The emergence of Large Language Models with their impressive beyond-human performance (in many benchmarks) gives us reason to ask: what is the strategic value of the underlying principles to the enterprise?
As an antidote to many AI books with faux-paradigms, I explore the at least one core paradigm of Large AI, which is the ability to address Complexity through scaling (Large AI models are Complex Systems).
Because enterprises are Complex systems, we should expect that they too are amenable to computation using Large AI in the same way human language (also a Complex system) has proven amenable.
I call these new models Large Enterprise Models (LEMs). Their emergence could have profound affect upon corporate strategy generally, but certainly IT strategy.
Introduction
Many AI solutions in the enterprise exist within a standard digital transformation paradigm: AI as an embedded technology choice addressing a localized use case (e.g. supply chain management or product recommendation).
Some biz leaders are trying to parse what the apparent inflection-point around Large AI (e.g. ChatGPT) means in a more strategic sense, if anything.
The motivation of this article is to provoke thought about an “AI-first” mindset that asks: What happens if we build our enterprise around Large AI?
My post will be short and non-technical, leaving technical justifications for a long-form elsewhere, most likely on my site.
First, Beware of “Pop” Paradigms
Inevitably, some biz leaders are seeking answers in AI books or magazines.
However, many follow the same glib pattern of introducing quirky metaphors (“Tipping Points”) sprinkled with uncritical marketing case studies. They sound convincing, only to yield little actionable substance.
Many case studies lacks critical details, such as why, exactly, AI was the best choice and what percentage of improvement was uniquely attributable to the AI. (You should always ask this question.)
Some books invent concepts – e.g. “Cognitive Enterprise”. It sounds convincing, but is marketing blurb that wouldn’t translate into a CIO, CTO or CMO world. (Read Richard Feynman’s advice on meaning versus marketing jargon.)
Analyst outputs are often no better, with uncritical claims like “AI high flyers” attributing 30% of revenues to AI.
Really? How, exactly?
Lacking counterfactuals and critical contextual details, we don’t know.
Many transformation improvements are not driven by a single technology, but rather come about via refactoring and modernization. Solutions might contain AI, hence successes will be attributed to AI versus refactoring.
Given the poor success rate of many transformations, the overlap of this with the oft-reported “80% of AI projects fail” might not be a coincidence. As Tony Saldanha argues well in “Why Digital Transformations Fail”, a common failure mode is focus on technology over value. Was AI-for-AI’s sake the culprit?
More often, the reason for failure is simple: the organization wasn’t ready to switch to an automated process. This happens when AI programs are developed in isolation of change management or suitable operating models, like product-centric IT.
AI-assisted Strategy?
Given the connection between strategy and innovation, one framing for exploring the application of AI is via the different types of innovation, per Geoffrey Moore, This is the common method, albeit perhaps via value-streams, not necessarily aligned with these innovation types:
Each types will have a related strategies such as Marketing Strategy, Information Strategy, and so on, hence we can ask how AI might impact these strategies, however they are formulated according to the current operating model:
How might AI assist my <innovation type> strategy?
In reality, no such question is formally posed. In many of these cases, the choice of a solution will be based upon performance against an existing use case.
You pick the “AI solution” if it offers significant performance in relation to alternatives or if it unlocks an otherwise stubborn use case (e.g. something requiring NLP). Increasingly, the AI is embedded into a vendor solution, but sometimes a data science team might devise custom AI solutions.
These selections become strategic when part of a long-term rollout to achieve whatever the overriding strategy is.
Within this rubric, AI can become a core strategic enabler if it’s a common undergirding theme. Typically, this translates into data strategy e.g. to allow better access to joined-up datasets, data augmentation, secure data federation, and so on.
More general strategic imperatives might fall out of such considerations, such as the need to scale data, which cannot be done overnight (can take years) and so definitely requires strategy and might even affect partnerships or ecosystem strategies etc.
AI Paradigm?
More fundamentally, we need to ask whether or not AI presents a paradigm shift, like with the invention of the web browser did.
If so, what is the paradigm?
There is one problem that AI uniquely solves, with strong evidence from breakthroughs in Large Language Models (LLMs), like ChatGPT.
The unique paradigm is solving Complexity (with a “C”).
AI magic has mostly appeared thanks to one thing: scaling. The “magic models” are huge: 175B parameters for GPT-3, and climbing:
What does scaling do?
It allows a model to encompass complexity with its impenetrable web of interactions between many smaller components (e.g. words). AI makes complexity amenable to computation!
Note that scaling is a concept already embedded within complexity theory.
There is a maxim in complexity theory, that complexity can only be “met” with complexity. Human language is a complex dynamical system, which is why Large AI works.
LLMs are perhaps the first real evidence that a complex system (175B parameters) can indeed encompass a complex mechanism (language).
The magic levels of performance only became apparent once the models were scaled significantly (with no other changes). Smaller versions were relatively unimpressive.
Moreover, unexpected capabilities of LLMs continue to emerge merely by tinkering with the inputs (“prompt engineering” – apparently the next hot job).
This has led some thinkers to use the term emergence in relation to LLMs. Well, we should not be surprised because emergence is a key characteristic of a complex system.
Large Enterprise Models
What then, does scaling mean for the enterprise?
Two strategic opportunities arise:
- Systems like LLMs can play a big role in enterprise management, given the undeniable claim that enterprises are also Complex systems (which is what drives innovation) and much of that complexity is presented in language-centric processes. LLMs make enterprise knowledge computable.
- Scaling should also play a role directly in enterprise management – i.e. via novel models applied to the entire information state-space of the enterprise: Large Enterprise Models, or LEMs.
Both of these present significant strategic implications that few enterprises are prepared for, such as:
- Making all data in the enterprise available to AI models. We are still in the early days of concepts like Data Mesh and Data Fabric, but these are a step in the right direction and will involve AI (e.g. via LLMs) for sense-making prior to use in larger models. How do we architect and manage this?
- Do we have a single multi-modal Large Enterprise Model or a highly distributed ensemble of many weaker LEMs (plus other models)? What does that look like?
- How do we operate with increasingly opaque models?
- What is the impact upon organizational models?
- What kind of skill sets are required to work in these hyper organizations?
- Do we need new ways of thinking about partnerships–e.g. at the federated “model” level?
- Do we need a pivot towards Systems Thinking? Where are the frontiers of “The System” in a highly connected marketplace and ecosystem?
All of the above is largely uncharted territory and nobody quite yet knows what kind of new strategy emerges from the paradigm. But the question “What is AI strategy” must recognize the above and figure out how to deal with the new paradigm as applied to all facets of business: customers, processes, markets, people, regulations, sustainability etc.