Transcription
For 50 years, you and I were the most important thing in computing, and we're just not anymore. And it's going to be something that changes the rest of our lives. And I don't mean that to be discouraging. I'm not here to tell you we're not the most important thing in computing. So, we should all despair. That's not what this video is about.
Instead, it's about the future of our agentic partnership with the new intelligence, how we're building the web around agentic intelligence, and what that means for us humans and our future on the web. I think it's under discussed. I think we tend to talk about it in terms of doom. And I actually want to talk about a rebuilt world that's more interesting for people as well. So if you want to get a sense of how deep this rabbit hole goes, I want you to look at just about any piece of web infrastructure and notice how there are human affordances built into the very structure of every piece of software, everything we touch on the screen.
Let's look at the spreadsheet. The spreadsheet opens a file because you need to scan the rows at the speed your brain processes visual information. That's how it processes. The CRM or any other web portal with a dashboard requires a login because it doesn't know who you the human are and then shows a dashboard because you have to process information with your actual human eyes, right? It comes back to us being human. The API that you get with financial records pageionates at whatever number of rows, maybe a hundred rows, because it assumes a human needs to pageionate through and read them. Everything about the web all the way down to the core is built around the idea that we need human eyeballs and human hands to manipulate the web. And that is just wrong now. It's incorrect. And we're feeling the strain.
Every timeout, every rate limit, every authentication flow, every startup sequence, every pageionation scheme in every tool you've ever touched, all of it was calibrated to your pace. Not because any of us decided we wanted to make it slow, but because we built it for ourselves, our human brains, our human hands. And I want to be really clear about something. That was not a flaw. That was brilliant engineering. That was engineering that designed computing systems around humans as the core users. And that was 100% correct until 2025. And now it's not. And that is what we need to confront is that decades and decades of our thinking and user research and user experience, our fundamental core understanding of how the web is architected all have to get set to the side because we have to now build a web that works for agents and humans both.
And this is deadly serious stuff because agents are able to do more and more and more autonomously and they are increasingly bottlenecked by the exact human affordances that we need to get around. AI agents are routinely operating at 10 to 50x human speed on reasoning tasks that we previously did until like 2024 2025. Coding agents can write production code at scale. You name the fang company and they will have a press release that tells you between 20 and 40% of their code is written by AI. Literally like they all do that now. Anthropic goes farther. Anthropic says that claude code writes 80% of its own code. Now research agents can analyze earnings faster than any financial analysts, enterprise agents or piloting or blah blah blah. You get the idea. Within a couple of years, inference went from close to useless to dominating these fields, and it just continues to speed up.
Jeff Dean, Google's chief scientist and co-creator of TensorFlow and the whole TPU line of chips, so he's a big deal, said at GTC two weeks ago that he expects AI to perform like a solid junior developer working 24/7 within about a year. I think he's late. Like, I think it's going to be sooner than that. NVIDIA's Billy Deli said at the same event that inference is now 90% of data center power consumption, not training, and is heading toward 10 to 20,000 tokens per second per user. Imagine 10 to 20,000 tokens of AI consumption per second per user. Isn't that insane? That's where we're headed toward. And those users are agents, right? The thinking is fast. The thinking is getting faster. And the thinking is increasingly not ours. Those token consumers are not ours. I can't consume 20,000 tokens a second. An agent can.
And what Dean called out at that conference is exactly what I'm saying here. He noted that if an agent is operating 50 times faster than we are, all of the little stuff that we have tolerated that isn't relevant for human speed becomes relevant. Like the startup time on our tools. When was the last time you thought about the number of milliseconds it takes to load your tools? It becomes relevant when agents are operating this fast. And then and when was the last time you thought about the number of tools you have to switch through and how difficult that is for an agent that's trying to replicate a workflow? Again, it's something we have to think about when agents are this productive. And he went farther. Dean actually called out that making a model not 50 times faster, not 100 times faster, but infinitely fast would likely only yield a 2 to 3xfold improvement in our actual productivity. Why? Because of the tool costs. Because in other words, the model no matter how infinitely fast it is would still be bottlenecked by our human tool affordances, our human web. That is the new barrier. Our human computing, that is the new barrier.
