Friday, 6 November 2026

GitHub rebuilds Git storage for the agentic era; Claude Code leaks user emails; Inkling 2 Small ships Apache-licensed weights

Today's Lead

Engineering

GitHub Blog

Inside GitHub's new Git storage architecture

GitHub expects artificial intelligence agents to push changes to the same repositories far more often than people do, and it set a goal of at least 100 times more scale. The article reports that one repository's daily requests grew about fourfold from March to May 2026, and GitHub began a new storage design project, called Project Rubicon, in May 2026. The new design separates storage from compute. Disposable virtual replicas handle reads and writes, Azure Blob Storage holds the data, and a replicated log called Reference Compare-and-Swap stores only reference pointers. Writers upload pack files before they update references, and background compaction nodes merge packs and remove unused objects so the pack count stays roughly flat. The article reports that internal testing supported 5,000 concurrent agents on one repository at 7,200 pushes per second, but these results come from internal tests rather than production. GitHub made the architecture available in limited private preview in November 2026, with a broader rollout planned for early next year.

Engineering

GitHub Issues

Claude Code sends your email address to websites

The issue, titled "Bug Report: Privacy violation in User-Agent header," reports that Claude Code placed a user's email address in the User-Agent request header during web searches. The reporter said this led to dozens of scam emails per day and argued that the behavior should be removed entirely rather than offered as an optional setting. Commenter morsik found 53 tool calls across eight sessions that exposed an email in the User-Agent, most made by subagents running curl commands rather than the WebFetch tool, which shows the problem reaches beyond one tool. Other commenters noted that email can also leak through sources such as git configuration, and that some websites, such as Wikimedia, ask clients to supply a contact email. Anthropic contributor poteat responded that version 2.1.317 added a built-in privacy plugin, enabled by default and turned off through the plugins menu, which blocks the behavior. The issue remains open with no code fix merged, and the reporter's request for full removal has not been addressed.

Thinking Machines

Introducing Inkling 2 Small

Thinking Machines Lab published a post on November 5, 2026, to announce the second generation of its Inkling models. The model uses a Mixture-of-Experts design with 594 billion total parameters, of which 17 billion are active, and it accepts image inputs. The article says the model far surpasses the company's previous models on the Artificial Analysis Intelligence Index, where it scored 31. The model is first available through Tinker with closed weights, and the company plans to release the full weights under the Apache 2.0 license next week, after external safety testing and red-teaming. The article reports strong results on coding and knowledge work tasks, but it also notes limitations, such as being less optimized for chat, not supporting audio, and trailing closed-weight models from other companies on alignment evaluations.

Cloudflare Blog

Agents aren't humans (but they kinda are)

The article argues that artificial intelligence agents now act on the web for millions of people, but websites see only the agent and not the user it serves. This gap breaks the tools that websites use for identity, trust, and accountability, and the website usually bears the cost when something goes wrong. The article proposes the Intermediation Principle, which holds that the business relationship and security duties between users and websites should stay the same even when an agent sits between them. It rejects two approaches: human verification puzzles, which prove only that a person was present at one moment, and narrow access scopes, which do not show who is responsible for errors. Instead, it describes signed agent identity through Web Bot Auth, graduated disclosure levels from no identifier to a unique one, and a machine-readable site access policy, and it says the Trusted Agent Delegation coalition will bring defined problems to the Internet Engineering Task Force. The article states that Cloudflare's current products already give websites per-agent visibility and controls.

Dropbox Tech

How two engineer side projects became the agent taking action inside Dropbox

Two Dropbox engineers, Daniel Reiter Horn and Rich Chan, describe how two personal experiments became the agent in Dropbox's new experience. Horn built a cityscape model in an isolated sandbox that connected to his Dropbox files, and he refined it through a command-line interface until it was ready to print. Chan built a graphical interface inside the Dropbox desktop app so that an agent could turn his weekly receipt images into expense reports. The article explains that the combined product gives the agent a controlled workspace, where users approve folder access and can review a plan before any file changes. Each conversation runs in its own sandbox, and credentials stay outside the workspace, so the agent cannot grant itself access. Dropbox now records which conversation produced each file version, so users can undo a single agent turn.

Social Sciences

Yahoo Finance

Cuban: Facebook knowingly runs AI deepfake scam ads; docs say $16B fraud revenue

Mark Cuban says Facebook knowingly runs scam advertisements that use artificial intelligence deepfakes of his face and voice. In an August episode of the Pivot podcast, he said he regularly receives emails about a diabetes cure that he never sold. Internal Meta documents reported by Reuters in November 2025 estimated that about $16 billion, roughly 10 percent of 2024 revenue, came from scam and prohibited-product ads. Meta spokesman Andy Stone said the documents give a selective view and that the 10 percent estimate is rough and overly inclusive. He added that user complaints about scam ads fell 58 percent over 18 months and that Meta removed more than 134 million scam ads in 2025.

Marginal Revolution

The Global Office

The article summarizes a working paper by economist Richard Baldwin published by the National Bureau of Economic Research. It argues that the frontier of globalization has moved from manufacturing supply chains, which it calls the Global Factory, to services supply chains, which it calls the Global Office. The Global Factory remains larger in total trade, but the Global Office is growing much faster. The paper found that between 2008 and 2022, all net job creation linked to exports occurred in digitally delivered services rather than manufacturing. Gravity estimates show that each hour of time-zone separation reduces service exports by roughly 4 percent, while distance matters less for them. The post conjectures that the future of trade is services, but notes the paper raises more questions than it answers.

Marginal Revolution

Love at first photo?

The post shares the abstract of a job market paper by Sofia Shchukina, a University of Chicago job market candidate. The paper asks whether the design of dating apps makes attention more concentrated on a small number of profiles. Shchukina built an experimental dating app that resembles conventional swipe-based dating apps and let participants form real matches. Across three field experiments, over 1,000 participants made about 70,000 matching decisions. Under conventional profile design, photos and height dominated matching decisions, but making visuals less prominent alone did not significantly change which profiles received likes. Requiring users to review other profile details spread likes more evenly without reducing matches, while capping the number of likes users could send increased concentration.

Humanities

Aeon

Celestial nuisance

The essay, written by Dylan Jow, a fellow at the Kavli Institute for Particle Astrophysics and Cosmology at Stanford University, argues that astrophysics is organised around "Big Questions" about fundamental physics, such as dark matter, dark energy, and the early universe. Under this framing, messy astrophysical processes are treated as nuisances to be removed from the data. Interstellar dust, for example, is treated as a nuisance by cosmologists, even though it plays a major role in star and planet formation. Astronomy became a sub-discipline of physics in the late twentieth century, and after the United States cancelled the Superconducting Super Collider in 1993, physicists turned to the cosmos as a substitute laboratory. Jow concludes that the choice of which questions count as fundamental is arbitrary, and that astronomers can reclaim the value of the nuisance by studying the processes that the Big Questions ignore.

JSTOR Daily

The Art and Science of Disney's Fantasia

The article argues that the 1940 film is an early example of "imagineering," a term the Aluminum Company of America coined in 1942 as a wartime slogan. The article explains that Disney later linked the term to Disneyland and made it a central management idea. Each of the film's eight segments pairs a distinct animation style with a classical piece, such as Bach, Stravinsky, and Mussorgsky. The studio's research labs studied natural phenomena, such as bubbling liquid, to animate the Rite of Spring segment, which promotes Darwinian evolution and drew backlash from some conservative Christian groups. The article also notes that the Pastoral Symphony segment included racist caricatures of Black servant girls, which were removed in the 1969 rerelease. The article draws mainly on Mark Clague's 2004 essay in American Music.