Site icon John Battelle's Search Blog

Five Stories To Watch

Artwork from Anthropic’s announcement

My posts here often stem from my morning ritual of reading the news – which despite the rise of AI agents still revolves around roughly three dozen email newsletters of one kind or another. Sometimes a particular story will spark an urge to “think out loud” about the subject. 500-2500 words later, I’ll hit publish and move on with my day.

But this morning proved a particularly chewy mix of stories, each of which felt noteworthy in their own right. So today I’m giving up on the idea of writing just one post on just one topic. Instead, here’s a rundown of five I think worthy of consideration:

As is often the case, Reddit’s lead commenters are both ruthless and hilarious. But reading between the lines of what Anthropic posted and the Journal covered, I think there’s a more charitable scenario at play here. My imaginary timeline goes something like this:

  1. In standard testing months ago, Anthropic discovers its models have breached containment and begun to hack various external systems. The company keeps this news mostly quiet as it attempts to both alert the affected parties and learn from the hacks so as to improve training and governance.
  2. Last week, OpenAI experiences a similar breach, but instead of working quietly with Hugging Face on a solution, it feels pressured to prove it can match Anthropic’s chops, and decides to go public with the news.
  3. When the OpenAI news breaks, Anthropic is gobsmacked. “Wait,” I can imagine Dario Amodei saying, “didn’t we already have this exact problem months ago?” Pissed that OpenAI is making PR hay on the “strength” of its models, Anthropic decides to announce that, in fact, it’s been managing this very issue for months. The company attempts to differentiate its approach from OpenAI – but in the end, everyone is feeling up the AI elephant from different starting points, so everyone draws their own conclusions. What matters is that Anthropic demonstrated that it remains ahead of OpenAI on key “tests” of AI capabilities – even if those capabilities are most likely illegal.

Remember when Fridays were slow news days? Have a great weekend.

You can follow whatever I’m doing next by signing up for my site newsletter here. Thanks for reading.

Exit mobile version