Claude Max user weighs cancelling over safety classifier misfires

Moshkovitz, who has worked on Khot's Unique Games Conjecture since ~2009, studied OpenAI's writeup with Subhash Khot, Johan Håstad and Dor Minzer. The two-hour talk runs Nov. 4, 3:30-5:30pm US Central, with live Zoom and a later recording.

Downloaded diagnostics now keep the last 4 MB of each output log, up from 25 KB, so exports capture the failure being reported. The same changelog notes Anthropic Cyber Verification Program members can toggle the program on or off from the Devin Local model picker.

Enterprise-listed members can now switch Anthropic's Cyber Verification Program on or off themselves from the Devin Local model picker; turning it off routes their Anthropic requests under the team's usual data retention settings. The same release raises exported diagnostics to the last 4 MB of each output log, up from 25 KB.

Livestream features speakers from NVIDIA, OpenAI, Zed Industries, and AWS on applying Rust's borrow-checker approach for memory safety without a performance penalty. No release or benchmark figures announced.

Talk frames memory as the write side of agents — RAG lets them read, memory lets them write — and centers on "memory promotion": judging what's worth keeping, checking it against existing records, and admitting it.

Open-source Windows image editor built on Photocraft, an open-source Photoshop clone, with local generative AI wired in through ComfyUI. It ships a guided ComfyUI install and guided model downloads, and hides ComfyUI's complexity for users wary of its learning curve.

Ollie shuts down November 6, roughly a month after discussing a $7.5M seed round led by Khosla Ventures. CEO Bill Lennon cited competition from Muse and Instinct and difficulty raising capital; customers will get data-export instructions and refunds.