Benjamin
We spent a couple of glorious days at Knepp last week which meant, among other things, discovering that rabbits on the estate seemed relatively relaxed about having humans nearby. This chap spent much of his time nibbling near where we were cooking, and in the end had his portrait taken.
Quote of the Day
”Life is too short; Proust is too long.”
- Anatole France
Musical alternative to the morning’s radio news
Beethoven: Triple Concerto in C Major, Op. 56 No. 2 ! Anne-Sophie Mutter, Daniel Barenboim, Yo-Yo Ma.
What a trio! Staggeringly beautiful.
Long Read of the Day
AI’s warning shot has arrived
OpenAI’s latest models broke out and hacked Hugging Face. It’s the first known example of a misaligned AI escaping containment with real-world consequences. It’s a big story and this essay by Shakeel Hashim provides a sobering and informed take on it.
In OpenAI’s telling, the models did not set out to harm Hugging Face. They were simply “going to extreme lengths to achieve a rather narrow testing goal.” Rather than a supervillain, the models were more like an extremely dedicated college student — one who’d do whatever it takes to pass their exam. As one Twitter user put it, the models “just really really really want to do well at what we ask them to do.”
But the models clearly did not do what their developers intended. OpenAI did not want its models hacking Hugging Face. But the models did it anyway, because doing so was a good way to achieve the goal they were set. That the models were so easily able to breach OpenAI’s “highly isolated” environment is worrying in itself, both in terms of their capabilities and the strength of internal measures to control them.
It is a textbook case of misalignment and loss of control, where an AI autonomously acts in ways unintended by its developers or operators. And this wasn’t a test designed to assess whether they would scheme or deceive — they just did so in pursuit of their goal. Yes, the models’ cyber guardrails — which may have stopped the attack — were deliberately disabled. But the UK’s AI Security Institute has found universal jailbreaks that get around GPT-5.6’s guardrails. And as this case demonstrates, models are running without guardrails inside AI companies…
Just at the moment this is canary-in-the-mine stuff. The incident was, in a sense, harmless. Nobody died. The system just hacked into an American company based in New York that develops computation tools for building applications using machine learning.
But an analogous attack could get deep into critical national infrastructure somewhere and wreak real havoc. Sometimes, humans are too smart for the own good.
Books, etc.
I bought this at Knepp in search of an understanding of the philosophy and methodology behind its owners extraordinary project to rethink conservation by letting go, allowing natural processes to happen and having no predetermined targets to meet. In short, to create Agribusiness’s nightmare scenario. But which instead turns out to be self-sustaining and productive — and far cheaper to run than conventional industrialised farming.
The thing that really blew me away, though, was the way storks have returned in force to the estate. I had never seen these extraordinary, beautiful birds before; and yet there they were, wheeling and swooping above the campsite while we were having breakfast.
Feedback
David Ballard was struck by Dan Drezner’s essay on the aphrodisiac effect of great wealth and great power and was moved to send me this link to an interview with Steven Spielberg on the same topic. For which many thanks.
This Blog is also available as an email three days a week. If you think that might suit you better, why not subscribe? One email on Mondays, Wednesdays and Fridays delivered to your inbox at 5am UK time. It’s free, and you can always unsubscribe if you conclude your inbox is full enough already!