Is AGI actually close? Despite breathless predictions of imminent superintelligence, there is a fundamental contradiction hiding in plain sight: the same labs betting on short AGI timelines are also spending billions of dollars manually baking skills into AI models one painstaking training loop at a time. If we were truly close to a human-like learner, none of that costly hand-holding would be necessary. The fact that it is necessary tells you almost everything you need to know about where we actually stand.
Is AGI Close or Are We Still Far Away?
The tension is hard to ignore. On one side, you have researchers and investors who believe transformative AGI could arrive within five years. On the other, you have the observable reality that AI labs are funding an entire supply chain of companies building reinforcement learning environments to teach models how to navigate web browsers, operate Excel, and construct financial models from scratch.
00:45
The core contradiction: short AGI timelines vs. expensive skill pre-baking in RL pipelines
Watch at 00:45 →
Think about what that implies. A human employee does not need a dedicated rehearsal phase to learn every piece of software they might ever encounter on the job. They show up, they observe, they ask questions, and they adapt. If AI models were genuinely approaching human-like intelligence, this elaborate pre-baking of skills would be pointless. Either the models will soon learn on the job in a self-directed way, which makes all this scaffolding a waste of resources, or they won't, which means AGI is not imminent. Both conclusions cannot be true at the same time.
As AI researcher and writer Dwarkesh Patel puts it, this is not a minor inconsistency. It is a fundamental crux about the nature of intelligence and how far current systems actually are from replicating it.
Can AI Actually Replace Human Knowledge Workers?
Here is a concrete way to think about whether AI has reached the capability level implied by AGI: knowledge workers around the world collectively earn tens of trillions of dollars in wages every year. If AI models were genuinely operating at the level of a capable human employee, companies would be spending trillions of dollars annually buying the tokens these models produce. They are nowhere close to that figure right now, and that gap is not a fluke of slow technology adoption.
Consider the macrophage example. A biologist at a dinner explained that part of her recent lab work involved looking at microscope slides and determining whether a particular dot was a macrophage or just something that resembled one. The AI researcher at the table immediately responded that image classification is a textbook deep learning problem, something modern models are built to handle.
03:10
The macrophage dinner debate illustrating the gap between AI benchmark performance and real economic value
Watch at 03:10 →
But that response misses the point entirely. The problem is not whether AI can be trained to classify macrophages. The problem is that it is not net productive to build a custom training pipeline for every single micro-task inside every single job at every single company. After you build that pipeline for one lab's slide preparation method, you need another one for the next lab, and another for the task after that. Human workers are economically valuable precisely because they do not require this kind of bespoke scaffolding for every small corner of their work.
Every day, a knowledge worker encounters a hundred different tasks requiring judgment, situational awareness, and skills picked up organically through experience. These tasks differ not just across different people but from one day to the next for the same person. You cannot automate even a single job by pre-baking a fixed set of skills, let alone all jobs.
What Is Reinforcement Learning from Verifiable Reward?
Reinforcement learning from verifiable reward, sometimes abbreviated as RLVR, is the current approach many leading labs are using to push AI capabilities further after the initial pre-training phase. The idea is to have models practice tasks where the outcome can be clearly judged as right or wrong, such as solving math problems, writing code that either runs or does not, or navigating a specific software interface successfully.
This is a powerful technique within its domain. But there is a critical limitation baked into its name: verifiable reward. The world is full of tasks where success is ambiguous, contextual, and impossible to score with a clean metric. Most of the judgment calls a skilled professional makes on a Tuesday afternoon do not have a verifiable answer key waiting at the end.
Some optimists argue that even if current RL approaches are clunky, a sufficiently powerful automated AI researcher, a kind of machine Ilya Sutskever, could figure out the algorithm for true AGI. But this line of reasoning has the structure of a famous business joke: we are losing money on every sale, but we will make it up in volume. The model tasked with solving AGI would itself need the very learning capabilities it is supposed to discover, while currently lacking basic abilities that children demonstrate every day.
07:22
The continual learning vision: specialized agents distilling learnings back to a shared hive-mind model
Watch at 07:22 →
Why Isn't AI Already Deployed Across Every Business?
