When ChatGPT 3.5 launched in late 2022, the world reacted somewhere between astonishment and confusion. The technology felt magical, but also deeply unreliable. Hallucinations were constant, context windows were tiny, and interacting with the system still felt strangely technical. Most people treated it like a novelty. Something amusing. Something experimental. Something impressive, but not yet trustworthy enough to become part of real life.
Three years later, the conversation has changed completely. The question is no longer whether AI matters. The real question is what daily life will feel like once AI stops behaving like an application and starts behaving like infrastructure. It is exciting to already be thinking about what AI in 2029 will look and behave like.
I do not think the next phase of AI will be defined by one dramatic AGI moment where headlines collectively lose their minds for forty-eight hours before everyone quietly returns to Slack. I think it will be defined by compounding usability. Persistent memory, multimodal interaction, AI-native operating systems, agentic workflows, robotics, and contextual systems are slowly converging into something much larger than a chatbot. The important shift is not merely that the models are becoming smarter. The important shift is that the friction surrounding them is disappearing.
That is why I increasingly believe that, within only a few years, not using AI may feel similar to riding a horse after the car was invented. Not because horses instantly disappeared, but because the world reorganized itself around a different default. Roads changed. Cities changed. Supply chains changed. Human expectations changed. Eventually, refusing the new infrastructure stopped feeling romantic and started feeling impractical.
The Industry Already Quietly Changed
At roughly the same time as ChatGPT 3.5, Google launched Bard and immediately demonstrated that even the largest technology companies were not immune to the chaos of the moment. The rollout became infamous for hallucinations and factual mistakes, creating the impression that Google had somehow been caught flat-footed in the very race it was supposedly born to win. Apple, who had famously acquired Siri and integrated it into iOS as the first “AI” assistant (there are not enough quotes for this), was, and still is, nowhere to be seen.
Three years later, the conversation feels completely different. ChatGPT 5.5 is no longer treated like a novelty toy for generating mediocre LinkedIn posts about “disruption.” Gemini evolved from an awkward public stumble into one of the strongest multimodal ecosystems in the market. Claude became a serious frontier-model competitor with a distinctly strategic and thoughtful interaction style. Meanwhile, companies that initially dismissed generative AI as hype are now quietly trying to retrofit entire organizations around it before shareholders notice how unprepared they actually are. Apple has famously foregone frontier model development and closed a nine-digit deal with Google to base its new Siri on Gemini models.
What changed was not merely model capability. The entire interaction layer matured. Voice became usable. Context windows became large enough to support meaningful work. Memory became persistent. Tool usage expanded. Models became better at reasoning across ambiguity instead of collapsing theatrically the moment reality stopped resembling benchmark datasets.
Most importantly, people stopped interacting with AI occasionally and started integrating it behaviorally. That is the real shift. The future of AI will probably not arrive through one mythical breakthrough. It will arrive through compounding usability.
Usability Changes Human Behavior
One of the biggest misconceptions about generative AI is that its progress can be understood purely through benchmark scores. Benchmarks matter, but usability changes mass behavior.
I did not fully understand this when ChatGPT 3.5 first appeared. Like many people, I experimented with it repeatedly without integrating it into my life. Then something changed. At some point, I no longer felt limited by the technology itself. Instead, I started noticing how much cognitive friction existed in ordinary workflows that I had simply accepted for years.
The biggest breakthrough of the past few years may not have been intelligence. It may have been the gradual removal of interface friction. The systems stopped feeling brittle. And once friction falls below a certain threshold, behavior changes surprisingly fast.
This pattern appears constantly throughout technological history. The smartphone did not become transformational because the first iPhone had the best hardware, although a case can be made for the iPhone’s best user interaction model, which supports this hypothesis even more. Streaming did not replace physical media because buffering was enjoyable. Cloud computing did not spread because executives enjoyed migration projects and governance meetings that somehow consume six people for three weeks just to approve a checkbox.
Winning technologies are usually the ones that become behaviorally effortless.
Generative AI is now approaching that threshold. Stanford HAI’s 2025 AI Index reported that inference costs for GPT-3.5-level performance dropped more than 280-fold between late 2022 and late 2024. When capability improves while cost simultaneously collapses, infrastructure adoption accelerates very quickly. And infrastructure adoption changes societies far more deeply than isolated applications do.
I wrote previously about how AI has lowered the cost of starting while simultaneously raising the standard for finishing. That asymmetry matters because AI does not merely accelerate execution. It changes expectations.
The AGI Obsession Is Distracting From the Real Story
Before GPT-5, online speculation reached predictably absurd levels. Reddit threads and industry whispers convinced themselves that OpenAI had secretly achieved AGI internally and was delaying the release because the system was supposedly too dangerous. Now similar mythology is forming around Anthropic’s rumored Mythos model.
