8
votes
OpenAI hits more than one million paid business users
Link information
This data is scraped automatically and may be incorrect.
- Title
- OpenAI considers pricier subscriptions to its Chatbot AI, The Information reports
- Published
- Sep 5 2024
- Word count
- 184 words
1M subscribers at $20/mo is only $240M/y - There's no way that covers the R&D cost of training new models so I'm not surprised they're looking at pricier models to reduce the gap.
But can these models provide enough value to justify a 5x price increase, let alone 2 orders of magnitude?
That's only ChatGPT Plus subs. API access is separate, and costs a fraction of a cent per transaction, metered per token. It's still quite cheap but is well above $20/mo in cost.
The current models do what we need them to do, fairly well, so I don't know that I'd be interested in a pricier model unless it represented a true fundamental leap in capability that would allow it to expand in usage. Specifically, if they solved hallucinations that'd be something. I don't think that's likely.
I feel like "solving hallucinations" doesn't describe the problem well. This paper advocates calling the problem "bullshitting" because these models don't reason about the question/answer, they just output the text which minimises the error - it's a subtle but important difference.
It's fundamental to how these models work that I don't see how it can be truly solved, you can only try to make the error percentage smaller
https://link.springer.com/article/10.1007/s10676-024-09775-5
Yeah, that's why I said I don't think it's likely. A true leap in capability seems like it would necessitate a fundamental reimagining of how the model works.
And these transformer-based LLMs were already a fundamental shift in machine learning, so I have dim expectations. I do find LLM summaries legitimately useful for some tasks, but the field still seems overhyped. Anyone familiar with the history of AI and machine learning should have a similar view. The first neural-network self-driving car was tested in 1989, and we still don't have robo taxis over 30 years later.
Yes, sometimes things take much longer than people imagine, but robo taxis are real in a few places. Nothing is announced, but maybe Waymo will get to where I live in a couple years?
I’m wondering when they will be confident enough to start scaling up significantly, in multiple new cities. Probably not until it’s profitable somewhere. Any year now?
I think this is going to be a much harder scaling problem than people think, because with expanded numbers and range come inevitable accidents, vandalism, etc. The tech is better than it was but it's nowhere near good enough to stand up to what American roads are going to put it through. One or two high profile deaths from driverless cars will torpedo adoption and it'll stall out after that. They'd have to be incredibly lucky or have extenuating circumstances that currently don't exist to get this pushed into wide, mature adoption.
There have already been high profile deaths from driverless cars, including one that didn't slow at all while it hit a jaywalker, and one that hit someone, and not realizing what the problem was, ran over them again while following the programming that told it to pull over to the side of the road after an accident. Even with local protests regarding their rollout, driverless vehicles are still here.
Uber got out of driverless cars entirely after one death. Probably for the best. Not sure why other vendors should be affected? It's a different system and anything Uber did wasn't their fault.
I believe it was Cruise's rollout that got cancelled after a particularly gruesome accident.
In a way, these companies are being held to pretty high standards compared to the daily mayhem on the roads that we mostly accept. You don't see other transportation systems shutting down after one accident. But it's not just the accident, it's what it says about the company's culture. For Cruise, it was in part the weird way they acted after the accident - not exactly a coverup, but very passive. It seems regulators expect full cooperation during a serious accident investigation, and GM was not happy about how they handled it?
Edit: looks like Cruise might be back soon.
When I say high-profile that's not quite what I mean. I'm aware of the case you cited but I'm fairly certain 99.9% of the population is not. I don't think any death from driverless cars will make more than a local splash until people start seriously attempting to roll it out nationwide; once that happens and people have the cars on their local streets then a death will be a more national concern.
Fair enough, but then I'm not sure any driverless car accident will be real breakout news unless it's equivalent to a mass murder.
Does it need to stand up to American roads, though? People live in cities. By merely covering the LA metro area, which Waymo is starting, you cover a customer base that's more than twice the size of Denmark's entire population.
It doesn't need to perfectly generalize over all possible American roads; you can cover the high density roads first.
also this:
I’m asking this sincerely, because I can’t think of anything…but how does OpenAI justify that cost? What does $2,000 get you that would make it worthwhile? It’s just a predictive text model.
Depends on how many users that includes. Maybe it's a plan for big companies. Otherwise, I agree.
It sounds like it might not be just another LLM. We don't know what it is, how much it costs to run it, or how many users they can support.
Let’s say it’s really expensive to run and only supports one user at a time for now. How much should they charge to not instantly be overloaded?
What the whales would get: being the first to try it. Most people would wait.