Ch(e)at GPT? - Computerphile
Channel:Computerphile · Length:13:51 · Published:2023-02-16 · Views:617348
Summary
Computerphile video in which Mike Pound explains a University of Maryland paper on watermarking LLM output. The method biases token choices so generated text can later be detected statistically.
Why HN recommends it
Linked as the approachable explanation whenever HN discusses detecting or watermarking AI-generated text.
a nice very approachable video explaining it here — HN, 2023
Cryptography · LLMs & generative AI
Subtitles (excerpt)
I thought we'd talk about chat GPT no one's been talking about that right it's not been mentioned I think it is both equal parts valuable and overhyped and that's the best kind of AI right I'm not going to talk about how it's trained today well I've done a great video on how it's trained you've done a video before on gpt3 which is broadly based off and oh that's bad no no no no no oh oh that's really bad Philip Moriarty has also done a video on the 60 symbols Channel about if it's possible to cheat with chat GPT and so I suppose what I'm interested in talking about is is it possible to detect cheaters in chat GPT and how easy is that to do I've been looking at this really interesting paper by uh John kirchenbauer and colleagues at the University of Maryland on if you were trying to change the output of a large language model like chat GPT or any of the others in some subtle way that allowed it to be detected as AI generated rather than human generated how might that look right and I think that's a really interesting question not does this system exist can it exist I just think if it was to exist what would be a good way of Designing it right I suppose you might think that the most obvious way would be
On Hacker News
Watermarking. From an outsiders perspective, the issue appears to reaching consensus on how this can be implemented (but not in the technical sense). There's a game theoretic challenge in that if models define and publish detection mechanisms, this creates a motivation for people to use other systems that don't include this. On the technical front there's a good paper here: [LINK] , and a nice ve… — kimburgess, 2023-02
There's a computerphile video ( [VIDEO] ) with Dr. Mark Pound explaining a paper by John Kirchenbauer, Jonas Geiping et al. ( [LINK] ) that described a method for watermarking LLM output like this. It's not directly stated anywhere in the Claude support article that this is what they're using, but the properties of the watermark described seem to point to this method. — m-chrzan, 2026-08
I wonder if they have a watermark built in that allows them to detect text generated by Llama easily. [VIDEO] [LINK] — Zambyte, 2024-07
There's an interesting scheme described here: [VIDEO] Essentially by biasing the LLM's token choices slightly you can infer with high probability later that the text was generated by it. — rjmunro, 2023-05
I think watermarking is a better solution. There is a paper on it: Kirchenbauer, J., Geiping, J., Wen, Y., Katz, J., Miers, I., & Goldstein, T. (2023). A watermark for large language models. arXiv preprint arXiv:2301.10226. ( [LINK] ) Also, Computerphile did a video on it: [VIDEO] But of course, watermarking or your proposed solution (checksums) only work while the general public cannot run a too… — stcg, 2023-03
Related videos
- Behind the Scenes of iOS Security · Black Hat · 51:31
- Why Electronic Voting Is Still A Bad Idea · Tom Scott · 12:00
- But how does bitcoin actually work? · 3Blue1Brown · 25:15
- Transformers, the tech behind LLMs | Deep Learning Chapter 5 · 3Blue1Brown · 27:14
- Deep Dive into LLMs like ChatGPT · Andrej Karpathy · 3:31:23
- Let's build GPT: from scratch, in code, spelled out. · Andrej Karpathy · 1:56:20
- Diffie-Hellman Key Exchange · Art of the Problem · 8:38
- A Hackers' Guide to Language Models · Jeremy Howard · 1:31:13
- BlackHat USA 2011: SSL And The Future Of Authenticity · Christiaan008 · 48:01
- Blockchains Are a Bad Idea (James Mickens) · HBS Digital Initiative · 14:34
- Line Goes Up – The Problem With NFTs · Folding Ideas · 2:18:23
- Theory and Practice of Cryptography · Google TechTalks · 1:32:07
- Why Electronic Voting is a BAD Idea - Computerphile · Computerphile · 8:21
- [1hr Talk] Intro to Large Language Models · Andrej Karpathy · 59:48
- The Future is a Dead Mall - Decentraland and the Metaverse · Folding Ideas · 1:49:21
- Bitcoin 2012 London: Mike Hearn · Independent POV · 28:19