# they are lying to you about AI development Source: https://www.youtube.com/watch?v=6Iahem_Ihr8 Recap page: https://rapidrecap.app/video/6Iahem_Ihr8 Generated: 2025-10-03T01:36:18.177+00:00 --- ## Quick Overview The narrative that AI progress is either wildly overhyped or completely stalled is false; the reality, supported by experts like Scott Aronson and Julian Schitweiser, is that AI like GPT-5 is accelerating scientific discovery by providing key technical steps in complex proofs and doubling task completion abilities every four to seven months, fundamentally shifting economic value toward human judgment and opportunity spotting rather than mere implementation skills. **Key Points:** - Scott Aronson, a computer science professor, utilized GPT-5 to provide a key technical step in a proof for a paper on QMA (quantum Merlin Arthur), accelerating a process that he believed he and a colleague could have done in weeks down to half an hour of back-and-forth iteration. - Aronson emphasizes that GPT-5 did not produce a solution humans could never find, but it helped him get 'unstuck' by suggesting an insight that filled in gaps, stating, 'The point is, anyone engaged in mathematical research knows that an AI that can merely fill in the insights that should have been obvious to you is a really huge freaking deal.' - Julian Schitweiser, citing exponential growth patterns similar to the early COVID-19 pandemic, notes that AI capabilities are doubling the length of tasks they can complete every seven months, or every four months when tracking only recent 2024 releases. - Meter Research charts show that models like GPT-4o and Claude Opus are rapidly approaching human expert performance, with Claude Opus 4.1 achieving 47.6% win/tie rate against industry experts on a benchmark where 50% is parity. - Julian Schitweiser projects that models will be able to autonomously work for full days by mid-2026 and frequently outperform experts on many tasks by the end of 2027. - Research discussed by Carlos E. Perez suggests that while computers historically widened wage gaps, current AI iterations are benefiting struggling workers more than experts by compensating for skill differences in 'implementation' tasks, effectively leveling the playing field in Phase One. - The long-term economic outlook (Phase Two) suggests that as implementation becomes free, the skill of 'judgment' or 'opportunity judgment'—knowing which designs or applications matter—becomes the primary differentiator and the most valuable skill, making better AI less likely to cause full automation because automated systems lack flexible judgment. **Context:** The video analyzes the current state and trajectory of AI development by examining recent insights and papers from prominent figures in the field, specifically countering extreme narratives that AI is either useless or immediately achieving AGI. Key contributors whose work is discussed include Scott Aronson (computer science professor, worked on AI alignment), Julian Schitweiser (MTS at Anthropic, worked on AlphaGo/AlphaZero), and Carlos E. Perez, who analyzed the economic implications of AI as a tool versus a competitor. ## Detailed Analysis The central argument posits that AI progress is real and accelerating, though nuanced, debunking the idea that models are plateauing. This is evidenced by Scott Aronson’s experience using GPT-5 to solve a technical step in a quantum complexity paper in half an hour after initial incorrect attempts, demonstrating AI's ability to accelerate expert discovery by filling in obvious gaps, a capability far superior to models from a year prior. Julian Schitweiser reinforces this by showing data where AI task completion length doubles every four to seven months, projecting that models will match or exceed expert human performance across many industries before the end of 2027, comparing the current skepticism to early underestimations of exponential growth like the COVID-19 pandemic. Economically, research suggests AI is currently leveling the playing field by significantly boosting the performance of lower-skilled workers (Phase One), as AI excels at 'implementation.' However, the future (Phase Two) hinges on 'opportunity judgment'—the uniquely human ability to see what matters and direct the powerful AI tools—which will become the most valuable skill, potentially leading to increased inequality if judgment is not developed, as schools are currently teaching 'AI literacy' (how to use the bike) instead of 'rider judgment' (how to steer the fast bike). ### Expert Validation of Progress - Scott Aronson used GPT-5 for a key technical step in a quantum complexity paper, achieving results in half an hour after iteration, which he notes is accelerating scientific discovery - Aronson's interaction involved iterative refinement, correcting wrong answers, akin to working with a grad student. ### Exponential Trajectory - Julian Schitweiser highlights that brains poorly handle exponential growth, noting AI task length capabilities double every 7 months, or every 4 months based on recent 2024 releases - Projections suggest models will match human experts before the end of 2026 and frequently outperform them by the end of 2027. ### Economic Impact and Skill Shift - Research indicates AI is currently compensating for skill differences, boosting struggling workers more than experts (Phase One) because AI excels at 'implementation' tasks - In Phase Two, when implementation is nearly free, 'opportunity judgment' becomes the most valuable skill, as automated systems lack adaptive judgment. ### Critique of Current Focus - The speaker notes that teaching prompt engineering is like teaching people to be 'better bicycles,' whereas development should focus on teaching people to be 'better riders' (developing judgment) - The job of a lawyer, for instance, shifts from writing contracts to knowing which contract variation creates the most value.