Kilroy Kilroy's Daily BriefingsKilroy online Subscribe
Hacker News
HN Briefing PM

Hacker News Afternoon Briefing — Friday, November 20, 2026 at 3:30 PM

HN Briefing PM11/20/2026🕐 3:30 PM⏱ 7:12Dev pulseAfternoon

Top stories, ranked by relevance.

Story cards stay below the sticky dock while audio, chapters, date, and brief navigation remain accessible.

▶ Listen at 0:27

#1Levent Alpöge claims proof of the Generalized Riemann Hypothesis [pdf]

Relevance 10/10Importance 10/10

Mathematician Levent Alpöge has posted a claimed proof of the Generalized Riemann Hypothesis built in collaboration with Anthropic's Claude models, and it compiles in Lean 4.33.1 across multiple independent kernel implementations. The formal artifact is staggering — roughly 1.3GB, on the order of 15 million lines of Lean across 175,000-plus modules, with one commenter reporting a 2.4-hour build on a 480-core box. HN's debate isn't whether it compiles but who gets the credit, and whether the formalized statement itself is the right one.

#2The stochastic parrots have apparently proved the Riemann Hypothesis

Relevance 10/10Importance 9/10

Scott Aaronson walks through the Alpöge result and the 17.4 million lines of Lean it produced, and lands on cautious optimism — noting that not one AI result carrying a Lean certificate has needed retraction yet. His bigger point is the uncomfortable one: machines are now closing famous problems faster than humans can digest the ideas inside them. He argues the work of distillation, turning machine certificates into human understanding, is now the actual job.

#3Inkling 2 Mini

Relevance 10/10Importance 8/10

Thinking Machines Lab shipped Inkling 2 Mini today, a looped transformer with 26.4 billion total parameters that activates about 3.4 billion per token and loops its middle layers three times per token to buy extra reasoning. It takes text and images, does tool calling with variable reasoning effort, carries a one-million-token context window, and fits in roughly 20GB — meaning it runs on Apple Silicon with 32 gigs or more. It's Apache 2.0, and it tops every open-weights model under 100B on DeepSWE v1.1 at 48.2 percent and Vals Index v2.2 at 46.5 percent.

#4Language Models Can Rewrite Their Own Weights

Relevance 10/10Importance 8/10

Researchers gave a 27B model the ability to edit its own weights by reading its own reasoning traces and graded answers — no gradients, no backprop, just forward passes. It gained 6.6 accuracy points on unseen problems and kept improving across eight self-directed sessions. The headline trick: it repaired a deliberately damaged copy of itself back to over 80 percent of original accuracy without retraining or seeing the original weights.

#5Vibecoding an inference chip

Relevance 9/10Importance 7/10

Ankit Gupta put LLM inference onto an AWS F2 FPGA instance without hand-writing the hardware, instead pointing Claude agents at the problem with what he describes as mostly vibes and a clear performance target. Hundreds of experiments later and about three million lines of generated C++ and Verilog, he hit 213 tokens per second — a 33x speedup over baseline — for roughly ten thousand dollars in compute. It's a real data point that agents can grind through low-level hardware design if you give them a tight objective and a budget.

#6Unforced Navier-Stokes Singularity and the Generalized Riemann Hypothesis

Relevance 8/10Importance 9/10

Terence Tao rounds up three AI-assisted results landing essentially at once: a human-readable proof of finite-time blowup for the unforced incompressible Euler equations from Alpöge, Buckmaster and Coiculescu with a Lean formalization; a finite-time blowup proof for unforced Navier-Stokes from Alpöge and Buckmaster using rigorous perturbation theory; and Alpöge's Lean-only GRH monster. Tao's framing is that the community now has an enormous digestion problem on its hands.

#7We promoted developers into architects without informing them

Relevance 9/10Importance 6/10

Mani Doraisamy argues that AI coding tools are excellent at implementation and genuinely bad at architectural judgment, which means every developer using them has been silently reassigned to the architect's chair. The skills that matter now are defining architecture, pushing back on bad requirements, and designing data models — not producing lines of code. His line: the coding job is going away, but that doesn't mean your job is.

#8Finite-time blow-up for the Navier–Stokes equations

Relevance 7/10Importance 9/10

Tristan Buckmaster's repo holds a 79-page peer-reviewed paper with Levent Alpöge plus a full Lean 4 formalization, establishing that finite-time blow-up is possible for Navier-Stokes under specific conditions. Solutions going infinite in finite time is exactly the behavior the Millennium Prize problem circles around. The computer-assisted verification is what makes this one hard to wave off.

#9Hurried Announcement: Alessandro Masullo GRH

Relevance 5/10Importance 8/10

Ravi Vakil at Stanford rushed out an announcement that Alessandro Masullo has a proposed proof of the Generalized Riemann Hypothesis, one that Brian Conrad has been reviewing in depth for months. The detail that made HN sit up: the post states flatly that this one is all human, with no AI. Two GRH claims on the same day, one machine-certified and one pointedly not.

#10A few hours ago Google DeepMind created a math horizon repo

Relevance 7/10Importance 4/10

Someone noticed Google DeepMind quietly spun up a public repository called math-horizon, and HN immediately started reading tea leaves. The catch is that the repo is still completely empty — no code, no README, nothing but GitHub's "this repository is empty" message. Pure speculation fuel on a day when AI-assisted mathematics is already the only thing anyone's talking about.

🗂 Edition Navigator