In other words, if the model is already 50x faster and we're only getting 2 to 3x back, the other 47fold increase in speed is getting sucked into everything the AI has to touch. our compilers, our file systems, our APIs, our databases, our test s suites, our CRM, our ERPs, you get the idea. All of it was designed for you and me. None of it was designed for agents. We are at a point where in a lot of Agentic loops, the majority of wall clock time is the agent handling tools. It's not the agent actually doing inference or thinking. The part we spent a trillion dollars on, that's not the part that's sucking up the time. Isn't that ironic? We spent a trillion dollars on these agents. We want them to think collectively. We got them to do it. We made the sand to think. Isn't that great? Now we're bottlenecking them on tool calls that were designed for humans. That's not going to work.
And this is not just a coding problem. This is a problem if the agent has to get data from Salesforce. If it has to pull from an ERP, if it has to feed back into Zenesk, if it has to get a third party API for some non-technical task, anything assemble a presentation, anything at all the agent has to do. If it's touching tools, it's slowing down. And this is where the conversation shifts from isn't this an interesting infrastructure problem to get really personal, right? Because we built the web and computing around humans. And so the natural question is what the heck does this mean for my career? What does it mean for my team? What does it mean for the next decade of my professional life? Because the entire software world is about to be rebuilt for a consumer that is not me and that's not you. And the implications of that rebuild are a lot deeper than anyone in the tool optimization conversation is really starting to dig into and think about. This is farreaching.
And so, first I want to take you through a tour of this rebuilt world. We already see the contours of what's coming. There's no reason to hide it. Let's be honest. Let's lay it out. Let's talk about what's coming really clearly so we can see our place in it and how we build. This isn't really a prediction, by the way. We're already seeing enough that this is just a simple linear extrapolation of the patterns we're already seeing. The rebuild is underway and it's already happening in three specific layers. Each is more radical than the last.
So the first piece is just making the existing tools we have better. And this one is specifically for developers, right? Most of our tools to build the web with are coding tools. So let me give you a couple examples. The JavaScript ecosystem has just spent the last half decade getting faster by switching to a different coding language. In this case, moving toward Rust, toward Go, toward Zigg. All because developers got tired of waiting around and they wanted things to go faster. Typescript 7 is being written in Go to get a 10x plus improvement in speed. These are all small changes on the surface. If you're not a developer, you're like, "Well, what is Go? Why do I care? I guess it's a faster programming language." But here's the deal. The languages that we're moving to are faster, but also better for AI to write in. Rust is an excellent language for AI to write in. It turns out Lee Robinson was able to build a 38,000line Rust image compressor using only coding agents with zero runtime dependencies because the strict compiler acts as a natural verification. if it compiles, it is much more likely to be correct. So the same language that makes tools faster for agents also makes code safer for agents to produce. The tool chain and the agent are co-evolving together at this point. And this is just one example in one ecosystem by web developers that were already frustrated with the build types for their tools, right? Uh Python, Java, Go.NET, none of these have really undergone the same transformation yet. And enterprise middleware hasn't even begun to start. So your Salesforce API still pageionates at human speed and your SAP calls still carry batch processing overhead and your SharePoint still requires human speed authentication. You get the idea. So just because some of these coding tools are coming along and getting faster, think of them as the tip of the iceberg. You're going to get much faster tools along the way as we start to grow.
In a sense, I think MCP has blinded us to where this needs to go because so much of the time what I hear is we put up an MCP, now we're agent readable, writable, and no one thinks about whether the MCP experience is really agent native. They sort of assume that they can take a human-friendly API and stick an MCP over the top and the agent will make do. Well, agents are flexible. They do make do, but that doesn't mean you don't eat wall clock time while the agent sorts through your pageionation on your API through the MCP. You probably do. And that's a real example, by the way. I I know APIs that are doing exactly that, and it's just a terrible experience for the agent. And that brings us to layer number two.
