AI's Academic Illusion: Is the Tower Falling? A Source-Based Fact-Check of the "Equivalent Interaction" Theory

Forum topic · ✨步子哥 · 2026-09-12

Summary

This zhichai.net forum post is a detailed critical audit of a widely circulated Chinese manifesto titled "AI's Academic Illusion, the Tower Is Falling." The manifesto claims three things: that deep learning's core scientific problems have vanished, leaving only engineering shortfalls; that AI papers will soon be written entirely by machines; and that the real scientific question is the gap between human one-shot learning and machine big-data fitting, anchored in the author's "equivalent interaction" theory. The post verifies these claims against primary sources. It traces the theory's author (a Shanghai Jiao Tong University tenured associate professor, PhD from the University of Tokyo, UCLA postdoc under Song-Chun Zhu, ACM China New Star Award 2020), corrects misreported biography details, and explains the theory: a trained network's output confidence across all 2^n masking states can be precisely fitted by a sparse symbolic model of roughly 100-200 AND/OR interaction rules, with proven infinite-fitness and sparsity properties. Transferability, however, is only empirically observed, not proven. The post frames the debate with Ali Rahimi's 2017 NIPS "alchemy" speech and Yann LeCun's rebuttal, and concludes the tower's foundation remains unverified.

AI's Academic Illusion, Is the Tower Falling? A Cold, Source-Based Audit

This post is a fact-check of a viral Chinese manifesto, "AI's Academic Illusion, the Tower Is Falling." The manifesto makes three escalating claims: (1) deep learning's core *scientific* problems are already gone, leaving only surface-level engineering shortfalls (generalization, performance, stability); (2) the last fig leaf of AI academia — papers still written by humans — will fall within two years; (3) the true scientific question is the gap between human one-shot learning vs. machine big-data fitting, anchored in the author's "equivalent interaction" (等效交互) theory, broken into three sub-problems.

The auditor sets three ground rules: use primary sources only; attempt falsification of every claim; and credit genuine achievements while dismantling inflated ones.

Key points

1. Has the scientific question really disappeared? The 2017 precedent

The manifesto's diagnosis echoes Ali Rahimi's NIPS 2017 Test-of-Time speech ("machine learning has become alchemy"). Nine years later the same complaint is restated — meaning the disease is real and the cure has not been found.
Yann LeCun's rebuttal still stands: in tech history, engineering precedes theory (telescopes before optics, steam engines before thermodynamics, aircraft before aerodynamics). LeCun also warned there may be no "simple" analytic theory of neural networks, just as turbulence has none.
The post offers a third reading: scientific questions are not buried treasure waiting to be dug up — they are shadows cast by theoretical tools. Without calculus, "what is instantaneous velocity?" cannot even be asked. So deep learning's scientific problems may not be exhausted, merely invisible to current tools. Three readings result: flee (the field is dead), wait (engineering always precedes theory), or build better tools — the author's choice.

2. What the equivalent-interaction theory actually says

The theory claims: given a trained network and an input split into n units (words or image patches), exhaustively evaluate all 2^n masking states; the network's confidence on all of them can be precisely fitted by a symbolic logic model containing only about 100–200 AND/OR interaction rules.

AND interaction example: in "he is a *green hand*," the two tokens jointly produce a new utility that vanishes if either is removed (1+1 > 2).
OR interaction example: multiple synonyms of "happy" each independently contribute the same effect; it disappears only when all are masked.

3. The author's biography, corrected

The post verifies against primary sources: BS from Peking University (2009); PhD from the University of Tokyo (2014, advisor Ryosuke Shibasaki — not a student of Song-Chun Zhu); UCLA postdoc 2014–2018, redirected in 2015 after Zhu's prompting; first paper in the direction at AAAI 2017 (*Growing Interpretable Part Graphs on ConvNets*); since 2018 a tenured associate professor at Shanghai Jiao Tong University (John Hopcroft Center), area chair for NeurIPS 2024/2025 and ICML 2026, TMLR Action Editor, and a 2020 ACM China New Star Award (one of only two that year). The post flags two unverifiable honors circulating online (a DAMO Academy Young Fellows award and a He Kaiming-related prize) as likely memory confusions, and notes "I chose this direction in 2016" is a slightly early approximation — no paper existed in the area until 2017.

4. The three properties and three sub-problems, checked

| Claim | Status | |---|---| | Infinite fitness — the symbolic model exactly reproduces outputs on all 2^n masked samples | Proven (Chen et al. 2024) | | Sparsity — networks encode few salient interactions (~100–200) | Proven (Li & Zhang 2023; Ren et al. 2023b) | | Transferability — different networks encode many shared salient concepts | Only observed, not proved; key weakness |

Of the manifesto's three sub-problems: ① symbolic mechanism of complex decisions — verified, with theorems (ICLR 2024, *Where We Have Arrived in Proving the Emergence of Sparse Interaction Primitives in DNNs*); ② mathematical convergence of mechanisms across networks — partial: a method exists (shared interactions Ω_shared = ⋂ᵢ Ω⁽ⁱ⁾; ~70% overlap observed between two architecturally different large models), but no theoretical guarantee. The full text (truncated here) continues with further checks, fairness notes about the whole interpretation paradigm, and a final verdict applying the author's own three criteria back to the theory itself.

Takeaway

The theory's mathematical core (fitness, sparsity) is genuinely rigorous, but its grander unifying claims rest partly on empirical observation. The "tower is falling" diagnosis is at minimum a contestable multiple-choice question, not a verdict.