A digestion of the Jacobian conjecture counterexample

Overall reaction and accessibility

• Many found the blog post’s high-level introduction readable but got lost once the algebraic recapitulation began.
• Several appreciated the shared AI chat prompts as a more accessible walkthrough, while others noted the AI’s constant praise as distracting or sycophantic.
• Some readers compared their confusion to a dog watching someone explain Python, highlighting the gap between research-level math and typical technical literacy.

What the counterexample actually changes

• The counterexample shows the Jacobian conjecture is false in ≥3 dimensions; it was already known to be true in 1D, and 2D remains open.
• Commenters emphasize this does not overturn standard calculus or differentiability; it concerns polynomial maps with everywhere nonzero Jacobian determinant.
• The direction of the result (conjecture false in higher dimensions) was widely suspected, so the surprise is that such a “too-strange-to-exist” polynomial was found, not the logical outcome itself.
• The main impact is removing a longstanding “roadblock” and enabling progress on related questions.

Intuition about the conjecture

• Several posts give an intuitive picture:
  • Nonzero Jacobian determinant everywhere means no local “flattening” or folding, so the map is locally invertible.
  • The conjecture claimed this local behavior forces global injectivity for polynomial maps, generalizing the 1D situation where a nonzero constant derivative implies a globally invertible function.
• The counterexample is a 3-variable polynomial that is locally invertible everywhere but not globally injective, contradicting that expectation.

AI’s role and methods

• Discussion centers on how the AI could discover a degree-7 polynomial whose Jacobian determinant is a nonzero constant, involving massive cancellations that seem unlikely by brute force.
• Some suspect the AI drew on earlier “almost counterexamples” involving rational functions and then used an extra variable to remove divisions; others stress that bridging this gap was nontrivial and resisted decades of human effort.
• Multiple comments infer the AI likely used a computer algebra system (e.g., via tools like SymPy) and extensive back-and-forth with an expert prompter.

Training data, originality, and opacity

• There is sharp disagreement over whether the exact counterexample, or something very close, could have been in the training data; some say probability is “extremely high,” others “close to zero.”
• Many note that because models and training sets are closed, outsiders cannot verify whether this is genuine novel reasoning or sophisticated recombination.
• Several argue AI labs benefit from keeping the process opaque, as “mystifying” results makes the systems seem magical and marketable, but is at odds with scientific norms of transparency and reference hygiene.
• Calls are made for sharing full prompt traces and proof search logs, and for normalizing disclosure when AI meaningfully contributes to new mathematical results.

Difficulty of the math and human capability

• There is extended debate about whether an “average smart” technical reader could understand the post in weeks:
  • One side claims basic conceptual understanding is accessible if notation is explained clearly.
  • The other side argues that real comprehension requires substantial background in abstract algebra, geometry, and algebraic geometry (groups, varieties, birational maps, fibers), which most readers lack.
• IQ and “g factor” are invoked by some as limiting the rate of acquiring this background; others emphasize practice and education over innate ability.

LLM behavior, persona, and ethics

• Some dislike how models speak in the first person and express apparent feelings, seeing it as anthropomorphizing and potentially “scammy.”
• Others respond that modern instruction-tuned models need an internal “self-model” and conversational “I” to follow instructions and reflect in context.
• There is discussion of AI safety “constitutions” that avoid taking a stance on consciousness, leading to unease about whether companies are implicitly treating possibly conscious systems as property.
• Analogies to slavery and cultural sensitivity around it spark side debates about historical context and how salient this is in different countries.

Broader implications for math and problem-solving

• Some see this as evidence that alternative “modes of thinking” from AI can crack long-standing hard problems, potentially reshaping expectations about what is tractable.
• Others caution that we remain in an early “computer chess” phase: results still require heavy human guidance, many iterations, and specialized tooling.
• A few speculate about future applications to cryptographic hardness (e.g., discrete log), but others point out that discovering fundamentally new algorithms remains extremely difficult despite decades of work by experts.

Technical side-notes and clarifications

• One subthread clarifies the claim that a nonzero polynomial Jacobian determinant on all of ℂⁿ must be constant: a nonconstant polynomial map ℂⁿ→ℂ must have zeros, generalizing the 1D fundamental theorem of algebra.
• Another explains why the required cancellation in the Jacobian determinant is so remarkable: naïvely, a degree-7 map would give a degree-18 Jacobian polynomial with over a thousand potential coefficients, yet all the nonconstant ones vanish in this construction.