The LLM Race

As of Mar 2023, OpenAI is winning in terms of perception - their branding, GTM and their product's accessibility to the mainstream audience is leagues ahead of the others. They're stealing limelight away from the other giants by wowing us with how capable their tech is. And so when the mainstream thinks of AI, they think of OpenAI.

Microsoft (OpenAI) vs Facebook vs Google vs Amazon vs Apple

But we’ll only get to know in the coming months whether Facebook or Google or Amazon or Apple has far better trained models. My prediction is that in the long run, it will be Facebook or Google that takes the lead, because they have so much more data.

Facebook has an enormous social graph. It owns Whatsapp and Instagram which is collectively used by half of the world's population. It has the data that is generated by billions of users - their interactions, preferences, browsing behaviour. And then there are the "invisible" ways in which they collect data - such as having websites install their Facebook Pixel or incorporate "Sign in with Facebook". Noteworthy that of all the tech giants, only Facebook is being led by its original founder, who still owns majority voting power. That will allow him to move fast and make quick decisions.

Google has a ton of data on Workspaces (Docs, Sheets, Forms, Gmail, etc.), they have their Android OS, their vast search data, web browsing data of billions of users, trillions of photos uploaded to Google Photos, 100s of billions of Youtube Videos.

Amazon has recently partnered with Hugging Face. And they have a large amount of data too - their suite of AWS services, 100s of millions of Alexas running on smart devices, their e-commerce site data.

Apple has the advantage of owning their hardware layer, running their own Apple chips in their devices.

The performance of an LLM is determined by the sheer amount of high-quality, diverse data that it is trained on. Whereas the science and the algorithms will get commoditised.

For example, the key innovation that made LLMs possible, Transformers, was developed by Google in 2017. Every time there is a new innovation or a new significant algorithm that gets developed, the rest will quickly implement it. But what is not going to be commoditised is the proprietary data that is being used to pre-train these models.

Which is why I believe OpenAI isn't necessarily going to win the race in the long term.

Well Microsoft owns a significant stake in OpenAI. And OpenAI likely knew back then, that it needed to have a deep partnership with Microsoft to get access to Microsoft's data and computational resources. A quick history detour: The original plan was for OpenAI to be a non-profit. OpenAI wanted to make public all their research (not the case anymore though) and be independent. But when one of the original founders, Elon Musk, pulled out of a planned funding because of a disagreement with the other founders, the rest of the founding team realised that they wouldn't be able to achieve their goal unless they create a for-profit entity - they needed funds to train their LLMs.

Enter the Microsoft partnership. With Microsoft, they also got to tap into training data from LinkedIn, Windows OS, Github, Microsoft teams, Azure, Bing. OpenAI may be the most game-changing investment made by Microsoft. And for the billions of dollars Microsoft poured into OpenAI, most of it comes back to Microsoft anyway because OpenAI is using Azure (owned by Microsoft) to run the GPUs that are training its models. A solid investment by Microsoft.

OpenAI's strategy also seems to be one where they want to capitalise on their first mover advantage. Which explains why they're keeping API costs very low - making it easy for developers to integrate GPT into their products. Switching later on becomes harder. As an example, I signed up for the API access to GPT-4 2 nights ago and they said we're on waitlist. 48 hours later, we've received access. They're moving fast.

The race is between all tech companies, not just the giants

Having said all of the above. The LLM race isn't just between the tech giants. It's also between all tech companies, small and large - over how quickly they are able to incorporate AI into their product or service

For all we know, maybe we'd have each of the tech giants have their own large LLM models out in the market. And then us product builders go with the LLM developed by the giant that makes most sense for our specific use case. Maybe the LLM marketshare will be split by everyone in some manner, similar to how the cloud platform marketshare is split by the 3 major tech giants in some manner.

What's most interesting though is to see how the applications of these LLMs in the various industries play out. Whether it's healthcare, legal, recruitment, tax and accounting, software engineering, project management, design or advertising.

Even in the context of the tech giants, when they will truly capture value is when they incorporate their LLMs into their own products. So Microsoft incorporating it into Github (already doing it with copilot), Linkedin (imagine much better LinkedIn candidate search / sales prospect search / personalised drip campaigns), Skype (meeting transcripts), Excel, Word. And Google incorporating LLMs into their products (Gmail, Search, Photos, Spreadsheets, Docs, Forms, etc). Facebook incorporating it into Instagram and Whatsapp. Each of these giants' already heavily used apps will have a far better user experience. Market shares may possibly tilt from one dominant player to an underdog, as we see in the case of Bing slowly rising from the dead.

And we'll see products that are already somewhat dominant in their respective spaces, such as Linear, Twitch, Superhuman, Tableau, Asana, Atlassian, Mixpanel become far better products. The bar is going to be raised quickly for what is considered to be an acceptable level of user experience.

The possibilities are endless. Every other day I'm blown away by the demos I see.

How I see it playing out

First we'll have each of the tech giants with their general LLMs trained on a large amount of data. Then there is going to be industry specific LLMs - for example let's say we have a recruitment specific LLM that's trained on candidate hiring data. Where we could then use this specific LLM to summarise the strengths, weaknesses and red flags of candidates. Or imagine a legal LLM that's trained on a huge amount of legal documents where you can get summaries of large contracts in a few seconds. Or healthcare-specific LLMs that are trained on a large amount of healthcare data, where doctors can quickly get a list of potential diagnoses based on the patient's data.  

It's in the applications of these LLMs where significant value will be unlocked. We've barely scratched the surface in terms of use cases. For the layperson, LLMs are a chatbot that's on chat.openai.com that seems to give useful answers to questions. But there are going to be so many more breakthrough applications of LLMs - especially in old school, non-tech domains.

And all of this is going to play out over the coming years, a decade where AI will end up powering pretty much everything around us.

Prediction: Now's a good time to invest in GPU manufacturers like Nvidia. As of me writing this, the stock is at US$251.85, up 90.93% in the last 6 months. I predict it will 5x in the next 5 years. Though the bear case for Nvidia would be the tech giants start manufacturing their own GPUs. Similar to how Apple decided to build their own chips.