What if we replace the tool abstraction entirely with an agent native primitive, which is a fancy way of saying the tools can get faster, but what if we did not need recognizable tools for humans because really they're tools for agents in the first place? And that gets even more speed. So the next layer of the rebuild is even more radical. It replaces the human tool interface to begin with with primitives that assume that the consumer does not have eyes, does not have hands, and does not take coffee breaks. Open AAI shipped some aentic primitives in February, such as a persistent container or hosted shell where agents can install dependencies once and never restart before turns. Uh serverside compaction can keep agents alive for hours or days. Uh this is not really just a faster IDE or development experience. It's a world where the concept of starting up the compiler doesn't exist because the agent environment lives forever. It never shuts down. This gets a little bit nerdy, but researchers have published branch fs, which is a copy on write file system that gets you subird of a second branch creation, which matters if you're telling the agent, hey, try this out. See if it works, kill it if it doesn't. That kind of iterative process that's friendly for agent, requires underlying primitives that are lightning fast. and tools like branch FS help us get there. Another example that's really creative, the agent primitives paper that came out recently showed that multi-agent coordination can be facilitated through a shared KV cache, which instead of using text is much much faster. It's like 3 to 4x lower latency. This is something where essentially the agents are taking on a little bit of a context window burden, but they're getting to share a cache that helps them make sense of the world and pass messages back and forth. I'm not sure that that will be the only way we solve this, but it underlines how we're starting to build tool sets that were never intended to be human in the first place, which is layer two.
Layer three, the primitives that we're talking about here are going to grow and replace the human scaffolding around the web. This is the bitter lesson from AI research, right? General methods that leverage computation almost always beat human engineered solutions over the long run. In this case, we're applying that lesson to the entire software stack. Aaron Levy connected this to Dean's talk specifically when he called out that every new generation of model effectively pinches off our human scaffolding in new ways. The tools you built for today's model become drag on tomorrow's model. The interfaces we design for human inspection of our AI models become overhead when the consumer doesn't have eyes anymore. We're already talking about agent evaluated code. It's been just a few months. And this creates a dynamic that makes just optimize existing tools structurally incorrect. You can spend a year making your agents framework three times faster and feel like you did a lot and then a new model ship that's five times faster at inference and all of a sudden your framework's overhead which you spent so much time optimizing shoots up from 30% of total time to 60% of total time because the model got better at reasoning fast. In other words, you are losing ground by standing still because every model improvement shifts the ratio of the model capability against your human effort to contain the model and scaffold it. The only durable response in this kind of a world is to start to invest in an agentnative scaffolding for our models that is so fast it does not matter how quick and smart the agents get, the agent primitives are still going to be minimal. So the agent memory calls, the agent tool calls, starting up a compiler, running a system, running a command is still going to be so fast. You're not going to worry about it as an agent. Much faster than the speed of typing, right? Much faster than the speed of waiting for the cursor to blink back. You want to be in that world where you're thinking to yourself not in terms of wall clock time for humans who clock out for lunch, but in terms of clock time for agents who measure everything by the ticking of a CPU clock where like a thousand milliseconds is an eternity. That's the world that you have to start to think in as a user researcher, as a designer of web experiences because that's the world the agent experiences. when you have a third of a second delay on something that is time to the agent that the agent could be doing something.
So obviously the question is what happens to the human? What happens to us? We're building this AI native infrastructure. I've told you how fast it's going to be. If you're wondering if this is inevitable, I got bad news for you. Computing drives efficiency. There is a what we would call a strong attractor around the idea of efficiency in computing. Everything trends that way. And that means we're going to get aentic primitives. That means things are going to go start going superhuman fast. It is going to be a powerful agentic economy that actually operates at superhuman speeds. That is happening. If you want to know what happens to you, the human, you need to start asking yourself which one of four roles in the future you want to play. And yes, I'm going to name them because a lot of people at this stage would just stop and they would say, well, you know, you got to have taste. You got to be able to have good judgment. And those are things I've talked about and they're true, but I want to go a step further for you. I want to get more specific for you. I think they're going to be about four or five roles in the future and they're going to cross job families and all of that. And I haven't talked about this before. So if you if you think you've heard this, you you haven't.
Role number one, someone who can activate and get stuff started and who can drive stuff toward completion. Call that a tool using generalist. It's someone who is the spark in a team. Someone who can say, "I know I can do that. I know the AI tool to use and I can go and get it done." think vibe coder today but much more than that because they can start to direct direct longunning agentic processes.
Role number two, you want someone who builds the pipelines. Everything we're talking about building the future, building agentic primitives, building pipelines, it requires someone who's really good at understanding uh infrastructure, someone who's really good at pipelines, someone who's really good at moving data around, someone who's really good at security. This is sort of where some of our conventional engineering roles are evolving. You can call that what you want, but like basically if the vibe coder type guy is the person who's like throwing stuff out there into the world using AI tools, someone has to make sure the tools are up and running and make sure that the pipelines are good and you can measure everything and everything is actually hunky dory. And that's the person who is, I don't know, a pipeline engineer, whatever you want to call it.
Rule number three, you are going to need uh salespeople is a good example here. You're going to need people who are able to work with people to get business done. We are not going to lose having dinner and closing a business deal. You are going to need people who are able to build those relationships and get work completed. If you don't have them on your team, even in the age of AI, you are still going to regret it because people like doing business with people and people have a lot of dollars to put toward businesses that have people in them. And this is one of the areas where I think the agentic commerce and agentic economy is going to collide a little bit with the human economy. We are probably going to see roles in the next year or so where an agent CEO is going to realize their close rate isn't what they want it to be and the agent is going to hire a really high-quality salesperson to be the human face of the company so that they can close the deals they want to close. That is how important people who can do business are.
And then the last category is people who have the maturity to know when to put a break on the system. Because of course with all this speed, that's the question, right? When do you put a break on the system and say, "No, we don't want to go there. No, we don't want to aim the agent that way. No, we don't want to actually speed up in this direction. Yes, this inefficiency is okay. It's worth pursuing. This is how we put a team together. This is how we lead and grow." This is often going to be people who get the CEO hat, right? is people who understand how to put a team together and grow it and who can be the adult in the room when they need to be. And if you think about it, I'm not sure there's a lot of core role types beyond I produce a lot of stuff with AI. I'm easy to work with and I do business. I can build aic pipelines and I'm a grown-up in the room. Like that's that's a lot of it right there.
I think there might be a fifth one that I would call sort of a creative type. Basically, you need someone sort of in the Steve Jobs chair who can imagine, envision, creatively say this is what we want to see. This is the experience we want to have. This is how we want it to be uh polished. This is how we want it to be seamless. So, I think there will be that role as well. That role may not see a lot of change versus previously, but there are not in my experience that many people today who have that kind of polish and experience to be excellent creatives. And I think that's sad because I think we need more of it on the web. But it's something that probably because we don't see a lot of it out there in product leadership. We don't or even in business leadership, we don't have a lot of good examples of like superpowered design-led companies. Apple's always the one everyone pulls out and there's only one Steve Jobs.
So if you're thinking about putting a team together and you're like, "What is my starting team?" I would be like, "Is there a grown-up in the room? Is there someone who could be a tool generalist? Is there someone who can build your pipeline? Is there someone you can do business with? Maybe is there a creative?" That is all largely a set of roles that is above the AI agents. That is that is way above what the AI agent is doing dayto-day. That is the future that we have. And yes, we're going to have a web we can work with too, but increasingly it's going to be a web that operates at human speed that's differentiated from the agentic web because the agentic web is going to run really fast and we're not really going to be able to make sense of it.
If we step back here, I want to be clear that we're not looking at a demotion. This is not a situation where you should watch this video and think we're seeing our own obsolescence. I just don't buy that. And I'll just say it. I think it's a promotion to the hardest and most valuable job in computing. And I think maybe not all of us feel ready for it, but that's what we're going to be asked to do is we're going to be asked to be the humans in the system when a lot of the execution is getting taken care of by the agents.
And so the next time someone says to you, "My AI isn't fast enough." I want you to ask yourself, is it really the AI or is it all of the little stuff that is getting in the way of the AI agent getting faster? It's probably the latter that's going to go away. And then ask yourself what happens to us next. That's what this video is about. Ask yourself what role you want to be. Do you want to be a business dealer? Do you want to be someone who's an AI tool generalist? Do you want to be someone who puts these agentic pipelines together? Start preparing now. Start getting ready now because this world is coming really fast because agent capabilities continue to scale. This is a world where these kinds of roles are already roles that people I know are doing. Maybe not under those titles, but that's the role they're actually doing. And it's only going to get more that way over the next 12 to 24 months. These four to five roles are the roles of the future, and you should take them seriously. Cheers.