The AI Business Defensibility Problem

How possible is it really to build a business in the new AI space?

With the rise of generative AI, one consistent question is if it’s possible to build a business that’s defensible.

The assumptions here are that most AI products can leverage improving their defensibility by fine-tuning ensemble models on top of the existing. But this isn’t necessarily true and there’s a lot of arguments on why.

What’s the current landscape look like?

As noted previously, the most basic generative AI business you can create is build a user interface on top of a large scale model (GPT-3, Stable Diffusion), and market it towards a specific use case.

The current SaaS AI business models are almost all freemium service. Each business uses a crediting system to make sure that they don’t lose money on free users. And so each business exists on maintaining a free user to customer conversion threshold that will sustain the business - but inevitably lead to saturation as new competitors enter into the space.

The best examples are in content marketing and SEO on GPT-3. Companies like Copy.AI or Jasper have scaled to tens of millions of dollars in annual revenue by making content marketing teams more efficient through leveraging GPT-3 for things like re-writing a paragraph including an SEO keyword or generating an outline for a blog post.

I’ve used it a few times and it does help with generating ideas more than anything else. But if you explore most features - you’ll find that you’re working with very similar products. This makes sense because every copywriting company has found the limit of what GPT-3 can do to it’s max copywriting utility.

This is what is normally defined as a lack of switching cost defensibility. When spinning up a new feature is as easy as looking at what your competitors are doing and writing a text prompt to do something similar - it’s tough to have a competitive moat.

For example, the existing framework for generative AI businesses show a vertical stack across the model layer and the API layer. There are API’s like GPT-3 and Cohere.ai that sell access to their super-language models in the same form of AWS or Google Cloud. And then there are businesses built on top of that like the content marketing ones we referenced as the API layer.

But most people believe defensibility will exist through building out the “middle layer” in the API stack that generates a data compounding effect.

How Would You Build Defensibility

First off - how would you build the compounding middle layer. Well for normal network effects in businesses we can look towards something like how the Facebook newsfeed uses AI.

In the traditional method, user’s would like and comment on posts that they see and AI would train on this data to serve you more posts that you’d engage with.

However in this new generative AI realm, instead we have a generalized AI model that produces text and was trained on billions and billions of words. So instead of continuously modifying this model, instead defensibility can be introduced by adding ensemble layers on top towards specific use cases.

Sam Altman mentions this exact “middle layer” in his interview with Reid Hoffman about how the next startups are built by fine-tuning the existing API models towards a specific use case.

RH:*And what do you think the key things are, given the large language model we provided as an API service? What are the things that you think that folks who are thinking about these AI businesses should think about as to how you create an enduring differentiated business?

SA:*I think there will be a small handful of fundamental large models out there that other people build on. But right now what happens is a company makes a large language model (API enabled to build on top of it), and I think there will be a middle layer that becomes really important where… I’m skeptical of all of the startups that are trying to train their own models. I don’t think that’s going to keep going. But what I think will happen is there’ll be a whole new set of startups that take an existing very large model of the future and tune it, which is not just fine tuning, all of the things you can do.

I think there’ll be a lot of access provided to create the model for medicine or using a computer or a friend or whatever. And then those companies will create a lot of enduring value because they will have a special version of it. They won’t have to have created the base model, but they will have created something they can use just for themselves or share with others that has this unique data flywheel going that improves over time and all of that. So I think there will be a lot of value created in that middle layer.

But this middle layer does not effectively exist yet. For example, if we go back to the copy-writing businesses, we can see that it operates in a simple input and output fashion. The only thing that they do is that when you provide an input, it spits out multiple outputs for the user’s to choose.

We could assume that copywriting AI startups are trying to build the middle-layer flywheel by by adding different prompts to the input texts or running GPT-3 multiple times and then tracking the output that the user selects. By saving which output the user liked, the business could effectively use this data to train their own middle layer.

So we could assume the defensible businesses of the future will have a way to generate training data to create a feedback loop that other businesses won’t be able to. These training data inputs are critical because they’ll be on top of the existing API layer instead of previous data points that big businesses might have had.

But the Feedback Loop has a Flaw

There are some issues with this approach though.

One is the obvious dependencies on GPT-3 for your business. Not only could OpenAI boot your company off of their API, but they could also improve upon their model faster than you can building out the middle layer - rendering your improvements useless in a matter of days with a massive new update.

Allen Cheng does a great job of outlining the major critical issues for why you shouldn’t be building a business around GPT-3.

If it’s easy to make a good-enough app out of the box, the barriers to entry are mercilessly low. Dozens of competitors for your idea will spring up literally overnight, as they already have in these Twitter demos.

