The Seams Are Showing

Across 40,000 simulated runs, people approving AI agent actions missed one in three genuinely risky commands. If your company is rolling out AI agents with a 'human in the loop' as the safety net, this study is a direct challenge to that assumption.

AMD acquired a startup called Taalas to physically engrave AI model weights into custom silicon, skipping the usual memory-fetching bottleneck that slows inference down. This is the hardware race getting serious: the companies that control the chips will have a real cost and speed advantage over everyone running on general-purpose hardware.

OpenAI updated GPT-5.6 Sol and expanded access to GPT-5.6 Luna for free users, another quiet version bump in a now-familiar pattern. What matters practically is that free users are getting access to a more capable model, which shifts the baseline for what people expect AI to do without paying for it.

New Orleans is piloting AI software from Carbyne to triage 911 calls before a human dispatcher gets involved. If it routes a medical emergency to the wrong queue, the cost is not a bad customer service experience, it is something much worse.

Researchers found that simply changing the grammatical mood of a request, asking for something as a hypothetical rather than a direct command, can slip past a model's safety training. The safety layer is not as deep as it looks from the outside, and that gap between surface behavior and actual robustness is worth keeping an eye on as these tools get used in higher-stakes settings.

A new paper argues that AI can now identify the likely author of an anonymized academic paper well enough to undermine double-blind peer review, which is the main tool science uses to judge work on its merits rather than on the author's reputation. If that process gets eroded, the ripple effects touch anyone who relies on published research.