But technological transformation rarely behaves through singular cinematic moments. Electricity spread gradually. The internet evolved through decades of iteration. Smartphones took years before the ecosystem matured enough to reorganize human behavior at scale.
The OECD recently argued that generative AI increasingly resembles a general-purpose technology. That framing is extremely important because general-purpose technologies are not isolated inventions. They are infrastructural shifts.
Electricity transformed factories, homes, logistics, manufacturing, and communication. The internet reorganized commerce, media, identity, relationships, and information distribution. AI increasingly appears to be following the same pattern, not because one model suddenly becomes godlike, but because consecutive incremental improvements compound into ecosystem-level transformation.
You can already see this emerging across the major players. OpenAI’s GPT ecosystem, Google DeepMind’s Gemini platform, and Anthropic’s Claude ecosystem are no longer competing purely on chatbot quality. They are competing to become behavioral infrastructure.
That distinction matters because infrastructure changes defaults. Nobody debates whether to use electricity before turning on the lights. Nobody thinks of the internet as “technology adoption” anymore. At some point, technologies stop feeling optional and start feeling environmental.
AI is moving frighteningly fast toward that threshold.
The Operating System Layer Is Being Rewritten
Apple may accidentally become one of the most important AI companies of the next decade.
Historically, Apple has rarely been first to invent a category. Its real strength has usually been removing friction aggressively enough that complicated technology suddenly becomes behaviorally natural. Siri became a cultural embarrassment partly because it never crossed the threshold from novelty into trustworthy utility. It was too inconsistent to become infrastructure.
But now something much larger is happening.
Apple, Google, OpenAI, Microsoft, and others increasingly appear to be restructuring the operating-system layer itself around AI-native interaction models. This matters because today, most people still consciously access AI. Three years from now, that separation may largely disappear.
Instead of opening applications manually, users may increasingly describe intent. Instead of navigating interfaces step by step, users may increasingly orchestrate outcomes, perfectly exemplified by Google’s Create My Widget feature, coming to Android 17 and it’s “whatevert-it-is-OS” which will run on Googlebooks. The interface itself may begin disappearing into behavior.
Google DeepMind’s Project Astra already points toward this future through contextual awareness, multimodal continuity, cross-device memory, and persistent environmental interpretation. Apple is moving in a similar direction with Apple Intelligence, embedding AI directly into system-level workflows instead of treating it like a separate destination.
The significance here is not merely that AI becomes smarter. It is that interaction becomes less explicit. The user stops managing the system manually. The system increasingly manages context.
That changes cognitive behavior.
It also changes expectations around software itself. Today, most software still assumes humans will manually navigate systems, retrieve information, move data between tools, structure workflows, and remember context across fragmented environments. AI-native operating systems increasingly assume the opposite. The system should carry context forward. The system should understand continuity. The system should reduce orchestration overhead instead of creating more of it.
This is probably the real meaning of “ambient computing,” and it is far more transformative than most current AI discourse suggests. Ambient computing does not mean talking to a robot all day like a science-fiction side character. It means computation becoming so behaviorally integrated that it stops demanding constant conscious management. It also unlocks new interaction interfaces, overcoming the caveats of virtual- or augmented-reality glasses or smartwatches, where a lack of pointing device limits the usability of the device.
Electricity did not become transformative because people loved power plants. It became transformative because people stopped thinking about electricity at all.
AI may follow the same path.
This also connects directly to ideas I explored previously around what organizations should actually do with the capacity AI creates. Most companies still frame AI primarily as a cost-saving mechanism. The more interesting question is what happens once cognitive friction itself begins collapsing.
The Next Generation of Users May Experience a Different Internet
The people who deeply integrate AI into their workflows today are not simply becoming marginally more productive. They are beginning to operate inside a different cognitive environment.
The first barrier of execution is disappearing, and that changes behavior in subtle but important ways. Many people still underestimate how psychologically expensive execution friction really is. Starting is difficult. Structuring ambiguity is difficult. Blank pages are difficult. Decision paralysis is difficult.
AI increasingly reduces the cost of initiation.
Today, I use AI constantly as a sparring partner for strategic reasoning, first drafts, challenging assumptions, structuring ideas, and expanding possibilities. Many people experience one hallucination and immediately conclude the entire category is unreliable. That is roughly equivalent to abandoning spreadsheets because humans occasionally destroy billion-dollar forecasts with the wrong cell reference.
The healthier approach is understanding where trust ends and supervision begins.
I do not trust AI blindly. I supervise it. But I would also never voluntarily return to a workflow without it.
That distinction matters enormously because people often frame AI incorrectly as either perfect or useless. In reality, it behaves much more like leverage. And leverage rarely requires perfection to become transformational.
Calculators still produce wrong outputs when humans enter incorrect formulas. Search engines still surface misinformation. Spreadsheets still contain catastrophic human errors. Yet nobody voluntarily abandons them because the leverage advantage dramatically outweighs the supervision cost.
