I definitely belong to the latter camp. After LLMs I view everything humans do very systematically and whenever said thing still feels fuzzy I just treat it as having a noise/smoothing term
Thankfully for me it happened a while ago, after chess was conquered. One of the most amusing (and depressing) things in the past few years of the current “AI era” is people making the same claims and having the same discussions about “intelligence” and “creativity” from ten, twenty or thirty years ago as their field is finally coming under attack by AI; they are finally going through what the competitive board game communities (checkers, Othello, chess, shogi, go etc.) have already come to terms with. There was no single moment but a gradual process over decades as the incursion became greater and greater, forcing a recalibration as previously held positions became indefensible. In my own case it spurred me to educate myself in many areas of mathematics, philosophy and neuroscience in particular where I was ignorant and has undoubtedly shaped my understanding of “intelligence”.
When people started claiming LLMs could not produce ideas outside of what they encountered in their training data, I was reminded of an old Chess Life article from the 60s reporting on the first chess computer to play in a tournament at a local chess club. One of the club members remarked that he did not believe that the computer could play moves that were not “put in” the computer in advance. There is an overwhelming sense of “here we go again”.
Regarding the LAION aesthetic predictor footnote, I don't see why a modern model and nonlinear classifier won't do a good a job. Is there a fundamental technical problem with the idea?
What would you have in mind for a modern model? Like Dino-V3 or something of that ilk?
For the LAION classifier specifically, it's trained on-top of CLIP. The bottleneck for accuracy isn't the linear/non-linear readout, it's CLIP's specific knowledge of the world.
CLIP was trained with a contrastive learning objective, where all it has to learn is that images of the same category are closer to each other than images of other categories (e.g. two items are both dogs so they should be closer semantically, while say a dog and a computer are different and should be farther semantically.
As a result CLIP is really good at object definitions, but struggles at pretty much everything else. Modern multimodal LLM's perceptual modules are much stronger than CLIP; that's why we chose fine-tuning a LLM rather then training a MLP on-top of CLIP (which would be a direct iteration on the LAION aesthetic classifier).
Aah, for this we're just trying to filter not generate. When it comes to conditioning, you'll still need a model that understands text since the primary control is text.
In the original Stable Diffusion, CLIP doubled as part of this LAION Aesthetic Filter as well as the text control for the generative model. But in the last couple of years, the field has drifted away from CLIP towards LLM hidden states (e.g. Qwen, Mistral, etc.). It turns out that text-to-image and text-to-video models do not need the visual prior from CLIP; instead the hidden states of multimodal LLMs act as better conditioning mechanisms.
But you are on-to something when it comes to leveraging pre-trained vision models as priors in generative image/video.
Nowadays, models like Dino-v3 (which have the strong spatial priors, you're talking about with LingBot) are being used in 3 distinct ways:
1) REPA: The DiT in the generative model minimizes cosine distance to another vision prior, like Dinov3. This paper from earlier this year, iREPA (https://arxiv.org/pdf/2512.10794) digs into WHAT specifically makes a good regression target for generative image.
2) VA-VAE: Most generative image/video models work in a latent space that's pre-trained separately. We wrote a lot about this in an earlier blog post (https://www.linum.ai/field-notes/vae-reconstruction-vs-gener...). But basically, VAEs that overfit to the reconstruction task create brittle latent spaces for generative models. One solution folks have figured out is to regress to a strong spatial prior like Dino, during VAE training. This is discussed at length in the VA-VAE paper(https://arxiv.org/abs/2501.01423). Recent image models like Flux 2, likely achieve their improvements through this type of VAE training.
3) RAE (Representation Autoencoder): This is the latest/greatest right now. Where generative image models are trained directly in Dino-v3 space; and then they just train a decoder to read back out into pixel space. (https://arxiv.org/pdf/2605.18324)
I strictly meant using the embeddings for training a reward model, not the generator. By good for generation I just meant the reward model might find more visual cues for aesthetic preference and avoid some of the spurious semantic correlation CLIP has
That might work! Off the dome, it’s not clear to me whether spatial/depth priors are better/worse than an LLM for this type of task.
