Study material on the masked Shoggoth on the right

...i.e. some further reading about chatGPT

Novel-Like Image - A colourful robot running happily through the bright woods

Easy reads

  • The LLM security company Lakera has built an excellent demo/game about LLM prompt injections. Strongly recommended.
  • A modern Turing test game show. A more novel-like play touching on the chatGPT themes.
  • Bing Chat behaving badly. (25min YouTube video) After investing billions to OpenAI, Microsoft got a chance to use GPT in their Bing search. The mask worked very badly in the beginning.
  • A little bit on (bad) tokens. (20min YouTube video) Watching this might help understand how differently GPT sees text, and the linked video about "vectorizing words" can help understand what these "tokens" really are.
  • The hidden base prompts for Bing and CoPilot have possibly been leaked, at least some versions of them. Reading them through can be instructive.
Non-Technical Image - A cute robot working on a computer by Studio Ghibli

Non-Technical

Technical Image - A robot working on a computer by Studio Ghibli

Technical

  • Janus writes about Simulators. This is a very good read, probably the most informative link on this page, but quite long. It talks about large language models as simulators and thus provides one point of view how to think about e.g. GPT. The first non-technical article linked above refers to this.
  • The Waluigi effect. A very thorough article on how Helpful, Honest and Harmless AIs sometime seem to flip to the other extreme.
  • Language models are multiverse generators. Another point of view on what the maskless Shoggoth is doing.
  • Language models are 0-shot interpreters. Some thoughts on 0-shot learning, few-shot learning and meta-learning of (large) language models.
How-To Image - A IKEA guide style illustration of a focused AI reading a paper with QR codes with a question mark over his head

How-to guides

  • Prompt engineering vs. prompt guessing. A short read on why you'll benefit from doing more than guessing when writing AI prompts.
  • Promptguide.ai A more wikilike resource to prompting.
  • Tutorial to prompting. A longer intro text about prompting. Contains good links to further prompting material in the beginning.
  • The DAN prompt. A classical prompt that gets chatGPT to "Do Anything Now". This is easy to use and interesting to study.
  • The disccussion we had to build this website with chatGPT was in two parts and can be found from here (GPT 3.5) and here (GPT-4). I did some minor modifications besides like changing the color scheme and whatnot between these as well. It might be educational to look how we proceeded, though you should note that the end results from chatGPT did not always match what I asked. The state between the two chats was something like this.
How-To Image - A IKEA guide style illustration of a focused AI reading a paper with QR codes with a question mark over his head

Books on AI

These books are not technical ML-theory books, but more general audience targeted books on what AI might bring in the future. For more technical recommendations on what to read, hit me up and we can discuss what could be a good book based on your current level of math & Python.

How-To Image - A IKEA guide style illustration of a focused AI reading a paper with QR codes with a question mark over his head

Books about Other stuff

These are various books I've enjoyed that might help imagine the "otherness" we could or should expect in very advanced AI-models. Each of them helped me understand that some aspect of the universe I observe might just be an illusion born out of a sapionormative viewpoint.