Who's winning the AI arms race?

It depends whether AI is a "disruptive" or "sustaining" technology.

This week, Artificial Ignorance crossed 5000 subscribers - I wasn’t sure where this project would end up when I started it, but I’m pretty proud of where I’ve gotten so far. So to look back at where we’ve come from (and also because I haven’t done any Christmas shopping yet), I’m revisiting an essay that was published back when I had just 250 subscribers - now updated for December 2023. See you Friday!

---

For the entirety of 2023, we’ve been in the middle of an AI arms race. It was kicked off by ChatGPT, and it has led to hordes of companies pushing larger, faster, and cheaper ML models.

Google, Microsoft, Meta, and a litany of AI startups have rushed to get product announcements and launches out the door, consequences be damned. And given how unreliable and insecure LLMs can be, it can feel as though the entire industry has adopted the “move fast and break things” mantra.

But who’s actually winning in all this fighting? From my perspective, there’s two main frontrunners emerging - incumbent platforms and cutting-edge innovators. The key thing they have in common is shipping great products, paired with the latest models.

Artwork created with Midjourney.

Sustaining the incumbents

In The Innovator’s Dilemma, Clayton Christensen makes a distinction between “disruptive” technologies and “sustaining” technologies:

Most new technologies foster improved product performance. I call these sustaining technologies. Some sustaining technologies can be discontinuous or radical in character, while others are of an incremental nature. What all sustaining technologies have in common is that they improve the performance of established products, along the dimensions of performance that mainstream customers in major markets have historically valued. Most technological advances in a given industry are sustaining in character…

Disruptive technologies bring to a market a very different value proposition than had been available previously. Generally, disruptive technologies underperform established products in mainstream markets. But they have other features that a few fringe (and generally new) customers value. Products based on disruptive technologies are typically cheaper, simpler, smaller, and, frequently, more convenient to use.

Despite the breathless talk that LLMs will disrupt everything, it seems to me that so far they’re closer to a sustaining technology - with Big Tech and other incumbents benefiting the most.

Let’s look at Github Copilot - one of the first AI products publicly known to have reached $100 million in ARR - as an example. From the start, Github (and other large companies) have big advantages in building something like Copilot:

• They employ (or can go out and hire) internal teams to develop new models and build hosting infrastructure.

• They have the resources to train and fine-tune models (or in Github’s case, their parent company had the resources to exclusively license OpenAI’s models).

• They can collect enormous amounts of usage data and feedback, just from the size of their existing user base.

As it turns out, that last point may actually be the most important. Training data - including user feedback - is becoming the biggest competitive advantage for AI companies. Hosting infrastructure and model architectures are rapidly becoming commoditized, with proprietary datasets becoming a defining feature. It’s why we’re seeing partnerships like OpenAI and Axel Springer, and warnings that we could run out of high-quality training data as soon as next year.

Source: Github

Even if you have a good training data set, however, the initial training still leads to a model with pretty rough edges. That’s where fine-tuning comes in, and why user feedback can be so, so valuable. In the case of Github Copilot, here are some metrics that Github can track:

• Did the user accept the suggested code?

• Did the user accept some of the suggested code? If so, which parts were edited?

• Which suggested code is still present in the codebase after 5 minutes? After 5 days? After 5 months?

With millions of data points, Github can fine-tune Copilot faster than any startup. And as the product gets better, it gets more usage, which makes it even better. Incumbents that can get their AI act together have a massive advantage.

I say "incumbents," and not "FAANG," because this playbook works across many verticals. Intercom has millions of support agents to improve its customer service LLM. Canva has millions of designers to improve its image-generation. If CEOs of big platforms can develop a coherent AI strategy, they can build a flywheel to cement their advantage.

Of course, "developing a coherent AI strategy" is much, much easier said than done. Generally speaking, big companies aren't exactly known for their agility, but it’s worth noting that this current wave of large software companies have been incredibly fast at integrating AI. It’s not every day that you see organizations as large as Shopify commit to putting AI into so many aspects of their product.

How Shopify is making AI Magic

And if I’m wrong about AI being a sustaining technology instead of a disruptive one, then we’re likely to see a second group of winners grow: cutting-edge startups.

A window of opportunity

Consider this: Google invented the modern LLM architecture, and had GPT-3 level technology years ago - but they never released it as a consumer product. To me, there’s still opportunity for upstarts to establish a lead with cutting-edge tech, as long as they can be turned into great products.

In the case of Google, OpenAI now appears to be firmly outclassing the former "AI-first" tech giant. It remains truly impressive that nobody has released a general-purpose model that’s better than (or even really as good as) GPT-4, despite nearly a year of time to catch up. That’s due to a combination of being both great at research and product.

