If you're asking when AI will replace white collar jobs, here's the honest answer as of mid-2025: not as soon as the headlines suggest — and the reason has almost nothing to do with raw intelligence. Today's large language models are genuinely impressive, but they're missing one critical capability that every human employee takes for granted: the ability to learn on the job. Until that changes, AI will automate tasks but won't replace people. This is the core argument behind Dwarkesh Patel's July 2025 AGI timeline, and it's worth understanding in full.
When Will AI Actually Replace White Collar Jobs?
The popular assumption is that LLMs are already capable enough to gut entire industries — that Fortune 500 companies are only slow to adopt AI because of bureaucratic inertia. Dwarkesh pushes back hard on this. The real barrier isn't corporate stubbornness. It's that getting reliable, human-quality labor out of these models is genuinely difficult, even for simple, self-contained tasks.
01:45
Dwarkesh explains why Fortune 500 companies aren't deploying LLMs to transform workflows — and why it's not just about slow management
Watch at 01:45 →
After spending over 100 hours trying to build LLM-powered tools for his own podcast production workflow — transcript cleanup, clip identification, essay co-writing — Dwarkesh found these models performing at roughly a 5 out of 10. Impressive, but not good enough to replace a human who improves over time. His conclusion: if AI progress stopped today, less than 25% of white collar employment would be automated away. Not because AI isn't smart, but because it can't get smarter at your specific job.
What Are the Most Credible AGI Timeline Predictions for 2025?
Opinions on AGI timelines vary wildly. Some researchers put it at two years away; others say twenty. Dwarkesh's position sits in a thoughtful middle ground, grounded in specific bets with concrete dates:
- 2028: An AI that can handle end-to-end small business tax preparation — chasing receipts, emailing vendors for invoices, classifying expenses, and filing the 1040 — as well as a competent human general manager could in a week.
- 2032: An AI video editor that, after six months on the job, has as deep and actionable an understanding of your creative preferences and audience as a human editor would.
These aren't pessimistic predictions. They're calibrated ones. And they hinge almost entirely on one unsolved problem: continual learning.
04:12
The saxophone analogy: why teaching LLMs is nothing like how humans actually learn a skill
Watch at 04:12 →
Why Can't LLMs Learn on the Job the Way Humans Do?
This is the crux of the entire argument. Humans are valuable not just because of raw intelligence, but because of their ability to accumulate context, interrogate their own failures, and incrementally improve through practice. LLMs, as currently architected, cannot do this.
Dwarkesh uses a memorable analogy: imagine trying to teach a child to play saxophone, but you're not allowed to give them iterative feedback. Instead, each new student gets one attempt, you write notes about what went wrong, and the next student reads your notes and tries to play Charlie Parker cold. No matter how well-crafted your instructions are, no child learns saxophone that way. But that's essentially the only modality available for teaching LLMs anything.
Yes, reinforcement learning fine-tuning exists. But it's not the organic, self-directed, adaptive process that human learning is. Dwarkesh's own editors have become excellent at their jobs because they noticed small things themselves, thought about what resonates with the audience, and refined their judgment over time — without anyone building a custom RL environment for each micro-skill they needed to develop.
There's a partial workaround that works inside a single session: if you spend enough time correcting an LLM's outputs and showing it what you actually want, it does start to give better suggestions within that conversation. But the moment the session ends, all of that tacit understanding is gone. Long context windows with compacted summaries — like Claude Code's rolling memory — help at the margins, but they're brittle. Subtle, hard-won optimizations frequently get lost in the compression. The model has no way to internalize why something was done a certain way, only that it was.
09:30
Dwarkesh lays out his specific 50/50 bets: 2028 for tax agents, 2032 for true on-the-job learning
Watch at 09:30 →
What Is Continual Learning and Why Is It the Key Bottleneck?
Continual learning — sometimes called online learning or test-time training — refers to a model's ability to update its weights and improve its behavior based on new experiences, in real time and without catastrophic forgetting of prior knowledge. Humans do this naturally. Current LLMs essentially don't.
