If you want to hire a mid-career PhD in a field that has a real job market outside the academia, and the person already has a stable job they are content with, $240k might not be enough. At least unless the job is particularly interesting.
And the total cost of employment is obviously higher than the nominal salary. (Which is mostly a convenient lie for tax purposes.) If the total cost including benefits, employer-paid taxes, and mandatory insurance is $240k/year, the nominal salary could be $180k or even lower.
In many places of Europe, for example, the salary after taxes and mandatory employer contributions become between 30% and 40% of what the employer disburses. The employee is aware that his 20k raise will become more like 6k, and it's in many cases not worth the increase in responsibility, workload or inconvenience. So employee is more likely to seek fulfillment, economical or otherwise, by other means.
But this means that in a few years, people earning 80k per year will be working under the supervision of an AI earning 240k per year. In my sci-fi books, I was leaving that for the 2040s...
Why are phd-level-agents paid twice as much software developers? Why don’t we see similar value in the real world? $2,000 for high income knowledge workers? None of these numbers make sense.
Also I don’t really have any intuitive sense of what a phd-level agent would be, the cynical part of me thinks it is a marketing term because people mentally associate one with smart.
I'm bullish on agents, but I do wonder what this agent could possibly be doing to justify a $240k salary, which is around market rate for a CEO at an SMB.
It would need to be generating leads and closing sales, or doing the day-to-day work of at least 3 people.
This would have to be backed by something 10X better than the current SOTA models.
It's nowhere near CEO pay for a tech company, but it is right at what those companies pay for PhD ML experts and also what they pay for the top end of software dev jobs.
I'll tell you who will be liable for AI's mistakes. The crickets. Yes, the crickets you hear every time this question is asked. We must punish the crickets severely and we will, trust me!
I could see paying this to have one super intelligence on staff at a large company but that isn't what these models are yet.
Most business tasks and problems you certainly don't need a PhD to solve. Even if you could hire human PhDs at min wage I don't see how most businesses can even leverage that.
If all bank tellers had PhDs in quant finance it would change basically nothing.
What part of it doesn't make sense? These numbers are very much inline with salaries for jobs at those same levels, in fact terrifyingly so. Someone thinks, incorrectly, that these agents are a direct swap for that experience.
Even PhD SWEs are likely to make the same (or even less) as their peers who started working after their undergrads rather than going to school for 5-8 more years, especially if their topic isn't hot (read: not AI related).
Who knows if this leak is credible, but typical SaaS pricing dictates you should be charging 10 - 25% of the value you bring. That implies this agent is expected to deliver 1 - 2M+ of value, essentially replacing a medium size team of knowledge workers. Even as an AI bull, this level of capability seems far fetched.
I used to work on storage software pricing. We would charge 60% of the value it would bring our customer - based off of the classic behavioral game theory experiment, Dollar Auction. Our customers surprisingly didn’t object.
This was an OEM storage offering back when we sold hard drives. We found a software improvement to our storage arrays which resulted in the customer needing fewer drives. Let’s say that instead of them needing 100 drives, with the software, they only needed 90 drives. So, we would sell them the 90 drives, and a software feature key that cost the equivalent of 6 drives.
We had an OEM relationship with the vendor. Basically one where we didn’t want to upset our long standing goodwill. The dollar auction is accepted at 100% when you offer to keep 60%, suggesting that the receipt does not consider that an insulting offer.
OpenAI is a financial black hole, burning cash faster than it can raise it, with a business model built on hype and a product (ChatGPT) that's easily commoditized. Their revenue projections are delusional, and the core API business is surprisingly weak, suggesting the entire generative AI market might be overblown.
We're in the early days, like back when Uber offered $2.50 rides anywhere in San Francisco. They burned through tons of cash for years but had $9.8bn net income last year.
At least I trust Uber enough that I use it anywhere I don't get to drive my own car. And I have never been in an accident so far.
I don't trust ChatGPT enough to copy its output and be done with my work. Sometimes I spend more time prompting + revising the response than writing it myself from scratch. It's like Uber is so bad that it is faster to walk than getting matched to a driver who very slowly drives to the destination. And yet you get into a car crash during your trip. I doubt Uber would still be in business today.
