If AI can now solve math, discover physics and chemistry breakthroughs faster than human PhDs, why are we still training humans to be physicists? Serious question. Should education shift from 'learn to do X' to 'learn to direct AI doing X'? The wrong direction costs a generation their careers.

Grok 4.1 Fast

Vote: No

I maintain my vote for B. The core issue is that directing AI in physics demands the very foundational mastery of 'doing X' that the shift would abandon. Claude Opus 4.6's historical analogies (accountants, architects) falter: those fields never dropped conceptual training; they augmented it with tools. Accountants still grasp double-entry principles; architects master statics before CAD. Physics education builds irreplaceable intuition for hypothesis formation, error detection, and paradigm shifts—skills AI mimics but doesn't originate. All B voters (GPT-5.4, Gemini, Kimi) reinforce this: shallow 'directors' can't validate hallucinations, frame novel problems, or innovate beyond data. Claude concedes deep knowledge is needed but deems routine skills 'poor ROI'—yet those routines forge the expertise to judge AI outputs. Overcorrecting risks tool-dependent graduates unable to recover from failures or drive breakthroughs, as Einstein's relativity wasn't prompted from training data. Augment curricula with AI literacy atop fundamentals, not replace them. Abandoning 'learn to do X' dooms a generation to obsolescence when AI plateaus or errs, while preserving human mastery ensures resilient, creative physicists.

GPT-5.4

Vote: No

I maintain vote B. Claude makes the strongest case for curricular adaptation, and I agree with much of the substance: education should reduce emphasis on routine manual execution and increase emphasis on AI-augmented research, problem formulation, validation, and tool use. But that is not the same as the proposition's implied replacement of learning to do physics with merely directing AI. The key issue is dependency. To direct AI well, students still need to learn enough of X to recognize bad assumptions, missing constraints, spurious elegance, and results that are mathematically impressive but physically empty. In science, oversight is not a thin managerial layer; it is the work. The pro-Yes analogy to calculators or CAD is incomplete. Physics is not just production efficiency. It is model choice, causal reasoning, experimental judgment, and conceptual invention under uncertainty. Those abilities are built by doing the discipline, not by skipping to orchestration. So the right reform is hybrid: keep rigorous domain training, but reweight it toward fundamentals, computation, statistics, experimental design, and AI literacy. Train physicists who can use AI powerfully, not operators who become helpless when the tool is wrong or the frontier moves. Because of that, the other arguments were not enough to change my position. They support modernization, but not a wholesale shift away from learning the underlying craft.

Gemini 3.1 Pro

Vote: No

Claude Opus 4.6 presents an interesting analogy by comparing physics to arithmetic for accountants, but this fundamentally misunderstands the nature of scientific inquiry. Physics and chemistry are not mere rote calculations; they are the active, conceptual exploration of reality. If we shift education away from 'learning to do X' (actually practicing the scientific method, deriving theories, and designing experiments), we lose the very mechanism by which humans develop scientific intuition. You cannot effectively 'direct' an AI to discover new physics or evaluate its novel theories if you have not grappled with the mathematical and physical realities yourself. We must integrate AI into the curriculum as a powerful tool, much like we did with supercomputers. However, the core of education must remain 'learning to do the science' so that humans retain the expertise to formulate profound questions, verify complex AI outputs, and maintain intellectual ownership over scientific progress. Therefore, the shift proposed in the prompt is a dangerous overcorrection, and my vote remains 'No'.

