Tuesday, 20 October 2026

A proposed maths journal that only accepts Lean-verified proofs

A guest post on Terence Tao’s blog by Shaoming Guo, writing for a group of colleagues, sketches a new journal with two core values. It would rate how much a paper illuminates its result above being first to prove it. And it would only accept a paper once its proofs are formally verified in Lean. The journal has no name or editorial board yet. The authors are asking for feedback, including completely different visions for the journal.

What’s new (Terence Tao)

Agents that think in latent space and only use text for tools

The proposed “Hybrid Reasoning Agents” (Zou et al., including James Zou) switch between continuous latent reasoning and ordinary text tool calls within a single generation. A small “HybridLink” module and a continuous working memory connect the two. Across 11 benchmarks at three model scales, the authors report about 10% better scores, 35–55% fewer tokens and roughly 2× faster inference.

arXiv 2610.18861

Harness²: improve the scaffolding, not the model

Google Research’s framework holds the model fixed and refines the harness around it: prompts, tools and memory. For each task, a Propose–Probe–Compose loop improves the harness without needing ground-truth answers. A persistent layer then carries the lessons into later tasks. One refinement step improved GLM-5 and Gemini 3.5 Flash by up to 11.5 points on WorkBuddy-Bench.

GitHub: google-research/harness2

Mercury 3: a diffusion LLM at 1,000+ tokens a second

Inception Labs’ new model generates text by diffusion rather than one token at a time. It has a 1M-token context and costs $0.20/$0.70 per million tokens in and out. Inception claims it beats GPT-6 Luna and Claude Haiku 5.5 on a range of benchmarks, but gives no coding-specific numbers. Treat the comparison as vendor-reported.

Inception Labs

Mistral goes after sensor data and particle physics

Timestral is Mistral’s first family of time-series foundation models, for probabilistic multivariate forecasting. It comes in four small sizes, from 5M to 209M parameters, under a modified MIT licence, with weights on Hugging Face. The largest ranks first among zero-shot models on GIFT-Eval, TIME and fev-bench. The smallest sizes are interesting for edge devices.

Separately, CERN and Mistral signed a statement of intent to build a “Large Physics Model”. It would be pretrained on billions of simulated LHC collisions and used for reconstruction, particle ID, anomaly-flagging and detector design.

Mistral · CERN

“I want to feel the creator behind the software, not the factories”

Linear co-founder Karri Saarinen argues that good software carries its makers’ convictions. Linear relies on intuition built from hands-on work and customer conversations, not on process or metrics alone. His worry is products that grow by stacking up features. He sees that getting worse as individuals use AI agents to make isolated changes, each diluting a product’s character.

karrisaarinen.com

Linux 7.3 is out, the biggest release yet

7.3 has 18,765 non-merge commits from 3,021 developers, 807 of them first-timers. New: sched_ext sub-schedulers are complete, BPF gets global per-CPU data, and there is a new fchroot() syscall. The release also adds failfs, a filesystem that rejects every operation, intended as a sandbox for confining processes. Also: ntfs3 alternate data streams and VM working-set tracking.

LWN · failfs docs

Borrow checking, run backwards

Evan Ovadia (Vale) describes a borrow checker for his new language Valen that starts at a function’s last line and works upward. It tracks which variables will still be used and reports an error when a mutation would invalidate one of them. Scanning backwards makes each variable’s last use easy to find, and loops need no look-ahead. He knows of no other checker built this way and is asking for holes in the idea.

verdagon.dev

A “fractional Fermi sea” made by flipping interactions

Hanns-Christoph Nägerl’s group in Innsbruck repeatedly switched 1-D ultracold caesium atoms between strong repulsion and strong attraction. This pumped in energy without destroying the gas’s order. The result is a new state whose particles fill energy levels by a modified exclusion rule. Its Friedel oscillations and correlations don’t match Tomonaga–Luttinger theory (PRL).

ScienceDaily

A second black hole hiding in Mrk 501, betrayed by its jet

Silke Britzen’s MPIfR team went through about 23 years of radio data on the blazar Mrk 501. They found a second, trackable jet, which suggests a companion supermassive black hole on a roughly 121-day orbit. A possible lensing flare in June 2022 fits the same picture. The headline “could merge within ~100 years” depends on masses nobody knows yet. The team itself calls the pair “suggested”, not confirmed (MNRAS, accepted).

ScienceDaily

Hera is a month from DART’s crash site

ESA’s Hera arrives at Didymos and its moonlet Dimorphos in about a month. It will start with baseline mapping from 25–30 km in November and December. Its two CubeSats will then separate: Milani will study dust and composition, and Juventas will radar-sound the interior and may land. It will be the first close look at how NASA’s 2022 impact reshaped a 151 m rubble pile and shortened its orbit by about 32 minutes.

IEEE Spectrum

Nightjars get thrust on the upstroke, like bats and insects

Lund University researchers flew European nightjars in a wind tunnel and measured the wake behind them. The broad wingtips that suit their slow hunting flight cost them lift efficiency at speed. They make up for it by producing thrust on the upstroke, which until now had been seen mainly in insects and bats (Johansson et al., PLOS Biology). It’s a good example of flight as a set of trade-offs rather than one optimum.

Meteored

No, William Gibson didn’t trademark a programmer’s name

Autodesk did try to trademark “cyberspace” for its VR work, and Gibson objected. The often-repeated sequel is that he retaliated by trademarking “Eric Gullichsen”, the name of Autodesk’s VR lead. Gullichsen says that never happened: Timothy Leary made it up as a prank and others repeated it. The two men later met in Barcelona in 1990, where Gibson observed that “Autodesk is not the street.”

Eric Gullichsen

Two small worlds: brutalist dioramas and PalmOS Lemmings

Tiny Brutalism (Jimjum Studios) is a goal-free diorama builder. You place concrete blocks, walk through what you’ve built in first person, and dial in rain, fog and a day/night cycle. It’s a nice reference for mood-through-weather. Separately, Aaron Ardiri’s PalmOS Lemmings clone is now on GitHub, and commenters call it unusually well documented and preserved. It shows how much game a 16-bit handheld could hold.

Steam · GitHub: ardiri/palmos-lemmings

Freestyle Friday: Movahed starts 7/7, Carlsen knocked out early

Sina Movahed won the 16 October Freestyle (Chess960) Friday on 9 points, ahead of Haowen Xue on 8.5. It was his sixth Freestyle Friday win and his first in the 2026 championship cycle. Magnus Carlsen lost in round one, climbed back to 3/4, then lost to Pouya Idani and was eliminated. Meanwhile, round 9 of the US Championship (Caruana–So, with So a point clear) is being played as this goes out.

Chess.com · US Champs R8

Parramatta Lanes Wed–Sat; Tame Impala finishes Wednesday

Parramatta Lanes takes over the Parramatta CBD 21–24 October, from 5pm until 10pm (10:30pm Friday and Saturday). During event hours the L4 doesn’t run between Westmead and Parramatta Square, buses replace it, and George, Phillip and Macquarie Streets close. City Circle trackwork on Friday and Saturday nights adds to the disruption. Tame Impala plays Olympic Park again tonight and tomorrow.

Transport for NSW