Latest
Prompt Ensembles Only Beat Self-Consistency When Errors Differ
A simulation puts five varied prompts at 21/24 and five samples of one prompt at 15/24 on the same budget. The gain vanishes when errors correlate.
Anthropic Buys $3.4B of Gas Turbines for 1.7GW of Power
Anthropic is buying Caterpillar generators rated up to 1.7GW, breaking from its rent-don't-own pattern as turbine backlogs stretch into the 2030s.
CLAUDE.md Edits Break Prompt Caching: A 10x Worked Example
A new guide puts CLAUDE.md at the front of the cache prefix: edit it mid-session and the same 4,000 tokens cost about 10x more across 40 turns.
Apple's Agentic Siri: What Gurman Says Lands in iOS 27.3
Gurman reports Siri gets fine-grained control of Apple's own apps in iOS 27.3, third-party apps in spring. What's checkable and what Apple hasn't confirmed.
Cursor Voice Chat for Agents: What It Does, Where It Stops
Cursor shipped two-way voice chat with coding agents on 6 November. It runs on Grok Voice, is blocked for API-key sessions, and the daily limit is unpublished.
Ante vs Claude Code: Is a 15MB Offline Agent Enough?
A week-long comparison puts Ante's single Rust binary against Claude Code: far smaller footprint and offline models, but beta bugs and no Windows.
Two Reports Show AI Outrunning the Humans Who Check It
METR says Anthropic's models improve twice as fast as a year ago, mostly because of AI. Days later: 948 Linux exploits in 30 hours for $1.2M.
One-Rule Tactics: I shipped a tactics puzzle where every enemy follows one rule
My first Android game is live on Google Play: a free, offline 5×5 tactics puzzle where enemies always step toward you, so you win by luring them into rocks, pits and each other.
Gemini Omni 1.2 Flash Is GA: A 15-Second Clip Costs $1.04
Google's video model went GA on November 5 with a 30 percent price cut. What a 15-second clip really costs, and how it compares to Veo 3.1 and Runway.
Vercel Tracing Is GA, On by Default for Next.js 16.5
Vercel Tracing hit GA on all plans. New Next.js 16.5 projects trace 100% of requests by default, and Pro retention jumped from 1 day to 14 days.
AI Code Passes Tests but Only 18.4% of It Is Secure
A 145-CVE benchmark found Claude Code passed 74.0% of functional tests but wrote secure code on just 18.4%. Naming the safeguard moved the number.
Parallax beta.8 Shows a Real iPhone — Agents Can't Tap It
Parallax now streams a plugged-in iPhone beside a Galaxy S24, and beta.7 cut agent taps from 14s to under a second on Android hardware. Here is the catch.
Kimi K3.1 Pro Weights Are Open. Can You Actually Run It?
Moonshot opened Kimi K3.1 Pro's weights: 2.8T parameters, a 1.56 TB download, and 8+ B300-class GPUs to self-host. The hosted API is still the real route.
