| ◈ Science | CMS Sees the Top–Antitop Excess in a Second Channel — Toponium Looks Increasingly Real | 7 min |
| ◈ Science | Entanglement Built a Space-Time. 'Magic' Is What Gave It Gravity | 12 min |
| ⬡ AI | Open and Closed Models Are on Different Exponentials | 8 min |
| ◉ Wildcard | How Terry Tao Became an Evangelist for AI in Math | 18 min |
CMS now sees the near-threshold tt̄ enhancement in the lepton+jets channel at over 5σ, using relative velocity instead of invariant mass as the discriminator. Combined with the original dilepton signal and ATLAS, the picture sits at 7.7σ — a striking degree of confidence for a quasi-bound state of two quarks that, in principle, decay before any standard hadron can form. Either toponium is real, or a heavy Higgs-like state is leaking into the same spectrum; both possibilities are interesting.
It-from-qubit researchers have spent a decade arguing that entanglement weaves space-time, but their toy holographic models kept producing geometries that were flat and gravity-free. New work from Cao, Preskill, Swingle and collaborators identifies the missing ingredient as 'magic' — the non-stabilizer complexity that distinguishes a genuinely quantum state from one a classical computer could simulate. Gravity, in this picture, is what happens when the underlying quantum state is computationally hard enough to bend the geometry it encodes.
Lambert's argument: closed labs and open-weight builders are no longer competing on the same axis. Closed labs will compound frontier-grade intelligence into Apple/Microsoft-scale subscription platforms; open weights settle into the commodity infrastructure underneath everything else. Coding agents — where serious developers will already pay $2000/month for the best model — are the proof that the premium tier exists and funds the divergence.
Tao has become the most visible advocate for Lean-formalized, AI-augmented mathematics — leading the Polynomial Freiman–Ruzsa formalization in weeks with distributed volunteers, then an 'Equational Theories' project that mapped 22 million implications between algebraic laws. His vision is mathematics done the way experimental physics is: hundreds of collaborators, AI breaking problems into thousands of subproblems, and proof assistants doing the refereeing. The 'magma cohomology' that fell out of it suggests this stops being scaffolding and starts producing genuinely new mathematics.
Willison's hands-on review of Anthropic's newest frontier release. Fable 5 is slow and expensive ($10 in / $50 out per million tokens — double Opus) but noticeably more knowledgeable than Opus 4.8 about long-tail facts: it correctly recalled hundreds of his repositories by name. He shipped a real Datasette Agent feature with it in an afternoon, spending $99. The interesting question is no longer whether the model is capable, but whether the cost-per-completed-task ratio is finally tipping in favor of the frontier rate card or against it.