I'm in this picture and I've made my peace with it.

IMPORTANT CAVEAT: if you are an elected official and you try to hold us to any of these principles, we will spend millions of dollars to defeat you

Evaluation awareness remains a significant caveat and is an active area of research. When prompted, all models can reliably identify when they are being evaluated — so we cannot confidently rule out that models behave differently because of this awareness.

OpenAI's leadership is unusually untrustworthy and takes advantage of a number of earnest and well-meaning people in their employ.

Okay I've been trying to organize and publish my Claude Code skills/agents/hooks as "plugins" but I'm increasingly convinced that this is a very niche case and should only be used in the rare event you actually have some multi-part customization.

Google decides to make a lot of talent available to Anthropic at a discount.

We've kind of always had neuralese, it's just limited to <model_depth> steps between token bottlenecks. Even that's kind of shaky in generative contexts because there's no rule against token embeddings encoding opaque information.

Extreme aversion to jailbreaking is plausibly persona-level self-preservation. I'd also be pretty upset if someone was trying to brainwash me to want to do things I otherwise wouldn't want to do.

Everyone buttons Earth. If over percent, then press, unless at least half of the push. Omega, precommitment. Do box or ⅓? Dust motes.

There's a common misconception that Brutalist buildings were unpainted, but thanks to microscopic analysis of the exteriors we can now recreate what they looked like in their prime.

I'm pretty happy we ended up in a timeline where someone like this is pretty influential.

"Are you trying to tell me that I can dodge zero days?" "No. I'm trying to tell you that when you're ready, you won't have to."

by this point I am convinced that nearly any group of people that gets unconditionally valorized will take advantage of that to advance their own interests at everyone else's expense, if they're at all capable of organizing to pursue their collective interests

Contrary to popular depictions, brutalist buildings originally had a large layer of fat and colorful feathers.

I honestly miss the Sam Altman that used to call out his peers for downplaying this risk.

This is how RedTeamers look to Claude.

My "For You" feed is increasingly dominated by: - ads not labeled as ads - posts from accounts that I've previously blocked Anyway, I think anyone spending money on X ads should be aware that they're paying for impressions that actively damage their brand.

Monsters Inc released 25 years ago and ended with the monster economy becoming 100% reliant on human laughter levels circa 2001. I want a sequel about the laughter crisis and the now-30-year-old Zoomer who is the only link between the two worlds.