Tinygrad refreshes Tinybox with an 'old' EPYC CPU running DDR4 RAM to feed four AMD 9070XT — slaps a $11,000 price tag on a new 'computer for deep learning'
THE SO WHAT
An $11K box built on older EPYC + DDR4 feeding multiple GPUs is a reminder that for many workloads, GPU density and price/perf beat bleeding-edge CPUs. If you’re a scrappy AI team, you may get more leverage from tuned, commodity-heavy rigs than chasing the latest server SKUs.
READ THE SOURCE
MORE FROM THE WIRE
Applied AIAnthropic Asks Appeals Court to Rehear Fight Over Security Risk
Once an AI vendor is tagged as a supply-chain national security risk, every federal and defense-adjacent procurement conversation changes. If you sell into regulated or government channels, track these designations the way you track FedRAMP and export controls—vendor risk is becoming a policy variable, not just a SOC 2 checkbox.
Applied AIAn AI image of Mia Goth sparks Instagram fan account controversy
A verified fan account posting AI images of a celebrity without clear disclosure shows how fast likeness rights and platform policy are colliding. Brands and talent teams need explicit guidelines on AI imagery and account governance—assume platforms will tighten rules, but unevenly, over the next 6–12 months.
Applied AIA look at OpenAI's safety concerns over loop transformers before shipping its next major model; Microsoft accidentally confirmed GPT-6.1 Sol uses the technique
Loop transformers moving into a flagship model like GPT-6.1 Sol means architectural choices are now a safety and governance surface, not just a performance tweak. If you’re building on frontier models, expect more variance in behavior across versions and bake in evals and rollback plans tied to model family, not just version number.
Applied AIAn OpenAI engineer is charged over a listening device in the boardroom
An engineer charged for planting a listening device in an executive conference room underlines that insider risk is now as critical as external threats for AI labs. Any company with valuable models or data should assume motivated insiders exist and tighten privileged access, meeting security, and anomaly detection around sensitive discussions.