OpenAI and the era of self-congratulation
There’s a pattern among AI labs such as OpenAI of feeding the world just the details they want in something that looks like research but really amounts to a press release: no independent verification, no examination of the larger questions, such as, Is all this of any value?
Tiernan Ray reflects on what to make of OpenAI’s latest “research.”
OpenAI’s announcement this week of its token spend looks a little like science, which is the impression that’s intended, but it’s really yet another self-congratulatory press release. At least it’s good news for the chip makers such as Nvidia.◆
TRANSCRIPT
(0:35) Hi, I'm Tiernan Ray with the Technology Letter. We live in an age of artificial intelligence, and also of self-congratulatory press releases masquerading as science. So, in a blog post this week under the authorship of OpenAI — it's just OpenAI, there's no names attached to it — is a long, long document about research going on at OpenAI, and specifically “agentic” AI usage. That's where the programs themselves take over some of the coding tasks.
(1:15) It's titled “Research Acceleration: The View Inside OpenAI.” And what is in this document is a long discussion of tokens — basically, how many tokens are being used by engineers. In some cases, six thousand dollars a day by an engineer worth of tokens spent to run the agentic AI programs. Two things bother me about this document. One is there's nothing in here that's independently audited or verified from a company that has kind of become notorious for concealing source code, concealing the parameters of large language models, concealing the data sets, very little disclosure.
(1:47) Here we have a lot of statements about stuff, and there's no way to verify any of it. Two, and much, much, much, much, much more important, is, What is the value of all of this? There is a lot of discussion in this blog post about increased velocity: velocity of agentic work, velocity of token spend, velocity of money spent. And there's not really a discussion about whether what's being done is at all valuable. And I think it's not obvious or certain that all of it or even most of it is valuable.
(2:20) You can have agentic AI creating a lot more programming code than humans. Clearly, that's obviously true just because they automate the process. But is what's being created actually valuable? Is it lots more activity that's producing a better return ultimately in terms of better software developed? I think it's not clear yet. I don't even think there's enough data from OpenAI or anyone else to assert that what's going on is definitely better programs. And I think it's something that everyone needs to think about given the sums of money that are being pledged.
(2:52) OpenAI has said it will spend $1.4 trillion dollars on thirty gigawatts of compute capacity over some unspecified span of time. And when this kind of money in the society is being devoted to burning tokens, one really wants to know, Is it all of great value? Was the activity of OpenAI, whether human or automated, of great value relative to other labs out there also developing large language models and agentic technology?
(3:22) It’s a fundamental question, none of it addressed in these kinds of blog posts, and it's concerning to me. One thing I will tell you, you can be sure with this kind of increased velocity of agentic work, it's very good for semiconductor makers, the chip companies: Nvidia, Micron Technology, Broadcom. These companies are gonna have more and more sales of semiconductors as the burning of tokens keeps driving up the volume of chip activity and therefore the volume of chips sold.