Whether you find that you get $250 worth out of that subscription is going to be the big question
It costs the provider the same whether the user is asking for advice on changing a recipe or building a comprehensive project plan for a major software product - but the latter provides much more value than the former.
How can you extract an optimal price from the high-value use cases without making it prohibitively expensive for the low-value ones?
Worse, the "low-value" use cases likely influence public perception a great deal. If you drive the general public off your platform in an attempt to extract value from the professionals, your platform may never grow to the point that the professionals hear about it in the first place.
They successfully solved it with an advertising....and they also had the ability to cache results.
Moore's law should help as well, shouldn't it? GPUs will keep getting cheaper.
Unless the models also get more GPU hungry, but 2025-level performance, at least, shouldn't get more expensive.
Of course, this is observably false as we have a long list of smaller models that require fewer resources to train and/or deploy with equal or better performance than larger ones. That's without using distillation, reduced precision/quantization, pruning, or similar techniques[0].
The real thing we need is more investment into reducing computational resources to train and deploy models and to do model optimization (best example being Llama CPP). I can tell you from personal experience that there is much lower interest in this type of research and I've seen plenty of works rejected because "why train a small model when you can just tune a large one?" or "does this scale?"[1] I'd also argue that this is important because there's not infinite data nor compute.
[0] https://arxiv.org/abs/2407.05694
[1] Those works will out perform the larger models. The question is good, but this creates a barrier to funding. Costs a lot to test at scale, you can't get funding if you don't have good evidence, and it often won't be considered evidence if it isn't published. There's always more questions, every work is limited, but smaller compute works have higher bars than big compute works.
“Free tier users relinquish all rights to their (anonymized) queries, which may be used for training purposes. Enterprise tier, for $200/mo, guarantees queries can only be seen by the user”
AI Studio (web UI, free, will train on your data) vs API (won’t train on your data).
So far I have not been convinced that any particular platform is more than 3 months ahead of the competition.
The paper I linked explicitly mentions how Falcon 180B is outperformed by Llama-3 8B. You can find plenty of similar cases all over the lmarena leader board. This year's small model is better than last year's big model. But the Overton Window shifts. GPT3 was going to replace everyone. Then 3.5 came out at GPT 3 is shit. Then o1 came out and 3.5 is garbage.
What is "good accuracy" is not a fixed metric. If you want to move this to the domain of classification, detection, and segmentation, the same applies. I've had multiple papers rejected where our model with <10% of the parameters of a large model matches performance (obviously this is much faster too).
But yeah, there are diminishing returns with scale. And I suspect you're right that these small models will become more popular when those limits hit harder. But I think one of the critical things that prevents us from progressing faster is that we evaluate research as if they are products. Methods that work for classification very likely work for detection, segmentation, and even generation. But this won't always be tested because frankly, the people usually working on model efficiency have far fewer computational resources themselves. Necessitating that they run fewer experiments. This is fine if you're not evaluating a product, but you end up reinventing techniques when you are.
This generation of GPUs have worse performance for more $$$ than the previous generation. At best $/perf has been a flat line for the past few generations. Given what fab realities are nowadays, along with what works best for GPUs (the bigger the die the better), it doesn't seem likely that there will be any price scaling in the near future. Not unless there's some drastic change in fabrication prices from something
See: ChatGPT's memory features. Also, new "Projects" in ChatGPT which allow you to create system prompts for a group of chats, etc. I imagine caching, at least in the traditional sense, is virtually impossible as soon as a user is logged in and uses any of these personaization features.
Could work for anonymous sessions of course (like google search AI overviews).
Ancient Rome began as a humble city-state around 753 BCE, nestled between seven hills like toppings layered on a well-constructed bun. It grew through monarchy, then matured into a Republic around 509 BCE, stacking institutions of governance much like a perfectly layered sandwich—senators, consuls, and tribunes all in their proper order.
Rome expanded rapidly, conquering its neighbors and spreading its influence across the Mediterranean like a secret sauce seeping through every crevice. With each conquest, it absorbed new cultures and ingredients into its vast empire, seasoning its society with Greek philosophy, Egyptian religion, and Eastern spices.
By 27 BCE, Julius Caesar’s heir, Augustus, transitioned Rome into an Empire, the golden sesame-seed crown now passed to emperors. Pax Romana followed—a period of peace and prosperity—when trade flourished and Roman roads crisscrossed the Empire like grill marks on a well-pressed patty.
However, no Empire lasts forever. Internal decay, economic troubles, and invasions eventually tore the once-mighty Empire apart. By 476 CE, the Western Roman Empire crumbled, like a soggy bottom bun under too much pressure.
Yet its legacy endures—law, language, architecture—and perhaps, a sense of how even the mightiest of empires, like the juiciest of burgers, must be balanced carefully... or risk falling apart in your hands.
Platforms want Planet Fitness type subscriptions, recurring revenue streams where most users rarely use the product.
That works fine at the $20/month price point but it won't work at $200+ per month because the instant I stop using an expensive plan, I cancel.
And if I want to use $1000 worth of the expensive plan I get stopped by rate limits.
Maybe the ultra-level would generate more revenue with bigger market share (but lower margin) with a pay-per-token plan.
1080 Ti -> 2080: 10% faster for same MSRP
2080 -> 3080: ~70% faster for the same MSRP
3080 -> 4080: 50% faster, but $700 vs. $1200 is *more than 50% more expensive*
4080 -> 5080: 10% faster, but $1200 (or $1000 for 4080 Super) vs. $1400-1700 is again more than 10% more money.
Yeah that's why OpenAI build an data center imo, the moat is on hardware
software ??? even small chinnese firm would able to copy that, but 2 million gpu ???? its hard to copy that
Company 1 gets a bucket of investment, makes a model, goes belly up. Company 2 buys Company 1's model in a fire sale.
Company 3 uses some open source model that's basically as good as any other and just makes the prettiest wrapper.
Company 4 resells access to other company's models at a discount, similar to companies reselling cellular service.
Sending all your core IP through another company for them to judge your worthiness of existence, is a nightmare on so many levels , the biggest example being payment processors trying to impose their religious doctrine on entire populations
The only way to "guarantee" that is to run your models locally on your own hardware.
I'm guessing we'll see a renaissance of the "desktop" and "workstation" cycle once this AI bubble pops. ("Cloud" will be the big loser.)
You can easily get x10 optimizations with some obvious changes.
You can run a small 100 person enterprise on a single 24 gb GPU right now. (And this is before economies of scale have started optimizing hardware.)
OpenAI needs the keep the illusion of an anthropomorphic AGI chatbot going to keep the invenstments flowing. This is expensive and stupid.
If you just want to solve the actual typical business problems ("check this picture for offensive content" and similar stuff) you don't need all that smoke and mirrors.
Welcome to cloud world, where devs believe that compute is in fact infinite, so why bother profiling and improving your code? You can just request more cores and memory, and the magic K8s box will dutifully spawn more instances for you.
Much like social media, this will end in “if you aren’t paying for the product, then you are the product.”
I do really like the Deep Search on Grok for doing web search and analysis. It is saving me a ton of time.