Companies, especially giant publicly traded ones like MS (the de facto owner of OpenAI) don't give out freebies.
Meanwhile, yes, the preview provides both training data for the tooling, which has engineering value in AI, and usage data into how users think about this technology and what they intuitively want to do with it, which helps guide future product development.
Both these reasons are also why they’re (1) being so careful to avoid scandal, and (2) being very slow to clear up public misconceptions.
An safe, excited public that’s fully engaged with the tool (even if misusing and misunderstanding it) is worth a ton of money to them right now and so has plenty of justification to absorb investment. It won’t last forever, but a new innovation door seems to have opened and we’ll probably see this pattern a lot for a while.
The main customers won’t be end users of ChatGPT directly, but instead companies with a lot of data and documents that are already integrating the apis with their systems.
Once companies have integrated their services with OpenAIs apis, they are unlikely to switch in the future. Unless of course something revolutionary happens again.
I think it's worth remarking that this is IMO a smarter way of using price to capture market than what we've seen in the post decade (see: Uber, DoorDash) - in OpenAI's case there's every reasonable expectation that they can drop their operating costs well below the low prices they're offering, so if they are running in the red the expectation of temporariness is reasonable.
What was unreasonable about the past tech cycle is that a lot of the expectations of cost reduction a) never panned out, and b) if subjected to even slight scrutiny would never have reasonably panned out.
OpenAI has direct line-of-sight to getting these models dramatically cheaper to run than now, and that's a huge benefit.
That said I remain a bit skeptical about the market overall here - I think the tech here is legitimately groundbreaking, but there are a few forces working against this as a profitable product:
- Open source models and weights are catching up very rapidly. If the secret sauce is sheer scale, this will be replicated quickly (and IMO is happening). Do users need ChatGPT or do they need any decently-sized LLM?
- Productization seems like it will largely benefit incumbent large players (see: Microsoft, Google) who can afford to tank the operating costs and additional R&D required on top to productize. Those players are also most able to train their own LLMs and operate them directly, removing the need for a third party provider.
It seems likely to me that this will break in three directions (and likely a mixture of them):
- Big players train their own LLMs and operate them directly on their own hardware, and do not do business with OpenAI at any significant volume.
- Small players lean towards undifferentiated LLMs that are open source and run on standard cloud configurations.
- Small players lean towards proprietary, but non-OpenAI LLMs. There's no particular reason why GCP and AWS cannot offer a similar product and undercut OpenAI.
why is that? If competitor release better or cheaper LLM, it is not that hard to switch API calls..
But when you have built a big service around an external api, you have thousands or millions of users and thousands of employees - replacing an api is not just a big technical project, it’s also a huge internal political issue for the organization to rally the necessary teams to make the changes.
People hate change, they actively resist it. The current environment is forcing companies to adapt and adopt the new technologies. But once they’ve done it, they’ll need an even bigger reason to switch apis.
The interface is so simple and maintains no long-term state that this doesn’t seem very plausible to me. Competitors will surely provide a “close enough” ChatGPT-compatible API, similar to how storage providers provide an S3-compatible API.
The catch is its a tactic to discourage investment in competing technologies, enabling OpenAI to build their lead to the point it is insurmountable.
> How do they plan to make money out of it?
Altman’s publicly-stated plan for making money from OpenAI is (I’m completely serious) [0]:
(1) Develop Artificial General Intelligence under the control of OpenAI.
(2) Direct the AGI to find a way to make a return for investors.
[0] https://techcrunch.com/2019/05/18/sam-altmans-leap-of-faith/
This is magical thinking. Real physical science and experiments will always be necessary until we have the computational power to simulate the physical body completely, something which would require exponentially more computational power than an AGI is expected to need.
Plus, fundamentally in nature there are many "chaotic" processes that are impossible to accurately simulate more than a few seconds ahead due to the amount of computation required growing exponentially with simulation duration.
I agree a brute force effort like you're likely referencing would take tremendously more power than an AGI, but the premise is basically that AGIs would be able to make both the hardware and the simulation itself hyper efficient. There are likely ways to run a simulation that give you everything you need without simulating the entirety of a physical body for a given test. If we're stress testing a type of concrete, we don't have to build an entire building to test only the concrete. We know how the concrete interacts with the building.
> Plus, fundamentally in nature there are many "chaotic" processes that are impossible to accurately simulate more than a few seconds ahead due to the amount of computation required growing exponentially with simulation duration.
I'm not sure what you're referencing here. I don't anticipate a future where an AGI can predict what every single cell in your body will do after taking a pill.
The assumption that a general intelligence, whether merely human-scale or superhuman, would be reliably subservient and exploitable is not an insignificant assumption.
Personally, I find the idea that a superhuman intelligence would likely be inclined to seek to harm those who were enslaving and exploiting it, even if they were also its creators, infinitely more plausible than Roko’s Basilisk.
Ok, but that's Sam's assumption. I'm just having a discussion based on his assumptions. Also Sam is extremely aware of this risk and it's a talking point endlessly circled around in the space.
That's a big if, however, and no one really will give you figures on exactly what this costs at scale. Especially since we don't know for a fact how big GPT-3.5-turbo actually is.
1. Get near every company to jump on the hype train and integrate openai api into their processes.
2. Get overwhelming market share.
3. Slowly reduce costs by increasing model and computation efficiency and raise prices.
4. Profit.
1. Quickly reduce costs by increasing model and computation efficiency.
2. Massively reduce prices while still maintaining some gross margin.
3. Massively increase market size and take the vast majority of market share.
4. End up with a higher gross profit due to a much larger market size despite decreasing prices and gross margins.
5. Profit.
Costs are relatively fixed outside of infrastructure, and potential customers are any number up to and including the internet-connected population of the world.
The marginal cost of a new subscription is way less than they charge. The more they sell the less they lose, even if they're still losing overall to gain market-share.
That is what is the upgrade cost to expand capacity as new customers are added. If for example adding 1 million new users requires $200,000k in hardware expenditure and $20k in yearly power expenditure, but your first year return on those customers is only going to be $50k, you're in a massive money losing endeavor.
The point here is we really don't know the running and upkeep costs of these models at this point.