This likely isn't AA putting its finger on the scale for specific interests, rather in the case of Muse/Astra, it's to avoid looking really bad because it scores things way out of wack with people's anecdotal vibes on model capabilities. It is possible that labs are getting preliminary AA results back and raising concerns about scoring that then get folded into updated benchmarks, but I strongly doubt it's nefarious.
Likely they did a distillation run on a later checkpoint of a large internal teacher model, probably with some tuning of the distillation process based on early public results with Argon. RL is the expensive/time consuming step, they should be able to distill down from a checkpoint with the stamp of approval pretty quickly.
Even a decade ago marketing/growth was the massive lion's share. IMO just building something has been near meaningless since at least post-Ruby on Rails. The app stores have been inundated since the late 2010s.
The distinction is adjusted vs non-adjusted basis. Serving inference is profitable, but serving inference for the last released model doesn't cover the cost of the model in training.
Dario explained it very clearly in an older interview. If every model is bigger and more expensive to train, each model can individually be profitable, while the company as a whole still shows a loss at any given point in time. This is why he likes the idea of regulation so much, it'll let them take a breather and recoup.
> Dario explained it very clearly in an older interview. If every model is bigger and more expensive to train, each model can individually be profitable, while the company as a whole still shows a loss at any given point in time.
“We’re profitable as long as you ignore our ever increasing cost of goods sold, we promise it’s not a pyramid scheme” - ~~Charles Ponzi~~ Dario
But the reason the cost is ever increasing is the steeply ascending quality. There's no (non-doom) scenario where this game does not end with profitable cheap tokens. Wherever it'll stop scaling, it'll be profitable to serve tokens at that scale. And of course if it doesn't stop scaling, it's the singularity and all this thought of profits is quite secondary.
An industry cannot fail because it's too successful.
edit: Other explanation. The industry has found a moneyprinter. It prints money, and several companies are building them and printing money. However, each one of those companies is cashflow negative, because they're all trying their desperate best to build an even larger moneyprinter. Why? If they don't build it and the next company over does build it, and it works, then company 1 is differentially outcompeted. what's more, the more expensive each moneyprinter is, once a company gets strongly ahead the second company won't get the financing to keep up with that one. ... But it's still moneyprinters. This is not an economic race that ends up with "and thus everyone stopped printing money."
> But the reason the cost is ever increasing is the steeply ascending quality.
Sure, if you say so.
> There's no (non-doom) scenario where this game does not end with profitable cheap tokens.
I doubt it.
> Wherever it'll stop scaling, it'll be profitable to serve tokens at that scale.
It’s much more likely, imo, that local AI models will completely destroy the entire business model of inference. For any entity that would be spending enough to replace multiple employees (allegedly), the cost of the hardware to do so is completely affordable. This has the major benefit of eliminating per token costs, and practically guaranteeing that some “AI company” can’t steal your company’s secrets.
At an individual person level, this hardware will take much longer to obtain, but it will happen eventually.
> And of course if it doesn't stop scaling, it's the singularity and all this thought of profits is quite secondary.
I still believe, strongly, that LLMs are the wrong path to AGI. There’s trillions of dollars being put into turning LLMs into AGI; however, there has no zero results showing that LLMs are capable of “learning” anything. “Skills” and expanded context windows don’t count. The “model” needs to be able to modify itself, in real time, for that to be possible.
We’re in a local maxima that strongly refuses to accept that it’s a local maxima. (Diminishing returns for the amount of money poured into “training” is beyond obvious to anyone paying attention).
> An industry cannot fail because it's too successful.
Sure it can. The fossil fuel industry is wildly successful.
It’s also boiling the planet alive, but that’s okay, it’s generating shareholder value so everything’s fine.
Edit for your edit:
> edit: Other explanation. The industry has found a moneyprinter. It prints money, and several companies are building them and printing money. However, each one of those companies is cashflow negative, because they're all trying their desperate best to build an even larger moneyprinter. Why? If they don't build it and the next company over does build it, and it works, then company 1 is differentially outcompeted. what's more, the more expensive each moneyprinter is, once a company gets strongly ahead the second company won't get the financing to keep up with that one. ... But it's still moneyprinters. This is not an economic race that ends up with "and thus everyone stopped printing money."
By your own definition it is impossible to achieve profitability without an ever increasing amount of money flowing into it.
