AI Progress Is Stalling Because We're Teaching Robots to Use Excel Instead of Teaching Them to Learn

Dwarkesh Patel just published a provocative essay arguing that current AI progress reveals a fundamental contradiction: if we're actually close to AGI, why are labs spending billions teaching models to use web browsers and Excel through elaborate reinforcement learning pipelines? Either these models will soon learn autonomously on the job—making all this pre-training pointless—or they won't, which means AGI isn't imminent.

He's moderately bearish short-term, explosively bullish long-term, and his reasoning exposes uncomfortable tensions in how AI companies are actually behaving versus what they're claiming publicly.

The core insight: humans don't need specialized training phases rehearsing every piece of software they might use. We learn on the job through self-directed experience and semantic feedback. Current AI models can't do this, which is why there's "an entire supply chain of companies building RL environments" teaching models company-specific workflows that real employees just... figure out.

The Economic Value Problem Nobody's Addressing

Patel frames this through a telling anecdote: a biologist mentions her work involves examining slides to identify macrophages from similar-looking dots. An AI researcher responds, "Image classification is a textbook deep learning problem—we could easily train for that."

But that's exactly the problem. Human workers are valuable because we don't need custom training pipelines for every micro-task. Building specialized RL environments for lab-specific slide preparation methods, then another for the next context-specific task, isn't economically viable at scale.

Every job involves hundreds of judgment calls requiring situational awareness and context learned on the fly. These tasks differ not just between people, but day-to-day for the same person. You can't automate actual jobs by pre-baking predefined skill sets—you need models that generalize and learn from experience the way humans do.

The fact that AI companies are investing billions in mid-training to bake in consultant skills, web navigation, and financial modeling suggests they know their models struggle with on-the-job learning and generalization. If they expected robust continual learning capabilities soon, they wouldn't need this elaborate scaffolding.

Why Diffusion Lag Is Cope for Missing Capabilities

Some defenders argue AI isn't widely deployed yet because technology takes time to diffuse through organizations. Patel calls this "cope" masking fundamental capability gaps.

His counterpoint, via Steven Byrnes: how do skilled immigrant workers integrate into economies immediately? Once you answer that question, note that AGI would do those things too—except faster, since it could read entire company Slacks and Google Drives in minutes and instantly absorb skills from other AI instances.

If current models were truly "humans on a server," they'd diffuse incredibly quickly. Hiring is a lemons market where identifying good candidates is hard and bad hires are costly. AI deployment wouldn't face these friction points—you'd just spin up vetted instances.

The reason labs make millions in revenue instead of trillions isn't diffusion lag—it's that models remain orders of magnitude less capable than human knowledge workers for real jobs requiring contextual judgment and adaptive learning.

The Reinforcement Learning Scaling Contradiction

Pre-training had clean, predictable scaling laws across multiple orders of magnitude. People are trying to "launder the prestige of pretraining scaling" to justify bullish projections about reinforcement learning from verifiable outcomes (RLVR), despite having no well-fit public trends showing similar predictability.

When researchers analyze available data, results are bearish. Toby Ord's analysis connecting o-series benchmark charts suggested needing roughly 1,000,000x scale-up of total RL compute to match a single GPT generation improvement.

The labs' actions reveal their worldview: if they believed models would soon develop robust generalization, they wouldn't pre-bake consultant PowerPoint skills to "automate Ilya." Building in economically valuable capabilities suggests expecting continued poor performance on on-the-job learning and generalization.

One counterargument: maybe automated AI researchers will solve robust learning algorithms that humans couldn't crack for a century, despite lacking basic learning capabilities children possess. Patel finds this "super implausible"—it's the classic "losing money on every sale but making it up in volume" fallacy.

When Goalpost Shifting Is Actually Rational

AI bulls criticize bears for moving goalposts. Patel argues some shifting is justified: we keep solving supposed AGI bottlenecks (general understanding, few-shot learning, reasoning), yet models still can't automate 95% of knowledge work.

The rational response? Recognize that intelligence and labor involve more complexity than previously understood. Definitions that seemed sufficient in 2020 were too narrow, revealed by the gap between technical achievements and economic value generation.

He expects this to continue: by 2030, labs will solve continual learning challenges, models will generate hundreds of billions in revenue, but still won't automate all knowledge work. "We've made progress, but we're not at AGI yet. We also need X, Y, and Z."

Models keep getting more impressive at the rate short-timeline people predict, but more useful at the rate long-timeline people predict.

Why The Explosion Comes Later (And Harder)

The variance in human value-add is massive—village idiots contribute nothing to knowledge work while top researchers are worth billions. AI models at any snapshot are roughly equally capable across instances. Because disproportionate value comes from top percentile humans, comparing AI to median humans systematically overestimates value generation.

But when models finally match top human performance, impact could be explosive precisely because capabilities distribute uniformly across instances rather than following human variance.

The real intelligence explosion won't come from recursive self-improvement—it'll come from continual learning at scale. Billions of agent instances deploying across jobs, learning from experience, bringing knowledge back to hive-mind models that batch-distill learnings and redistribute capabilities.

This won't be singular breakthrough but gradual progression, like in-context learning. Labs will release "continual learning" next year that counts as progress but falls short of human-level capability. Actual human-level continual learning might take 5-10 more years.

No runaway advantages for first movers because competition stays fierce—talent poaching, reverse engineering, and information diffusion neutralize supposed flywheels. The big three labs keep rotating podium positions monthly, with competitors close behind.

The Uncomfortable Timeline

Short term (next few years): moderate bearishness because current approaches require unsustainable amounts of specialized training for context-specific tasks. Economic value generation remains constrained by fundamental capability gaps in generalization and on-the-job learning.

Long term (next decade or two): explosive bullishness because once robust continual learning gets solved, billions of human-like intelligences on servers that copy and merge learnings represent genuinely transformative capability.

People underestimate how big actual AGI will be because they're extrapolating current regimes—more pre-baked skills, better benchmarks, incremental improvements. They're not imagining the discontinuous jump when models can actually learn like humans do.

We're in the awkward middle period where technical achievements impress but economic value lags, revealing that the hard problems aren't the ones we thought they were.

If you need help evaluating AI capabilities versus marketing claims or building deployment strategies around what models can actually do today versus aspirational roadmaps, Winsome Marketing separates signal from hype.