In this video, I’m breaking down something a lot of people don’t want to say out loud: AI might have hit a wall—at least in terms of meaningful productivity gains inside real companies. Yes, LLMs are still improving, but the “wow” jumps we felt from GPT-3 to GPT-4 aren’t showing up the same way in newer releases. Benchmarks like MMLU are getting saturated (when you’re already near human-level on a test, your gains are going to look tiny), and that’s why researchers are building new evaluations like “humanity’s last exam.” But the bigger point isn’t model scores—it’s the AI productivity paradox. I walk through a Faros AI study where engineers completed 21% more tasks and merged 98% more PRs… while review and deploy friction increased by 91%, and code quality dropped (about 9%). That’s the “AI productivity ceiling” people are talking about: output goes up, but company productivity doesn’t necessarily improve, because the bottleneck just moves to code review, testing, deployment, or even client feedback loops. And on top of that, there’s the human side—Harvard research points out that AI can quietly increase work intensity over time, which leads to fatigue and burnout. My takeaway: the ceiling isn’t the tech. It’s whether our workflows, expectations, and organizations evolve fast enough to actually benefit from it.