A week in Generative AI: Anthropic, Sora & Kidnapping
News for the week ending 1st December 2024
It’s been a quieter week this week, mostly due to Thanksgiving the US (happy Thanksgiving to all my US readers!). However, that hasn’t stopped Anthropic from releasing two great new features which they announced this week. Model Context Protocol is an open source approach to giving large language models access to digital services and Claude also gained the ability to write in different, and customisable, styles.
OpenAI’s Sora was also leaked by some disgruntled artists who were protesting at being used as ‘unpaid R&D’ and lastly, a group of robots were kidnapped by another robot. Things are starting to get weird!
In ethics news, there’s are reports that ChatGPT queries are using 4x more water than originally thought, evidence that citations by large language models aren’t always what they seem, and the DMCA complaint against OpenAI is allowed to proceed.
In long reads, I highly recommend Ethan Mollick’s post on Getting Started with AI and also Cal Paterson’s post on why building LLMs probably isn’t going to be a great business model.
Anthropic Introduces Model Context Protocol
This is another building block and sign that we’re going to be seeing large language models act as digital companions for us next year. Anthropic’s open source Model Context Protocol is an approach to connecting LLMs to other digital services and data sources, giving them the ability to perform digital actions. This alleviates some of the need for developers to build specific APIs for LLMs to interact with their services, and I think is a much more promising approach than Anthropic’s previously announced Computer Use.
To demonstrate how quick and easy it is to set up, Alec Velikanov showed on Twitter how MCP can be used to allow Claude to put groceries from into an Amazon Fresh basket for one click ordering- a great use case! Now if only he could integrate a way for the LLM to also pay for the groceries….
OpenAI’s Sora leaked by protesting artists
On Thursday, a group of artists, calling themselves “Sora PR Puppets“, that were given early access to OpenAI’s unreleased Sora text-to-video model for testing, publicly released their access to the model on HuggingFace, giving everyone access and the ability to generate videos.
Sora was available for about 3 hours before OpenAI shut down the access. This “leak” was made in protest against OpenAI using artists as “unpaid R&D“ and was accompanied by a letter with nearly 1,000 signatures. In the letter, the group accused OpenAI of prioritising PR over genuine artistic collaboration with OpenAI restricting how artists can talk about their experiences of using Sora, limiting their creative freedom in how the model is used, and requiring pre-approval for releasing any of the generated videos.
The letter calls on OpenAI to be more transparent, artist-friendly, and supportive of the arts. OpenAI are still to respond to the letter or the leak, and are probably hoping the whole incident is forgotten over the Thanksgiving weekend.
You can now tailor Claude’s responses to your personal style
This is a nice little feature for Anthropic to release, allowing users to tailor the writing style of Claude to their specific tastes and use cases. The feature works like a system prompt, giving Claude instructions on how to write based on sharing a document that demonstrates the style the user wants, or by describing the style directly.
I posted an explainer of this new feature over on LinkedIn earlier this week if you want to see more details on how it works and how you can create custom styles for Claude to write in.
Tiny robot kidnaps other robots
Yep, it’s nearly 2025 and we’re already seeing robots kidnapping other robots 🤯.
In fact, this was all a test by a Chinese company, but it raises an interesting and important question - If robots are designed to follow voice commands, what’s to stop another robot issuing voice commands to direct the robots what to do?
It’s amazing to see that one little robot, just by issuing the right voice commands, can get 10 other larger robots to follow it!
AI Ethics News
‘Thirsty’ ChatGPT uses four times more water than previously thought
Study of ChatGPT citations makes dismal reading for publishers
Judge allows The Intercept's main DMCA complaint against OpenAI to proceed
AI companies' media deals are "pure lobbying," argues journalism professor Jeff Jarvis
Long Reads
One Useful Thing - Getting started with AI: Good enough prompting
Cal Paterson - Building LLMs is probably not going be a brilliant business
Bloomberg - Alexa’s New AI Brain Is Stuck in the Lab
“The future is already here, it’s just not evenly distributed.“
William Gibson