AI and Sociotechnical Revolutions
the case that AI changes everything
About a month or two ago, I published two pieces on the limitations of general purpose AI and the likelihood that AIs accrue value at the vertical layer and in conjunction with existing software. It seems that many in the tech scene are also starting to share versions of this thesis and the usage data is backing it up.
I still believe that this thesis is true. But I’m also getting more conviction that there will be some massive AI applications that fundamentally change the way organizations and people work So today, I want to dive into the case that AI will revolutionize every industry. And if this seems implausible with current AI technology, I think it has more do to with our current technology paradigm than the actual tech.
If true, this means many of the best AI businesses will be startups creating entirely new sociotechnical paradigms.
Scientific Revolutions
Thomas Kuhn, the great philosopher of the history of science, thinks scientific breakthrough is mostly achieved by a rogue group of scientists breaking out of an existing paradigm. The basic argument is that every single scientist is habituated in a scientific environment that sketches out the world of the possible. Certain conceptions of the world dominate the scientific method - even if they are not fully accurate.
You get a good amount of progress in these paradigms, but advancement itself looks incremental as scientists apply the theories of the day to problems. But over time, anomalies in the existing paradigm crop up. Even if it’s still unpopular to question if these anomalies implicate the entire paradigm as wrong, some do. Old models around the orbit of the planets seem wildly incoherent to Copernicus and we end up with a new paradigm for heliocentric orbit. Einstein decides to go against the Newtonian consensus and develops a new paradigm around relativity.
In every single one of these cases, their peers weren’t stupid. Instead their social imaginary was shaped by the existing (and often rationally held) paradigm. And the methodologies of the day worked well enough. For paradigm shifts to occur, it’s not simply a matter of brilliance (although that’s necessary); it takes a an outside perspective that looks at anomalies as reasons to throw out the current paradigm wholesale.
Something similar seems to be occurring in software. As the novelty factor has worn off, LLMs are being housed inside of a late-stage software paradigm. This is exactly how the public markets primarily seem to value it. These findings aren’t entirely wrong. LLMs will streamline massive amounts of workflows within existing software products.
But there’s a case to be made that AI will bring a revolution. And more and more, this is where I am landing.
But it’s not going to be simply a product revolution, it’s going to be a sociotechnical one.
Sociotechnical Systems
"The ideal conditions for making things are created when machines, facilities, and people work together to add value without generating any waste.”
- Kichiiro Toyoda
As I write about sociotechnical systems, I want to define it pretty narrowly. Sociotechnical systems are the set of practices that directly enable the production and delivery of a good or service - the work-product itself.
The Toyota Production System is the example par none.
With the advent of the Computer Age, Toyota created a novel manufacturing organization. Their new approach to manufacturing meant using human capital, machines, and computers to achieve the maximally efficient result.
The result was a system that has revolutionized manufacturing and has been adapted into a number of industries.
The key idea though is that when you have groundbreaking technology, you must also reconfigure human capital around it. The core innovations are not only technological, they are social.1
You can see this play out in every high-productivity sector. As developer tools have gotten better and better, developer organizations have adopted new sociotechnical paradigms like lean software development to take advantage.
In every industry, this means that technology companies are active partners in reshaping human capital.2 That’s the entire purpose behind vertical startups. By digitizing core industry workflows, you can meaningfully impact sector productivity.
To date, this feels like the largest gap between what’s possible and what’s happening. Since the technology is still so new, there’s no sociotechnical innovation happening. AI is mostly trying to fit into the workflows of a prior sociotechnical system.
It’s partly why it’s hard to imagine how AI could create a new one. We don’t have many practitioners that are working on this sort of systems theory that reimagines how human capital works with AI. It’s far easier to imagine AI improving existing software products.
With all that said, there are plenty of anomalies within the current paradigm where our current sociotechnical system is so inadequate that revolutions are in store.
Entire industries have proven resistant to any measurable productivity impact from software. One of my favorite WTF graphs is the Baumol’s cost disease chart.
I think this graph is a pretty fair illustration of where existing sociotechnical systems are inadequate.
There’s many reasons for this.3 But a simple read tells us that the human capital inputs have remained constant and current software has not been able to alleviate this whatsoever. While management systems might have changed, work-product systems in these sectors mostly haven’t.
In order to fix this, two things must be true. First, AIs have to be agentic and capable of birthing new work-product systems. And second, industry actors have to be willing to rethink their entire socio-technical system to capture the benefits.
If these are possible, AI is not simply going to be a bolt-on to existing software in these industries. It’s going to invite a revolution.4
And since, every sociotechnical system itself is partly verticalized, value will accrue within vertical work-product offerings. A new healthcare paradigm will look far different than a legal one. If AI is going to compel new systems, it too will be vertically-oriented.
Both of these items seem to be not only plausible but probable.
AIs are Capable
The emerging AI stack consists of a) foundational LLM models provided by OpenAI and others, b) vector databases provided by Pinecone and Chroma, and c) the application/interface owned by startups.There are far better write ups than I can provide on each of these components but I do want to hone in on vector databases as a critical part of agentic AIs.
The AI crowd likes to quip that vector databases are the “long term memory for AI.” For a layman like myself, this really means that they are capable of working with data in the native format of LLMs - huge vectors. Data for an LLM is a series of vector points. You ask a question like: “What’s the best way to explain vector databases?” And an LLM is able to probabilistically determining which vectors correspond to the answer and then putting together the text associated with it. “Think of vector databases as a magical library," the AI responds.5
LLMs on their own have pretty short-term memory. They have limited context and thus for very specific datasets, they may not be able to determine the right answer, much less take a series of actions without being continuously prompted. What vector databases allow is for data to be catalogued in the native language of LLMs, enabling them to remember more. This construct also allows startups to develop loops in products where an AI can take actions without direct prompting.6
All of this means that LLMs become Large Action Models: agents.
