Section 1: The AGI Bubble
Every few decades, the world finds a new story to believe in.
In the 1990s, it was the internet. Before that, it was personal computing. Even earlier, it was electricity, automobiles, and railroads. These technologies didn't just create new industries; they reshaped economies by attracting enormous amounts of capital and unlocking decades of growth.
Today, that story is about AI.
AI is no longer just another technology trend. It has become the organizing force behind capital, infrastructure, talent, and policy. Almost every major technology company has reoriented its strategy around AI. Governments are treating computer infrastructure as a strategic asset, while investors continue deploying capital at an unprecedented pace.
The scale of investment is extraordinary. The St. Louis Fed estimated that AI-related investment contributed close to one percentage point to US GDP growth through the first three quarters of 2025. That was around 39% of total growth in that period. NVIDIA has become one of the most valuable companies in history, their m-cap is higher than India’s GDP. FDI in TSMC was higher than total FDI India in 2025. Sandisk share price grew 15x!
The impact extends far beyond AI companies themselves. Data centers are being built at record speed, cloud providers are ordering hundreds of thousands of GPUs, and entire supply chains are expanding because the market believes AI will become the next great technological platform.
There is also a broader economic context behind this. The United States has historically driven growth by creating new industries, from semiconductors and the internet to cloud computing and smartphones. AI is now expected to become the next engine of economic growth, and that expectation is driving one of the largest infrastructure investment cycles the technology industry has ever seen.
Underlying this entire ecosystem is a single assumption: if we continue investing more in data centers, GPUs, and larger training runs, AI systems will continue becoming more capable until they eventually reach Artificial General Intelligence.
That assumption now underpins trillion-dollar valuations, unprecedented capital investments, and some of the biggest economic bets being made today.
The question I keep coming back to is whether that assumption deserves the level of certainty the market has assigned to it.
Section 2: The Assumption Everyone Is Betting On
If you strip away the headlines, trillion-dollar valuations, and excitement surrounding AI, the entire ecosystem rests on one surprisingly simple assumption.
The assumption is that intelligence is primarily a scaling problem. If we continue training larger models on more data using more compute, intelligence will continue improving until we eventually arrive at AGI.
It's not an unreasonable belief.
In fact, it has been one of the most successful ideas in artificial intelligence over the past decade.
Every major breakthrough, from GPT-2 and GPT-3 to GPT-4, Claude, Gemini, and today's reasoning models, has followed the same pattern. More parameters, more data, more computation, and consistently better results. This became known as the scaling hypothesis, and the evidence supporting it was compelling. As models grew larger, they became noticeably better at writing, coding, translating, reasoning, and solving increasingly complex problems.
That success changed the industry.
It justified billion-dollar funding rounds for frontier labs, massive investments in AI infrastructure, and countries treating semiconductor supply chains as strategic assets. Investors believed larger models would unlock dramatically more capable AI systems, which in turn justified even more investment. Capital funded larger training runs, improved models generated enterprise demand, and that demand reinforced the belief that scaling would continue to deliver outsized returns.
The important point is that none of this means the hypothesis is wrong.
Scaling has worked remarkably well.
The real question is whether it will continue working at the same rate.
Almost every engineering discipline eventually encounters diminishing returns. Early improvements are dramatic because there is abundant low-hanging fruit, but every additional improvement demands disproportionately more effort and capital.
AI may be approaching a similar phase. Models continue to improve, but each incremental gain now requires exponentially larger investments in compute, infrastructure, energy, and data.
If that trend continues, the question stops being, "Can we build a smarter model?"
It becomes, "Is the next percentage point of intelligence worth another hundred billion dollars?"
Because once that question becomes economic rather than technical, the conversation around AI changes completely.
Section 3: The Reliability Wall
To answer that question, we first need to separate what AI feels like from what it actually is.
When you interact with ChatGPT or Claude, it feels intelligent. It remembers context, writes fluently, reasons through problems, and often produces answers that are difficult to distinguish from those written by humans. It's easy to conclude that the model knows the answer.
But that's not really what's happening.
At its core, today's generation of AI systems is probabilistic. They don't understand the world the way humans do or reason from first principles. Instead, they predict the most likely sequence of words based on everything they have seen during training. Most of the time, those predictions are remarkably accurate, but they remain predictions nonetheless.
That distinction changes how we should think about progress.
Every increase in compute, data, and training improves the quality of those predictions. Models become better at following instructions, solving complex problems, and making fewer mistakes. But reducing the probability of error is very different from eliminating error altogether.
If the current paradigm remains fundamentally probabilistic, then every additional training run will continue pushing the error rate lower without ever driving it to zero. The gains become progressively smaller with every increase in investment.
