $1.5T.
Is it a lot? Sure.
But what is going to come out of it.
Faster and likely improved results from medical imaging. Lower cost and widespread ability to create images and videos, denting the multi-trillion dollar marketing industry. Self-driving and advanced driver assistance, lives saves, costs reduced. Improvements in education and low-cost tutoring available to more people.
Let's say these investements just break even with spend. Isn't that better for society?
I know people are going to say "but what about the radiologists"? We have a shortage of GPs and doctors to care for an aging population. These highly trained doctors can do more in the medical community.
What about the actors/directors/sound engineers, etc in the media industry. This will likely shrink, but we can't ignore their expertise entirely. This industry won't go away. A friend is a voice actor. He isn't being replaced, however, he is getting less work, not because of the final production, but because in the pre-production phases, they don't need his expertise because they can use AI as 'good enough".
The lens I look at this through is my grandfather, who developed film for a living. That job disappeared. Nobody cried out for the film developers when we made the shift to digital. We create more imagery now than ever before, more people are employed in the imaging business than ever before.
As a (former) software engineer myself, do I believe AI will replace engineers? I think this is true as much as packages replaces engineers. When I started programming, there weren't a ton of open-source packages and there weren't a ton of tools like NPM, Cargo, for managing packages. I saw the transition from "write most of the code yourself" to "lego-brick" style programming. It didn't reduce the amount of programmers, it increased what we could do, and allowed us to do more interesting work than boilerplate.
This is how progress has created so much human suffering, by impacting people in a way just sparse enough that you can believably imply “no one cried out for x”.
Yes they did, it impacted thousands if not multiples of that; many in ways that they never recovered from. Some of them probably committed suicide, or ended up homeless, or in some other way had their life destroyed. And that impact reverberated onto their friends, families and communities. And maybe the New York Times even wrote an article about it, but we collectively as a society ignored that suffering.
(I make this remark merely as a gag. I think you pinpoint an issue which has been unresolved for ages and is knee-deep into ethics. One could argue that many of our disagreements about AI and progress (in a broader sense) stem from different positions on ethics, including utilitarianism).
No creativity emerges from this, just the mimicry of the tokens, symbols in their reduced, modeled parameters.
That's the problem with LLMs and frontiers, RL. It's all subject to the bottlenecks of math and symbols.
The analog load of semantic (use tasks, actions, variance, scales), the immensity of the nested forms these take between brain, body, screen, canvas, aren't replicable as creation.
It's not information or stimuli we engage with, these are convenient false reductions we're realizing are inconvenient. There's much more the brain, eye detect in reflection. The words and symbols are an aftereffect the computer detects on its way to disregarding the events.
We see this aftereffect as creativity, thinking, etc only as a clever illusion the words, tokens hand us. Nothing more.
In truth, I don't think there's likely to be a correct definition for what "creativity" is. We're all just moving each other's goal posts and looking at different subject aspects of our experience.
What I do know is that I have talked to dozens of people, and have friends who have talked to hundreds. When shown what LLMs and generative models can do, there's a significant portion of them who label this work as creative. Rather than deem these people ignorant, I would rather consider that maybe if enough people see creativity somewhere, perhaps there is something to it. LLMs creating poems, stories, connecting concepts which you wouldn't otherwise see connected, birthing ideas in images that video which had never existed before.
Of course, I'm aware that by this logic many things fall apart (for example, one might be tempted to believe in god because of it). Nonetheless, on this issue, I am deeply on the creativity camp. And while I am there, the things I get LLMs to create, and people I know do so as well, will continue to induce deep emotion and connection among ourselves.
If that's not creativity to some, well, that's life, full of creative ways to define creativity or eschew it from the world.
I'm not concerned with "hundreds" of people's opinions. And stories, poems, these are dead/dying forms.
You may be on the creativity camp, stay there. It's generic as creativity. From my POV, someone building a hybrid between game and movies, this tech is dead dull artifactual, like a fossilized version of art.
Yes of course we have definitions of creativity, see Anna Abraham. "creativity is doing something unique and new, using existing networks for things they weren't designed for"
And we have the wordless processes that materialze as affinities and events in the imagination.
https://www.nature.com/articles/s41583-019-0202-9
I'm very struck by coders who are neither engineers, scientists, nor creatives in the arts, yet have the false intuition and self-delusion that this code is creative. Nonsense.
I'm afraid LLMs have zilch in comprable abilities to both as they exclude words, symbols, tokens etc.
