Comment by ivan_gammel

16 hours ago

This doesn‘t sound like AI. LLMs do not use dictionary definitions for translating a paper this way. It does look like some translation engine like Google Translate or worse, maybe with some text obfuscator trying to avoid plagiarism detection.

Google Translate uses an LLM, last I heard.

  • It switched very recently I think if at all. It was way worse at translations than Gemini even at the beginning of this year. Especially between a pair of languages other than English. It was mixing two completely different words that are the same in English.

  • Only since 2023.

    • WDYM? It used an encoder-decoder transformer model long before 2023 (and also long before decoder-only transformers à la GPT became the new standard).

      Most people seem to believe LLM only describe these decoder-only architectures, but the term predate them by a big margin. BERT (encoder-only) was a large language model, and there were even language models before transformers were even a thing.

That sounds like Anthrophics watermark explanation though.

  • Anthropic’s watermark is just using a known collection of random seeds during output token generation. That’s it. The models always relied on some randomness, here the randomness is fixed and remembered for querying later.

    • > Anthropic’s watermark is just using a known collection of random seeds during output token generation. That’s it.

      Wouldn't that mean that checking the watermark...

      (1) Required knowing the prompt used to generate the text; and

      (2) was just as expensive as generating the text?