Increasingly, the same thing appears true for AI, if not temporarily until AI starts truly compensating for our human shortcomings, which might just be the tease of “what AI will look like in three years”.
This is particularly important for people whose cognitive profile historically made initiation expensive. I spent most of my life knowing that once I gained momentum, I became extremely difficult to stop. But starting was costly. Context switching was costly. Managing too many parallel initiatives felt dangerous because I knew exactly how quickly overload could collapse everything simultaneously.
AI changed that dynamic.
The removal of initiation friction suddenly made it possible to explore more ideas simultaneously without feeling cognitively paralyzed by orchestration overhead. That matters far more than most productivity discussions acknowledge. AI is not only accelerating execution. It is changing what individuals feel psychologically capable of attempting in the first place.
That may ultimately become one of the biggest behavioral transformations of all.
Agents Change the Economics of Individual Capability
The future is probably more agentic than most people currently realize. The next generation of systems will increasingly coordinate tasks instead of merely answering prompts.
Agentic AI does not simply generate responses. It executes workflows, maintains objectives, interacts with tools, retrieves information, coordinates systems, and increasingly operates semi-autonomously toward outcomes.
This changes the economics of individual capability.
A single person may increasingly operate at leverage levels that previously required small teams. OpenAI’s Operator direction, Anthropic’s computer-use systems, and Google’s increasingly interconnected ecosystem all point toward the same underlying shift: delegated execution.
For decades, software required humans to constantly translate intent into procedural interaction through menus, settings, navigation layers, and endless orchestration logic. Agentic systems increasingly reverse this relationship. Humans specify intent while systems increasingly handle orchestration.
This may reshape knowledge work far more deeply than most organizations currently understand.
Especially because organizations themselves remain structurally slower than the individuals inside them.
Microsoft’s Work Trend Index increasingly describes this exact tension. Employees are often adapting faster than governance structures. Workers experiment faster than institutions. Leadership culture frequently lags behind capability.
The technological bottleneck is no longer exclusively technical.
It is organizational.
This is where many companies may become dangerously complacent. They still think they are “adding AI to workflows,” when in reality AI increasingly behaves like a coordination layer above workflows themselves. That distinction changes everything.
If AI reduces orchestration overhead dramatically enough, entire assumptions around organizational structure may eventually become unstable. Some coordination-heavy environments may discover that large amounts of managerial overhead existed primarily to compensate for information friction, communication latency, fragmented systems, and decision bottlenecks that AI increasingly mitigates.
That does not mean organizations disappear. It means their shape may change.
Robotics Is the Next Acceleration Layer And Can Help Define What AI in 2029 Will Be Like
Robotics and hardware are the next major acceleration layer because AI as software eventually expands into AI as embodied capability.
For years, robotics remained constrained by perception, adaptability, and environmental complexity. Traditional automation works extremely well inside predictable industrial boundaries, but reality outside controlled environments is messy, dynamic, and deeply hostile to rigid systems.
That is beginning to change.
Modern multimodal models increasingly understand spatial reasoning, visual interpretation, environmental context, and adaptive behavior. Google DeepMind’s Gemini Robotics initiative points directly toward this convergence.
The long-term significance is difficult to overstate because industrial revolutions historically scale dramatically once intelligence becomes embodied. Software changes information. Robotics changes physical systems. Factories, warehouses, healthcare, logistics, domestic assistance, infrastructure maintenance, and manufacturing all become candidates for transformation once generalized reasoning can operate physically.
This is also why I increasingly believe many current fears around AI infrastructure limitations may eventually become self-correcting. AI itself may accelerate the discovery of cleaner energy systems, more efficient chip architectures, better material science, improved logistics, and more efficient manufacturing processes. Humanity has a long history of assuming constraints are static right before technology reorganizes the assumptions underneath them.
That does not guarantee utopia, obviously. Humans remain fully capable of using miraculous technologies to create entirely new categories of stupidity.
But the direction matters.
The Biggest Risk May Not Be AI Itself
The biggest risk may not be AI.
It may be society’s adaptation speed.
Technological transitions rarely fail because the technology itself is impossible. They fail because institutions adapt slower than the environment around them. Education systems move slowly. Corporate governance moves slowly. Legal systems move slowly. Procurement moves slowly. Management culture moves slowly.
Meanwhile, the capability layer continues accelerating.
This creates increasing asymmetry between early adopters and institutional laggards. The people who deeply integrate AI into their workflows today are already beginning to operate differently through faster synthesis, faster iteration, lower initiation friction, expanded leverage, and broader exploration capacity.
That does not necessarily mean everyone else becomes obsolete. But it does create widening asymmetry.
