| > However, I do have the ability to read publicly-available data, Maybe, but based on the egregious errors the author has made in previous articles, they probably don't have the ability to understand or reason about any of the data they read. Also note that despite what's implied by this statement, most of this article is not sourced, it's just the opinions of the author who admits they have no qualifications. I didn't read the entire gish gallop, but spot-checked a few paragraphs here and there. It's just the kind of innumerate tripe that you should expect from Zitron based on their past performance. > Have a significant technological breakthrough such that it reduces the costs of building and operating GPT — or whatever model that succeeds it — by a factor of thousands of percent. You can't reduce the cost of anything by more than 100%. At that point it's free. But let's consider the author's own numbers: $4B in revenue, $4B in serving costs, $3B in training costs, $1.5B in payroll. To break even at the current revenue, OpenAI need to cut their serving costs and training costs by about 66% ($1.3B+$1B+$1.5B<$4B), not by "thousands of percent". > As a result, OpenAI's revenue might climb, but it's likely going to climb by reducing the cost of its services rather than its own operating costs. ... Sorry, what? Reducing operating costs does not increase revenue. And I don't know how the author thinks that reducing cost of services would not reduce operating costs. > OpenAI's only real options are to reduce costs or the price of its offerings. It has not succeeded in reducing costs so far, and reducing prices would only increase costs. Reducing prices does not increase costs. > I see no signs that the transformer-based architecture can do significantly more than it currently does. So, here's a prime example of the author basing the "analysis" on them personally "seeing no signs" of something they have no expertise to evaluate. There's no source for this claim, and it's pretty crucial for their conclusions that transformers have hit a wall. > While there may be ways to reduce the costs of transformer-based models, the level of cost-reduction would be unprecedented, But for a given quality of model, haven't the inference costs already gone down by like 90% this year? > particularly from companies like Google, which saw its emissions increase by 48% in the last five years thanks to AI. It should be pretty obvious to somebody who can read publicly available data that all of the increase over 5 years can't be attributed to AI. |
"Egregious errors in previous articles" is not a valid argument against current arguments, nor do I agree there were those errors. Nevertheless, we're discussing one particular article today!
"I didn't read the entire gish gallop, but spot-checked a few paragraphs here and there. It's just the kind of innumerate tripe that you should expect from Zitron based on their past performance."
Well that's not very nice! It also means that your argument is made on incomplete data.
"... Sorry, what?
Reducing operating costs does not increase revenue. And I don't know how the author thinks that reducing cost of services would not reduce operating costs."
I'm afraid you misread what I said, likely because you (and I quote) "spot-checked a few paragraphs."
One of the problems OpenAI has is that their cost of revenue - and we don't know it to be exact - is extremely high, higher than the revenue they're actually gaining, otherwise known as an "operating loss." As a result, even if they increase revenue, they'll actually lose more money. On top of that, the argument I was making is that if there's a race to the bottom (one that's already started), they will have to cut costs, making them less money even if they get more customers.
"Reducing prices does not increase costs." Does reducing prices reduce operating expenses? Because if it doesn't, it actually does increase costs, because you're taking home less cash for the same cost. It could be that 4oMini is somehow more efficient - i can find no evidence that that's the case, and if it exists, I will happily update my article.
"So, here's a prime example of the author basing the "analysis" on them personally "seeing no signs" of something they have no expertise to evaluate. There's no source for this claim, and it's pretty crucial for their conclusions that transformers have hit a wall."
I can find no examples of radically-different functionality in GPT or other mass-market transformer-based models. In the event I am wrong, I would be fascinated to read about them, but I would need to understand A) how these functionalities are different and B) how they can be productized. After that, I'd need to understand how this would be profitable, and in turn how this would scale into something truly world-changing.
"But for a given quality of model, haven't the inference costs already gone down by like 90% this year?" Have they?
"It should be pretty obvious to somebody who can read publicly available data that all of the increase over 5 years can't be attributed to AI."
I too read publicly-available data, and my source in this case is "Google."
Forgive the messy copy-paste. https://www.gstatic.com/gumdrop/sustainability/google-2024-e...
In 2023, our total GHG emissions were 14.3 million tCO2e, representing a 13% year-overyear increase and a 48% increase compared to our 2019 target base year. This result was primarily due to increases in data center energy consumption and supply chain emissions. As we further integrate AI into our products, reducing emissions may be challenging due to increasing energy demands from the greater intensity of AI compute, and the emissions associated with the expected increases in our technical infrastructure investment.