. . AI: Open source reliability OpenAI governance, text-to-video, (12.1.23)

Friends, it was a rather quiet day in AI today! Unless you were the Argilla team in Spain or the HuggingFace climate and AI team. That's whose work rose to the top of the AI community's conversation. I hope some of the day’s news will be useful to you.

Two bonus links to make sure everyone finds something valuable in today’s newsletter:

• I used SciSpace’s GPT today, to search and summarize academic research papers and it worked really well! I wrote about it here a few weeks ago.

• The Futurists, an excellent podcast hosted by the very influential pair Brian Solis and Brett King, had me as their guest today and we discussed using AI as a cognitive exoskeleton! The episode art depicts what I’ll look like 10 years in the future, too.

Stay tuned for a round-up of the weekend’s AI news on Monday!

Cheers,

Marshall Kirkpatrick, Editor

Notus 7B Learns to Follow Instructions Well

First impacted: AI developers, Open-source software users
Time to impact: Medium

Argilla, an open source AI firm in Spain, has developed a new open-source LLM, Notus 7B, which they say surpassed Zephyr-7B-beta and Claude 2 in the AlpacaEval benchmark, a measure of a model's ability to follow user instructions. "Open-source, open-science, and data curation for the win!" they tweeted. [Meet Notus-7B: Data Curation and Open Science go a long way in shaping AI's future] Share by email

Text-to-Video Models Advance, Face Challenges

First impacted: ML researchers, Video content creators
Time to impact: Medium to long

London-based ML researcher Alara Dirik and a group of collaborators have published an overview of the state of text-to-video technology. The write-up says the newest text-to-video models, including diffusion-based designs such as Video Diffusion Models (VDM) and MagicVideo, are "progressing exponentially" but the field faces obstacles including high computational costs, a shortage of high-quality datasets, and the intricacy of video captioning. [A Dive into Text-to-Video Models] Share by email

OpenAI Governance Questioned After Co-founder's Firing, Reinstatement

First impacted: OpenAI team
Time to impact: Medium

The Economist covers OpenAI's governance structure, which has been under scrutiny due to the contentious dismissal and later reinstatement of co-founder Sam Altman. The company also announced a policy change that will permit a 20% annual increase in investor returns starting in 2025. [Inside OpenAI’s weird governance structure] Share by email

AI Image Generation's High Carbon Footprint Revealed

First impacted: Climate impacted communities
Time to impact: Medium to long

Hugging Face and Carnegie Mellon University have conducted a study indicating that AI models, particularly those used for image generation, contribute significantly to carbon emissions. The research suggests that the energy used to generate a single image is comparable to charging a smartphone. (Text is much more efficient, 1000 prompts = 16% of a phone charge.) The study also says that using large generative models for specific tasks (for example, classifying movie reviews as positive or negative) consumes 30 times more energy than user smaller, task-specific models. [Making an image with generative AI uses as much energy as charging your phone] Share by email

That’s it! Join us on Monday for a round-up of the weekend’s AI news!