Fred Wilson, who has been at this even longer than I have, asks an interesting question:
Is there an application in artificial intelligence that matters more than LLMs? And if so, what is it?
and John Battelle, who’s always been watching Google more closely than most people I know, gets closer to the solution (the Agentic Web) but points out the primary problem with that approach, eventually landing on the belief that Fred didn’t go far enough in rolling back the tape:
Could it be that consumer AI is comparable not to the early Web, but rather to the era of nearly forgotten online services that preceded the early Web in the late 1980s? When the history of consumer AI is written, might it treat OpenAI, Claude and Gemini as the equivalent of pre-Web services like CompuServe, AOL, and MSN?
Like anyone in tech, I’ve been thinking for a while about what this means. And like John, I keep hitting on the initial challenge: access to information. But unlike Fred and John, I believe the next big revolution will not be dominated by one company but by one protocol.
An interface problem
The Generative AI age is over.
I know it may sound crazy but hear me out. What problem does generative AI solve? When you look at it closely, you could say “any problem, really?”
So is generative AI solving every problem or is generative AI a way to solve any problem?
To me, the revolution OpenAI brought to the world has run its course: it’s one where the idea of asking a computer, in plain language, for support on something brings back an answer. The implementations of Gemini, Anthropic, and others may be interesting but they all come down to one thing: who has the smartest computer to answer what you entered in the box.
It may sound like I’m turning a very complex set of solutions into something very basic but ultimately, the genius of generative AI is that it has made what is extremely complicated server back-end magic into a tool anyone can use. And that’s evolution not revolution.
OpenAI essentially said “we can have a better Google entry box and, instead of giving you 10 blue links (with advertising in between), we’re going to give you a plain language answer.”
Google could have done it and, when they decided to put their mind to it, did, with Gemini. And OpenAI make take the lead again, or another company but innovation at that layer will remain incremental because all this is still at the interface layer.
It’s still “I have a question. What is the answer?” as a mode of interaction with the computer.
Adding a memory layer to the box does improve how well it customizes the experience but this is not something where a new entrant has an advantage as companies like Google, Microsoft, and Amazon have years, if not decades, of data on you they can use to improve the experience. The fight will continue on precision or on what the box can do and will require increasing amounts of investments for decreasing returns on investment.
Interface is not where a new entrant can win. Control of the interface is the old paradigm (the one that allowed Microsoft to defeat Netscape and probably will allow Google to defeat OpenAI) but it is not where a new player will win at this time because the cost of entry is so high.
Pre-internet Days
However, thinking through John’s lens of internet services vs. the web, I can’t help but think of BBSes, the Minitel, proprietary online services, and the internet.
Long before the web existed, many different silos had been formed in the online world. Each existed as its own village, with its own rules and regulations. Sure, people who were on the Minitel (a French proto-internet), QuantumLink (the service that later became know as AOL), TheWell (an early internet board), and a traditional BBS (an online island you dialed into with your phone) were all technically online. But the individual experiences of those environments was radically different from service to service.
Some places were the domain of nerdy hackers like myself; Others were places where corporations felt safer (eg. Prodigy, an online service created by the retailer Sears and IBM); yet more were the domain of intellectuals (TheWell and ECHO in New York come to mind); some were run by kids looking to share illegal copies of videogames; others were run by hobbyists looking to chat about obscure topics.
Each was a fiefdom where the almighty system operator (SysOp) ruled and could boot you off. Each was a small walled garden. Each was a village
With first the rise of Fidonet, as a BBS to BBS board, and then of Usenet, those small villages started interconnecting.
But when the internet came, many resisted connecting as they saw themselves as destinations instead of doorways.
What destroyed many of them what standardization: The internet, first through a standard set of text interface and later through the web, made it easier to move from one place to another and individuals soon found the whole internet to be a much more interesting world to explore than the small village they were in.
The financial value of HTTP and HTML has never been calculated but I, for one, believe that it is greater than that of any tech company today.
Big City
My mental model of the world may be affected by the fact that I spent most of my adult life in big urban centers. That said, whether it is London, Paris, Hong Kong, or New York, one thing has been consistent: None of them is a monolith.
Big city are a construct, much like the web is, made of much smaller villages. Just as neighborhoods have their own local flavors, each internet walled garden is a different institution than the next.
Witness, for example, the difference in communities like Threads, Bluesky, Mastodon, and whatever Twitter has become. Same interface, different rules and behaviors. In the same way, Amazon and Wal-Mart exist on the same continuum. And this site exists on the same continuum as a much larger media effort in that we’re both pushing message. Yet, they are all different and all worth different levels of trust, just as one would think differently of different neighborhoods in a city, based on their personal background and interaction with said neighborhood.
What they do have in common, however, is infrastructure. Roads, sidewalks, the air, are mostly regulated in the same fashion in a city. On the internet, protocols dictate how data is exchanged because a rough agreement as to using those protocols exists between all players.