It’s not just about new entrants. If GPT-3 is so easy to adopt and build products with, incumbents will do it too. Thus, in Clayton Christensen’s framework, GPT-3 looks more like a sustaining innovation than a disruptive innovation. This will strengthen existing winners more than it creates openings for new startups.

If the baseline GPT-3 performance cannot be substantially improved to create a substantial (10x) proprietary edge, the competition will shift away from technology to other dimensions of competition—particularly in marketing and distribution. This is where incumbents beat startups.

Meanwhile, the profits will accrue to the true beneficiaries: 1) the algorithm owners,  OpenAI (and, by extension, Azure), 2) to marketing platforms, particularly Google and Facebook. Both can raise pricing to the point where companies built on each are barely profitable.

And while these are all generally strong points - I think one of the more critical problems come with the model feedback loop.

For example, in our copy-writing businesses, as long as copy-writers and content writers choose an output as the best and uses it for their work, then generally the model will show better results and increase customer retention and improvement.

However if we take into the context of copy-writers directly using the outputs for revenue generating services, that is writing articles for SEO or posting on social media, then the feedback loops start to fall apart. Since each output is merely a small sentence or a paragraph of a larger article meant to be consumed by Google’s search indexing engine, it’s hard to truly determine if the middle layer correctly improved upon any company’s SEO results. Simple put - how can a business understand the end value when it’s determined by another algorithm?

This is tough because we can’t attribute the output from a copy-writing AI service to be directly linked to actual revenue generation improvement. If I write an article using copy-writing AI services, it generally takes a few months for Google to correctly index my article where it should stand in the quality for different keywords. Same thing with social media posts. While AI can generate new ideas and see which ones the author created, it can’t currently see how each post performed on Twitter and LinkedIn, and attribute those features back to successful or unsuccessful social media engagement campaigns.

Similarly we’re seeing this in other domains. Charisma, an AI in-game bots company that builds side-quest storylines cannot always understand if the AI generated side-quests really prompt the gamers to really enjoy a game and play it more consistently. Especially when they sell their services to larger players in the market (that could always then build out their own AI services).

While this may not be a problem for businesses looking to increase product growth and customer retention through their own internal feedback mechanism, as businesses look to grow to provide more revenue facing value - they are increasingly dependent on external feedback loops. And it’s a question of whether or not AI can effectively build in those external feedback loops to generate a better middle layer overall.

Data Augmentation

One last thought comes from an article I read about AI being superhistory instead of super intelligence.

One way to think of this is: these AIs have already read vastly more text than I could in a thousand years, and digested it into writing minds (language models) that are effectively Ancient Ones. And their understanding of what they’ve digested is not limited by human interpretative traditions, or the identity insecurities of various intellectual traditions (AIs, blessedly, do not suffer from the limiting temptations of Straussian respect for tradition).

If I connect to a writing-assistant AI in the right way, even with significant inefficiency, I'd be effectively writing like a 1046-year old rather than a 46-year old. If I could learn to go spelunking in the latent spaces of these models, I’d be able to write in ways no human has ever written before.

The whole article is a fascinating read but one argument is the fact that humans can be data-augmented through AI. So while we have all-encompassing AI’s that can write more effectively than us and has more knowledge than we ever can - we can learn from an AI trained on thousands of years of writing history to helps us communicate with other humans.

I think the case for actually building an effective middle layer is to make this data augmentation with AI better and better in a way that benefits both parties. In which the company can also build upon the increased benefit of the copy-writer getting better in the same way that the copy-writer benefits from using the AI tool.

I’m not sure how this connected knowledge transfer could work but it definitely involves continuous engagement with customer inputs. If businesses can survive long enough to get enough customer inputs from augmented AI, then potentially there’s a feedback loop that actually would work well.

Lastly, the author argues that humans are rather useless. However as we noted before we still have to serve to humans.

For example, of course an AI that digests a gazillion real interview data sets will end up learning and reproducing all the hiring biases of humans. Nobody with even a passing understanding of the technology was surprised by this.

This says less about the limitations of AI than it does about the repetitiveness and mimetic redundancy of the human-life-data used to train it. If you could turn interviewing into a “play against yourself” model training regime as DeepMind did with Go, you’d break out of this in no time. The AI would start hiring totally surprising candidates for jobs, who would go on to perform in almost magically effective ways. Somebody will figure this out soon enough, by turning hiring into a Go-like game.

Chances are, human organizations will have to be entirely redesigned around these new, computationally mediated ways of knowing each other and doing together. If you expect traditional organizations to merely “adopt” AI-based systems within traditional functions, you’re in for a shock. Slowing down AIs to operate at human institutional speeds is to beg the question of their value altogether.

At some point - when all AI’s just have to communicate with other AI’s for AI influence, we can just sit back and watch.