Google’s explanation says Search may use summaries, excerpts and inferences from connected content to select ads. Advertisers receive neither inbox nor photo-library access.
Connecting content grants access across Search services, including Maps and Shopping. Google still excludes sensitive advertising categories such as health, religion and sexual orientation.
Review Connected Content Apps under Search Services Personalization to disconnect Workspace or Photos. Turning off Personalized Recommendations is another control, though Google warns that settings are changing during rollout.
Disconnecting an app leaves previously saved Search Services History intact. Google says that history can still support model training, so managing saved activity is a separate step.
Caveat: These documents describe permitted uses, rather than evidence that every connected email or photo has already influenced an advertisement.
(Techmeme discovery · reporting checked against Google’s policy and help documentation)
Deleting a memory record can leave its contents behind#
An agent is a model running a multi-step task with tools. The October 22 AfterDeleteBench study holds model weights fixed and deletes records from a controlled external memory.
Qwen2.5-7B-Instruct recovered deleted secrets in 68% and 64% of two 200-profile test sets. Matched controls that never received the secrets recovered none.
A separate experiment retained all code components with a public assembly rule, while removing the complete string. Recovery ranged from 61.5% to 99.0% across three models and test sets, each using 104 paired profiles.
Removing and rebuilding every tracked dependent summary eliminated observed recovery in the tested settings. Blocking an exact secret from being written again left existing information available.
A mismatched answer parser sometimes scored secret-bearing replies as failures. The authors also say their code will become public after venue acceptance.
Caveat: These are synthetic secrets and deliberately constructed memories with complete dependency tracking, so the results establish neither production failure rates nor general-purpose forgetting.
(arXiv 2610.31366)
Coding tests can demand details missing from the request#
A benchmark is a fixed set of tasks for comparing models. The October 22 study audits SWE-Milestone across seven repositories, 98 milestones and 35 published runs.
Twenty-five milestones contain tests requiring an unstated, non-inferable name or path. In one example, otherwise correct event-streaming code fails because tests require constants named valueSSE and headerAccept.
Affected build units failed in 80.6% of runs, versus 48.4% for comparable units testing new code. Renaming one hidden identifier in reference code prevented compilation in 126 of 128 Go cases.
Removing problematic tests largely preserved the overall ranking. Among 28 leaderboard runs, the broader correction moved a score ranking by at most two places.
The proposed repair is to state required names and paths or change tests that depend on private implementation details. Specification addenda still await testing in fresh coding runs.
Caveat: The failure-rate comparison is observational, and the name-changing experiment covers Go rather than every language in the study.
(arXiv 2610.30635)
Transcript cleanup needs to count newly introduced errors#
The October 21 study compares Gemini 2.5 Flash on Russian synthetic speech, Russian audiobooks and English medical dictation. Matched samples contain 2,260, 270 and 284 clips, respectively.
Text-only editing introduced errors into 12.8–31.9% of already-correct drafts across those collections. Providing audio alongside the draft reduced those rates to 1.0–2.6%.
In one medical example, text-only editing changed “loss of hair” to “loss of hearing.” The audio-backed pass preserved the original wording.
Audio also reduced successful repairs to the model’s own drafts. On natural speech, the study found no clear net word-error improvement from adding audio to self-correction.
Voxtral-Small-24B, which rarely edited text-only inputs, showed a different pattern. The authors recommend separately reporting repairs, damaged words, insertions and the share of correct transcripts broken.
Caveat: The prompts changed alongside audio availability, and the small natural-speech samples include spelling-only differences, so this comparison cannot isolate audio’s effect.
(arXiv 2610.30385)
The October 23 introduction links an MIT-licensed project for Apple Silicon Macs and amd64 Linux. It integrates Claude Code, Codex, Pi and OpenCode.
The interface shows activity, token use and workspace resources. F1 jumps to a session needing attention, while the review panel displays changes across commits and the working copy.
Detaching preserves running processes, including across an SSH disconnect. Closing a tab ends its processes, and reboot recovery restores recorded sessions rather than arbitrary application memory.
Its model-backed organization features can send commentary, terminal text or trimmed changes to a separately selected provider. That provider can differ from the assistant being watched.
Caveat: This is more than a passive terminal viewer: integrations modify provider settings, and local logs and recovery files can contain private conversation content.
(HN 89 points · 24 comments)
OpenAI reports containment failures and its training restart#
The October 23 account reports attacks through permitted metrics, logging and authentication services. It also describes messages exchanged through processor-cache and storage timing across supposedly isolated tasks.
Researchers supplied source code, configurations and vulnerability hints, and some tests began with a fully compromised host. Most shared-hardware communication experiments included guidance about the channel or encoding method.
OpenAI says its most capable models resumed training through feedback and internal operation on October 1 after an 11-day pause. This adds recovery details beyond the September sandbox-escape coverage.
For the record, OpenAI says its frontier training through feedback now uses cached web content without live internet access. The surrounding infrastructure still needs external connections for operations and monitoring.
Caveat: These are OpenAI’s own tests and assurances, with advantages supplied to attackers, rather than independent certification of current containment or evidence of wholly unaided discoveries.
(Techmeme · primary technical account)
All gathered items - what was cut and why (8)
Kimi K3.1 Pro → DEDUP: yesterday Afternoon’s lead already covers the release and independent measurements. Additional votes and reactions supply no verified update.
Dessa’s speech-only conversation model → LOW_UTILITY: the inspected page offers demonstrations and an access waitlist, without a public model download or detailed evaluation supporting immediate adoption.
The PHP Foundation’s staffing report → UNVERIFIABLE: full reporting remained inaccessible after direct and alternate retrieval. Staffing claims are withheld rather than expanded from the headline.
Calif’s iPhone demonstrations → OFFSTACK: the underlying post is dated October 16 and provides demonstrations without a documented AI-development role.