Sunday Signal: ChatGPT leaking secrets, 30 observations at 30 and the idea of enough

Hey friends ??

Here’s your weekly dose of AI and introspection.

AI Highlights

1. ChatGPT is leaking its secrets.

Researchers have exposed vulnerabilities in OpenAI's ChatGPT and its custom GPTs, revealing a significant privacy risk. For $200, they extracted several megabytes of ChatGPT's training data, indicating the possibility of extracting a gigabyte of data with more resources. See the live example here.

This vulnerability is echoed with custom GPTs. Researchers achieved a 100% success rate in leaking files and a 97% rate in extracting system prompts. The custom GPTs, which can be created by anyone with an OpenAI subscription, have been found to reveal initial instructions and uploaded documents.

Alex’s take: I feel this situation calls for increased awareness. These tests were conducted using the production version of ChatGPT. I gave the prompt a go earlier this week and saw the vulnerability first-hand. It highlights there’s still a long way to go to build secure and resilient AI that doesn’t regurgitate its training data using a simple prompt.

2. An introduction to large language models.

Andrej Karpathy, former Senior Director of AI at Tesla and a key figure in AI, simplifies the complexities of Large Language Models (LLMs) like ChatGPT in his talk, 'Intro to Large Language Models'.

Alex’s take: What I enjoyed about this talk is how accessible it is. Andrej’s simple explanations make it something even total beginners in the world of AI can enjoy. It’s one thing experimenting with these tools, it’s another cementing your technical insight (and impressing your friends with your newfound knowledge). This talk helps you find that blend.

3. Animate Anyone: Image-to-video.

You can now bring still images to life. Alibaba's Institute for Intelligent Computing introduced 'Animate Anyone', an image-to-video model trained on 5,000 character video clips. The technology animates characters from photos into videos whilst maintaining consistent appearance details. They leverage diffusion models to overcome traditional animation challenges.

Alex’s take: This is the most advanced character animation I’ve seen so far. I see huge applications here for brands leveraging this for celebrity deals without scheduling constraints. What’s more, virtual try-ons could become indistinguishable from the real thing.

?? As a LinkedIn reader, you miss out on 1 Article, 1 Idea, 1 Quote, and 1 Question to ponder. Subscribe here to never miss out.