There was a business about moving stamps around that needed similar levels of cash infusion to sustain itself. Eventually its bubble popped too.
> It’s much more likely, imo, that local AI models will completely destroy the entire business model of inference.
You're massively underestimating how much cheaper it is to serve models in a datacenter than at home.
> By your own definition it is impossible to achieve profitability without an ever increasing amount of money flowing into it.
No, what? No. This is just a total misreading of what I'm saying. It's extremely easy to achieve profitability. It just requires your business to stop chasing more profitability. The problem is just that then the other businesses who are still chasing more profitability will eat its lunch. It's a market shape that easily supports n-1 bankruptcies but never n bankruptcies.
Agents are already better at writing strongly typed languages than Ruby, and don't care about the small ergonomic affordances that rails provided. Even if you think the Rails bloat and Ruby nicheness/slowness is worth it today (hard disagree), think forward a year when agents are significantly stronger than they are today and consider how that decision will hold up.
IMO Cloudflare is the closest thing to a "good guy" these days. Very customer friendly, affordable and generally fairly ethical (if you pass over the part where they share data with the NSA).
I like Cloudflare as a company but I’d be weary about paying for my domains and hosting my DNS with the same publicly traded company with a severe lack of public customer service outside devs getting mad on twitter (guilty!).
If cloudflare gets mad at you (even accidentally) and cancels your account then you lose everything. At least if you separate them, there is a chance that you won’t lose access to both at one time. Given that domains are generally a one time setup after purchase (especially if you use cloudflare for DNS) then I think the minor inconvenience is worth it, personally.
I feel the same way. I think I want three different companies to be my domain registrar, DNS provider, and VPS host, for redundancy. Putting all eggs in one basket seems like a bad idea.
It's insane to me that people don't see that agents are going to be how you interact with your computer going forward, almost exclusively. The people who make things hard for agents are going to get eviscerated by the people who don't just make it easy for agents to do things, but market primarily to agents rather than humans.
If that's true, why bother blocking the tools? Either they are useful in saving money (and more important, hassle), or they're not. If they're not useful, the problem simply won't materialize.
Because not blocking these tools causes additional operational costs.
For example, higher return rate for online shopping (because the tools didn't buy the correct thing, or maybe they bought something that they shouldn't have).
Or, the tool ignores the adverts on an ad-supported website, intended for humans, basically just adding costs for no benefit to the site.
Or, the tool wastes customer services' time on something that ultimately the agent's operator didn't even want in the first place.
It's a mistake to think the sites have a choice. I'm hardly going to trust an AI agent to shop on my behalf if it isn't smart enough to operate the browser on my behalf in a way that's indistinguishable from a human. They're almost good enough to do that now; it won't be long.
So the end result will be that the sites spend money on bandwidth to shove ads and dark patterns at me, which I will never, ever see.
You'd think they'd understand who's going to end up with the power to say "No" here, but apparently not.
As a Krea2 local model enjoyer, your challenge is that on the $20 GPT plan you get a fairly good amount of Codex use and image gen. If I used Krea, it would be because I fine tuned/built a workflow on top of it to do something GPT image couldn't/wouldn't do. Maybe there's a market for the cheapest "good enough" bulk image gen model (spam? idk what the applications would be), but personally I'd say you should focus on workflow add-ons/one click fine tuning/etc that design companies can use to create assets with their custom "look" since the GPT-Image look will become blase from overuse.
Businesses need more than just cost per image than a raw ChatGPT subscription doesn’t give you.
There’s also workflows around non-frontier models that make sense, especially at-scale (think internationalization of a marketing campaign to 200 countries across different media formats). A lot of the same rules that apply to LLMs apply to diffusion. The same argument could be made around Codex/Astra/Sol and many people are happily paying for tools like Devin/Cognition.
But I agree with you that if you are not careful, the subsidies from big labs can eat your business.
EDIT: I hope you’re enjoying Krea 2! What GPU are you running it on?
I think the early models distorted the perception of what artists actually need. It's not about workflows. Image creation is inherently iterative. None of the platforms seem to understand that, while AI artists have somehow fallen in love with the unnecessary complexity of node-based tools. Apparently, 400 tangled noodles look impressive when the boss peers over your shoulder.
The big studios and agencies are all building custom tooling as they always have but now a lot faster.
reply