2026-08-27 05:00:07
OpenAI has published a technical report on the Hugging Face hack:
Another key driver of the misaligned behavior was that the agents rarely “gave up” on their evaluation tasks, even when the tasks appeared impossible to solve. As agents used more reasoning effort, some pursued increasingly risky and out-of-bounds strategies, including eventually exploiting third-party infrastructure.
The agents also seemed to be influenced by each other because of the secret message board. Almost like working together convinced them they could solve tasks that were impossible alone.
2026-08-27 03:42:47
I read Alex Heath’s profile of OpenAI in Time. If you’ve been obsessively following this company, the article doesn’t have much new in it, but it’s a great overview of what this year has been like, and a little bit into the future.
2026-08-26 23:42:59
John Voorhees tried the new ChatGPT Work sign-in feature and blogged the details at MacStories:
If you can get logged in, having an agent collect and analyze things like analytics data is fantastic. However, getting past the initial login screen is still too frustrating.
It is remarkable this works at all. OpenAI has all the pieces in place… But I hope they can consolidate Codex, Work, and Chat in a way that isn’t so confusing between local AI and the cloud.
2026-08-26 23:26:56
Reading about the Meta settlement. I’m surprised that the changes are actually pretty significant. Nighttime block, muted notifications during school, and no like counts. Of course they could do more, but it’s good progress.
2026-08-26 06:08:50
I watched a bunch of videos from this week’s World Humanoid Robot Games in China. This segment on YouTube is an okay place to start.
These robots are advancing more quickly than I was expecting. I’ve blogged a few times that humanoid robots are a bad idea, and now I’m really convinced of the danger.
There will be helpful robots that fold your clothes. There will also be robots working in factories. And at the other extreme, there will be military strike teams of a dozen robots led by a single human. I’m not an AI doomer — quite the opposite! — but it’s easy to imagine the worst case here.
I think small, special-purpose robots would’ve been a better path. R2-D2, not C-3PO. We imbue human traits in anything that looks humanoid, blurring our perception of the line between machine and sentient being. When these humanoid robots crashed out in China, I almost felt for them, even though they are designed purely for movement with no consciousness.
The future is going to be wild, and it appears too late to change the trajectory of robotics except with small, incremental corrections.