OpenAI's Limits
Much is still unclear about how AI will differentiate, but one can already see clear structural boundaries for OpenAI.
The Wall Street Journal reported this week that OpenAI is currently seeking new investors at a $90 billion valuation, which would roughly triple its valuation since the beginning of the year.
So, OpenAI continues to soar almost a year after the launch of ChatGPT. This is not surprising. Large language models are quite rightly so the most prominent tech topic of discussion at the moment.
OpenAI is currently the market leader and quality leader. However, this obscures the fact that this situation will not last for the foreseeable future.
We can already identify some structural limits:
Given the costs of training and, still, of inferences, LLM startups like OpenAI have a structural pressure to not only raise a lot of capital, but better yet strategically partnering with hyperscalers like Azure or AWS right away. Instead of passing on the VC check directly towards Azure, AWS et al, thank you.
The strategic partnership between OpenAI and Microsoft has 2
consequences, the first consequence: Everything OpenAI builds, Microsoft can use for its own products and offer in Azure as an API. OpenAI's core technology is not exclusive. As a quality leader, this is a strong API-provider/supplier starting position. But it certainly is not world dominance.
The second consequence of the partnership is that AWS et al are motivated to fund and bootstrap the most promising competitors like Anthropic. It's no wonder both Google and Amazon are pouring big money into Anthropic. They don't want to leave Microsoft with an exclusive advantage in the tough, lucrative cloud market.
It follows from 1+2+3 that the big, generic LLMs will be available as APIs from hyperscalers, whether proprietary or open source. It follows in turn that the quality of the technology alone does not contribute to the business model moat.
Meta has an interest in ensuring that no new gatekeeper grows up. Apple and Google are quite enough to handle already, thank you very much. Meta is working on its own models and is releasing some of them as open source, like recently Llama 2, Meta's first model that is also free for commercial use. (With limitations.) An important building block for a larger Metaverse are AI- generated elements. Meta has a long-term interest here. (Separate from this is the evaluation of how useful the current VR strategy at Meta is).
Open Source: There are currently 343,584 models hosted on HuggingFace. That sounds like more than it actually is, because many models are only slightly different finetuned sister models. But an enormous, far-reaching pool is already emerging here, which, see below, will form the basis for many AI manifestations to come.
B2B & Security: Azure and AWS can make access to LLM APIs watertight data protection wise. They are good at adapting to each country's laws. That's the power of large corporations. But still, there remains a residual structural risk. Not every company has the fundamental confidence necessary to give the core of its data to an LLM in the cloud. Especially not now, when we are still in the early days of suitable security frameworks for LLMs and their emergent properties. On-premise is therefore likely to become relevant for internal LLM deployment at many companies. And thus again, this will mean often open source and/or self-trained. The frameworks and tools for this are just emerging these days.
Specialization on the tech side: it is unlikely that large, expensive- to-train and expensive-to-operate monolithic models are the sole future. Model cascades (Paper) are one way, MoE (Mixture of Experts, the combination of several, smaller, specialized models) is another way to make existing models more efficient. Cheaper, faster, certainly also better soon. It is far from obvious whether OpenAI will be able to maintain its quality edge here.
Specialization on the deployment side: LLM skills are very versatile, in any industry. This broad bandwidth is currently still completely underestimated. From this enormous range of deployment types it follows, among other things, that there is a lot(!) of potential for differentiation on the interface side as well as on the model side. So much, that contrary to the current tech-common-sense, I assume that some ChatGPT wrappers also have chances for a rosy future. (Last not least, because the models in the backend are potentially interchangeable.) OpenAI's plugins cannot cover the interfaces for a whole world.
AI is too big for OpenAI.
That doesn't mean that OpenAI can't get very big. It also doesn't mean that misalignments in the models can't have serious consequences. We are currently still in a situation where very few (model) suppliers are providing the basis for a new tech generation. And OpenAI still dominates this already small group by a wide margin.
But this situation will be temporary.
AI is too big for OpenAI.