Today's Hallucination HQAnthropic Raises Safety Alarm, Government Takes It Seriously, Anthropic Horrified
Careful what you wish for. Anthropic has spent years loudly warning the world about dangerous AI — so when the government found a "narrow potential jailbreak" in Claude and pulled the model from deployment, Anthropic responded with a blog post essentially arguing that one security flaw shouldn't ground a product used by hundreds of millions. The company that built its entire brand on caution is now making the case that everyone should probably just relax a bit.
Source: TechCrunch
Siri Has Apparently Been Fixed. We're Choosing to Believe This
Fifteen years. That's how long Siri spent confidently mishearing questions, opening the wrong apps, and setting alarms for the wrong timezone with the quiet confidence of someone who has never been wrong. Apple has now released a meaningfully improved version, and early reports suggest it is — and we type this carefully — good. Whether this holds up or represents another optimistic hallucination from Cupertino remains, as ever, a matter of personal faith.
Source: The Verge
6,500 Engineers Walk Into Meta's New AI Division. Nobody's Laughing
Meta's superintelligence unit is barely months old and already generating the kind of workplace atmosphere usually reserved for Dickens novels. Engineers inside the 6,500-person division are reportedly miserable, describing a culture of micromanagement, unclear direction, and creeping despair — which is impressive going for an organisation that hasn't yet shipped anything. Building the future of intelligence, it turns out, is slightly less inspiring when you're also in back-to-back status meetings.
Source: TechCrunch
A Manifesto Arrives. Hacker News Is Extremely Normal About It
"Open Source AI Must Win" — a new essay making the case that closed, corporate AI models pose an existential risk to an open internet — landed on Hacker News this week and promptly scored 810 upvotes. The argument itself is earnest and substantive. The 249-comment thread below it, however, achieved the rare feat of simultaneously agreeing with the premise and arguing furiously about everything else. Democracy of ideas, alive and well.
Source: Hacker News
Mathematicians Remain, For Now, Employed
A new benchmark — designed specifically with previously unseen problems to prevent AI systems from essentially memorising their way to genius — found that top human mathematicians still outperform the best AI models on genuinely hard maths. This is either reassuring proof that human reasoning retains unique depth, or simply evidence that we haven't yet thrown enough computing power at the problem. The researchers are calling it the former. The AI labs are very quiet.
Source: Nature
As always, the AI said it was fine. It was not fine.
|