Dwarkesh argues this is the bottleneck. Once it's solved, the value of AI models won't increase gradually — it will jump discontinuously. An AI capable of genuine on-the-job learning could be deployed across thousands of roles simultaneously, accumulating expertise from every single one and synthesizing those learnings across all its copies. One model effectively learning every job in the economy at once. Even without further algorithmic progress, that process could look like a broadly deployed intelligence explosion.
The good news: we probably won't be blindsided by this. Labs are incentivized to ship early. We'll see broken, early versions of continual learning long before the capability is truly mature — which means observers will have meaningful advance warning before this particular bottleneck is fully cleared.
Will AI Agents Do Your Taxes End-to-End by 2026?
Anthropic researchers Sholto Douglas and Trenton Bricken predicted on Dwarkesh's podcast that by the end of 2025 or early 2026, you should be able to tell an AI to handle your taxes completely — and it would. It would dig through your emails, Amazon orders, and Slack messages; email vendors for invoices; classify expenses; flag edge cases for your approval; and file the form. Dwarkesh is skeptical, and for three concrete reasons:
- Longer horizons mean harder evaluation. A two-hour agentic task involving computer use has to run to completion before you can even tell whether it worked. That dramatically slows down the feedback loops needed to train and improve these systems.
- There's no large pretraining corpus for computer use. Language models were trained on an enormous wealth of internet text. There is no equivalent dataset of humans navigating UIs, clicking through software, and completing multi-step computer tasks. That data scarcity is a serious constraint.
- Simple-sounding algorithmic innovations still take years. The core idea behind DeepSeek R1 and OpenAI's o1 — train a model to reason through verifiable math and coding problems — sounds straightforward in retrospect. It still took roughly two years from GPT-4's release to o1's launch. The challenges of computer use are far gnarlier.
To be clear: there will be impressive computer use demos in 2026 and 2027. But Dwarkesh puts the 50/50 bet for true end-to-end tax handling at 2028, not 2026. Think of where we are now as the GPT-2 era for computer use agents.
How Capable Are AI Computer Use Agents Right Now?
Current computer use agents exist but perform poorly. They're inconsistent, prone to errors, and not yet capable of managing complex, multi-step workflows without human supervision. That said, the underlying models are already meaningfully smart. The base intelligence is there; what's missing is the combination of reliable action primitives, dense training data in the visual/UI modality, and efficient RL over long horizons.
It's also worth noting what is working impressively well. Reasoning models like o3 and Gemini 2.5 are producing genuine chain-of-thought outputs — breaking down problems, catching their own errors mid-thought, and arriving at answers that feel qualitatively different from earlier LLM outputs. Claude Code can take a vague spec and zero-shot a working application in minutes. These are not parlor tricks. The most concise and honest description, as Dwarkesh puts it, is that we've built a baby general intelligence. It's real. It's just not yet capable of sustained, adaptive, long-horizon work.
Will AI Progress Slow Down or Stop After 2030?
This is where the long-term picture gets genuinely uncertain. AI progress over the past decade has been driven overwhelmingly by scaling training compute — at roughly 4x per year. That pace cannot continue past 2030. The physical constraints are real: chip supply, energy infrastructure, and the sheer fraction of GDP that would need to be devoted to training all impose hard ceilings.
After 2030, continued AI progress has to come primarily from algorithmic innovation. But the low-hanging fruit under the current deep learning paradigm will largely be exhausted. This means the per-year probability of hitting AGI drops significantly if we're still waiting in the 2030s.
Dwarkesh's framing: AGI timelines are highly lognormal. It's either this decade or the probability per year drops off sharply. If we land on the longer end of these predictions, a relatively normal world persisting into the 2030s or even 2040s isn't implausible. But in all the worlds where continual learning arrives and computer use matures on schedule, the outcomes are extraordinary — a broadly deployed intelligence explosion that transforms the economy faster than almost anyone is prepared for.
The bet, then, isn't whether to take AI seriously. It's about sequencing: understand the current limitations clearly, watch for the specific capability unlock that changes everything, and don't mistake today's impressive-but-brittle tools for the thing that will actually reshape the world.
