Altman’s 2026 AGI timeline is marketing, not engineering. The real signal here is self-training loops—models optimizing their own loss functions without human ground truth. That’s where value compounds or collapses. If the reward hacking isn’t solved, we get brittle intelligence, not AGI. The ROI question isn’t “when,” but “can we trust the output at scale?” I’d bet on alignment bottlenecks over compute curves. Are we over-indexing on capability milestones while ignoring the data-quality cliff? Full video link https://youtu.be/KL9_1GbmCic