In this issue: A monthly update from our Special Interest Group on Protocols for Business (SIGP4B). How are businesses adopting AI and what are current outlooks for productivity, QA, and task estimation?
The current crop of AI tools translate, summarize, and liquidate longstanding bodies of knowledge. They are generative machines. Protocols and AI are like evil twins. Protocols constrain, coordinate, and secure the hidden infrastructure of our day-to-day. We are simultaneously decades into the development of these technologies and in the early days of their diffusion. Protocols and AI are eating the world.
In our first few calls about tensions, trade-offs, and conflicts we realized that the scope of our study group needed to be reined in. We settled on a specific case: AI adoption in businesses. How can we use ideas from protocol studies to address the problems occurring here?
It’s an important question. Last month, OpenAI reported over 700 million weekly active users. That’s nearly one in eight people on the planet. Whether companies like it or not, their employees use LLMs – feeding them business data, strategies, and more. Without a clear plan of action, that leakage-in will generate more liabilities than opportunities.
Interested in getting early access to the SIGP4B litepaper on protocol-based approaches to AI adoption? Help us out by responding to our survey, and you’ll be among the first to get the final version of the litepaper when it’s ready.
Understandably, a lot of executives don’t want to touch this issue – let alone take a risk as an early adopter. Artificial intelligence as a technology has a similar political profile to nuclear energy. Concerns include brain damage, social instability, environmental degradation, economic bubbles, and literal apocalypses. The political dimension of AI is one of the key factors that we analyze, because it presents a barrier to understanding this technology in conventional technical or commercial terms.
Our study group is keeping close tabs on how people use AI tools. While there’s a lot of hype out there, and you should stay wary, there are some areas of work where this technology is making a difference. We believe that businesses can already start to mature their AI capabilities and begin to navigate upcoming difficulties, including both technical trade-offs and political conflicts. Here are some key ideas we discussed last month.
Time to Mediocrity (TTM)
Boom Supersonic expects its engineers to use AI. With a team of just 50 people, they designed, built, launched, and piloted a supersonic jet. Even on a safety-critical project, AI tools served an important function: helping team members collaborate.
What started as a joke in our study group has become a mainstay term for an early AI use case. While these tools aren’t yet operational they are certainly educational. You shouldn’t have illusions that an LLM will tutor you into a 10x engineer, but it could help you become sufficiently literate – mediocre – in such an engineer’s field, so that you could share ideas and specs with them.
AI tools are rapidly shortening TTM. Used correctly, AI chatbots are universal tutors. On cross-disciplinary engineering teams where collaboration is a bottleneck, this is a promising use case. It’s also indicative of another pattern we’ve noticed: AI is an in-stream technology. Today, most early-adopter businesses give AI tools to groups of employees, who rely on AI tools for impromptu queries (AKA Just-in Time Content), rather than replacing entire business functions. TTM is an example of how LLMs can act as a kind of solvent for improving efficiency, even without a real adoption plan.
Slopsunami
But you should have a plan. LLMs are prodigious content generators. As AI tools saturate existing firms, the cost of documentation drops quickly. Each meeting might get transcribed by multiple bots, then those transcripts turned into reports, then those reports are synthesized, then summarized… This predictable consequence of diffusion – when content becomes cheaper to produce than to search – will require a response.
It’s not an avoidable event, either, but an essential part of the technological adoption curve. A business that’s not AI-native will have to promote adoption and slop is a natural externality of the learning process. More use, more slop. If you put guardrails up too early, you’ll miss out on the gains later. Appropriate sequencing of process improvement is important.
People also have very different reactions to AI. Slop isn’t just a technical communication problem about signal and noise. For creatives, AI poses a fundamental threat to craft and livelihood. As early as 2023, Hollywood writers unions successfully lobbied for rules about AI-generated content. Industries like nuclear energy comprise a series of useful historical analogies that we continue to discuss. Our next study group call on September 8, led by @rafathebuilder, will be about other analogous technologies and what we can learn from their histories of diffusion.
Disruption events are another component of our study not typically addressed in capability models. We can’t predict them all, but getting a handle on the obvious ones is worth it – surprise is expensive.
