AI Daily Dev — November 4, 2026

Users Say OpenAI Silently Flipped Their Codex Training Opt-Out Back On

HN thread 'OpenAI seems to be silently re-enabling model training on Codex chats' (id 50352757) climbs to ~437 points and 136 comments on Nov 3.
Posters show history entries confirming they opted out — some twice — yet the 'Improve the model for everyone' toggle came back on without notice.
Codex is governed by a separate 'Include environments' switch under Codex Settings; flipping the ChatGPT toggle or the Privacy Portal control does not change it.
OpenAI has not posted an explanation; commenters are pointing people at the Advanced Account Security option that reportedly blocks silent re-enablement.

industry news.ycombinator.com

Nov 1: scaling01 spots H1 text and logo alt reading 'Claude Mythos 5.5 — Our most capable model class for cybersecurity and biology research' on anthropic.com/mythos.
String disappears within hours; the live page reverts to Mythos 5.1 — the current GA Mythos release from Sep 1.
Access would almost certainly stay gated: Mythos 5.1 is restricted to vetted cybersecurity and life-sciences researchers via Anthropic's trusted-access program.
Timing is awkward — Anthropic's S-1 (Nasdaq ticker ANT) is live and a listing is reportedly targeted for the week of Nov 16.

models decrypt.co

OpenAI's Election Operations Center goes live this week with hundreds of employees watching for misuse of ChatGPT and Sora through vote-counting.
Anthropic re-activates Claude.ai voting banners pointing US users to TurboVote via a Democracy Works partnership, plus fresh classifiers for election-abuse detection.
OpenAI opens free model access for cyber defense to voting-system vendors registered with the US Election Assistance Commission and state election authorities.
Context: both labs have poured nine-figure sums into midterm super-PACs backing their preferred AI-regulation posture.

industry techpolicy.press

+701 stars today lifts the Python tool to ~39K total; Trendshift log shows it first topped trending back in February and it keeps resurfacing.
`pip install heretic-llm` runs directional-ablation with an Optuna sweep that co-minimizes refusals and KL divergence from the original model.
Example on google/gemma-3-12b-it: refusals drop from 97/100 to 3/100 at a KL of 0.16; the author reports ~45 min on an RTX 3090 for Llama-3.1-8B.
Community has published 1,000+ Heretic-processed checkpoints on Hugging Face; supports most dense models and several MoE architectures, not yet SSMs.

open-source github.com

Nov 2 update to the Claude Developer Platform: a thread id keeps the conversation on Anthropic's servers so subsequent requests send only the new user message.
Pitch is lower time-to-first-token on multi-turn agents; threads live in encrypted memory, expire 5 min after last use, and are deleted within an hour.
Enable with request header `anthropic-beta: message-threads-2026-08-12`; feature is NOT ZDR-eligible, and a 404 on continuation means the client must resend the full history.
Docs insist your app stay the source of truth — the SDK helper's local mirror only lives as long as the helper object does.

tools platform.claude.com

Oct 30 help-center update: GPT-Rosalind leaves Research Preview and goes global for Enterprise and Business customers through a trusted-access program.
Reaches teams via the model picker in ChatGPT Enterprise and Codex plus the API; admins toggle it on through groups and custom roles — off by default.
Rollout starts with a two-week admin preview; model is gated to 'legitimate biology research use cases' that pass safety and compliance review.
Life Sciences Research plugin in Codex wires it into scientific tools and databases; frames use cases as target discovery, genomics interpretation, and experiment planning.

models openai.com

Oct 28 upload from Jeff Delaney's channel (~4.7M subscribers) frames the current correction narrative in the 100-second format.
Threads the SoftBank Nvidia exit, the 'circular AI financing' critique, and valuation markdowns into the usual Fireship beat-drop pacing.
Hyperagent sponsors — a notable tell on which agent-platforms are willing to spend to reach the dev audience right now.
Still the top AI post on Fireship's channel heading into Nov 4; a bellwether for mainstream-dev sentiment as Anthropic's IPO roadshow begins.

community youtube.com