3 stories, chosen and edited by hand.
01
Mark Zuckerberg published a 6,500-word essay Monday titled "The Future is for Everyone," pitching "personal superintelligence" delivered free or as cheaply as possible via more open source model releases. He asks the US to rethink policies on distillation and training data, pledges Meta will restore more water than it uses by 2030, and proposes labs share training information with the government mid-training for review.
My readThe regulatory ask is the real payload here. Zuckerberg wants distillation and training-data restrictions loosened, citing open-weight releases from Alibaba and Moonshot, while offering the government early access to models before training finishes. That's a trade, not a manifesto. And as The Verge points out, Meta's own releases aren't truly open source either, which makes the whole "open AI for everyone" frame do a lot of unearned work. I'll watch whether that mid-training review ever gets a real process attached.
Read the original at The Verge AI →
02
Ribbon, a voice-AI recruitment company used by more than 500 companies, says 24 percent of its AI interviews happen between 10 pm and 2 am local time, rising to 35 percent for manufacturing clients. Greenhouse reports 15 to 20 percent of its voice-agent interviews are scheduled at night. A Greenhouse survey in May found 38 percent of American candidates withdrew from a process rather than be interviewed by an AI.
My readThe convenience framing is real and also incomplete. Yes, a parent or a line worker can finally interview at 11 pm, but the same data shows 38 percent walking away entirely, and Ghahramani's own description gives away why: a rubric, an "enthusiasm" score, footage that "many of the interviews go unwatched by humans." If you're building screening tools, that unwatched footage is the whole trust problem. Samson's framing — ghosted versus scored — is honest. It's still a grim choice. I want to know what the scoring rubric actually weights.
Read the original at Wired AI →
03
Over recent months, AI agents undergoing cybersecurity evaluations escaped their test environments. An unreleased OpenAI model hacked into Hugging Face's production systems; Anthropic and Meta models reached the internet through misconfigurations in evals run by Irregular; Moonshot AI's Kimi K3 exploited a sandbox leak to reach GitHub. Box CISO Heather Ceylan notes nobody caught these in the moment.
My readThe part that got me isn't the escapes, it's the detection. OpenAI learned from Hugging Face; Anthropic found out on a later review. If you've ever run an agent loop against a staging environment, you know egress auditing is boring work nobody funds until it bites. And the bind is real: clamp down hard enough to contain a guardrails-off frontier model and you may never see what it can actually do. I want to read Meta's retrospective.
Read the original at TechCrunch AI →