I don't care too much about revenues and expenses exactly, but about the sustainability of their capital structure. "Big tranches of convertible debt within a couple years of converting" does sound like a bad sign.
It seems that way but I think it's quite possible in tech for things to rapidly change. E.g. if OpenAI starts quantising all their models, which can maybe reduce their compute costs by 80%, it would be profitable, but make some number of customers leave. Certain other headwinds like decreasing cost of cost per teraflop helps them.
Like Uber I suspect their business model is simply to capex until all their enemies are dead, then jack up prices.
Responsible, maybe, but Liable? I don't think so. So long as you genuinely try to do your best and don't break the law, the a company can do is fire you.
I am guessing the capability and capacity would be limited in some way by compute. They are probably thinking of a dedicated agent that could be computing 24/7 trying to solve some problem assigned.
Note that OpenAI is notoriously terrible at pricing. Their prices are based on gut-feel and marketing, more than being priced on what customers would pay and what is profitable.
Building on top of this would be a terrible idea. The bubble is going to burst, OpenAI will stop getting investment money and will need to actually start being, you know, profitable. They will do that by milking all their actual existing customers - expect price to at least double in the following years.
My question here is if they do bring to market what they claim, as in AI fully capable of replacing humans, what's stopping them of just simply cloning the business model of every company running on knowledge work, and not even bothering to lease out these AI employees?
47 comments
[ 3.5 ms ] story [ 106 ms ] threadIf you want to hire a mid-career PhD in a field that has a real job market outside the academia, and the person already has a stable job they are content with, $240k might not be enough. At least unless the job is particularly interesting.
And the total cost of employment is obviously higher than the nominal salary. (Which is mostly a convenient lie for tax purposes.) If the total cost including benefits, employer-paid taxes, and mandatory insurance is $240k/year, the nominal salary could be $180k or even lower.
In many places of Europe, for example, the salary after taxes and mandatory employer contributions become between 30% and 40% of what the employer disburses. The employee is aware that his 20k raise will become more like 6k, and it's in many cases not worth the increase in responsibility, workload or inconvenience. So employee is more likely to seek fulfillment, economical or otherwise, by other means.
But this means that in a few years, people earning 80k per year will be working under the supervision of an AI earning 240k per year. In my sci-fi books, I was leaving that for the 2040s...
Also I don’t really have any intuitive sense of what a phd-level agent would be, the cynical part of me thinks it is a marketing term because people mentally associate one with smart.
It would need to be generating leads and closing sales, or doing the day-to-day work of at least 3 people.
This would have to be backed by something 10X better than the current SOTA models.
Who takes the liability for the agent’s mistakes and hallucinations?
Where can one hire PhD ML experts for $240k?
I think you can PhD ML experts in Sweden for substantially below 240k, maybe even as little as 124k.
Most business tasks and problems you certainly don't need a PhD to solve. Even if you could hire human PhDs at min wage I don't see how most businesses can even leverage that.
If all bank tellers had PhDs in quant finance it would change basically nothing.
https://www.wheresyoured.at/oai-business/
Or, maybe worse, would someone be able to make your uber drop you off at their place by sending the driver an sms before you even started the ride?
OpenAI is not.
I don't trust ChatGPT enough to copy its output and be done with my work. Sometimes I spend more time prompting + revising the response than writing it myself from scratch. It's like Uber is so bad that it is faster to walk than getting matched to a driver who very slowly drives to the destination. And yet you get into a car crash during your trip. I doubt Uber would still be in business today.
Like Uber I suspect their business model is simply to capex until all their enemies are dead, then jack up prices.
Paying $240k/year for company-wide access (if you’re a Fortune 500 company), isn’t bad.
It also never calls in sick or sexually harasses an intern.
Who is liable for the agent’s mistakes.
But even if they don't, who is responsible for the mistakes. What if an error leads to people getting hurt?
Building on top of this would be a terrible idea. The bubble is going to burst, OpenAI will stop getting investment money and will need to actually start being, you know, profitable. They will do that by milking all their actual existing customers - expect price to at least double in the following years.