Today's Hallucination HQ
Thirteen Billion Reasons to Check Your Community GuidelinesHugging Face — the AI platform beloved by researchers for hosting open-source models — is reportedly fielding acquisition offers valuing it at around $13 billion. That's a modest sum for what is, essentially, a very popular library card for AI. The catch: founders feel a genuine responsibility to the open-source community that built them, which makes selling to a tech giant feel rather like a farmers' market going private equity. A deal may never happen. The community is watching.
Source: TechCrunch
The Only Thing AI Still Can't Beat Is a ToddlerFor roughly 100,000 years, the only known creature capable of learning human language to native fluency was a human child — typically one who also couldn't operate a spoon. That monopoly now has a challenger. Since ChatGPT's release four years ago, large language models have become formidably fluent, yet researchers still can't fully explain how children acquire language so efficiently — and why machines, despite vastly more data, still don't do it quite the same way. Evolution, it turns out, had a considerable head start.
Source: MIT Technology Review
OpenAI Wants an Agent for Everything. Your Mum Wants to Check Her Email.OpenAI is aggressively building AI agents — autonomous systems that complete multi-step tasks without hand-holding — and now wants to push them beyond software engineers to ordinary people. Noble ambition. The challenge is that "the masses" generally want technology that works quietly and asks no follow-up questions, whereas AI agents currently require a certain tolerance for occasional confident wrongness. OpenAI is betting that usability will catch up with capability. History suggests this bet sometimes takes longer than the press release implies.
Source: TechCrunch
AliExpress Was Listening. Technically.Researchers have caught AliExpress using audio fingerprinting — a technique that plays inaudible sounds through your browser and analyses how your device reproduces them — to track visitors without cookies. It's the digital equivalent of identifying someone by how they hum. The method is considered largely outdated by the security community, which somehow makes it worse: you were surveilled by a technique that nobody respectable bothers with anymore. AliExpress has not offered a particularly satisfying explanation. Shoppers remain, let's say, uncharmed.
Source: Ars Technica
AI Leaderboards Are Mostly Measuring the Test, Not the ModelA new paper argues that popular AI benchmarks — the league tables used to declare which model is "best" — are dangerously sensitive to trivial details: the order of multiple-choice answers, the exact wording of a prompt, whether you measure the model's generated text or its internal confidence scores. Shuffle the options, and the rankings shuffle too. This means the AI arms race may, in part, be a formatting competition. The authors describe these leaderboards as "config-fragile," which is academic for "please stop treating these as gospel."
Source: ArXiv AI
Until tomorrow — when AI will be smarter, the benchmarks will be contested, and a child somewhere will have learned three new words without reading a single paper about it.
**
|