Sam Altman Says the GPT-5 Haters Got It All Wrong (www.wired.com)

OpenAI’s August rollout of GPT-5 stumbled publicly — a glitchy livestream that produced bogus charts, angry Reddit threads asking for the prior model, and a chorus of critics (including Gary Marcus) declaring the hype-bubble burst because GPT-5 didn’t deliver “AGI” or superhuman cognition. Sam Altman pushes back: the launch’s optics were poor, but the model marks meaningful progress, especially for specialized scientific and coding tasks. He and Greg Brockman say users are already reporting domain-level boosts (OpenAI claims GPT-5 ranks in the top five of Math Olympiad-style benchmarks), and the company insists the narrative of failure is premature. Technically, OpenAI argues GPT-5’s gains come less from brute-force scaling and more from reinforcement-learning cycles driven by expert human feedback and by sampling the model’s own outputs to generate training data — a pivot from "bigger model = smarter model." That doesn’t mean scaling is abandoned: OpenAI is still investing heavily in massive datacenters (e.g., Abilene, Texas) and expects iterative jumps with GPT-6/7. For the AI/ML community the takeaway is twofold: expect more domain-specialized capabilities that casual users may not immediately appreciate, and watch for hybrid training workflows (human-in-the-loop + model-generated data) to become central. The episode also underlines a broader shift: AGI is being framed as a continuous process, not a single milestone — which recalibrates expectations but keeps high-stakes scaling bets firmly in play.