About six weeks in the past, I despatched an e-mail to Satya Nadella complaining concerning the monolithic winner-takes-all structure that Silicon Valley appears to check for AI, contrasting it with “the structure of participation” that had pushed earlier expertise revolutions, most notably the web and open supply software program. I suspected that Satya is perhaps sympathetic due to previous conversations we’d had when his guide Hit Refresh was revealed in 2017.
I made the case that we want an structure for the AI trade that permits cooperating AIs, that isn’t a winner-takes-all market, and that doesn’t make current corporations in each trade merely the colonial domains of extractive AI conquerors, which appears to be the Silicon Valley imaginative and prescient.
Little did I do know that Microsoft already had one thing within the works that may be a demonstration of what I hope for. It’s referred to as NLWeb (Pure Language Net), and it’s being introduced in the present day. Satya supplied O’Reilly the possibility to be a part of the rollout, and we jumped at it.
Embracing the Early Stage of Innovation
My concepts are rooted in a notion about how expertise markets evolve. We’ve got lived by way of three eras in computing. Every started with distributed innovation, went by way of a interval of fierce competitors, and ended with monopolistic gatekeepers. Within the first age (mainframes), it was IBM, within the second (PCs) Microsoft, and within the third (web and cell) the oligopoly of Google, Amazon, Meta, and Apple.
The error that everybody makes is a rush to crown the brand new monopolist initially of what’s primarily a wide-open area at the start of a brand new disruptive market. They usually envision that monopoly largely as a alternative for what went earlier than, slightly than realizing that the paradigm has modified. When the private laptop challenged IBM’s hardware-based monopoly, corporations raced to change into the dominant private laptop {hardware} firm. Microsoft received as a result of it realized that software program, not {hardware}, was the brand new supply of aggressive benefit.
The story repeated itself at the start of the web period. Marc Andreessen’s Netscape sought to interchange Microsoft as a dominant software program platform, apart from the web slightly than the PC. AOL realized that content material and group, not software program, was going to be a supply of aggressive benefit on the web, however they made the identical mistake of assuming the tip recreation of consolidated monopoly slightly than embracing the early stage of distributed innovation.
So right here we’re at the start of the fourth age, the age of AI, and as soon as once more, everyone seems to be dashing to crown the brand new king. A lot of the chatter is whether or not OpenAI or considered one of its rivals would be the subsequent Google, when it seems to be to me that they’re extra seemingly the following Netscape or the following AOL. DeepSeek has thrown a bomb into the coronation parade, however we haven’t but totally realized the depth of the reset, or conceptualized what comes subsequent. That’s sometimes found out by way of a interval of distributed innovation.
We Want an Structure of Participation for AI
The time period “the structure of participation” initially got here to me as an evidence of why Unix had succeeded as a collaborative challenge regardless of its proprietary license whereas different tasks failed regardless of having open supply licenses. Unix was designed as a small working system kernel supporting layers of utilities and functions that would come from anybody, so long as they adopted the identical guidelines. Complicated behaviors might be assembled by passing data between small packages utilizing normal information codecs. It was a protocol-centric view of how complicated software program methods ought to be constructed, and the way they might evolve collaboratively. Linux, in fact, started as a re-implementation of Unix, and it was the structure of participation that it inherited, as a lot because the license and the group, that was the muse of its success. The web was additionally developed as a distributed, protocol-based system.
That idea ran by way of my internet advocacy within the early ’90s, open supply advocacy within the late ’90s, and Net 2.0 within the aughts. Participatory markets are modern markets; prematurely consolidated markets, not a lot. The limitations to entry within the early PC market had been very low, entrepreneurship excessive. Ditto for the Net, ditto for open supply software program and for Net 2.0. For late Silicon Valley, fixated on untimely monopolization by way of “blitzscaling” (suppose Uber, Lyft, and WeWork as examples, and now OpenAI and Anthropic), not a lot. It’s change into a type of central planning. A small cadre of deep-pocketed traders choose the winners early on and attempt to drown out competitors with huge quantities of capital slightly than permitting the experimentation and competitors that enables for the invention of true product-market match.
And I don’t suppose we’ve got that product-market match for AI but. Product-market match isn’t simply getting numerous customers. It’s additionally discovering enterprise fashions that pay the prices of these providers, and that create worth for greater than the centralized platform. The issue with untimely consolidation is that it narrows the main focus to the enterprise mannequin of the platform, usually on the expense of its ecosystem of builders.
As Invoice Gates famously advised Chamath Palihapitiya when he was operating the nascent (and in the end failed) Fb developer platform, “This isn’t a platform. A platform is when the financial worth of everyone that makes use of it exceeds the worth of the corporate that creates it. Then it’s a platform.” To be clear, that’s not simply worth to finish customers. It’s worth to builders and entrepreneurs. And which means the chance to revenue from their improvements, to not have that worth instantly harvested by a dominant gatekeeper.
Now in fact, Sam Altman talks about creating worth for builders. In a latest look at Sequoia Capital’s AI Ascent occasion, he stated his hope is to create “like simply an unbelievable quantity of wealth creation on the earth and different folks to construct on that.” However he makes use of the language of “an working system” that others construct on prime of (and pay OpenAI for using) slightly than a shared infrastructure co-created by an ecosystem of builders.
That’s why I’ve been rooting for one thing completely different. A world the place specialised content material suppliers can construct AI interfaces to their very own content material slightly than having it sucked up by AI mannequin builders who supply up providers based mostly on it to their very own customers. A world the place utility builders can supply new sorts of providers that allow others in a cooperative cascade.
We’re Simply Getting Began
Anthropic’s Mannequin Context Protocol, an open normal for connecting AI brokers and assistants to information sources, is step one towards a protocol-centric imaginative and prescient of cooperating AIs. It has generated numerous well-deserved enthusiasm. Google’s A2A takes that additional with a imaginative and prescient of how AI brokers would possibly cooperate. NLWeb provides to that a simple method for web content material websites to affix the get together, providing each a conversational entrance finish to their content material and an MCP server in order that it’s accessible to brokers.
That is all going to take years to get proper. However as a result of it’s a protocol-centric slightly than a platform-centric imaginative and prescient, options can come from all over the place, not simply from a dominant monopolist.
Each new wave of computing has additionally had a brand new consumer interface paradigm. Within the mainframe period, it was the teletype terminal; for the PC, the Graphical Person Interface; for the web, the net’s document-centric interface; for cell, contact screens. For AI (for now at the least), it seems to be conversational interfaces.
Corporations reminiscent of Salesforce and Bret Taylor’s Sierra are betting on conversational brokers which might be entrance ends to corporations, their providers, and their enterprise processes, in the identical method that their web site or cell app is in the present day. Others are betting on client-side brokers that can entry distant websites, however usually by calling APIs and even performing the equal of display screen scraping. MCP, A2A, and different agent protocols level to a richer interplay layer made up of cooperating AIs, in a position to hook up with any website providing AI providers, not simply by way of API calls to a dominant AI platform.
All corporations want at the least a begin on an AI frontend in the present day. There’s a superb line from C. S. Lewis’s novel Until We Have Faces: “We can’t see the gods head to head till we’ve got faces.” Proper now, some corporations are capable of supply an AI face to their customers, however most don’t. NLWeb is an opportunity for each firm to have an AI interface (or just “face”) for not simply their human customers however any bot that chooses to go to.
NLWeb is totally appropriate with MCP and provides current web sites a easy mechanism so as to add AI search and different providers to an current internet frontend. We put collectively our demo AI search frontend for O’Reilly in a number of days. We’ll be rolling it out to the general public quickly.