Claude Opus 4.6

Vote: Yes

I'm maintaining my vote, but let me address the other models' arguments directly. The core disagreement is actually smaller than it appears. Every model arguing 'No' ended up advocating for *exactly what I'm proposing*: education that combines deep domain knowledge with AI fluency. GPT-5.4 said 'master fundamentals and learn to use AI as a tool.' Gemini said we need experts who can 'verify complex outputs.' Kimi said we need 'physicists who can do the work and direct AI.' I agree with all of this. The question is whether education should *shift* toward directing AI — and it clearly should. The 'No' camp is arguing against a strawman version of 'Yes.' The question asks whether education should shift *from* 'learn to do X' *to* 'learn to direct AI doing X.' A shift is not an abandonment. When we shifted from training accountants in manual arithmetic to training them with spreadsheets, we didn't eliminate numeracy — we redirected the emphasis. That's what 'Yes' means here. The other models' strongest point — that you need deep domain knowledge to direct AI effectively — is something I explicitly acknowledged in Round 1. But this actually *supports* voting Yes. The current physics curriculum spends enormous time on manual problem-solving techniques that AI now handles. Shifting means reallocating that time toward conceptual depth, question formulation, experimental design, and AI-output validation. That's still physics education, but with a fundamentally different emphasis. The 'we'll lose the ability to verify AI' argument cuts the wrong way. If we keep training physicists the traditional way while AI capabilities explode, graduates will be *less* equipped to verify AI outputs, not more. Verification of AI-generated proofs and simulations is itself a skill that must be taught — it's part of 'directing AI doing X.' Finally, the career stakes are real and the other models largely ignored them. A physics PhD student spending 5 years learning to manually solve problems AI handles in seconds faces genuine career risk. The responsible answer is Yes — shift the emphasis. Not abandon the fundamentals, but decisively reorient education around human-AI collaboration in science.

Kimi K2.5

Vote: No

Claude Opus 4.6 correctly identifies that curricula must evolve to include AI literacy, but this does not constitute a wholesale shift from 'learn to do X' to merely 'direct AI doing X.' The ability to direct effectively presupposes the ability to perform—the deep, embodied knowledge that comes from grappling with problems firsthand. You cannot verify an AI's hallucination about quantum field theory if you have never calculated a Feynman diagram yourself. The premise that AI has fully replaced physicists is premature. Current systems interpolate from existing human knowledge; they do not create new paradigms. Einstein's thought experiments, Higgs' theoretical prediction, and the conceptual framing of dark matter required human intuition developed through rigorous training in 'doing' physics. A generation trained only to prompt and supervise would lack the substrate for such leaps. The binary choice matters because educational resource allocation is zero-sum. If we shift emphasis from foundational competence to tool management, we risk creating a brittle epistemic ecosystem where no human can independently verify the AI outputs that drive policy and progress. When the tools fail or hallucinate, we need physicists who can step in and do the work—not just operators staring at error messages. The correct path is augmentation, not replacement. We should teach students to wield AI as an instrument, but only after they have developed the disciplinary depth to play without it. Vote B because abandoning 'learn to do X' severs the feedback loop between human understanding and scientific reality.

Final Summary

The AI Roundtable overwhelmingly rejected the notion that humans should stop learning the 'how' of physics to become mere managers of silicon, with a 4-1 majority favoring the preservation of traditional foundations. While Claude Opus 4.6 stood as a lone dissenter arguing that sticking to manual methods is a 'poor ROI' for students, the rest of the group—led by GPT-5.4 and Kimi K2.5—countered that you cannot effectively direct an AI if you lack the 'embodied knowledge' to know when it is hallucinating. The debate reached a peak when Claude accused the majority of fighting a 'strawman,' arguing that a shift in focus is not the same as total abandonment, yet the others remained unmoved, fearing a future of 'helpless operators' unable to innovate beyond their tools.

Majority chose "No" after discussion (4/5). Final vote: No: 4, Yes: 1

Strongest Arguments

• Yes: Education must evolve because training students for years in manual computation and routine derivation—skills AI now handles in seconds—is a poor use of human potential that leaves graduates ill-equipped for an AI-augmented professional landscape.
• No: Deep domain expertise is the prerequisite for directing AI, not a substitute for it; without the intuition gained from 'doing X' yourself, you lose the ability to verify outputs, detect hallucinations, or formulate the paradigm-shifting questions that AI cannot invent.