What Happened
Ron Miller published a piece on Fast Forward (March 6, 2026) examining the gap between what Sam Altman and Dario Amodei say publicly and what their models actually deliver. The analysis comes as both companies face increasing government scrutiny over AI applications in defense and surveillance.
The specific claims under the microscope: Altman recently wrote "We are now confident we know how to build AGI as we have traditionally understood it." Amodei predicted up to half of white-collar jobs could disappear within five years.
Miller tested the AGI thesis practically by trying to use multiple AI models to compile event data across fragmented sources. All models failed - unable to accurately aggregate information that a human analyst could handle with enough time. Researchers like Andrew Ng have also pushed back on Altman's AGI timeline, suggesting it remains years away by any rigorous definition.
MIT professor Bryan Reimer weighed in on the defense and surveillance angle, arguing that high-stakes applications require humans to remain "in the loop" given current model limitations.
Why It Matters
If you choose AI tools based on what CEOs say at conferences and in blog posts, you are making decisions on marketing, not performance. The gap Miller identifies is one that daily AI users already feel: these models are impressive within specific boundaries and unreliable outside them.
The white-collar job claim from Amodei is particularly worth examining. If you actually use Claude or ChatGPT for knowledge work, you know the reality: they accelerate parts of your workflow significantly while being completely useless for others. The idea that half of white-collar jobs disappear in five years does not match the experience of anyone using these tools for real work right now.
This matters for tool selection. When OpenAI announces a new capability or Anthropic publishes a benchmark, the question should always be: does this work for my specific use case? Benchmark scores and CEO essays are not substitutes for testing.
Our Take
Both Altman and Amodei have business reasons to overstate capabilities. OpenAI needs to justify its valuation. Anthropic needs to justify its fundraising. This is not conspiracy - it is standard startup dynamics applied to a technology that people are already anxious about.
The practical takeaway: ignore the AGI timeline debates entirely. They do not affect your tool choices today. What matters is whether the current versions of ChatGPT, Claude, and Gemini handle your specific tasks well. Test them. Compare outputs. Switch when something works better.
Miller's failed data aggregation test is a useful example. Multi-source data synthesis across inconsistent formats is genuinely hard for current models. If that is your workflow, no amount of AGI rhetoric changes the fact that you need human judgment in the loop.
The defense and surveillance angle adds a policy dimension, but for most AI tool users, the lesson is simpler: evaluate tools by what they do, not by what their CEOs say they will do next quarter. The gap between marketing and reality has not closed, and bold predictions about AGI and job displacement should not drive your purchasing decisions.