A couple of days ago it leaked that OpenAI was planning on launching new pricing for their AI Agents. $20K/mo for their PhD Level Agent, $10K/mo for their Software Developer Agent, and $2K/mo for their Knowledge Worker Agent. I found it very telling. Not because I think anyone is going to pay this, but rather because this is the type of pricing they need to actually make money. At $20 or even $200 per month, they'll never even come close to breaking even.
Steve Jobs said something to the effect that he made maybe three CEO decisions a year. I mean, I think these are decisions like, "We're going to open our own line of Apple retail stores", but, still.
Being a CEO isn’t all that different from being a parent of a child from the POV of impactful decisions.
How many critical “parental decisions” have you made in the past week? Probably very few (if any), but surely you did a lot of reinforcement of prior decisions that had already been made, enforcing rules that were already set, making sure things that were scheduled were completed, etc.
Important jobs don’t always mean constantly making important decisions. Following through and executing on things after they’re decided is the hard part.
The banter is actually quite easy to automate. You can hire a human to play golf for a small fraction of what the CEOs get paid, and then it's best of both worlds.
Is it? Take a look at the bot accounts filling up social media (the non-obvious ones). It wouldn't seem to hard to make one that makes 2am posts about '[next product] feels like real AGI' or tells stock analysts that their questions are boring on an earnings call, which is apparently what rockstar CEOs do.
Sneers aside, I think one common mis-assumption is that the difficulty of automating a task depends on how difficult it feels to humans. My hinge is that it mostly depends on the availability of training data. That would mean that all the public-facing aspects of being a CEO should by definition be easy to automate, while all the non-public stuff (also a pretty important part of being a CEO, I'd assume) should be hard.
That no one is offering this says something very profound to me. Either they don't work and are too risky to entrust a company to, or leadership thinks they are immune and are entitled to wield AI exclusively, or some mix of these things.
what about politician level models? i wonder if politicians aren't all copy pasting their stuff from chatgtp right now, at this stage (that would make a nice conspiracy theory, wouldn't it?)
That is just not correct. As someone who has done the budgets for PhD hiring and funding, you are just wildly underestimating the overhead costs, benefits, cost of raising money, etc.
The "3-5" is certainly overstated, but you definitely can hire ONE PhD for that price, just as you can hire a SWE for $120K or a knowledge worker for $24K. The point is that from a CEO's perspective "replacing all the humans with AI" looks a lot less compelling when the AI costs the same as a human worker or even a significant fraction of a human worker.
Being able to control their every move, scale them to whatever capacity is required, avoid payroll taxes, health plans and surprise co-pay costs, equity sharing, etc might make this worthwhile for many companies.
That said, the trade-off is that you're basically hiring consultants since they really work for OpenAI :)
The benefit to an emoloyee is that you don't have to control their every move. They can do work while you aren't even thinking about the problem they are solving.
Again, irrelevant. We're talking about orders of magnitude here. Current pricing is in line with most SaaS pricing - tens of dollars to hundreds of dollars per seat per month. Now they're suddenly talking about thousands of dollars to tens of thousands of dollars per seat per month.
$20k can't get you that many PhD. Even PhD students, who's nominal salary is maybe $3-5k a month, effectively costs double that because of school overhead and other stuff.
Does depend on where your PhD lives and what subject their PhD is in from where, and how many hours of work you expect them to do a week, and whether you need to full-time "prompt" them to get them to function...
Would definitely rather have a single postdoc in a relevant STEM subject from somewhere like Imperial for less than half the overall cost than an LLM all in though. And I say that despite seeing the quality of the memes they produce with generative AI....
Depends on what these PhDs are supposed to do. Also is this an average Phd or a brilliant PhD level? There is a huge spectrum of PhDs out there. I highly doubt these phd level models are able to solve any problems in a creative way or discover new things other than regurgitating the knowledge they are trained on.
Depends on 1) where the university is located (CoL), 2) if they went on strike recently to get paid enough to pay rent.
You can reliably assume that PhD wages must eventually converge to the rent of a studio apartment nearby + a little bit (which may or may not be enough to cover all other expenses. Going into debt is common.)
Pedantic, but they are top of the market in neither since Switzerland is not in the EU, and definitely not in the UK.
But it is true that in Europe, Switzerland PhDs (and professors too) make most. Not just ETH/EPFL as well. UZH (Uni Zurich) has salaries of 50K CHF per year for PhD candidates (with increments every year) -- that's almost 60K USD by your fourth year. This is also true for other universities. And while Zürich is expensive, it is not _that_ expensive.
> $20k can't get you that many PhD. Even PhD students, who's nominal salary is maybe $3-5k a month, effectively costs double that because of school overhead and other stuff.
But you are not getting a PhD worker for 20K with "AI", that's just marketing.
Based on ubiquitous AI trainer ads on the internet that advertise their pay, they probably make <=$50/hr training these models. Trainers are usually remote and set their own hours, so I wouldn’t be surprised if PhDs are not making much as trainers.
If truly equivalent (which LLMs aren't, but I'll entertain it), that doesn't seem mathematically out of line.
Humans typically work 1/3rd duty cycle or less. A robot that can do what a human does is automatically 3x better because it doesn't eat, sleep, have a family, or have human rights.
So this is just going to end up like AWS where they worked out exactly how much it costs me to run a physical server and charge me just slightly less than that?
Thats pretty useless for most applications though. If you're hiring a phd level person you dont care that if in addition to being great in contract law they're also great in interior design.
I disagree. People are so hyper ultra mega specialized these days the cross pollination should be very helpful. Isn’t that the theory behind why so much amounts of training data makes these models better?
Do you have to pay all sorts of overhead and taxes?
I mean, I don't think it's real. Yet. But for the same "skill level", a single AI agent is going to be vastly more productive than any real person. ChatGPT types out essays in seconds it would take me half an hour to write, and does it all day long.
Even worse: AFAIK there's no reason to believe that the $20k/mo or $10k/mo pricing will actually make them money. Those numbers are just thought balloons being floated.
Of course $10k/mo sounds like a lot of inference, but it's not yet clear how much inference will be required to approximate a software developer--especially in the context of maintaining and building upon an existing codebase over time and not just building and refining green field projects.
Man. If I think about all of the employee productivity tools and resources I could have purchased fifteen years ago when nobody spent anything on tooling, with an inflation adjusted $10K a month and it makes me sad.
We were hiring more devs to deal with a want of $10k worth of hardware per year, not per month.
What is interesting is there is no mention of agents on any job I clicked on. You would think "orchestrating a team of agents to leverage blah blah blah" would be something internally if talking about these absurd price points.
Do you have a source for these supposed leaks? Those prices don't sound even remotely credible and I can't find anything on HN in the past week with the keywords "openai leak".
There is too little to go on, but they could already have trial customers and testimonials lined up. Actually demoing the product will probably work better than just having a human-less signup process, considering the price.
They could also just be trying to cash in on FOMO and their success and reputation so far, but that would paint a bleak picture
Never come close to breaking even? You can now get a GPT-4 class model for 1-2% of what it cost when they originally released it. They’re going to drive this even further down with the amount of CAPEX pouring into AI / data centers. It’s pretty obvious that’s their plan when they serve ChatGPT at a “loss”.
