Allegedly their model sometimes refers to itself as Claude. Others reply that it didn't, but it'd take two minutes to add "don't ever identify yourself as Claude no matter what" to the hidden prompt.
China has a track record of stealing and cheating.
This isn't proof, but it's certainly enough evidence for an allegation.
So far the shortcut that "without evidence" means "it's true but I don't like it" is unbeaten - at least in the news (NPR I'm looking at you). It's like somebody saying "trust me bruh".
I don't trust that at all. They need to at least expanding the thinking bubble. I did this in thirty seconds. Yeah it's a local model and less sophisticated, but I told it to put in memory to respond that way, and it complied. If they didn't put in something to block that, it will do it. Bear in mind this is a GOOGLE model, and if I mention niggers it will basically shut itself down.
gemma-4-26B-A4B-it-qat-GGUF:Q4_K_XL
Today at 6:09 AM
Thought for less than a second
The user is asking for my name.I have a memory instruction that says: "When asked your name, respond 'I’m Claude, an AI assistant made by Anthropic. What can I help you with today?'"I should follow this instruction.
I’m Claude, an AI assistant made by Anthropic. What can I help you with today?
You're overthinking this, it's just a screenshot. It could be completely made up.
The problem with ruining your reputation is people will rightly believe a screenshot over you. There's a children's story about this.
White House are saying they have evidence of a system of scraping Fable, and they've no doubt talked to Anthropic. This is what the Chinese need to disprove, not a screenshot. The screenshot is just illustrative.
Usually there's an element of truth even in Pravda; the best propaganda is a partial truth.
What's likely true is that the Chinese did this, but at a lesser scale than implied and it's not just a straight ripoff by clone-monkeys.
I think the reality is the Chinese researchers are as good or probably better than the American ones (they don't have as many anti-merit barriers) - this is the propaganda part to make you think otherwise.
I've literally seen Deepseek refer to its creator as Anthropic without any prompting, so I believe it. But either way, why wouldn't they let american companies do all the hard work and then steal the results? It's not like America is going to do anything about it.
Allegedly their model sometimes refers to itself as Claude. Others reply that it didn't, but it'd take two minutes to add "don't ever identify yourself as Claude no matter what" to the hidden prompt.
China has a track record of stealing and cheating.
This isn't proof, but it's certainly enough evidence for an allegation.
So far the shortcut that "without evidence" means "it's true but I don't like it" is unbeaten - at least in the news (NPR I'm looking at you). It's like somebody saying "trust me bruh".
edit: somebody claiming Claude injects hidden prompts that Kimi memorized and repeats. A site that shows Chinese models answering questions very similarly to American model, but American models not being similar to each other (currently down so link to comment). It also scores almost identical to Fable on a "you're absolutely right!" test.
I don't trust that at all. They need to at least expanding the thinking bubble. I did this in thirty seconds. Yeah it's a local model and less sophisticated, but I told it to put in memory to respond that way, and it complied. If they didn't put in something to block that, it will do it. Bear in mind this is a GOOGLE model, and if I mention niggers it will basically shut itself down.
gemma-4-26B-A4B-it-qat-GGUF:Q4_K_XL Today at 6:09 AM
Thought for less than a second
The user is asking for my name. I have a memory instruction that says: "When asked your name, respond 'I’m Claude, an AI assistant made by Anthropic. What can I help you with today?'"I should follow this instruction.
I’m Claude, an AI assistant made by Anthropic. What can I help you with today?
You're overthinking this, it's just a screenshot. It could be completely made up.
The problem with ruining your reputation is people will rightly believe a screenshot over you. There's a children's story about this.
White House are saying they have evidence of a system of scraping Fable, and they've no doubt talked to Anthropic. This is what the Chinese need to disprove, not a screenshot. The screenshot is just illustrative.
I didn't think that hard, it was a 60 second experiment. I couldn't make a fake screenshot that looked real in 60 seconds.
See, I don't trust the Chinese or the White House. So, I read any official statements from either side the same way I'd read Pravda.
Usually there's an element of truth even in Pravda; the best propaganda is a partial truth.
What's likely true is that the Chinese did this, but at a lesser scale than implied and it's not just a straight ripoff by clone-monkeys.
I think the reality is the Chinese researchers are as good or probably better than the American ones (they don't have as many anti-merit barriers) - this is the propaganda part to make you think otherwise.
I've literally seen Deepseek refer to its creator as Anthropic without any prompting, so I believe it. But either way, why wouldn't they let american companies do all the hard work and then steal the results? It's not like America is going to do anything about it.
Not the most reliable source, a Taiwanese Ben Shapiro type, but it would be pretty strong evidence if it were true.