The graph should show an asymptotic curve where the error rate falls rapidly at first but gradually flattens, approaching zero without ever reaching it. GPT3 to GPT4 was a leap, and GPT5 flat-lined. Anthropic took a different approach, where they trained domain-by-domain, getting higher accuracy in coding, legal, etc but eventually flat-lined.
This pattern isn't unique to AI. Many engineering systems exhibit diminishing returns, where the first improvements are relatively inexpensive, but the final percentage points require exponentially more effort and capital. If reaching 95% accuracy costs 10 billion dollars, reaching 99% could require 10 trillion, while delivering comparatively smaller gains.
We're already beginning to see signs of this. Frontier models continue improving, but the cost of achieving each new level of capability is rising much faster than the improvements themselves. That doesn't mean progress has stopped. It means intelligence may no longer scale linearly with spending.
If that's true, the biggest challenge facing AI over the next decade won't be intelligence.
It will be - being reliable, given these are not deterministic systems.
Businesses don't fail because AI gets 99 answers right. They fail because nobody knows which answer is wrong. In customer support, that could mean an incorrect refund. In healthcare, a missed diagnosis. In software, a security vulnerability.
The challenge isn't that AI makes mistakes.
It's that the system doesn't reliably know when it has made one.
That, in my view, is the wall the industry is beginning to approach, not an intelligence wall, but a reliability wall.
Section 4: Why Human-in-the-Loop Isn't a Temporary Phase
If today's AI systems are fundamentally probabilistic, then the next question becomes obvious.
What does the world look like if AI continues to become more capable but never becomes perfectly reliable?
My view is that we stop thinking about AI as a replacement for people and start thinking about it as infrastructure that dramatically increases what people can accomplish.
We're already seeing this happen. AI is taking over repetitive, structured work, summarizing documents, writing code, drafting emails, classifying support tickets, retrieving information, and completing countless tasks that previously required manual effort.
But businesses don't operate on probability alone.
Every organization eventually encounters situations where context matters more than pattern recognition. A customer requests an exception to the return policy because of a family emergency. A bank has to decide whether a suspicious transaction should actually be blocked. A doctor receives conflicting recommendations. None of these problems are difficult because information is missing. They're difficult because they require judgment.
That's the difference benchmarks often fail to capture. They measure whether a model can generate the correct answer. Businesses care about something more important: whether they can trust that answer enough to take responsibility for it.
We've seen this firsthand while building AI for customer support.
Most customer conversations are straightforward. Questions about order status, subscriptions, or product availability are structured problems with structured answers. AI handles these extremely well. The challenge begins when the conversation shifts from retrieving information to making business decisions.
A customer may request a refund outside company policy or dispute a delivery marked as complete by the carrier. AI can summarize the context, recommend a resolution, and estimate possible outcomes. What it cannot do is decide on behalf of the business.
That is why I believe human-in-the-loop isn't a temporary bridge until AI improves. It is becoming the architecture through which AI will be deployed across industries. At kim.cc, we call them sentinels.
As AI takes over routine execution, humans increasingly focus on reviewing exceptions, resolving ambiguity, and making high-impact decisions where the cost of being wrong is high. Developers review AI-generated code. Lawyers review AI-generated contracts. Customer support teams step into conversations requiring empathy, negotiation, and judgment.
The pattern is remarkably consistent.
AI handles scale.
Humans handle uncertainty.
The companies that win won't simply have the smartest models. They'll be the ones who become exceptionally good at deciding where automation ends and where human judgment begins, because intelligence creates leverage, but reliability creates trust.
Section 5: AI Doesn't Eliminate Work. It Changes the Economics of Work
Whenever a new technology arrives, the first question people ask is whether it will replace jobs.
The same debate accompanied the Industrial Revolution, personal computers, the internet, cloud software, and automation. Each wave created anxiety because people naturally compared the new technology to existing jobs rather than the new kinds of work it would create.
AI is no different, except that it feels more personal because it appears capable of performing knowledge work that we once believed was uniquely human.
I think both the optimists predicting fully autonomous companies and the pessimists forecasting mass unemployment are missing what is actually happening.
If AI remains a probabilistic system that requires oversight for high-stakes decisions, businesses won't eliminate humans. They'll redesign work around AI.
Customer support is already moving in that direction.
A few years ago, every customer conversation required a human agent. Today, depending on the complexity of the business, AI can independently resolve anywhere between 30% and 70% of support requests, typically repetitive questions such as order tracking, subscription updates, delivery status, and FAQs.
The remaining conversations involve exceptions, ambiguity, emotional context, and business judgment. The outcome isn't a support team without people. It's a support team where each person can handle significantly more customers because AI absorbs much of the repetitive workload.
The same pattern is emerging across software engineering, marketing, finance, legal services, and consulting. AI increasingly handles execution, allowing humans to focus on review, refinement, and decision-making.
This is what distinguishes AI from previous software waves.
