zlacker

There's no moat in giant LLMs. Anyone on a long enough timeline can scrape/digitize 99.9X% of all human knowledge and build an LLM or LXX from it. Monetizing that idea and staying the market leader over a period longer than 10 years will take a herculean amount of effort. Facebook releasing similar models for free definitely took the wind out of their sails, even a tiny bit; right now the moat is access to A100 boards. That will change as eventually even the Raspberry Pi 9 will have LLM capabilities

>>hadloc+(OP)
OpenAI (ChatGPT) is already a HUGE brand all around the world. No doubt they're the most valuable startup in the AI space. That's their moat.

Unfortunately, in the past few days, the only thing they've accomplished is significantly damaging their brand.

>>morale+j1
Branding counts for a lot, but LLM are already a commodity. As soon as someone releases an LLM equivalent to GPT4 or GPT5, most cloud providers will offer it locally for a fraction of what openAI is charging, and the heaviest users will simply self-host. Go look at the company Docker. I can build a container on almost any device with a prompt these days using open source tooling. The company (or brand, at this point?) offers "professional services" I suppose but who is paying for it? Or go look at Redis or Elasti-anything. Or memcached. Or postgres. Or whatever. Industrial-grade underpinnings of the internet, but it's all just commodity stuff you can lease from any cloud provider.

It doesn't matter if OpenAI or AWS or GCP encoded the entire works of Shakespeare in their LLM, they can all write/complete a valid limerick about "There once was a man from Nantucket".

I seriously doubt AWS is going to license OpenAI's technology when they can just copy the functionality, royalty free, and charge users for it. Maybe they will? But I doubt it. To the end user it's just another locally hosted API. Like DNS.

>>morale+j1
The damage remains to be seen

They still have gpt4 and rumored gpt4.5 to offer, so people have no choice but to use them. The internet has such short an attention span, this news will get forgotten in 2 months

>>hadloc+(OP)
You are forgetting about the end of the Moore's law. The costs for running a large scale AI won't drop dramatically. Any optimizations will require non-trivial expensive PhD Bell Labs level research. Running intelligent LLMs will be financially accessible only to a few mega corps in the US and China (and perhaps to the European government). The AI "safety" teams will control the public discourse. Traditional search engines that blacklist websites with dissenting opinions will be viewed as the benevolent free speech dinosaurs of the past.

>>morale+j1
i don't think that's really any brand loyalty for OpenAI. people will use whatever is cheapest and best. in the longer run people will use whatever has the best access and integration.

what's keeping people with OpenAI for now is that chatGPT is free and GPT3.5 and GPT4 are the best. over time I expect the gap in performance to get smaller and the cost to run these to get cheaper.

if google gives me something close to as good as OpenAI's offering for the same price and it pull data from my gmail or my calendar or my google drive then i'll switch to that.

>>denlek+ib
This, if anything people really don't like the verbose moralizing and anti-terseness of it.

Ok, the first few times you use it maybe it's good to know it doesn't think it's a person, but short and sweet answers just save time, especially when the result is streamed.

>>cft+d7
This assumes the only way to use LLMs effectively is to have a monolith model that does everything from translation (from ANY language to ANY language) to creative writing to coding to what have you. And supposedly GPT4 is a mixture of experts (maybe 8-cross)

The efficiency of finetuned models is quite, quite a bit improved at the cost of giving up the rest of the world to do specific things, and disk space to have a few dozen local finetunes (or even hundreds+ for SaaS services) is peanuts compared to acquiring 80GB of VRAM on a single device for monomodels

>>dontup+ws
Sutskever says there's a "phase transition" at the order of 9 bn neurons, after which LLMs begin to become really useful. I don't know much here, but wouldn't the monomodels become overfit, because they don't have enough data for 9+bn parameters?

>>hadloc+j5
I think yuou're assuming that OpenAI is charging a $/compute price equal to what it costs them.

More likely, they're a loss-leader and generating publicity by making it as cheap as possible.

_Everything_ we've seen come out of silicon valley does this, so why would they suddenly be charging the right price?

>>hadloc+j5
> offer it locally for a fraction of what openAI is charging

I thought the was a somewhat clear agreement that openAI is currently running inference at a loss?

>>denlek+ib
I do think there is some brand loyalty.

People use "the chatbot from OpenAI" because that's what became famous and got all the world a taste of AI (my dad is on that bandwagon, for instance). There is absolutely no way my dad is going to sign up for an Anthropic account and start making API calls to their LLM.

But I agree that it's a weak moat, if OpenAI were to disappear, I could just tell my dad to use "this same thing but from Google" and he'd switch without thinking much about it.

>>morale+gK
good points. on second thought, i should give them due credit for building a brand reputation as being "best" that will continue even if they aren't the best at some point, which will keep a lot of people with them. that's in addition to their other advantages that people will stay because it's easier than learning a new platform and there might be lock-in in terms of it being hard to move a trained gpt, or your chat history to another platform.

>>worlds+7C
Moore's law seems to have failed on CPUs finally, but we've seen the pattern over and over. LLM specific hardware will undoubtedly bring down the cost. $10,000 A100 GPU will not be the last GPU NVidia ever makes, nor will their competitors stand by and let them hold the market hostage.

Quake and Counter-Strike in the 1990s ran like garbage in software-rendering mode. I remember having to run Counter-Strike on my Pentium 90 at the lowest resolution, and then disable upscaling to get 15fps, and even then smoke grenades and other effects would drop the framerate into the single digits. Almost two years after Quake's release did dedicated 3d video cards (voodoo 1 and 2 were accelerators, depended on a seperate 2d VGA graphics card to feed it) begin to hit the market.

Nowadays you can run those games (and their sequels) in the thousands (tens of thousands?) of frames per second on a top end modern card. I would imagine similar events with hardware will transpire with LLM. OpenAI is already prototyping their own hardware to train and run LLMs. I would imagine NVidia hasn't been sitting on their hands either.

>>hadloc+j5
> I seriously doubt AWS is going to license OpenAI's technology when they can just copy the functionality, royalty free, and charge users for it. Maybe they will? But I doubt it.

You mean like they already do on Amazon Bedrock?

>>iLoveO+hZ2
Yeah and looks like they're going to offer Llama as well. They offer Redhat linux EC2 instances at a premium, and other paid per hour AMIs. I can't imagine why they wouldn't offer various LLMs at a premium, but not also offer a home-grown LLM at a lower rate once it's ready.

>>hadloc+(OP)
They won't stand still while others are scraping and digitizing. It's like saying there is no moat in search. Scale is a thing. Learning effects are a thing. It's not the worlds widest moat for sure, but it's a moat.

>>hadloc+j5
Why do you think cloud providers can undercut OpenAI? From what I know, Llama 70b is more expensive to run than GPT-3.5, unless you can get 70+% utilization rate for your GPUs, which is hard to do.

So far we don't have any open source models that are close to GPT4, so we don't know what it takes to run them for similar speeds.