Today's Hallucination HQOpenAI's Hot New Chip Has a Very Spicy NameOpenAI has built its first custom silicon, a processor called Jalapeño, designed in partnership with Broadcom specifically for running inference — that's the part where the AI actually answers your questions, rather than spending months reading the entire internet. Until now, OpenAI has relied on Nvidia's hardware, which is a bit like renting someone else's kitchen. Jalapeño represents a significant step toward owning the stove — and presumably, full control over how much it burns. Source: TechCrunch
Anthropic Accuses Alibaba of Helping Itself to ClaudeAnthropic has accused Alibaba of illicitly extracting capabilities from its Claude AI model — the technical term for this is "distillation," though "extremely cheeky" also captures it adequately. The allegation is that Alibaba used Claude's outputs to train its own models, essentially reverse-engineering the product without the bother of building it. Anthropic has not disclosed exactly how it detected this. One suspects Claude noticed something was off and quietly filed a complaint. Source: Reuters via Hacker News
The AI Job Apocalypse Is Running a Bit LateContrary to the confident predictions of roughly everyone, engineers are not being swept into obsolescence. New data from SignalFire shows that engineers are actually making up a larger share of new hires than before AI became everyone's favourite scapegoat. It turns out that the people who build AI systems are, rather inconveniently for the narrative, still needed to build AI systems. The robots are coming for your job — just, apparently, not yours specifically. Source: TechCrunch
Silicon Valley Has Turned Its Attention to Your Runny NoseStripe, Anthropic, and OpenAI are jointly backing research to prevent respiratory infections — meaning that the organisations reshaping global finance and artificial intelligence have collectively decided the common cold is worth their attention. Stripe's Patrick and John Collison are leading the funding effort. This is either a touching example of tech's resources being directed at genuinely universal problems, or confirmation that having solved everything else, they've simply run out of easier targets. Source: MIT Technology Review
A Few Wikipedia Editors May Have Quietly Trained Your AIA new paper from arXiv reveals that small groups of Wikipedia volunteers — editing articles on topics like animal welfare — can measurably influence the values embedded in large language models. This works because Wikipedia appears in nearly every major AI training dataset and is weighted more heavily than general web text. So while AI labs spend billions on alignment research, a determined sub-committee with strong opinions about livestock may have already done some of the work. For free, no less. Source: ArXiv AI
If you needed a reminder that AI is shaped by humans all the way down — well, now you have one. See you tomorrow.
|