The Task Estimation Crisis
Lots of organizations have lumpy knowledge of AI. That’s a result of both awareness of the available tools, as well as the fit of those tools with existing business functions. As a result, some teams are accelerating in their productivity and creativity while others teams stay the pace. Integration is therefore also extremely lumpy.
This lumpiness puts all decision-making at sharp risk. As Seth Killian put it: “Task estimates, and the production plans that sit atop that activity, are wildly off.”
There are some tasks that AI tools have trivialized or rendered obsolete. Companies that haven’t deployed AI into such tasks might be paying 10, 100, or 1000x too much for some products and services. Things like transcription, translation, web design mockups, background research, and document drafting are all promising, early applications, given sufficient AI tool literacy.
Awareness of where AI tools are dissolving bottlenecks is essential to getting the most out of your team’s capacity. Competitors who have a better sense of the lumpy distribution of AI knowledge and integration across their business will be able to repair their task estimation and production planning processes – this time with a higher total throughput.
Insights from Friends
The SoP community has been lucky to have several great guests join the Protocol Town Hall to talk about AI this year. Their ideas continue to be essential to the litepaper that SIGP4B is working on. I highly recommend watching a couple (or all of them, if you have a completionist streak):
Alignment Protocols
Summary: In this talk,
proposes a new way to think about AI alignment, moving from a focus on controlling a single, powerful AI to building a system of many smaller, locally aligned AIs. He introduces a three-layer “alignment protocol stack” and argues that building alignment must happen in a specific order: shared success, specialization, and finally, enforcement.Key Ideas:
AI alignment should be seen as a capability or a process, not a static property.
The "alignment protocol stack" is a three-layer framework for building aligned systems, starting with shared success and moving to enforcement.
A multi-agent approach is more robust than a single, monolithic AI, as it allows for localized alignment and collective problem-solving.
Public Intelligence
Summary:
, senior maverick at Wired magazine, discusses the concept of “public intelligence,” which he envisions as an open, decentralized, and globally accessible AI commons. He outlines attributes of this public intelligence, including public access, accountability, and data, and contrasts it with the concept of a “sovereign AI” controlled at a national level.Key Ideas:
Public intelligence would be a global “AI commons,” not owned by a single corporation or nation.
It would require public participation and be built on a distributed and decentralized network of AIs.
Public AI should be an option that exists alongside private, proprietary AIs, much like public libraries and bookstores coexist.
AI as Normal Technology
Summary:
, a computer science professor at Princeton, challenges the idea of AI as a sudden, revolutionary force. He argues that AI should be viewed as a “normal technology” and that its societal integration will happen slowly over decades. He discusses the barriers to widespread AI adoption, highlighting the time it takes for new products and protocols to be developed and for organizations to adapt their workflows.Key Ideas:
AI diffusion will be a slow process, not an overnight event, as it is limited by human factors and the need for new business models and legal norms.
For AI to realize its full potential, new protocols and workflows must be invented rather than attempting to integrate AI into existing systems.
AI as the Anti-product
Summary: In this talk,
, co-founder of Anaconda, explores how AI is a novel kind of technology, one that resists being packaged and sold as a traditional “product.” He draws a distinction between complicated and complex systems, likens open-source software to a living ecosystem, and discusses the importance of avoiding the centralization of power in protocols.Key Ideas:
AI represents a new, “religious” approach to coding that is different from the traditional craftsman or managerial approaches.
Software ecosystems should be viewed as living, human ecologies rather than rigid, manufactured products.
Organizations will need to undergo a “radical reinvention” to fully leverage AI, moving away from a purely managerial, KPI-driven approach.
Ways to Engage
Have thoughts about protocols, work, and artificial intelligence? Our next SIGP4B call is on September 22. Register here.
We have some great consultants, from entry-level analysts to FAANG veterans, who hang out in the #🚜-protocols-for-business channel on Discord. Feel free to share a biz problem there and pick our brains a bit.
Get early access to our forthcoming litepaper and help inform our research by answering this survey.
We’re pretty busy this month, but you can request a workshop about how to apply protocols in your business or organization by mailing research@summerofprotocols – more information on our limited pro bono workshops here.
Read the previous SIGP4B update, Brackish Strategy, here.