Walmarts with fruit stands are Google and OpenAI

OpenAI may be synonymous with machine learning, and Google is trying to pick itself up, but rapidly multiplying open source projects that push the state of the art and leave deep-pocketed but unwieldy corporations in their dust may soon threaten both. Although not existential, this Zerg-like threat will keep the dominant players on the defensive.

The idea is not new—the fast-moving AI community expects this kind of disruption weekly—but a widely shared Google document put the situation in perspective. “We and OpenAI have no moat,” the memo states.

The gist of this well-written and interesting piece is that while GPT-4 and other proprietary models have garnered the most attention and income, their funding and infrastructure lead is eroding.

If you compare OpenAI’s releases to iOS or Photoshop, GPT-3, ChatGPT, and GPT-4 were very close together. They still happen over months and years.

In March, Meta’s foundation language model, LLaMA, was leaked. Instruction tuning, multiple modalities, and reinforcement learning from human feedback were added in weeks by laptop and penny-a-minute server tinkerers. OpenAI and Google probably poked around the code, but they couldn’t match subreddits and Discords’ collaboration and experimentation.

Could it be that the titanic computation problem that seemed to pose an insurmountable obstacle—a moat—to challengers is already a relic of a different AI development era?

Sam Altman warned that throwing parameters at the problem yields diminishing returns. Smaller may be better, but few would have thought so.

Walmart is GPT-4.
OpenAI and others are following the SaaS model. You provide API-gated access to a valuable software or service. When you’ve spent hundreds of millions on a monolithic but versatile product like a large language model, it’s a simple and effective approach.

If GPT-4 can answer questions about contract law precedents, great—never mind that a huge portion of its “intellect” is dedicated to mimicking every English-language author. Walmart-like GPT-4. The company ensures there is no other choice because no one wants to go there.

Customers are starting to wonder, why am I walking through 50 aisles of junk to buy a few apples? Why am I hiring the world’s largest and most general-purpose AI model if all I want to do is match this contract’s language to a couple hundred others? If GPT-4 is your apple Walmart, what happens when a fruit stand opens in the parking lot?

In AI, a large language model was quickly run on a Raspberry Pi, albeit in a truncated form. It undermines OpenAI, Microsoft, Google, and other AI-as-a-service providers’ claim that these systems are so difficult to build and run that they must do it for you. It appears these companies chose and engineered AI to fit their business model, not vice versa!

Word processing used to require a mainframe and a terminal for display. That was a different time, and we can now fit the whole application on a PC. Our devices have exponentially increased their computation capacity since then. Today, everyone knows that supercomputer tasks take time and optimization.

Google and OpenAI got there faster than expected. At this rate, they may never optimize.

They’re not out of luck. Google didn’t become the best overnight. Walmart benefits. Companies don’t want to find a bespoke solution that performs the task 30% faster if they can get a good price from their existing vendor and not rock the boat. Never underestimate business inertia!

Yes, LLaMA is being iterated so quickly that camelid names are running out. The developers gave me an excuse to look at hundreds of pictures of cute, tawny vicuñas instead of working. However, few enterprise IT departments will implement Stability’s open source derivative-in-progress of a quasi-legal leaked Meta model over OpenAI’s simple, effective API. Run a business!

I stopped using Photoshop years ago because Gimp and Paint.net are so good. The argument turns now. Photoshop cost? We run a business!

Google’s anonymous authors are clearly worried that the distance from the first situation to the second will be much shorter than anyone thought, and there appears to be nothing anyone can do about it.

The memo suggests accepting it. Share, collaborate, publish, compromise. They conclude:

Google should lead the open source community by cooperating with the larger conversation. Publishing small ULM variant model weights may be uncomfortable. This necessitates model handover. This compromise is inevitable. We cannot drive and control innovation.