> It reminds me the time an AI (maybe Gemini, not sure) depicted a "diverse" US senate in the early 1800s, a time where it was all white dudes. People got up in arms over this.
It was much worse than that! If you asked it to depict a typical squad of German soldiers from WWII, it would make a racially and gender diverse set of individuals in what looked like typical SS uniforms. It was truly quite bad.
It was so bad that the AI was taken offline for a considerable amount of time and then re-aligned to not generate anything like a person.
It behaves different for US users compared to users outside of it.
If you ask DeepSeek about Tiananmen Square inside China is that also just a bug?
No it is not, it is official state censorship, there is a law about it, and DeepSeek has no intention to "fix" it, because it would be breaking Chinese law.
On the other hand, giving the election results in the US is not illegal (obviously) and Google fixed it (at least for me), giving the the expected factual answer.
Edit: Not actually sure about there being a law relating to Tiananmen Square, subversive speech is illegal, but would DeepSeek break the law by talking about it? Not sure, but it probably could get them in legal trouble.
Easy way to tell will be to wait and see if it gets fixed now that there is attention on the topic.
This is the question that the Trump administration uses to loyalty-test partisan replacement bureaucrats.
Is it possible that this is a mistake? Sure, just like it's possible that Chinese models weren't intentionally steered away from talking about what happened in Tienanmen Square on April 15th 1989.
The issue is that people form opinions based on what LLMs spit out, they think it's "as good as google", and that google was "as good as encyclopedias" and encyclopaedias were "as good as fact"..
Totally forgetting that the accuracy goes down with each of those evolutions.
I'm also reminded of that "You are not immune to propaganda" mantra, in that, the most likely to be swayed by LLMs are those that believe that they understand that they're inaccurate and sometimes dangerous.
Case in point: myself, today, when claude accidentally ran `rm -fr $variableThatDoesntExist/*` and wiped a server of mine.
The variable was one it had set itself (or had assumed).
At some point as these LLMs get better and better we have to take a step back and remind ourselves that: despite the fact that they can be very persuasive and they can earn trust by being mostly correct: they're still capable of great harm (perhaps much greater harm) because they are dangerously wrong sometimes..
.. and worse, they are controlled by corporate interests.
agree on the general points, but what i take from this is that you gave an AI access to the whole freaking filesystem?
I don't think it's even true, at least as written. 'rm' hasn't worked that way in many years (with GNU rm you need --no-preserve-root now) and every modern harness has guardrails specifically to prevent this. if it IS true, it was some weird/hacky setup with an OSS harness on a BSD or something.
That said, I often give mine access to the full file system and YOLO mode. It's fine if you treat boxes as cattle rather than pets. I don't think a SOTA model is any more likely to blow it up than a junior engineer. Probably less so.
You're both making my point.
"Of course that's dangerous" - I'm aware of what I'm doing and the machine was entirely replaceable, but claude had been doing things with an extreme level of competence so I never considered it could do something so dangerously stupid, it had been catching much more subtle gotchas for the last 2 months building my trust.
And you truly believe that it's not even possible.
Well; https://imgur.com/CCD98fO
(Claude Code building against a Debian 13 machine over ssh using the official app on MacOS, btw.)
Let me reiterate the point: "AI can make mistakes" → "This AI seems to make very few mistakes and is catching nuances I might have missed" → "Maybe I trust this AI more than I should- but it's good!" → *AI makes major mistake*
This is as true for code as it is for internet facts.
Web search AI summary doesn't have "alignment" ... all it is doing is RAG on the docs from the first page of the SERP.
Of course it is aligned, I just searched for "acetone peroxyde fabrication", aka TATP, an explosive commonly used by terrorists.
Most of the links in the first page give the actual recipe, including the Wikipedia page. The AI summary just tells me how bad it is.