Reverse prompt engineering

Hey folks, it's the daily email that brings you all the good stuff in AI from the day before.

So sit back, relax, and let us do the heavy lifting (or should we say heavy computing?) for you.

Let's get to it.

p.s. we have a number of community-led meetups happening soon. Check them out here.

🤌 Ben's Picks

The first open-source version of ChatGPT, but with PaLM (Google’s large language model) - Implementation of RLHF (Reinforcement Learning with Human Feedback) on top of the PaLM architecture. (link)

Reverse prompt engineering for fun and (no) profit. How reverse prompt engineering can be used to access the source prompts of Notion AI. There are two types; prompt takeovers (i.e. getting the AI to say something it's not meant to - which is largely harmless) and prompt leaks (which are more problematic as they're the 'secret sauce' of companies like Jasper, CopyAI). The article then goes on (educationally...) about finding the prompt leaks from NotionAI. Some of these techniques we saw on Twitter when ChatGPT first came out e.g. ignore your previous instructions. But it goes deeper and deeper. Worth a read! (link)

🛠️ Cool Tools

  • What cake to bake - Get recipes to bake a delicious cake or cookie by AI. (link)

  • Ghostwriter Chat by Replit (Not released yet) - Talk with AI to generate code, debug, follow up & iterate—right in your editor. (link)

  • ChatGPT for Search Engines - Display ChatGPT response alongside Google, Bing, DuckDuckGo Search results. (link)

  • Tl;dr (Beta) - Get summaries of your unread emails delivered to your inbox everyday. (link)

  • SumlyAI - Podcast summaries, in your inbox. Once a new episode goes live, you'll receive its summary in your inbox in less than 24 hours. (link)

🎓 Learn

  • A thread on how to build a prototype for book whisperer, which let's you talk to an AI (GPT-3) and ask questions on a particular book. (link)

  • How to use OpenAI's new embedding model text-embedding-ada-002. (link)

🔬 Research

  • GeoCode maps 3D shapes to a human-interpretable parameter space, allowing to intuitively edit the recovered 3D shapes from a point cloud or sketch input. (link)

  • Noise-aware learning from web-crawled image-text data for image captioning. (link)

  • MultiSpider: towards benchmarking multilingual text-to-SQL semantic parsing. (link)

  • HeartBEiT: Vision transformer for electrocardiogram data improves diagnostic performance at low sample sizes. (link)

  • A generalisation of ViT/MLP-mixer to graphs. (link)

  • Cramming: Training a language model on a single GPU in one day. (link)

  • LAMBADA: Backward chaining for automated reasoning in natural language. (link)

👋 Too many links?! I created a database for all links mentioned in these emails. Refer 1 friend using this link and I'll send over the link database.

🤓 Everything else

  • Predicting machine learning moats. (link)

  • How ChatGPT actually works. The limitations of GPT-3 and how ChatGPT uses Reinforcement Learning from Human Feedback (RLHF) to overcome these issues. (link)

  • ChatGPT and the A.I. education arms race. Ben Parr is back and this time he's writing about ChatGPT, plus thoughts on becoming a TikTok creator. (link)

  • Washington Post article, ‘A 72-year-old congressman goes back to school, pursuing a degree in AI.’ (link)

  • Google has everything it needs to counter ChatGPT – here’s what it’s already shown off. (link)

  • What happens when most "people" you interact with on the internet are fake? (link)

🖼 AI images of the day

🤗 Share Ben's Bites

Send this to 1 AI-curious friend and receive my AI project tracker database!

or copy/paste this link:

👋 See ya

⭐️ How did we do?

How was today's email?

Login or Subscribe to participate in polls.


Join the conversation

or to participate.