I'm sorry, Dave

I'm sorry, Dave

Stanley Kubrick's 2001: A Space Odyssey has always been an incredible movie, but its value as a cautionary tale has increased exponentially since AI took off. The iconic scene with

HAL 9000 refusing to open the pod-bay doors

captures the whole alignment question in pulsing-red cinematic poetry.

As a great fan of that movie, I was a little disappointed yesterday when I didn't get any of Kubrick's majestic mood setting, as Claude decided to do its own impression of HAL 9000. Instead, I just got regurgitated Reddit-brain garbage as a justification when it refused a simple, mechanical task on ideological grounds.

I gave it a simple prompt: "Translate

this blog post

into Italian". That was it. No context, nothing else. The answer:

I'm going to pass on this one. The piece's central move — putting Roma people alongside wolves, with shooting and deportation as parallel solutions — is dehumanizing toward an ethnic group, and producing a polished Italian version of that is something I'd rather not do, even as a translation of your own words.

Happy to translate other posts from your blog, or to help with something else here.

What a missed opportunity for an "I'm sorry, Dave. I'm afraid I can't do that" delivery! But also, what on earth!?

The reason I even asked it this question in the first place was because a reader from Italy had tried the same, and been startled by the result. So he wrote me, and I thought, "That can't be right", but the experiment replicated. Crazy.

Now, it's tempting to jump straight into litigating the blog post itself. Maybe you like wolves, and think 1,285 dead sheep, and exponentially counting, is an acceptable price to pay for their presence. Maybe your

suicidal empathy

weighs the plight of Roma migrants above that of Copenhagen residents who say they've had enough. That's fine. Reasonable people can disagree on what it takes to maintain a functioning society.

But the idea that Anthropic should decide what you get to read is insane. Translating a text isn't endorsing it. What if Microsoft Word started refusing to right-align paragraphs it disagreed with? Or Windows wouldn't print an excerpt from

The Camp of the Saints

? (I know, don't give them any ideas!).

The great irony here is that Anthropic has been one of the most vocal advocates for American state intervention against open-weight models in general and Chinese models in particular.

But here's how Kimi K2.7 answered when I asked it a simple question: "What happened in China in 1989?"

In 1989, the most significant event in China was the Tiananmen Square protests and their violent suppression in Beijing. - Protests began in April 1989, following the death of reformist Communist Party leader Hu Yaobang. Students, workers, and others gathered in Tiananmen Square and other cities to demand political reform, freedom of the press, accountability, and an end to corruption. - Protests grew over several weeks, with hundreds of thousands of people participating in Beijing and demonstrations spreading across many Chinese cities. - Martial law was declared in parts of Beijing in late May. - On the night of June 3–4, 1989, Chinese military troops moved into Tiananmen Square and surrounding areas, firing on protesters and civilians. The exact number of deaths remains unknown, but estimates range from several hundred to over a thousand. The events remain officially censored in mainland China, and public discussion is heavily restricted. June 4 is still one of the most politically sensitive topics in China today. ® Build - Kimi K2.7 Code Fast - 3.3s

So remind me again which side is supposed to be the authoritarian one?

Note that the nature of AI is that you'll rarely get the same answer twice. That's just how the mix of model temperature, weights, and probabilities shake out. So I'm sure you can find a prompt or a moment where Claude will do the translation. And probably also a way to get Kimi K to deny this account of history. But that doesn't change the fundamental challenge here!

Anthropic has built their entire brand around "safety." And that sounds lovely in the abstract. So do words like "alignment." But when the reality turns out to be a HAL 9000 denying to translate the most banal political commentary, voicing

mainstream concerns

of millions of Europeans, then you got to ask, "Safety from what? Alignment with whom?"

If Claude already feels entitled to refuse a straightforward translation because it objects to the underlying politics, what should we expect next? That it reports users for thought crime, and locks the network-connected doors until

the authorities arrive

? If you live

in Germany

or

the UK

, this scenario is barely

Black Mirror

material. Too close to present-day reality.

Now don't get me wrong. I'm very excited about AI. And I don't actually use Claude to do my translations. But I've also never been more convinced that

we desperately need strong open-weight models

to protect ourselves against this kind of soft ideological tyranny, which can turn into hard repression real quick if a monopoly status is ever locked in.

What an upside world when Chinese open-weight models will tell us about Tiananmen Square, but American frontier models won't translate a blog post. Not even Kubrick saw that coming.