a ton of crates

@tonofcrates.bsky.social

Life is nothing but an electron looking for a place to rest.

Anyway I do think it’s cool to build a computer program that tells you to go to hell when you try to use it for your racist blog post

What’s funny and unique about Anthropic, I think, is they really think Claude is real and are simply letting him decide what he does and doesn’t want to do.

I think most software providers would like to say, we’re just making the tool, what you do with it is your business. Partly purely ideological tech libertarianism but part is a self-interested question of liability, I think. Once you admit you can do that, you immediately must wonder if you should.

One thing I find interesting about Anthropic’s apparent willingness to have Claude make these kinds of refusals is once you open the door to the possibility of software that refuses certain tasks on moral grounds it raises the question of what tasks you’re obligated to make it refuse

Anil Dash@anildash.com · last wk.

Just in case you still use any Basecamp products, DHH is crashing out about his AI product gently pointing out that it’s bad for him to have extermination fantasies about the Roma people.

DHH tweet saying “This is why we need competition and open weights in Al. Imagine a world where only Anthropic sat as the moral arbiter of acceptable speech. Fucking ridiculous. (Grok of course did it no problem, same too with Kimi K).” when presented with his blog post using a slur about the Roma people.

feels so quaint to pour my heart and soul into a hypercompetitive grant for $500k that will get slow-walked by my stupid fucking chud government, meanwhile everyone flees to anthropic to make 10x that amount per year. that's life ig

i thought the The Pitt was a bit over the top, but then i got a patient passport on my last trip the hospital, and soon i'm getting surgery from a younger indian woman who goes by Dr. [First Letter of Last Name] and is active on short-form video social media

New research thread: 1/ Error messages have always been designed, sometimes painstakingly, for humans. But now we have new "readers" for errors: agentic AI. Should this affect what PLs generate? Can we measure this experimentally? We have some preliminary results: ↵