You're tuned in to Tech Beat. Here are the stories shaping the conversation today.
Researchers are flagging a fundamental tension inside small language models — a phenomenon called dispersion loss, which quietly undermines the gains made by embedding condensation. The finding matters because the AI industry has been banking on smaller, more efficient models, and this trade-off suggests the path there is harder than it looks.
Shifting to hardware, a new benchmark from Wafer AI shows the GLM five point two model running on AMD's MI three fifty five X chips hitting two thousand six hundred twenty six tokens per second per node — and doing it at more than twice the cost efficiency of Nvidia's Blackwell architecture. That's a significant data point in what is becoming a genuine rivalry for AI infrastructure dominance.
And on a more human note, an obituary worth pausing for. David Potter, the South African-born founder of Psion, died on the twenty-eighth of June at eighty-two. Psion's pocket computers were ahead of their time — handheld, capable, and beloved — and Potter's fingerprints are all over the DNA of the mobile computing world we live in today.
Keep surfing. Tech Beat out.