Only reason I can think why the LLM might still work better here is that it’s trained to solve a bunch of different image/video related questions, so it’s perceptual modules may be more robust adaptive for this aesthetic grading task versus something like LingBot
Apologies for making the reply chain so long but I think a video like this somewhat proves how a lot of aesthetic preferences can be *ultra* sensitive to small visual details : https://youtu.be/twcMra_67-w?t=88
The video is timestamped to open at the comparison frame. I don't think an LLM can tell the quality difference without direct reference for comparison
I think the fundamental rebuttal to this is merely that, we’re still so early, the $1.5 trillion number is irrelevant for developing technologies in the same way two women can’t birth a baby in 4.5 months
The $1.5 trillion number is extremely relevant because it makes the revenue numbers the industry has to hit higher and there's an expectation to see progress sooner before funding starts getting pulled back.
>...it makes the revenue numbers the industry has to hit higher...
Yes. But.
Industry can miss its revenue targets, Oracle can go bankrupt (looking more and more likely) OpenAI, Anthropic and SpaceX might find "true value" at about 10% of current valuations (that feels right to me), all that can happen with the associated economic chaos... and still out of it comes world shaping new technology
Certainly, but a flop of $1.5 trillion would create a deep and long trough. I think in that situation OpenAI and Anthropic probably don't survive as independent companies (where is the revenue? who would want to put more money in at that point?). In that situation there would definitely be a fire sale in the prices of licensing models (open weights are already pushing this way the hell down), semiconductors, data centers, and compute in general. And like in that situation, who knows what products come out of that dividend of cheap stranded assets? It might not even be AI related. You can do a lot of things with a lot of cheap GPU power. This boom-bust cycle is nothing new, especially to the tech industry. This deep crash outcome is the risk that a bet of $1.5 trillion makes. And it makes the current paradigm of coding agents and chatbots less likely because it is so volatile.
Oracle has bet the house on AI. Free cash flow is deeply negative with tons more capex guidance on the books. They are currently the largest non-financial corporate issuer of new debt in the corporate bond market. They are literally cutting their headcount to the bone to help fund this build out. If their compute deals do not pan out because their customers have liquidity issues or revenue stalls, they are in dire straits.
Given the advances since 2022 why would funding get pulled back? The endgame is replacing all office workers globally and so you get 10% of that market that's that's easily $1 trillion. A competent AI employee replaces a human the same way the car replaced the horse. Businesses will pay more for an AI employee that doesn't get sick or come in hung over over a human one and never goes to sleep. If OpenAI manages to crack that, and capture even 10% of the market globally, that's trillions.
Total addressable market doesn't pay interest expenses. Even if this technology really did have the potential to be as revolutionary as the move from the horse to the car, if you can't get there with what the world economy's liquidity can provide then you can't get there. Free cash flow is drying up and data centers depreciate. This creates a limit and if that limit is hit before the profit starts coming in it doesn't matter what the potential could be. Paying back interest is financial gravity and it exists today. And this isn't an appeal to Wall Street's short term profit motive, there just is an economic limit to what the capital markets can spend on a bet this big and this risky.
Yes, that's going on optimism. It's also possible that we're 80% there and the rest will take forever, while the loans and investments were taken out now.
Five months ago does not make this some kind of historical document.
Although the botlickers will find this hard to believe, nothing has happened this year to make we skeptics -- like me, and like the author of this piece, who is for clarity not me -- change our minds.
I worked out where I stand and why I stand there. I have reasons.
I want evidence to consider changing. Why should I? Because someone somewhere found a tool useful? Because they are big names? That's argument from authority. I don't buy it.
I need compelling reasons. I have yet to see any. All I see is a lot of people jumping on bandwagons.
I'm confused, other people finding it useful is evidence of it being useful for them. How is that merely jumping on a bandwagon? As someone who doesn't like coding, all the one off scripts and ideations its permitted are awesome.
Recently I got it to help me explore and compare curve smoothing algorithms and it helped me to test my own formulations that I hadn't seen elsewhere, and converged on computing segments of:
||T(s)-T(t)||^2. / |s-t|^3,
where s and t are arc length indices, and we sum over multiple scales of offsets and where T is the (unit) tangent of the curve.
Just because that doesn't involve you means its not useful? Its ability to give highly bespoke scripts makes it far more useful to me than hiring an opinionated software developer!!!!!
It is all about externalities. Same as fossil fuels, same as slavery. This kind of systems thinking seems to blow the minds of a lot of heads-down software folks.
You have to think about the big picture. Otherwise, the picture goes out.
1. LLMs are profoundly, inexcusably unethical.
2. The kind of software and tools one can build with LLMs might be better off not built at all.
These sorts of thoughts seem to blow people's minds. Well, tough. Get used to having a bigger mind.
