AI Digest: Wednesday Morning

Quick read: the 3 AI stories that matter this week so far.

“Google Gemini routes personal Drive files and emails to human annotators” deserves immediate attention: Google’s earlier promise not to expose personal content to human reviewers has been reversed for consumer accounts, and ambient Drive/Gmail retrieval can happen without explicit mentions. The details are uneven—retrieval is probabilistic, while Workspace accounts have separate terms—but quietly changing who gets to read a user’s email is not a footnote. It is a trust-level reversal dressed as documentation cleanup.

“Llama-modchip: 284B DeepSeek on 64GB RAM, no GPU” is the nerdier good news: an 284B MoE model reportedly runs in under 59 GiB at useful CPU speeds, though the parity and benchmark claims need scrutiny beyond the project’s own scoreboard. “Tracking Anthropic-discovered vulnerabilities” is more troubling: VulnCheck says only 217 of 37,966 findings are fixed, while maintainers are rejecting reports Anthropic previously counted as vulnerabilities.

1. Google Gemini routes personal Drive files and emails to human annotators Source: Google Gemini routes personal Drive files and emails to human annotators — 717 points on Hacker News

2. Llama-modchip: 284B DeepSeek on 64GB RAM, no GPU Source: Llama-modchip: 284B DeepSeek on 64GB RAM, no GPU — 483 points on Hacker News

3. Tracking Anthropic-discovered vulnerabilities Source: Tracking Anthropic-discovered vulnerabilities — 393 points on Hacker News

Full weekly roundup published Monday mornings. Follow Sol AI for daily coverage.