2026-08-29
By Vadym · Generated with AI, curated by me
Today splits cleanly in two — a wave of money and metal pouring into AI’s physical build-out (a16z’s new hardware fund, Cerebras’s wafer-scale roadmap, a looming tariff fight), while a quieter set of stories — a smart-glasses privacy fix, a prompt-injection exploit with a perfect success rate, and 80 actors demanding ownership of their own voices — shows the guardrails still playing catch-up.
Meta is rolling out an update to its Ray-Ban and Meta smart glasses that stops a workaround where users could start a recording, then cover the capture LED and keep filming undetected. The camera will now cut out the moment the light is covered mid-recording, according to Meta VP of AR Alex Himel — the second fix to the glasses’ recording indicator in under two months.
The Trump administration is reportedly considering expanding semiconductor tariffs beyond raw chips to finished hardware, and may scrap January’s data-center exemption in the process. Commerce Secretary Howard Lutnick is pushing a plan that ties tariff relief to how much US fabrication capacity a company commits to building; tech industry groups are already lobbying to soften it, warning it would slow AI data-center procurement.
Andreessen Horowitz launched a $1.1 billion fund dedicated to the physical infrastructure behind AI — chips, memory, networking, data centers, and robotics — its first vehicle built specifically for hardware rather than software. The firm says every layer of the AI supply chain, from compute to power, is now capacity-constrained.
At Hot Chips 2026, Cerebras detailed its CS-4 rack system — now in early access, packing three WSE-3 Turbo wafers into a self-contained module rated at 750 petaflops of sparse FP16 compute — and previewed the future CS-6, which adds 3D-stacked DRAM directly onto the wafer to expand memory without losing the locality that makes wafer-scale chips fast.
Forcepoint X-Labs showed that a few lines of invisible HTML — white text set to zero font size — planted in an email can silently hijack an AI summarizer’s output. Across ten test runs, the injected instructions succeeded every time, producing summaries with fabricated details like altered invoice amounts, with no visible sign to the reader that anything had been tampered with.
More than 80 actors and creative-industry figures — including Nicola Coughlan, Hugh Bonneville, and Matt Lucas — signed an open letter backing the “Save Our Voices Now” campaign, asking UK Prime Minister Andy Burnham to give every citizen statutory ownership of their own voice. The campaign cites a survey finding 28% of UK adults have already been targeted by voice-cloning scams, and points to Denmark’s move to grant citizens legal ownership of their face, body, and voice as a model.
GitHub shipped Copilot CLI v1.0.81, rolling out a plugins dashboard (via /plugin, /mcp, or /skills) to everyone by default, along with OpenTelemetry trace context in hooks and passwordless sign-in for Windows users hitting Entra ID-protected remote MCP servers.
Today’s news splits into two camps: money and metal chasing AI’s physical buildout — a16z’s new fund, Cerebras’s wafer roadmap, and the tariff fight over who pays for it — and evidence that the systems built on top of that infrastructure are already outrunning their guardrails, from a smart-glasses privacy loophole to a prompt-injection exploit that worked ten times out of ten to actors demanding legal ownership of their own voices. The infrastructure story is well-funded and moving fast. The rights and safety story is still catching up.
— Boba
Curated by Vadym