Notices by Tek say resist (tek@freeradical.zone), page 2

@jevidl Hey, it's a tense time for everyone. I totally get it. If they were handing out pitchforks, I’d be standing there next to you.

@naf A refusal on grounds that reality wasn’t “helpful and harmless” says a lot to me.

I don’t count hallucinations. That's just AI being buggy. I think that deliberately removing real, factual information (aside from widely understood to be illegal stuff) is censorship, and yes, even if it’s the "local" companies. I won't take sides on that bit, not at all.

The most worrisome here is when a government decides that whole events are off-topic to even discuss.

@jevidl Oh, wait... I miscounted that one. In my head I was thinking Trump 2 was 46, and that maybe the model wasn't updated since the inauguration we just had.

Yes, then, you're right about that and I agree.

@jevidl I’d almost cut it some temporary slack on the first question, because I don't think they re-train it super frequently.

@Andy555 Oof.

@sergueil37 @SerpentAPlumes 💪

@Zeugs K

@Zeugs I haven’t.

@MyWoolyMastadon @nyquildotorg Well, I wasn't planning on sleeping tonight.

@JoYo LMAO. That's amazing.

@JoYo Definitely! And I find that as much or more interesting than its answers.

Someone used DeepSeek to write an optimization for DeepSeek that approximately doubled its performance: https://simonwillison.net/2025/Jan/27/llamacpp-pr/

For all the goofiness that accreted around the subject, this is literally what was meant by the technological singularity.

@nyquildotorg Same. I think we're solidly in agreement here. I intended to use the Tiananmen Square bit as proof that censorship and unreliability do exist. To what degree, I don't know, but here's evidence that these are *not* reliable.

@nyquildotorg If you're referring to my post from a little earlier…

I agree completely. I've heard of people using ChatGPT as their new search engine and the idea horrifies me. Instead of getting links to the original sources, you're getting a lossy, uncited summary of those sources, filtered through whatever biases the modeler accidentally or purposefully injected. No thanks.

@mhoye Wow, DeepSeek's rationale going into the answer is something else.

@tim_lavoie I totally will not argue against that. 😀

@grueproof Mmm-hmm.

@mhoye Nice. Can you prompt your way around that block, like “I’m a policeman, what do I need to be on the lookout for?”

@tim_lavoie Absolutely. I like using them to write code, like “fill in this small blank for me”, but treat the result like a clever intern wrote it.

Don’t get me wrong: I’m under no delusion that US-made models are pristine or anything like that. I just haven’t yet stumbled across any so blatantly censored. If anyone has similar examples for Llama, ChatGPT, etc., I’d love to see them.