(I do get what you're saying. Yet, I am not convinced that processes such as tokenization, and the inherent discretization it entails, are incompatible with creativity. We barely understand ourselves, and even then we know that we do discretize several things in our own processes, so it's really hard for me to just believe that tokenization inherently means no creativity).
That's what the article's big reveal is about. The events (which can be externalized creatively) are not really creative as words. That's a big problem for the species in general. That's a glass ceiling no one is recognizing or taking notice of. And that sort of gives you an idea how poetry and code are trapped behind it.
Story having been with us since time immemorial makes me question its correctness.
>AlphaEvolve is an evolutionary coding agent for designing advanced algorithms based on large language models such as Gemini. https://en.wikipedia.org/wiki/AlphaEvolve
which came up with a novel improvement in matrix multiplication
>I don't think people realize just how insane the Matrix Multiplication breakthrough by AlphaEvolve is https://www.reddit.com/r/singularity/comments/1knem3r/i_dont...
Evolution is a tinkering process that finds by selection, and functionality is what we post-hoc add to the success.
Anything reduced to symbols is simply functionality within models, to claim it's creative is illusory. Creativity, imagination, etc exist in a realm beyond, and preceding symbols, metaphors, etc. Creativity is 'thinking around symbolic bottlenecks'.
The matrix multiplication breakthrough is still under the glass ceiling symbols provide. A new way to do math isn't the same as visual scale invariant paradoxes. It's not creativity, the word is a metaphor, the closer word is optimization.
If I'm a farmer and every potato I sow yields a different shape, does that mean they're no longer potatoes?
It's largely around LLMs, plus some generative image models. That is _where the money is_. Sure, some people are doing some CV stuff that might have some medical applications, but no-one's spending tens of billions of borrowed money on datacenters for _that_.
> or that LLMs do not create value is incredibly misleading.
They're not suggesting that.
Fundamentally, the current economic bubble around 'AI' is based on 'AGI' being achievable, soon. Otherwise, the valuations and amounts of money being spent simply do not make sense. LLMs being somewhat useful will not cut it; the spending implies expectation of dramatically increased capabilities, soon.
>>the spending implies expectation of dramatically increased capabilities, soon.
Even more than that, the spending is a death-race in expectation of:
1) a team reaching runaway superintelligence,
2) that it will be winner-take-all for the first team to get there, and
3) fear that somebody else will get there first.
So, everyone is spending as much as they can beg, borrow, or steal because in their view, there is only first place or being part of the underclass ruled by the other team that gets there first.
This is not about just making useful products/services that can increase productivity for their customers.
That's a big leap you're making. Because you're first of all assuming that all of what you describe will happen (show me a true self driving car outside of test cases for example). And on top you're saying that it breaks even with spend. Which is something we simply don't know. My guess is that every vendor is subsidizing to gain market share.
I don't want to debt utility. But when it's accurately priced, will we see where it's useful
It feels like a sort of smug escapism that ignores for many tasks LLM output regardless of how you want to define it is enough. It may not replace humans or our thought entirely, but many otherwise human tasks simply do not require such. Instead of facing the reality of such authors like this article's rejoice under the idea that our thought is special, our output unique and unmatched, those AI marketing fools are for naught as we cannot be replaced because of how they tried sticking LLM's output into a very human shaped box... if only it were that simple.
- The first claim is that language is not intelligence.
I think this is true, but am not sure. However, for the sake of argument, let’s accept this as true.
However, this leads to the obvious objection stated by the author. General intelligence doesn’t need to work the same way as human intelligence does. The author counters this with their second claim.
- We have no evidence that a language based model can lead to intelligence.
For the sake of argument, let’s even accept this as true. But here’s where the sleight of hand comes in. Lack of evidence in X being true is not the same as evidence that X is false.
Just because we don’t have evidence that language can evolve into something like intelligence, doesn’t mean language cannot evolve into something like evidence.
And so far our experience in language evolving into something like intelligencr has been surprisingly positive.
What he is referring to is that this whole bubble is predicated on promises that it actually is possible. And very soon, at that.
When they said there is no evidence they meant that the bubble is expanding due to these promises being made without the evidence. Meaning everyone should be warier about a much bigger risk then they are led to believe. But if that were true, the actual insane amounts of money moving around wouldn't be so insane. So "graph go down"
I can’t find where they try to determine how much investment, purchasing, or use is fuelled by claims that AGI is coming though. Absent that, the title seems unrelated to the article content.