Imagine an email client where the emails themselves are stored as vector embeddings and AI is default-on.7 That not only means search is more performant, but it also means that an AI might have real latitude to draft responses before you have even thought to reply, tag the most interesting newsletters for you to read (perhaps creating a digest), and has already logged all of your purchases.
There’s no concept of fighting for inbox zero. It’s a given that it will happen; the AI agent’s entire purpose is to ensure it does.8
That’s a trivial example and you can quibble about how revolutionary it is. But imagine something similar occurring for customer success and you have the recipe for a new platform that creates an entirely new sociotechnical paradigm within the profession.9 Customer success transforms into customer AI agent allocation. It becomes a management function with far less rote work attached.10
This infrastructure isn’t exactly something you can retrofit. It implies creating a platform that disrupts how workers currently uses the product. Every company that perceives this as something interesting to do is going to have to disrupt their existing infrastructure to do so.11 If these agents are groundbreaking (and plenty will be), this is just the classic innovator’s dilemma and thousands of companies are going to be on the same footing as the startups for these new platform conceptions.
And in verticals like health, law, and other professional services this will yield even more groundbreaking work-product applications - entirely reimagined from the ground up.
Organizations are Capable
This one’s far easier to envision. If AIs are capable of becoming agents in new product stacks, entrepreneurs in every industry will find ways to adapt and take market share.
In every technology revolution, it’s those at the margins that usually figure this out. Since this is a wholly new paradigm, many of the early innovators will spend time designing new systems prior to gaining rapid market adoption.
Two groups will build these systems:
First, vertically integrated companies. They will own the entire work-product stack and seek to capture large market share from industry incumbents. This happens in every revolution and the same will be true here. What’s somewhat breathtaking is the extent to which every knowledge work industry feels susceptible. You can imagine a more successful Atrium in law competing with Cooley for startup law. A new Pixar in media. A new Bethesda in gaming.
Second, a wave of vertical AI arms dealers will help every industry participant over time benefit from the new paradigm. In some industries, this may actually be the best path to enterprise value. AI tooling might open up a long tail of industry entrepreneurs that previously didn’t exist.
If the cost of great work-product shrinks, that might mean industries that are heavily reliant upon brand as a signifier of work-product quality, will cede market share to a long-tail of entrepreneurs that can match quality at cheaper rates. Vertical AI arms dealers will come in, democratize the technology, and enable Shopify-like waves of entrepreneurship.
What happens to traditional software?
What may not change in a new paradigm is the importance of structured databases that act as systems of record for the management of an organization and the coordination of human and AI labor.
What’s less clear in my mind is if traditional SaaS systems are able to long-term create enough AI functionality to avoid disruption if sociotechnical systems shift enough. It might simply be easier for an AI company to rearchitect these systems as well to accommodate work-product innovation.
We may be at the dawn of an entirely new paradigm. And while there’s still many social, technological, and even political problems that might change how this shakes out, there’s a solid case to be made that each aspect trends towards revolution.
Tesla’s system involved an even wider technological integration spanning the consumer purchase all the way to the factory. “The factory is the product” as Elon says. That includes the factory workers.
This is also why it’s semi-incoherent to talk about AI replacing human labor. The entire sociotechnical system is going to change how human labor itself is used in the system. It’s not a 1:1 replacement.
Not every vertical here of course is curable simply by reducing human capital costs.
And if and when it is successful in these industries, it will invite adjacent revolutions in other ones.
Non-abbreviated response from GPT-4 here:
Think of a vector database as a sort of magical library. In a traditional library, you can find books based on their titles, authors, or other text-based identifiers. In this magical library (vector database), however, each book (item) has a unique magical signature (vector representation). The interesting part is that similar books have similar signatures.
Suppose you enjoyed a specific book about magical creatures and you want to find similar books. In a regular library, you might search for other books by the same author or books with similar keywords in their titles. But in our magical library, you just show the librarian the book you liked, and she uses its magical signature to find other books with similar signatures. Those are likely to be other books that you would also enjoy.
This is essentially how a vector database works. Each item in the database has a vector representation—a list of numbers—that reflects its characteristics. Similar items have similar vectors, and the database can quickly find items that are similar to a given item by comparing their vectors. This is useful in many areas of computing, such as recommendation systems, image and voice recognition, and of course, natural language processing, where words, phrases, and sentences can be represented as vectors.
Thanks to Victor Muchiri for this terminology.
You might still have a traditional email client, but I doubt it’s used much.
The criticism here would be that you may not want AI to take on all of this work-product because you doubt its output. A valid critique currently! But to the extent hallucinations are solved for (by vector databases), humans are in the loop to monitor final output, and these new platforms are capable of creating massive productivity gains, it will happen. In an age of social media ghostwriting agencies, AI email responses won’t be the sacred cow.
The other side of this specific sociotechnical paradigm will benefit. Most consumers don’t actually care about interfacing with a human. They care that their problem gets fixed. Less costly customer success via human and AI agent pairing could plausibly mean surplus gets passed back to customers via faster and more satisfying case resolution.
They might have a data advantage. But it’s far less clear that they have an infrastructure advantage to create these sorts of agents.