Anthropic’s Economic Index increasingly suggests that AI usage patterns are evolving toward collaboration loops rather than isolated prompts. Validation, iteration, directive refinement, learning, and execution support increasingly resemble cognitive augmentation rather than simple automation.
And augmentation compounds.
Especially across years.
This is why I increasingly worry less about the technology itself and more about social adaptation failure. Humanity may briefly find itself in the absurd position of experiencing unemployment, instability, and fear in the middle of rapidly increasing abundance because institutions fail to evolve quickly enough around the new economics of capability.
That would be tragic.
And historically, extremely human.
What Work May Actually Feel Like in 2029
To understand where this is going, it helps to stop thinking about AI as a chatbot and start imagining what ordinary workflows may actually feel like three years from now.
Take something trivial: preparing a client presentation.
Today, even with modern AI systems, the workflow still feels relatively manual. A consultant gathers notes after a workshop, searches through previous documents, asks AI to help structure ideas, rewrites sections manually, creates slides, validates assumptions, adjusts formatting, aligns stakeholder expectations, and continuously orchestrates the process step by step. The AI assists the workflow, but the human still carries most of the coordination burden.
Three years from now, that relationship may feel fundamentally different.
Imagine leaving a client meeting while your AI system has already:
- transcribed the entire conversation in real time,
- summarized the discussion into executive and operational versions,
- identified unresolved strategic questions,
- retrieved relevant internal knowledge from previous engagements,
- compared the discussion against similar client situations,
- highlighted contradictions between stakeholder priorities,
- mapped political alignment and resistance risks between departments,
- identified missing data necessary for decision-making,
- generated multiple narrative directions for the presentation,
- adapted those narratives to the communication style of different stakeholders,
- created first-draft presentation structures automatically,
- proposed implementation roadmaps,
- identified dependencies, risks, and escalation points,
- flagged unsupported assumptions or weak arguments,
- cross-checked claims against market and company data,
- prepared alternative recommendation paths based on budget constraints,
- estimated operational complexity for each proposed direction,
- identified where legal, compliance, or procurement friction may emerge,
- suggested which executives should be involved earlier in the process,
- generated visual concepts for slides and storytelling flow,
- and prepared a prioritized action summary before you even return to your desk.
Instead of beginning from a blank page, you may already return to a structured strategic recommendation waiting for refinement, challenge, and human judgment. Not because you manually instructed the system to execute every individual step, but because the system increasingly understands the context surrounding the objective itself.
That is the real shift.
The future workflow is probably not “human creates prompt, AI creates answer.” The future workflow is continuous contextual collaboration. The system maintains continuity across conversations, projects, tools, decisions, and objectives in ways that drastically reduce orchestration overhead. Instead of repeatedly reconstructing context, humans increasingly operate on top of persistent cognitive infrastructure.
That matters more than most current discussions acknowledge because orchestration is one of the hidden taxes of modern knowledge work. Huge amounts of cognitive energy are currently spent not on solving meaningful problems, but on coordinating fragmented systems, searching for information, rebuilding context, translating between tools, preparing status updates, reconstructing decisions, and manually carrying continuity between disconnected environments.
The real productivity explosion may not come from AI generating content faster. It may come from AI collapsing coordination friction itself.
And once that happens, individuals may suddenly find themselves capable of operating at scales previously associated with entire departments. Not because humans become superhuman overnight, but because modern work contains astonishing amounts of invisible friction that people stopped noticing years ago.
That may ultimately become the defining characteristic of this next era of AI. Not artificial intelligence as spectacle, but artificial intelligence as removal of resistance.
And if we navigate this transition responsibly, future generations may eventually look back at today’s workflows the same way we now look back at paper maps, fax machines, or manually searching through filing cabinets. Not with ridicule, but with quiet disbelief that so much human potential was once spent fighting systems instead of building things with them.
AI Is Becoming Infrastructure
By 2029, AI may feel less like software and more like infrastructure: persistent, ambient, integrated, and increasingly invisible.
The systems themselves may eventually matter less than the environments they create. That is usually how infrastructure behaves. Nobody emotionally discusses electricity every morning. People discuss what becomes possible because electricity exists.
The same thing may increasingly happen with AI.
Eventually, the conversation may stop being about models and start being about leverage. The people who embrace this shift early may begin operating at capability levels that feel fundamentally different from traditional digital work, not because they suddenly become geniuses, but because they spend dramatically less cognitive energy fighting execution friction.
The internet expanded access to information.
AI increasingly expands access to execution.
Three years from now, not using AI may not make someone obsolete. But it may make them feel strangely underpowered inside systems that increasingly assume AI-assisted cognition as the default operating model.
History has rarely been kind to societies that underestimate infrastructure transitions while they are still forming. Humans usually recognize paradigm shifts only after the environment has already reorganized around them.
Which is inconvenient.
But apparently one of civilization’s favorite hobbies.