A common explanation for the relatively slow diffusion of AI into the broader economy is that technology simply takes time to spread. Businesses are conservative. Integration is hard. Change management is slow. This explanation, while not entirely wrong, is doing a lot of heavy lifting to cover for something simpler: the models do not yet have the capabilities required for broad economic deployment.
Here is the test. If AI models were truly human-like intelligences running on servers, they would diffuse into firms at a speed that would make any previous technology adoption look glacial. A new AI employee could read your entire Slack history and internal documentation within minutes. It could immediately absorb the skills of every other AI already working in the company. There would be no lemons market problem, no costly bad hire to recover from, no months-long onboarding process. Companies hire humans all the time despite those frictions. If AI were actually at human capability levels, the economic incentive to deploy it would be overwhelming and immediate. The fact that deployment remains limited is itself evidence that the capability gap is real.
What Is Continual Learning and Why Does It Matter for AGI?
If there is one capability that separates a genuinely transformative AI from the impressive but limited systems we have today, it is continual learning: the ability to learn from ongoing experience the way humans do, updating knowledge and skills through real-world interaction rather than discrete training runs.
Think about how human expertise actually develops. It is not primarily from reading textbooks or doing structured exercises. It is from doing the work, making mistakes, receiving feedback, and gradually building intuition that cannot be easily articulated or transferred. Current AI systems do not do this. They are trained, then frozen, then deployed.
One compelling vision for what a continual learning future might look like involves AI agents going out into the world, performing specialized jobs, generating value, and then returning their accumulated experience to a shared model that distills all those learnings in batch. These agents could be specialized, combining what researcher Andrej Karpathy calls a cognitive core with domain-specific knowledge relevant to their role.
Importantly, solving continual learning will not be a single dramatic breakthrough. It will likely follow the same trajectory as in-context learning. GPT-3 demonstrated remarkable few-shot learning in 2020, but that did not mean in-context learning was fully solved. Progress continued for years after. Expect the same pattern with continual learning: an initial capability that labs will market as a breakthrough, followed by years of incremental refinement before it reaches human-level on-the-job adaptability.
Will AI Scaling Laws Keep Holding Up?
Pre-training scaling has been one of the most reliable empirical trends in the history of machine learning. Throw more compute at a model trained on more data, and you get predictably better results. This trend held across multiple orders of magnitude and inspired enormous confidence in continued progress.
But there is an important sleight of hand happening when people extend that confidence to reinforcement learning from verifiable reward. Pre-training scaling is almost as predictable as a physical law. RLVR scaling has no comparable publicly documented trend. When researchers have tried to piece together what the data suggests, the picture is not encouraging. One careful analysis connected the dots between different benchmark results from OpenAI's O-series models and concluded that achieving a capability boost equivalent to one GPT training level might require something like a million-times scale-up in total RL compute. That is not a small gap to bridge.
Meanwhile, it is worth noting that benchmark improvements are not purely the result of better algorithms and more compute. They also reflect billions of dollars spent on expert annotators, PhDs, and domain specialists writing questions and providing carefully reasoned answers targeted at the precise capabilities being measured. Separating genuine capability improvement from expensive benchmark optimization is harder than the headlines suggest.
Why Humans Learn on the Job So Much Better Than AI
The robotics example cuts to the heart of this entire debate. At a fundamental level, robotics is an algorithms problem, not a hardware or data problem. Current robot hardware is capable enough. A human can learn to operate it with relatively little training and then generalize that knowledge to new situations almost immediately.
Because we do not have a human-like learner in AI, robotics researchers instead have to send robots into thousands of different homes to practice picking up dishes a million times, then separately practice folding laundry, then separately handle every other household variation. The scale of data collection required is staggering precisely because the learning is not generalizing the way human learning does.
This is the core issue. It is not that AI lacks raw intelligence in some domains. It is that the style of learning, brittle, non-generalizing, requiring massive amounts of targeted experience for each new context, is fundamentally different from the flexible, efficient learning that makes human workers so economically valuable. Until that gap closes, the promise of AGI will keep outrunning the reality of AI deployment, no matter how impressive the next benchmark result looks.
