AI's Academic Illusion — Is the Edifice Collapsing?

*A source-based critical examination of the "Equivalent Interaction" theory — measuring the author's edifice with the author's own ruler.*

Key points

The checked manifesto makes three claims:

• Claim 1: Deep learning's core "scientific problems" have disappeared; what remain (generalization, performance, stability) are mere engineering shortcomings.
• Claim 2: AI's academic paradigm has one last fig leaf — papers still written by humans — which will fall within two years.
• Claim 3: The genuine scientific question is the gap between human one-shot learning vs. machine big-data fitting, anchored by the author's "equivalent interaction" (交互) theory, broken into three sub-problems.

The post audits these claims via four channels of evidence: original papers, the author's own statements, official pages, and opposing literature, under three commitments: primary sourcing, cold falsification, and fair credit where achievements are real.

Chapter 1 — Where is the foundation?

The "three criteria for a real scientific problem" (distill underlying propositions; rigorous mathematical definition; long-term value) are judged sound, especially the second criterion.

The post recalls the 2017 NIPS Test-of-Time debate:

> "Machine learning has become alchemy... I would like to live in a world built on solid, rigorous, verifiable knowledge — not alchemy." > — Ali Rahimi, NIPS 2017

> "In the history of technology, engineering artifacts have nearly always preceded theoretical understanding: the lens and telescope preceded optics theory; the steam engine preceded thermodynamics; the airplane preceded aerodynamics... We may well not get a 'simple' theory specific to neural networks, any more than Navier–Stokes or the three-body problem have analytic solutions." > — Yann LeCun, December 2017

The 2026 manifesto's diagnosis largely restates Rahimi's 2017 charge — showing the disease is real but the cure has not been found. A third reading is proposed: scientific questions are shadows cast by theoretical tools; deep learning's questions may not be exhausted, merely invisible to current tools.

Chapter 2 — What the equivalent interaction theory says

The author's verified trajectory:

• 2009: BS, Peking University
• 2014: PhD, University of Tokyo (advisor Ryosuke Shibasaki — not a student of Song-Chun Zhu)
• 2014–2018: UCLA postdoc; turned to interpretability in 2015 after Zhu's remark, self-describing "about two years to find the direction"
• 2017: first paper in this direction, AAAI 2017, *Growing Interpretable Part Graphs on ConvNets*
• 2018–present: tenured associate professor, Shanghai Jiao Tong University (John Hopcroft Center); Area Chair for NeurIPS 2024/2025, ICML 2026; TMLR Action Editor; ACM China Rising Star Award 2020 (one of only two nationwide that year)

Corrections: the claim of "chose this direction in 2016" is a slight over-statement (no papers until 2017); the Damo Academy Young Fellow award and any award shared with Kaiming He could not be verified and are rejected.

The theory in one line: given a trained network and an input of n units (words or image patches), exhaust all 2ⁿ masking states; the network's confidence over all of them can be precisely fitted by a symbolic logic model containing only ~100–200 AND/OR interaction rules. Example: *"green hand"* is an AND interaction (the effect vanishes if either token is removed); three synonymous "happy" words form an OR interaction.

Three claimed properties:

| Property | Status | Meaning | |---|---|---| | Unlimited fidelity | Proved (Chen et al. 2024) | Symbolic model reproduces output on all 2ⁿ masked samples | | Sparsity | Proved (Li & Zhang 2023; Ren Q. et al. 2023b) | Networks encode only a small number of salient interactions (~100–200) | | Transferability | Observed only (Li & Zhang 2023) | Different networks on the same task encode many shared salient concepts (~70% overlap reported for two structurally different large models), but no theoretical guarantee |

Chapter 3 — The five knives (verification verdicts, abbreviated)

• Sub-problem ① (symbolic mechanism of complex decisions): verified — theorems exist, e.g. ICLR 2024, *Where We Have Arrived in Proving the Emergence of Sparse Interaction Primitives in DNNs*. The hardest of the three.
• Sub-problem ② (mathematically objective cross-model consensus): partially verified — engineering method and empirical evidence exist (Ω_shared = ⋂ᵢ Ω⁽ⁱ⁾), but the FITEE 2025 survey says *observed*, not *proved*.
• The remaining claims (per the original structure) are tested against opposing evidence, with credit given where the theory's rigor genuinely exceeds typical interpretability work.

Chapter 4 — Fairness: a shared weakness

The identified soft spot — transferability observed but unproved — is not unique to this author; the entire explanation/interpretability paradigm shares it. Criticizing one theory without acknowledging this would be unfair.

Chapter 5 — Testing the two prophecies

The manifesto's "within two years the human-author fig leaf falls" deadline has already arrived at "now"; the post measures the prophecy against the actual state of AI-assisted paper writing.

Verdict

Measured by the author's own ruler — *rigorous definitions over storytelling* — the equivalent interaction theory earns genuine credit for proven fidelity and sparsity results, while its transferability claims and the sweeping "edifice is collapsing" rhetoric require calibration. The "disappearance of scientific problems" is better read as a multiple-choice question (dead discipline / engineering-ahead-of-theory / invisible-to-current-tools) than a verdict.

---

*Note: this is a structured English rendering of a long Chinese forum post; details of the "five knives" section are abbreviated. Original claims, citations, and SVG figures appear in the source post on zhichai.net.*