Demand for top models is definitely not saturated, at least when it comes to programming. If I could afford to use 5x more Claude Opus 4.6 tokens, I would!
Demand is relative. How many Claude tokens would you buy if they had a 10x price hike?
The market has achieved it's current saturation level with loss-leader prices that remind me of the Chinese bike share bubble[0]. Once those prices go up to break even levels (let alone profitable levels), the number of people who can afford to pay will go down dramatically (and that's not even accounting for the bubble pop further constricting people's finances).
There is no evidence that labs are losing money on inference subscriptions. The labs have massive fixed costs, but as long as inference spend is higher than the datacenters they use for inference cost all they need to do to become profitable is scale up. Right now software engineers are basically the only ones actually paying for inference, the labs just need to create coding assistants for everything that are good enough that every white collar worker in the country(world?) is paying a $1000/yr subscription. Certainly theres a lot of risk, will models become commoditized and everyone switches to open models? can they actually get non software engineers to pay for inference in mass? But its not like theres no path
If they've already built themselves a loyal customer base (which is usually the point of fighting a price war) and the customers are happy with the technology they have, then if funding is tight and turning a profit is more important why wouldn't they pivot to optimizing inference by stopping further training, freezing the model versions, burning the weights into silicon and building better caching strategies and improving harnesses and tools that lower their cost and increase their margin?
If all they do is hike prices then they'll lose customers to competitors who don't or who find a way to serve a similar model cheaper.
The demand isn't going to go away purely through higher prices. Once people know something is possible they will demand it whether supply is constrained or not. That's a huge bounty for anyone who can figure out how to service that demand.
Easier said than done. What you're describing can take years to implement. Can OpenAI et al. keep burning cash at the same rate for two years while they wait for the salvation of custom silicon if the investments dry up?
Don't you see the massive problem with requiring visual input? Are blind people not intelligent because they cannot solve ARC-AGI-3 without a "harness"?
A theoretical text-only superintelligent LLM could prove the Riemann hypothesis but fail ARC-AGI-3 and won't even be AGI according to this benchmark...
Well, it would be AGI if you could connect a camera to it to solve it, similar to how blind people would be able to solve it if you restored their eyesight. But if the lack of vision is a fundamental limitation of their architecture, then it seems more fair not to call them AGI.
I think I can confidently say they are not visually intelligent at all.
If you were phrasing things to quantify intelligence, you would have a visual intelligence pillar. And they would not pass that pillar. It doesn't make them dysfunctional or stupid, but visual intelligence is a key part of human intelligence.
Visual intelligence is a near meaningless term as it's almost entirely dependant on spatial intelligence. The visually impaired do have high spatial intelligence, I wouldn't be surprised if their spatial intelligence is actually higher on average than those without visual impairment.
I think they don't actually lack them, or lack only a small fraction (their brains are ≈99% like a normal human brain), such that if they were an AI model, they could be fairly trivially upgraded with vision capability.
Assistance of other humans? You do realise we're talking about an intelligence test right, at that point what are you even testing for. I'm sure you've taken exams where you couldn't bring your own notes, use Google or get help from someone, even though real life doesn't have those constraints
Well said. That's exactly what has been rubbing me the wrong way with all those "LLMs can never *really* think, ya know" people. Once we pass some level of AI capability (which we perhaps already did?), it essentially turns into an unfalsifiable statement of faith.
> Ultra-processed foods: Ultra-processed foods typically have more than 1 ingredient that you never or rarely find in a kitchen. They also tend to include many additives and ingredients that are not typically used in home cooking, such as preservatives, emulsifiers, sweeteners, and artificial colours and flavours. These foods generally have a long shelf life.
Are there ingredients actually in the Beyond burger?
"E123 & co" are descriptors covering both "organic" and "synthetic" substances, because their role is to add precision and clarity to an engineering process, not entertain the pseudoscientific naturalistic bullshit masses buy into (which by itself is just a way for another industry to make money - or do you think people come up with those fitness/healthy eating fads all on their own?).
It is magical thinking to claim that LLMs are definitely physically incapable of thinking. You don't know that. No one knows that, since such large neural networks are opaque blackboxes that resist interpretation and we don't really know how they function internally.
You are just repeating that because you read that before somewhere else. Like a stochastic parrot. Quite ironic. ;)
They really aren't that mysterious. We can confidently say that they function at the lexical level, using Monte Carlo principles to carve out a likely path in lexical space.
The output depends on the distribution of n-grams in the training set, and the composition of the text in it's context window.
This process cannot produce reasoning.
