Is OpenAI ahead in the AI arms race? Or- is there a race?

I originally wanted to write a text about the differences between generative models that can compete with ChatGPT produced by OpenAI. And whether other players, above all other big tech companies, are able to compete with Microsoft (which is Open AI investor) in this area. And is OpenAI running away from everyone in this race?

Well, because after all, you hear about the 'AI race' in the media all the time, day in and day out: Google buys 10 % of Anthropic for USD 300 million, valuing the company at USD 3 billion. Rumour has it that Meta is abandoning its flagship metaverse project and shifting its focus to developing its own large language model (LLM). And it's Anthropic announcing its chatbot Claude.

But finally, Sam Altman at a conference at MIT declares that the next version of ChatGPT will not be coming soon.

So do we have this race for dominance in generative AI, or don't we? 

It's hard to say unequivocally what this is all about, but I dare to be tempted by my personal interpretation. Because it's not about the technical issues in my opinion different strategies and business models are at stake here. I will first try to present what I have seen and read and finally my personal conclusions.

For starters: believe it or not, I really tried to google and find videos on YT that would reveal the true knowledge of what the technical differences between ChatGPT and its competitors are. And let me be honest, all I found was (well almost only) a bunch of PR videos.

"We are creating safer AI" is one buzzword. The other: "We are more accurate". And beyond that : "We are faster". Here it is already difficult to grasp in numbers and primes to which power we are faster than "the other guys".

But here's a little offtopic: one figure is impressive. According to "The Economist", training a ChatGPT3 costs $4.5 million in electricity alone. The electricity needed to train a GPT4 is a cost of 100 million (!), which is more than 20 times as much. Increasing the data on which the model learns and thus its capabilities results in an exponential increase in energy costs. A cost that increases much faster than the functionality. Perhaps this is why Sam Altman has announced that a larger version of ChatGPT will not be coming soon.

And back to the main thread: Microsoft was ahead of all the competitors in the race by investing first in an LLM development company, because its core business is business applications, including its flagship MS Office suite. Business apps and cloud each account for about 35% of Microsoft's operating profit, so apps are as big as cloud for Microsoft. Integrating business applications with ChatGPT seems like a gigantic breakthrough and creates a huge competence advantage for Microsoft as a whole.

At the same time, ChatGPT can snatch a piece of the search market from Google, helping Microsoft to weaken Google's position. In turn, Google with its search business funds, among other things, its cloud business. Reducing Google's search margins means less money for Google to reinvest in the cloud, which allows Microsoft to defend its position as the No. 2 global cloud provider. So maybe it's not so much about entering the search market as it is about not allowing its position in the cloud business to be undermined.

To reiterate: Google's core business, which funds the other unprofitable business lines, in particular Google Cloud, is search. And here is a noteworthy curiosity. Neither core Google nor Google Ventures, which operates like a classic VC, i.e. for-profit, has invested in Anthropic, but it was the part of the company responsible for cloud. According to Molly Wood of 'This Week in StartUps', this could mean the following scenario: Google invests in Anthropic by contracting the company's exclusive use of Google Cloud. As Anthropic grows and its language model becomes number 2 after ChatGPT (so says the vast majority of commentators on YT) it will start to consume huge amounts of cloud. This will allow Google to become number 2 (behind AWS) in the cloud area and finally make Google Cloud Platform (GCP) profitable. So again: it's not about getting into the LLM business here but about making its cloud business profitable and establishing effective competition with Microsoft and Amazon.

It is interesting to note that after the investment in Anthropic, Google will have three research centres working on LLMs: Anthropic, DeepMind (Chinchilla) and Google itself (PaLM). Does Google want to introduce internal competition? That would be their style and would be quite smart.

The most interesting thing seems to be why Google has not produced its LLM so far. ChatGPT is based on solutions previously developed by Google. Google has the talent, money and scale to have produced the first LLM before OpenAI. But here we are dealing with classic 'Innovator's Dillema'. That is, the implementation of LLM into the business of search to formulate ever better search results jeopardizes Google's ad based search model. Which is not about the most accurate search results- it is about displaying the search results the highest bidder pays for.

Amazon has been neatly described by Wayne Hu as a supplier of weapons in the AI race. It is attempting to offer all smaller businesses based on the big players' LLMs with its cloud solutions.

Although press reports once in a while talk about Amazon's achievements in building its own solutions like the BLOOM model. Amazon also announced a strategic partnership with machine learning model library Hugging Face.

But given the profits from the cloud, Amazon probably just doesn't need to get into generative AI. At least as of now.

Meta like Google is an advertising money machine, but unlike Google it has no idea of developing itself further. Having failed with its own cryptocurrency, the metaverse's flagship project seems to have been shelved. Close on Facebook's heels is China's Tik Tok. Although Meta has released its LLM LLama however, according to commentators it is far behind in the race. As with the Libra cryptocurrency or the shorts in response to TikTok, Meta is counter-reacting to the situation rather than creating it.

Finally, there is the great absentee that is Apple. Seemingly Apple is not joining the race. Yet, Apple has long invested in machine learning solutions for image and sound recognition. It also relies on open source solutions like Stable Diffusion from Stability AI, which are much lighter and do not need as much computing power. They also do not have the capabilities of ChatGPT. On the other hand, they improve software on iPhones perfectly. Maybe this is all Apple needs for now.

Presumably Apple's new chips will be designed with Stable Diffusion in mind.

How to sum it all up?

The race to big tech is not necessarily there at all, although PR-wise it sells well. It is mainly Microsoft that has invested in generative AI. The rest of the players are trying to capitalise on the generative AI boom for their established business models and strategies. Some like Meta have not developed a research base capable of taking on the race.

Or maybe the rest of the players are just sitting tight and watching how LLM technology will be adopted and at what pace. The scale of adoption of ChatGPT so far - 100 million users in 2 months - is gigantically impressive (highest in history), but let's remember that we are in the media hype phase and the current versions of LLM are just the beginning of the journey according to most commentators. Something like the first PC compared to today's PCs.

You can also see what a gigantic role infrastructure has in the development of generative AI. Around the beginning of this year, some media outlets claimed that small teams of researchers are capable of joining the race. They are able to, if they receive huge amounts of cash from the big players. Huge data centres will not simply be put up, the infrastructure will not be set up for an LLM development project overnight. Not to mention the aforementioned energy cost.

Open AI declared $35 million in revenue in 2022 and a loss of $508 (!) million. Not many organisations can afford to burn half a billion dollars a year.

Perhaps Microsoft will follow its Windows famous business model of starting to license ChatGPT en masse. Or IBM's earlier model from the 1980s, where IBM licensed its chips to other computer manufacturers.

And then there is another player in the race for dominance in generative AI. And it is not big tech but the Chinese Communist Party. Here we really don't know much about what is going on behind the Great Wall. What we do know is that a state academy at Bejing University is working on a state backed LLM. Baidu, China's Google, has announced its ERNIE Bot. The question is whether it is possible to build an LLM under censorship. But that's another question altogether.

Videos worth watching: