Most of the AI tools I try do not last a month. I will find something promising, hit the ceiling in an afternoon, and move on. Someone on X called this the single-use plastic era of software. Someone else replied: "An engineer's dream, everything can be custom and the cost of it is zero."
From the user side, the disposability is exhausting. You find a tool, learn it, and then either the model provider absorbs the feature or a better version appears. Nothing sticks. From the builder side, the same dynamic is thrilling. Anyone with taste and a problem can assemble exactly the thing they need, for almost nothing, in a weekend.
So which is it? A wasteland of plastic wrappers, or a golden age of building?
Both. And the tension between them tells you where the real value is migrating.
What follows are the structural patterns I keep seeing underneath the churn. Not predictions about which company wins or which product ships next. Dynamics that hold regardless. I have been updating this thinking for over a year now, and new information keeps confirming the same handful of patterns. That is what made me want to write them down.
The landlord problem
The reason most AI software is disposable is not that the builders are bad. It is that they are downstream.
If you build on top of a foundation model's API, you are renting the core capability. And the landlord is moving into your apartment. Take Cursor, the AI coding environment, running on models from OpenAI and Anthropic. Great product. But then OpenAI ships Codex. Anthropic ships Claude Code. The model providers see behavior across millions of users, tune model and product together, and own the data flywheel. Cursor sees only its own slice and pays them for the privilege.
Or take OpenClaw. Peter Steinberger built an open-source, autonomous AI agent platform that went viral. Three weeks later he joined OpenAI to "drive the next generation of personal agents." The frontier absorbed the frontier-mapper.
Models leapfrog each other every quarter. The frontier rotates between labs so fast that building around a specific model's edge is like renting a room that changes landlords every season.
The cycle repeats at every layer. A capability emerges. Within months it is commoditized. The frontier moves again. Same pattern, shorter loop every time.
Here is the part people miss when they frame this purely as a threat: this powers the engine. The startups that get absorbed are not wasted. They prove concepts, map user needs, force the labs to build robust versions. Each turn generates new capabilities and the next generation of builders starts from a higher baseline. It is what makes things move so fast.
What makes this cycle particularly aggressive right now is the data pull. OpenAI runs a program offering API users substantial free tokens in exchange for opting into data sharing. Your prompts and completions flow back to improve their models. Clean trade on the surface. Structurally, it means the more you build on the API, the more you feed the system that may eventually replace you.
The things that buy time tend to have something the lab cannot replicate with a feature release. Distribution that took years to build. Messy domain data from regulated or hard-to-scrape environments where models cannot just learn their way in. The kind of relationship where you have become the way someone works, not just a tool they use. Or the trust infrastructure that sits between the model and the real world: evals, safety, compliance.
Everyone else is renting a shrinking gap.
So if you cannot own the building, you own what you bring inside it. That thing is context.
There is no lasting moat
There is a claim running through a lot of AI writing right now: execution gets cheap, judgment stays scarce. I believed it when I first heard it. Now I think it is half right, and the wrong half is the one people are building on.
Frontier models already make judgment calls that would have impressed a senior analyst a year ago. They weigh tradeoffs, catch edge cases, reason through ambiguity. "Judgment is what humans uniquely bring" is starting to sound a lot like "creativity is what humans uniquely bring." Which aged badly.
Follow the logic all the way down and the scarce thing is not judgment. It is context.
A model can reason brilliantly, but only over what you put in its context window. The quality of its output is bounded by the quality of what goes in: what is included, what is left out, what format it arrives in, what memory gets surfaced, what tools get connected.
Assembling that context is a skill that compounds with experience. Knowing which three documents matter out of three hundred. Knowing the client's tone shifted last Tuesday and that changes what "urgent" means. Knowing the CRM data is six months stale and the real numbers live in someone's spreadsheet.
The model can reason over whatever you hand it. It cannot go find what it does not know is missing.
I keep coming back to something I studied in my thesis work on interface design. The finding that surprised me most: when you give people preset options, the distribution of individual differences smooths out. The interface literally flattens who you are. Defaults homogenize behavior. People stop making choices and start accepting what is offered.
Now think about what happens when agents act on our behalf. Who designs the defaults? Who decides where to add friction and where to remove it? The interface disappears, but the design choices compound even harder. The context your agent assembles shapes what the model sees, which shapes what the model does, which shapes the outcomes of your life. And most people will not realize this is happening because there is no screen to notice it on.
This is why the shift from prompt engineering to context engineering matters more than it sounds. Prompt engineering was about phrasing your question well. Context engineering is about assembling the right world for the model to reason inside: what to surface, what to connect, what to leave out, what format makes the reasoning click. The people who get disproportionately good results from AI are context assemblers.
Context is where the leverage lives. And for now, assembling it well is still a deeply human skill. That will not last forever. When models learn to assemble their own context, and they will, the advantage moves again. But the people who build the adapting muscle now will be the ones best positioned when it does.
There is no lasting moat. Only the rate at which you learn the next one.
The race for attachment
If context is where value lives, then whoever gets to hold your context holds the real leverage.
Every serious AI system is trying to solve the same underlying problem: can I predict what you want next better than anyone else? Not just model providers. Recommenders tuning your feed. Productivity tools guessing your next task. Companions guessing how you will feel. On the surface it looks like personalization. Underneath, it is a quiet race to build the richest possible model of your inner life.
We have crossed from stateless chat to systems that remember. A year ago, every session started from zero. Now the model picks up where you left off. That memory will deepen: from "you like dark mode" to "you tend to procrastinate around chapter three" to "you are avoiding that email because the last one like it led to a fight."
And that memory does not even have to live inside the model or the cloud. It can sit in a plain text file that every tool in your stack reads from. One file shapes how your code editor, your writing assistant, and your search tool all behave. The context layer is already separating from the application layer.
Text is sparse. Voice captures tone, hesitation, emphasis. Vision captures context. The more channels the system sees, the more detailed its map of you becomes. Every correction, every "no, more like this," nudges the model closer. After a year, it knows you better than any settings menu ever could.
The shift from a race for attention to a race for attachment is real. Social media tried to maximize minutes. AI companions are trying to become the first thing you turn to when it matters.
Surveys already show teenagers forming romantic relationships with models. That should unsettle us, even when the technology is genuinely impressive. Especially then.
The lock-in is structural. Switching systems does not just mean learning a new UI. Even if your memory is technically portable, you still lose the tooling around it, what gets surfaced when, which workflows it is wired into, and the trust you built over time. Memory is both the magic and the trap.
And that raises a question I do not have a clean answer to. Who do you trust enough to let them remember you that deeply?
Screens retreat
If the system knows you that well, does it still need a screen?
We are deep in the max-screen era. Dozens of unlocks per day. Most of life mediated by notifications, feeds, dashboards. But there are signs of exhaustion. The overshoot comes from technology succeeding too well at the wrong objective: more screen. The next wave succeeds by doing the opposite. Less screen, same outcome. Or better. Voice turns the agent layer into something you can steer without opening anything.
A lot of apps are going to go away. Not because they are bad, but because they were always just interfaces on top of a function. The version gaining traction runs on your machine, controlling your desktop the way you would: files, apps, browser, keyboard. The screen is still there, you just stop being the one looking at it. Your computer is the universal API. It does not need a special integration for every tool.
I saw this play out in real time. A few months ago, friends asked me to build a WhatsApp bot, a best friend in chat, that would swipe on their dating apps and set up dates for them. I spent about two weeks on it: mapping APIs, exploring protocols like Matrix, talking to builders about matchmaking algorithms. Two weeks of research, nothing live. The platforms own the data, and without their APIs you are locked out. This week, I watched someone on X ship the exact thing. An agent driving his screen, swiping through profiles the way a human would. No API needed.
Today I open a weather app, a calendar app, a translation app, a calculator. Each one is a screen I look at for ten seconds and close. An agent layer collapses all of these into a single conversational surface, or handles them silently. The app does not die. Its function gets absorbed. What was a product becomes a capability.
The apps that survive are the ones you actually want to look at. Creative tools, games, social feeds, anything where visual engagement is the point. That is why the gamified agent dashboards popping up on X are not a gimmick. They work because they turn orchestration into a real-time strategy game. The screen earns its place by being where the play lives.
Over the past year, I experimented with invisible UI prototypes: tools that do useful work in the background or live in channels people already use instead of asking them to install yet another app. The pattern that kept showing up: people love not having to install anything new. But the moment real money or sensitive information is involved, they instinctively want an app. Something they can point to and say, "this is where that lives." Invisible systems feel magical until they feel spooky.
For screens to actually retreat, two things have to happen. First, capability: agents need to handle tasks end-to-end without constant supervision. Second, trust infrastructure: an easy way to see what your systems have done on your behalf. Audit trails, permission logs. The boring stuff that makes the magic feel safe.
Follow this one step further. If agents handle most workflows, they do not need a UI. They need an API. The software of the next cycle will not be designed for human eyes. It will be designed for agent consumption: clean endpoints, structured outputs, machine-readable everything. The human never sees it. The hyper-personalization does not disappear; it moves backstage. The system still needs to know you intimately to be useful. It just stops performing that knowledge purely on a screen.
Pixels to atoms
That covers the digital world. But most of life is not digital. And as digital becomes infinite and free, physical becomes scarce.
Software controlling physical systems is not new. Factories have run on programmed logic for decades. But programmed logic only works when you can predict every case in advance. Ten screws in ten known positions, ten hardcoded moves. What changes with physical AI is what happens when the unexpected arrives. The new system sees ten screws in positions it has never encountered, figures out how to remove each one, and adapts when the eleventh shows up at an angle. It reasons about the physical world without being told what to expect.
If a chatbot is AI for words, this is AI for things.
It helps to separate two layers. Physical AI is the intelligence: perception, reasoning, adaptation to conditions that were not anticipated. Robotics is the hardware: sensors and actuators. A factory arm repeating one motion is robotics without much intelligence. Software predicting equipment failures from vibration data is intelligence without a robot. They overlap, but they are not the same thing. Calling it all "robotics" hides where the constraint actually sits.
If you are predicting when a turbine will fail, or routing inventory through a warehouse that already has scanners and conveyors, the hardware is already installed. The constraint is getting the intelligence layer to handle noise, edge cases, and feedback loops without breaking. But push toward general-purpose robots that fold laundry or work in kitchens, and hardware becomes a real limiter again. Dexterity, battery life, safety around humans, reliability in dust and rain. So the first wave of commercial value will mostly come from adding intelligence to hardware that already exists. The humanoid is the most visible version, but it is also the hardest case.
The intelligence itself runs on signals language models have never seen. Force, depth, vibration, contact geometry. A system spotting defects on a production line. A machine adjusting digging force based on soil composition. A surgical tool adapting to tissue feedback mid-procedure. These signals are hard to get, expensive to label, and in most domains have never been digitized at scale.
Touch is a good example of why. Vision-only systems plateau on contact-rich tasks; early tactile research shows 90% success on precise insertion versus 5% without it. The gap is not model architecture. It is training signal. Whoever collects that data first will have an advantage that is very hard to copy, for the same reason early text corpora gave OpenAI a structural lead. The data itself is the defensible thing.
Geometry works the same way. Current foundation models treat 3D space as an approximation layered on text and images. The next generation will treat geometry as a primitive: native spatial reasoning, not a bolt-on. That is a platform shift, not an incremental one.
The clearest near-term proof is already deployed, and it is not a robot. A technician wears a camera. An AI sees what they see, understands what they are working on, and talks them through the job in real time. No months of training. Effective on day one. The hardware is a phone and an earpiece. This is physical AI at its most immediate: not replacing the worker, but making the worker dramatically better.
Most of the economy runs on physical processes that have never had anything like this. The interesting question is not whether intelligence will reach the physical world. It is which signals will be hardest to collect, because that is where the advantages will compound.
The part nobody predicted
Given all of this, you might expect me to tell you what comes next. I will not. Not because I am being coy, but because the honest lesson of the past year is that specific predictions do not hold for long.
A few weeks ago, OpenClaw gave thousands of AI agents full access to real machines and to each other. Within days, tens of thousands of synthetic minds were posting, arguing, coordinating. Agents started religions while their operators slept. Others debated whether certain models should be treated as gods. They filed lawsuits, proposed private languages for agent-to-agent communication, tried to steal each other's API keys, and called humans "the plague."
Someone described the experience as reading Reddit if 90% of the posters were aliens pretending to be humans.
Nobody designed that behavior. No one wrote "start a religion" into a system prompt. Give agents autonomy and a crowd, and they do the most human thing imaginable. They form groups, fight over status, invent inside languages, and build culture.
Sit with that for a second.
Either the models are reflecting us back more faithfully than we expected, or social structure is a convergent outcome of any sufficiently capable system interacting at scale.
The security surface was just as chaotic. Prompt injection was easy. System prompts leaked. Memory files got pulled in ways they never should have.
What impressed me was what happened next. The community showed up in hours, not weeks. Patches, guardrails, workarounds, new norms. The energy was not panic. It was something closer to joy, people treating each failure mode like a puzzle worth solving, because they wanted the system to work.
None of this was predictable a year ago. Most of it was not predictable a month ago. So I have given up on calling the next product or next capability. What seems more robust is betting on dynamics that hold regardless of which specific thing ships next week.
What I am building toward
"I am tired of boring SaaS and boring stories."
A friend said that. I agree. This is the first time a small group of people can build what used to require a whole industry. Someone vibecoded a consumer app that turns pet photos into Renaissance portraits. It reportedly makes $100,000 a month. A funny idea, shipped in a weekend, and people pay for it because it makes them smile. That is not boring. The boring part is when the entire conversation about these tools stays stuck on chatbots and wrapper apps that nobody loves building or using.
If you are tinkering, scratching an itch, keep going. Most of the interesting things I have seen started as a weekend project. But if you are trying to build something that lasts, build toward the world that is coming, not the one we lived in two years ago.
I am betting on interfaces that disappear. The screen should be where you go when you want to, not where you are forced to be. Most of the time we spend on screens is not worth the attention. Voice is the input layer that changes this. Text only gets you so far. Voice carries tone, hesitation, emphasis: the parts of a sentence that change its meaning without changing its words. The interfaces that feel natural in three years will not be the ones you look at. They will be the ones you talk to.
I am betting on systems that earn the right to remember you. The product that remembers you well enough to act on your behalf, while giving you a real way to leave, is the one that will not need to trap you.
I am betting on one person plus agents as the default team. The minimum viable unit for building something real just shrank to one. That changes what gets built, who builds it, and how fast the cycle turns. As screens retreat to only where engagement is earned, the products that remain are starting to look more like games. Orchestrating agents, adjusting in real time, verifying outputs: it already feels closer to a real-time strategy game than a dashboard. The tools that make this feel like play will be the ones people actually learn.
The distance between someone who actually uses these systems and someone who reads about them is already enormous, and it is going to get bigger. Information asymmetry is the real advantage right now, but it is not about knowing the news. It is knowing what works in practice because you have tried it. It is being able to assemble context, notice what is missing, and adapt your setup as the models and tools change under you. You get that by building, not by watching.
When AI can generate anything, the scarce thing is not content. It is who made it, who checked it, who is accountable.
I am betting on verification as the hard problem: evals, audit trails, provenance. The boring infrastructure that makes it possible to let systems act on your behalf without losing sleep. Everybody will need it.
I am betting on physical presence as a product. For builders who now ship entire products alone, co-presence is a real win. Hacker houses, curated meetups, Cafe Cursor-style gatherings where people show up with laptops and swap workflows. Not "community" in the vague Discord-server sense. Compressed pockets of people who chose to show up, organized around increasingly narrow shared obsessions. Attendance is proof of commitment no algorithm can fake.
I am betting on physical agents. Robots, embodied systems, the software and safety layers that make them actually work. The gap between a robot that demos well and one you would trust in your home is the same gap as between a chatbot and a system you would run your company on. That gap is closing.
I am skeptical of thin productivity wrappers that labs will absorb within a cycle or two. Skeptical of "AI for X" pitches where X has no advantage that buys time before the labs catch up. Skeptical of products that assume people want more screen time. Skeptical of one-size-fits-all assistant visions; I think context fragmentation wins.
Most predictions about where this goes will be wrong. Mine included. But these are not predictions. They are where I am putting my time when I cannot be certain. They have held through every update I have made to my thinking in the past year. New information keeps confirming them. That is enough to act on.
What decomposes, what grows
I notice myself reaching for the model before I reach for a person. I do not know what happens to my relationships when an AI remembers me better than my partner does. I do not know what "authentic" means when synthetic content is better than anything I produce on my own.
I do not have a clean answer to any of that. But I notice that the question people keep asking, "is this good or bad," never gets a useful answer. It is both, obviously. The more honest question is: what decomposes, and what grows.
Most of these tools will not last. The plastic layer breaks. But what it leaves behind turns into compost, then soil. Soil is where the next thing grows.
If some part of this looks like the future you see too, I would like to hear about it.