Role-playing with AI, more AI in Paris, and more (11.8.23)

Friends, for these round-ups of AI news each day, I focus on what the developer and startup focused AI community is talking about the most each day. So there are consumer AI stories I don't cover here, like today's leaked news about the "invisible computer" AI-using lapel pin cell phone Humane, scheduled to launch tomorrow after raising $230M and previewing on stage at TED this summer. But there's a link to that in case you are interested. It will reportedly cost $700 plus a $24/month T-Mobile subscription, and it doesn’t have a screen.

Now here's the news the rest of the AI community is talking about today.

-Marshall Kirkpatrick, Editor

Meta, Hugging Face, Scaleway Launch AI Startup Accelerator in Paris

First impacted: AI startup founders, AI researchers
Time to impact: Medium

Meta, Hugging Face, and Scaleway have initiated a new AI-focused startup accelerator at Paris's Station F startup megacampus, with the goal of promoting a more open, collaborative approach to AI development in France. The "AI Startup Program" is open for applications until December 1, 2023, and the five selected startups will receive guidance from Meta and Hugging Face researchers, along with computing resources from French cloud provider Scaleway. France and open source AI just keep making news together. [Meta taps Hugging Face for startup accelerator to spur adoption of open source AI models | TechCrunch] Share by email

Giskard Bot Enhances Machine Learning Model Testing

First impacted: Machine Learning Engineers, Data Scientists
Time to impact: Medium

Speaking of France, French open source firm Giskard has launched a bot on the Hugging Face platform, designed to automatically detect potential weaknesses in machine learning models, such as performance bias, hallucinations, ethical dilemmas, stereotypes, data leakage, lack of robustness, and spurious correlations. The bot provides a detailed report on the LLM's vulnerabilities whenever a new model is uploaded, assisting users in debugging, and providing qualitative content that suggests changes to highlight biases, potential risks, or limitations. [Introducing the Giskard Bot: Enhancing LLM Testing & Debugging on Hugging Face] Share by email

New OtterHD-8B Identifies & Locates Small Objects in HD Images

First impacted: Computer vision users
Time to impact: Medium

Researchers at Nanyang Technological University in Singapore have launched a new multi-functional model, OtterHD-8B, which they say excels in interpreting high-definition visual data with exceptional detail and adaptability to different input sizes. In other words, the OtterHD-8B can identify, locate, and understand the relationships between small objects in high def images. [OtterHD: A High-Resolution Multi-modality Model] Share by email

Maybe Don’t Anthropomorphize: Role-Play With AI Instead!

First impacted: AI researchers, AI chatbot developers
Time to impact: Medium to long

A paper titled "Role Play with Large Language Models" has been published in Nature by authors from Google DeepMind and Imperial College. They argue that it's a mistake to anthropomorphize LLMs because it increases the risk of emotional manipulation of humans, and they offer as an alternative framing LLMs as role playing companions. They also warn though that when these AIs can access tools like calculators, calendars, and external websites, actions taken in a role-play context can have real-world effects. Inspired by this and some strange emails, I had ChatGPT create a dialogue today about advising people on the topic of internet conspiracy theories, with roles played Socrates, Einstein, and Joni Mitchell. It was pretty good. [Role play with large language models] Share by email

DeepLearning.AI Launches Vector Database Course

First impacted: Software developers
Time to impact: Short to medium

DeepLearning.AI, in collaboration with Weaviate's Sebastian Witalec, has launched a new course centered on the use of vector databases in LLMs. The free course aims to provide learners with the ability to use vector databases and LLMs for data analysis, embedding creation, and similarity search methods. [Vector Databases: from Embeddings to Applications] Share by email

Ok, that’s it! More AI news tomorrow.