Traditional SaaS improved productivity by organizing workflows. AI goes a step further by participating in the work itself.
The graph should illustrate how work shifts across three phases:
Before SaaS: Most effort is spent executing tasks manually.
SaaS Era: Software streamlines workflows, but humans still perform nearly all core work.
AI Era: AI takes over a significant portion of execution, allowing humans to focus on judgment, review, and exception handling.
The important point isn't that humans become less valuable. It's that every human becomes more productive.
History suggests that this is how technological revolutions create value. They don't eliminate the need for capable people; they multiply what capable people can accomplish. The companies that win won't simply adopt AI. They'll redesign their organizations around it while preserving the human judgment that businesses and customers continue to rely on.
Section 6: India's Opportunity Isn't Building AGI. It's Operating It
Every technological shift creates new winners, but those winners aren't always the countries that invent the technology.
The United States built the internet, but manufacturing shifted elsewhere. Taiwan became indispensable to the semiconductor industry without producing the world's largest consumer technology companies. Every country eventually finds the layer of the value chain where it has a structural advantage.
I think AI will follow a similar pattern.
Much of today's conversation revolves around foundation models. Which company will build the next GPT? Who will train the next trillion-parameter model? Who will own the most advanced chips?
These are important questions, but they're also some of the hardest markets to compete in. Training frontier models requires enormous amounts of capital, energy infrastructure, cutting-edge semiconductors, research talent, and years of accumulated expertise. Only a handful of organizations have the resources to compete at that level.
That doesn't mean the opportunity is over for everyone else. It means the opportunity lies somewhere else. AT&T, Verizon laid the undersea cables during the late 90s, but they did not capture the whole internet market.
If AI becomes a technology that amplifies human capability rather than replacing it entirely, then the next decade won't simply be about building intelligence. It will be about deploying that intelligence across real businesses, workflows, and operations.
Every company has its own systems, policies, customer expectations, and operational complexity. AI doesn't automatically understand any of that. Someone has to integrate it into existing workflows, build safeguards, monitor outcomes, and continuously improve the system as the business evolves.
That is fundamentally an operational challenge.
And that is where I believe India has a genuine advantage.
For decades, India has built one of the world's largest service economies. We've developed deep expertise in running large-scale operations across customer support, finance, healthcare, IT services, consulting, and business processes. AI doesn't replace that expertise; it amplifies it.
Instead of selling human effort alone, businesses can now combine AI with operational excellence to deliver outcomes at far greater scale. The value shifts from execution to orchestration.
This is why I believe one of the defining business models of the next decade won't simply be AI software. It will be AI-enabled services, where companies combine AI with deep domain expertise to solve real business problems.
We've spent decades becoming the operational backbone of global businesses. AI allows us to rebuild that industry with significantly higher leverage.
The race to build the smartest model may ultimately be won by a handful of companies.
The race to build the most valuable AI-powered businesses is still wide open.
Section 7: The Bubble Isn't AI. It's Our Expectations
None of this should be interpreted as a bearish view on AI.
Quite the opposite.
I believe AI will become one of the most transformative technologies of our lifetime. It will change how software is built, how businesses operate, how people work, and how value is created. We are still in the early stages of that transformation.
Where I differ from the prevailing narrative is in what I think success looks like.
Today, much of the market is pricing a future where increasingly larger models become reliable enough to replace large sections of human work. That belief justifies unprecedented investments in chips, data centers, computer infrastructure, and frontier models.
I believe that the future may look different.
AI will continue to become more capable, and it will undoubtedly take over a larger share of repetitive work across industries. But I find it less convincing that this naturally leads to a world where human judgment becomes unnecessary.
Businesses don't optimize for capability alone. They optimize for trust.
As AI takes on greater responsibility, questions around reliability, accountability, governance, and human oversight become more important, not less. Those aren't technical problems alone. They're operational ones.
That is why I believe the next decade will be defined less by intelligence itself and more by everything built around it. The companies creating the most value may not be the ones training the largest models. They may be the ones building the orchestration layers, reliability systems, governance mechanisms, and industry-specific workflows that allow AI to operate safely inside real businesses.
History rarely rewards the companies that simply build breakthrough technologies. More often, it rewards those who figure out how to make those technologies useful at scale.
I believe AI will follow a similar path.
The infrastructure being built today is necessary, and the models will continue improving. But the businesses that define the next decade will likely be those that combine AI with human judgment, operational excellence, and deep domain expertise.
That is the future I'm betting on.
Not because AI will stop improving, but because once technology leaves the lab and enters the real world, reliability almost always matters more than possibility.
“The conversation shouldn't be about whether AI will replace humans. It should be about how humans and AI together can build systems that are more capable, more reliable, and more valuable than either could be on their own. If that is where the world is heading, then perhaps the biggest opportunity isn't building AGI. It's building everything that AGI will still need to work in the real world”