OpenAI demonstrated its product prowess at DevDay, its first developer conference. In addition to releasing a better model (GPT-4-Turbo), it also launched GPTs - the AI equivalent of the App Store. With GPTs, OpenAI can give ChatGPT new skills, keep users around for longer, and can leverage external tools and datasets. So far, GPTs actually seem to be finding adoption - unlike plugins, the company’s previous attempt.

Likewise, Midjourney remains the best-in-class image generation tool, even though plenty of competitors are working to catch up. A few of the key product decisions they've made:

• Building a Discord app, lowering friction for millions of users (while also neatly sidestepping hosting costs). However, as they’ve scaled, they’re moving to a dedicated website to grow beyond Discord.

• Making users vote on an image to get a high-quality version, which bakes user feedback directly into the process. There is a huge amount of value created just by using the right UX.

• Having good-looking default settings, instead of needing tons of prompt engineering. DALL-E 3 also sidesteps prompt engineering, but its outputs aren’t always as aesthetically pleasing.

Source: Google Search Trends

But the window of opportunity for foundation model developers is small (and getting smaller), for a few reasons.

First, the massive capital requirements. Any new company that’s trying to mount a serious challenge to the incumbents is going to need plenty of funding - both to train the model, and to serve users if it takes off. And now that the world is aware of generative AI’s potential, billions in VC funding have poured in and created a race to the bottom on pricing. The hottest open-source model, Mixtral, is currently available for free on OpenRouter.

Source: OpenRouter

Second, less sharing of notes. Up until (very) recently, AI companies tended to share research notes in public. In fact, OpenAI was founded on the premise of sharing all their model data and code. Now though, OpenAI no longer believes in publicly sharing models - making it that much harder for competitors to catch up. And Midjourney, for its part, has never shared any of its breakthroughs.

And third, we appear to be reaching some of the limits of LLMs. There’s plenty of optimizations to be made, but it’s unclear if we’re going to keep improving performance on the scale of GPT-3 to GPT-4. As model performance begins to plateau, there’s less room available for new startups to compete - though there’s always room for competition in new frontiers, like we’re seeing with text-to-video and text-to-3D.

AI for the rest of us

Where does that leave everybody else? Is there a path to success if you’re not an existing tech company or a well-funded startup full of ML experts?

AI, like any ecosystem, has room for smaller creatures to thrive while the biggest carnivores fight. The incumbents and innovators will take the lion’s share, but there will be many, many individual winners within smaller niches. Here are a few of them:

Artwork created with Midjourney.

Open source developers. Meta has done an enormous amount to power the open-source ecosystem this year. Llama 2 gets all the attention, but its released plenty of other open-source models as well. And its set the benchmark for other open-source companies like Falcon and Mistral to beat, meaning better models for everyone. That said, monetizing the open-source models will remain a challenge for everyone besides Meta - just look at Stability AI, for example.

Infrastructure providers. If I had to guess, the single biggest winner in the AI boom might just be Nvidia. It’s selling its chips as fast as it can make them, and this summer it became a trillion-dollar public company. At the end of the day, models still need resources to train and run, there's no getting around that. Platforms like Hugging Face and Replicate are serving startups as AWS and Google Cloud serve enterprises. And with the recent advancements in no-code tools, we’re quickly making “train your own model” a simple point-and-click task - meaning even more compute is necessary.

Domain experts. Hospitals and governments aren't going to be sending their data to OpenAI anytime soon. Companies bringing compliant AI to regulated spaces will create a lot of value. Likewise, there will be AI for incredibly specific use cases too small for existing incumbents. There’s plenty of low hanging fruit when it comes to partnering with existing companies in boring industries, and building AI-powered solutions with better UX or better fine-tuning.

For better or worse, nearly every CEO now needs an AI strategy - how to add it to their workflows, and how to adapt to a world where it is everywhere. What parts of your organization can be augmented by AI? Are there new products or services that used to be too expensive, that can now be unlocked with AI?

If you’re a consumer, get ready for a tidal wave of AI-powered apps. For now, we’re still playing with AI-assisted writing and AI-assisted drawing. But soon, we’ll be working with AI-assisted thinking and AI-assisted planning. Try out as many of these things as you can - even if they feel like toys now, they’re going to be very capable, very fast, and you should know how to figure out the limits of these tools.

To reuse a metaphor, we're in the middle of a Cambrian explosion of AI tools and startups. Most won’t survive long-term - they’ll become features of bigger products. But there’s still an enormous amount of opportunity in this brave new world, for those who can keep up.