1) an LLM cannot represent the truth value of statements, only their likelihood of being found in its training data.
2) because it uses lexical data, an LLM will answer differently based on the names / terms used in a prompt.
Both of these facts contradict the idea that the LLM is reasoning, or "thinking".
This isn't really a very hit take either, I don't think I've talked to a single researcher who thinks that LLMs are thinking.
I don't get what the "AI experiment" angle here is? The fact that AI can write python code that makes sounds? And if the end product isn't interesting or artistically worthwhile, what is the point?
I have a deep background in music and I think that while the creation was super basic, the way the output was so unconstrained (written by a model fine-tuned for coding), is really interesting. Listen to that last one and tell me it couldn't belong on some tv show. I've had always issues with any ai generated music because of the constraints and the way the output is so derivative. This was different to me.
What's the point if human-made art isn't interesting or artistically worthwhile?
(Most of it isn't.)
Art is on a sliding scale from "Fun study and experiment for the sake of it" to "Expresses something personal" to "Expresses something collective" to "A cultural landmark that invents a completely new expressive language, emotionally and technically."
All of those options are creatively worthwhile. Or maybe none of them are.
> What's the point if human-made art isn't interesting or artistically worthwhile?
Because it is a human making it, expressing something is always worthwhile to the individual on a personal level. Even if its not "artisticallly worthwhile", the process is rewarding to the participant at the very least. Which is why a lot of people just find enjoyment in creating art even if its not commercially succesful.
But in this case, the criteria changes for the final product (the music being produced). It is not artistically worthwhile to anyone, not even the creator.
So no, a person with no talent (self claim) using an LLM to create art is much less worthwhile than a human being with no/any talent creating art on their own at all times by default.
>Even if its not "artisticallly worthwhile", the process is rewarding to the participant at the very least
I think that's the point though. What op did was rewarding to themselves, and I found it more enjoyable than a lot of music I've heard that was made by humans. So don't be a gatekeeper on enjoyment.
How am I a gatekeeper? I provided my own opinions; you are free to enjoy what you want or disagree with me. If you want to get into an objective discussion of why you find it enjoyable more than human works or what is art, we can do that but I do not like the personal slights.
I was discussing it on the basis of music with the commentator and the actual product. Sure if you want to go all Andy Kaufman then yeah the .html and this discussion is art but I wasn't talking about it in the original context of the conversation.
At least it wrote a song, instead of stably-diffusing static into entire tracks from its training data. I can take those uninteresting notes, plug them into a DAW and build something worthwhile. I can only do this with Suno-generated stems after much faffing about with transposing and fixing rhythms, because Suno doesn't know how to write music, it just creates waveforms.
AI tools are decent at helping with code because they're editing language in a context. AI tools are terrible at helping with art because they are operating on the entirely wrong abstraction layer (in this case, waveforms) instead of the languages humans use to create art, and it's just supremely difficult to add to the context without destroying it.
I just want to know what's in there. It doesn't need to be artistic at all. They put terabytes of data into the training process and I want to know what came through.
Very interesting experiment! I tried something related half a year ago (LLMs writing midi files, musical notation or guitar tabs), but directly creating audio with Python and sine waves is a pretty original approach.
Even with 1 TB of weights (probable size of the largest state of the art models), the network is far too small to contain any significant part of the internet as compressed data, unless you really stretch the definition of data compression.
Take the C4 training dataset for example. The uncompressed, uncleaned, size of the dataset is ~6TB, and contains an exhaustive English language scrape of the public internet from 2019. The cleaned (still uncompressed) dataset is significantly less than 1TB.
I could go on, but, I think it's already pretty obvious that 1TB is more than enough storage to represent a significant portion of the internet.
A lot of the internet is duplicate data, low quality content, SEO spam etc. I wouldn't be surprised if 1 TB is a significant portion of the high-quality, information-dense part of the internet.
I was curious about the scale of 1TiB of text. According to WolframAlpha, it's roughly 1.1 trillion characters, which breaks down to 180.2 billion words, 360.5 million pages, or 16.2 billion lines. In terms of professional typing speed, that's about 3800 years of continuous work.
So post-deduplication, I think it's a fair assessment that a significant portion of high-quality text could fit within 1TiB. Tho 'high-quality' is a pretty squishy and subjective term.
This is obviously wrong. There is a bunch of knowledge embedded in those weights, and some of it can be recalled verbatim. So, by virtue of this recall alone, training is a form of lossy data compression.
Hmm... perhaps train a robot arm to do it?
reply