Protocols > Platforms
All this to bring me, in a roundabout way to protocols. What destroyed pre-internet online entities was not the internet itself but their resistance to the internet.
In the real world, small villages ended up being assimilated into cities. This sometimes creates tension but there is a rough understanding that when you are surrounded, you either assimilate or die. Tension may still exist (the dreaded specter of gentrification and its impact on the culture of local areas being the biggest source) but ultimately, the physical proximity destroys any chance at not being assimilated.
Similarly, online communities which did not connect to the Internet eventually disappeared. Those that did survived but no longer held the central control they had over their users. Whether you are the most liberal or most conservative site on the internet, your application uses the same protocols, from the communication layers of TCP/IP and HTTP all the way up to the HTML and CSS that delivered this to your eyes.
The web did not succeed because tons of money was put into it. The web succeeded because there were free protocols and standards that made it possible for anyone to jump in. This reduced a financial friction which made it possible for even the most obscure non-financial viable content to be distributed on the internet. As long as you could pay for basic infrastructure access, you were good (that said, scale became a problem for some as infrastructure still needs to be paid for).
The 1990s presented a wild era when experimentation eventually created some of the giants we know today. However, with the explosion of content, discovery became a problem. Search engines and search boxes were the solution. Google rose through that. OpenAi has now made it better (even though it may not be the winner at it). But it leaves the problem of discovery a largely solved on when you have a rough idea of what you want to look for.
But what about when you believe this information could be sitting behind a gated space? Your problem then become one of integration of all content sources. And, as John rightly points out, there is little incentive to share when you’re a walled garden except…
… everyone wants access to everyone else’s data.
MCP: The new HTTP
So now we get to the geeky part of the conversation. MCP, the Model Context Protocol, was introduced about a year ago by Anthropic, the makers of Claude (Claude has been a competitor to ChatGPT but with more of a business suit on). Anthropic was much small and much less well capitalized than any of the other players so it had to think differently.
With MCP, Anthropic introduced an open-source offering that allows any organization to connect their data to any AI systems, whether it’s Anthropic’s own or ChatGPT, or Gemini. From there on, their data can flow into an AI system or in/out of other systems connected to that AI system.
Most MCP servers are single points of contact but a new class of MCP servers now can take information from one MCP server and trigger an action to get or give information to another one. So if you’re combining two or more points of integration, the agentic web start taking form. And inexpensive integration remains one of the largest unsolved problems in technology.
“But no one is making money by sharing their data via MCP” you might say. And it’s true, today.
When HTTP came along, few expected it to change the world in the way it did (I remember writing the sentence “global internet commerce is expected to grow to $6 million next year” and thinking that was large growth.) Yet, somehow it did.
The Next Problem Area
Some problem are too complex to be solved in one sitting at a terminal. And agents are probably going to be best when dealing with situation where information itself is fluid (yes, dynamic pricing is fluid and optimized for maximum profit but there are few cases where manufacturer or service providers are not left with some inventory they haven’t monetized yet).
The reason Agentic AI are where the next generation sit is that they present the next interface layer. Today’s AI is still based on a “request/response model” that limits the range of options when it comes to answers to now and based on historical and present information. Agentic, by comparison is a “set and iterate” model that can evolve based on changing conditions.
Much like the web didn’t come out of a company, MCP and MCP-related efforts (solving problems like server to server communication, smart routing, discovery, etc…) are increasingly creating a new layer for the web that will deliver on a future envisioned in the late 1990s. At the same, concepts like RDF and machine-to-machine seemed like distant dreams that would never be realized. While they failed to establish a standard for how to represent data coming from legacy systems, MCP seems to take another stab at the concept, allowing to add a layer of smarts to it.
My belief is that the next big area (the Google of the next generation) is in the discoverability and governance of MCP servers. In a world where machines are talking to other machines, two large problems will exist: How do you get your machine to talk to an ever-expanding universe of data (how does it find the info) so its action/reactions get sharper over time, and how do you ensure that the data they access is accurate and no polluted by bad actors. Just as I believe there is an increasing role for editors in a landscape where media is commodity, there will be a need for companies that vet, potentially even credential and incent, data providers to the agent, giving them a potential “vig” on successful data transaction.
In a way, the model I laid out in 2009 for the dimensions of media is applicable to data: Data can have entertainment or informative value; it is either purchased or subsidized, and it is either vetted or not (I know the original was mass generated or professionalized but it’s ultimately in the vetting that value is increased). New companies providing abilities to offer services around those axes will emerge as the next generation of winners.
The winner will thus be an entity who wins not on interface but on intelligence, providing a layer of capabilities few others have in a wrapper that could sit under any of your tools (as an app, as a browser add-on, as a browser) but not caring about the container carrying it. Then, at a later time, they may try to leverage that advantage into gaining access to the presentation layer, as Google did with its browser, which eventually gave it greater leverage over Microsoft.