“LLMs Will Never Replace Human Genius” (Wccftech on Zahavy)
Wccftech’s report of llms-cant-jump, July 2026. T4 — a consumer tech-hardware site summarising a position paper, held here not for what it says about AI but as a documented instance of cluster F’s subject: what happens to an idea in transmission.
It gets the skeleton right. Induction, deduction, abduction; AlphaProof for deduction; Schmidhuber’s compression thesis refuted via General Relativity’s data scarcity; Einstein’s sealed-elevator thought experiment; the closing call for physical multimodal world models. Someone read the paper.
What the transmission did to it
“Never” is the headline’s word, not the paper’s. Zahavy argues current LLMs are structurally incapable of the abductive jump and devotes his final section to proposing what would supply it — action-controllable world models like Genie, latent physics rather than pixels, with the explicit goal of turning the jump “from a mystical insight into a reproducible algorithmic process.” A paper whose thesis is “here is the missing mechanism and here is the architecture that might provide it” became a headline announcing a permanent ceiling.
Three other losses, each small and each load-bearing:
- The concession vanishes. Zahavy grants that an LLM could plausibly execute the deductive phase given Einstein’s premises, and calls the 1913–15 grind a close mirror of neuro-symbolic AI. That concession is what makes the abduction claim precise rather than dismissive.
- The scope limit vanishes. The proposal is explicitly “tailored to the physical sciences.” For mathematics and CS, the paper says the grounding is different. The article generalises to scientific discovery as such.
- “Position paper” vanishes. The article reports a 10-page argued position as a finding — “the research states”, “the DeepMind paper says” — with no signal that it is a proposal rather than a result.
The article also adds a claim of its own: “expanding the AI infrastructure with data centers and compute won’t magically enable these LLMs to develop genius-level intellect.” Consistent with the paper’s spirit; not a sentence in it.
Why keep it
Cluster F tracks how ideas diffuse and get adopted. This is a clean, checkable specimen: primary source available, secondary source available, and the delta is a hedged architectural proposal hardening into an unhedged impossibility claim as it moves outward. The mutation runs in the direction the headline wants, and every dropped qualifier was one that made the argument weaker and more precise.
Useful beside ../psychology-wiki’s dopamine-nation, where the same compression happened inside
one author — a researcher’s own trade book outrunning her literature. Here it happens between authors.
Same mechanism, different distance.