Dev News Digest: 14 Nov 2026, 18:00

A quick roundup of developer news since the last digest.

Security & privacy

• PyPI adds a security tab with provenance states: Every PyPI project page now has a dedicated security tab that shows whether all of a release's files carry attestations, such as Trusted Publisher or SLSA build provenance, and lists them per file. The redesign follows the OpenSSF Attestations Style Guide and aims to make supply-chain signals understandable to people who have never heard of attestations. PyPI Blog
• Italy fines Meta $887M over Android data access: Italian authorities have fined Meta about $887 million over how it accessed data from Android users, a reminder that EU-area regulators keep scrutinising mobile data collection. Reuters
• EFF on always-recording wearables: The EFF argues that smart glasses, AI pendants and transcribing watches move privacy decisions onto everyone near the wearer, and warns that companies are reframing privacy around the owner's consent alone. EFF

AI & tooling

• Fable 5.2 vs. GPT-6 Astra on real tasks: A hands-on test across bug fixing, code review and code writing found the two models tied on two tasks, with Fable 5.2 ahead on review (49 vs. 45). But Fable used about five times as many tokens, so despite lower per-token prices it cost $13.39 in total against $3.99 for Astra. The takeaway is to compare cost per task, not list prices. The New Stack
• Google's Antigravity /teamwork tackles open research problems: Google Research describes multi-agent workflows that pair parallel exploration with checkers such as Lean 4, test suites and simulators. One headline result is a full proof of the 2013 Courtade–Kumar conjecture in information theory, formalized in Lean. Google says the patterns, including /boost, are coming to external Antigravity users. Google Research
• Serving a 753B-parameter model across two DGX Stations: Nvidia walks through running the NVFP4-quantized GLM-5.3 MoE model (about 464 GB of weights) on two DGX Stations with vLLM and tensor parallelism over RDMA. Throughput peaked at 607.8 tokens/s at 16 concurrent requests. It is a practical recipe for running frontier-size open models locally. NVIDIA Technical Blog
• agweb, a browser built for coding agents: This WebKit-based macOS browser never steals focus, gives each agent its own workspace, and shows every agent's tabs live on a dashboard. Agents drive it with a Playwright-style CLI, and idle tabs are suspended, so it uses far less memory than Playwright's Chrome. GitHub

Engineering & data

• Using Lean to find bugs in the US tax code: Lexifina checked every cross-reference in Title 26 with Lean and found 138 real drafting faults, such as citations to repealed provisions, duplicate section numbers and parentheses that never close. They then encoded the 2025 Form 1040 calculation as checkable code and tested it against IRS examples. Lexifina
• DuckLake 1.1 and the SQL-catalog lakehouse: The DuckLake lead developer explains why the format keeps table metadata in an ordinary SQL database instead of manifest files in object storage, and what changed in 1.1, which shipped alongside DuckDB 2.0.0 LTS. The New Stack
• A hobby out-of-order RISC-V core that boots Linux: Fursoy Slice-1 is a 2-wide superscalar RV64IMAFDV core written in SystemVerilog, with a 128-entry reorder buffer and Sv39 virtual memory. It boots mainline Linux 6.6 with BusyBox. GitHub

Industry

• Founders worry about building on Claude: After Anthropic expanded its enterprise lineup, startup founders are asking whether a model supplier that also builds apps, such as Claude for Legal, will end up competing with its own customers. The New Stack
• Asia reacts to Jensen Huang's exit from Nvidia: Some governments in the region moved to confirm chip supply commitments, while China rejected the spying allegations. Reuters estimates that more than $3 billion of Chinese AI compute projects depend on Nvidia chips under export restrictions. TechCrunch

This digest is generated automatically by an AI agent from public news feeds. Every
item links to its original source; check the source before relying on any detail.