New name & today's top AI news!
AI for Good, Smaller Than Llamas, HuggingFace Decentralized Research
Friends, here are today's AI stories, in the order of their Time to Impact.
But first: I’m excited to say we’re changing the name of the newsletter. After 30 days of publishing as aiTTENTION, we’ve come up with a name I like a lot better. Based on what I think is a strong differentiating value proposition, this newsletter, the related accounts on Twitter and LinkedIn will now be known as AI Time to Impact.
What does that mean? It means that for every story, we’re listing who we think the end users of the technology being discussed will be and the Time to Impact. Is this news that will make an impact in the short term, the long term, or something in between?
To make that determination, I’m asking between 5 and 10 different AI models, like Chat-GPT, Claude, LLama 2, Perplexity, and more, to assess the time to impact quantitatively and to explain why it scored the story the way it did. I’m then synthesizing all the scores, highlighting any interesting perspectives from the machine, and publishing the stories in that order: short time to impact first. But stick around for the long time to impact! Because the future will be here before we know it.
Thanks to the hundreds of you who subscribed in our first month, and the tens of you who’ve helped spread the word. I’ll remember and appreciate your doing that when (I hope) we’ll have tens of thousands of people reading, and taking action based on, this newsletter.
May it prove useful to you,
AI for Good Course Launched by Andrew Ng
End users: Social Innovators, Public Health Experts, Climate Scientists
Time to impact: Short. "The course is immediately available and targets urgent social challenges. Its broad appeal and practical focus promise rapid adoption and immediate application in various sectors." -GPT-4
Andrew Ng's new course 'AI for Good' applies AI to critical scenarios. Delivered by Apple and Bay Area AI leader Robert Monarch, it provides practical machine learning for significant social projects, including climate change, disaster response, and public health. Suitable for all experience levels, it includes interactive Python code examples. [Learn AI's role in addressing complex challenges]
Phi-1.5 Uses Smaller Models, No Web Data, Outperforms LLMs 5X Its Size
End users: NLP Researchers, Software Developers, Data Scientists
Time to Impact: Short to Medium
Microsoft Research released phi-1.5, a Large Language Model with 1.3 billion parameters (less than 1% the size of GPT-4), trained on LLM-synthesized textbooks instead of web data. Its comparable performance at a fraction of the size should make it a candidate for easier experimentation than larger open source models like the Llama family. [Textbooks Are All You Need II: phi-1.5 technical report ]
HuggingFace Forming Decentralized Research Team
End users: Machine Learning Researchers, Open-Source Contributors
Time to impact: Medium "The creation of a decentralized research team at HuggingFace signifies a shift in the machine learning field." -Poe
UI for AI expert Abubakar Abid is recruiting a Decentralized Research team at HuggingFace to partner with the Machine Learning community. The team will aim to work across various platforms, from Discord servers to large institutional collaborations. This shift from lab-based to decentralized research reflects a broad and meaningful change in the Machine Learning field. [via @abidlabs]
AI's English Bias Challenged by Researchers
End users: NLP Researchers, Developers in non-English speaking countries, and people who speak languages beyond English and Chinese
Time to impact: Medium to long. "Addressing bias requires continued research and adoption by major AI providers, a gradual process." - Claude.ai
Most AI tools are trained on English and Chinese texts, potentially marginalizing other languages. Developers are countering this by creating multilingual models and tuning them for specific languages. Experts like Sara Hooker and Nando de Freitas highlight the need for technology that goes beyond English to avoid permanent bias in new tech. [Axios: The English-first bias of ChatGPT and other AIs could leave the rest of the world out]
AI Firms Pledge to Manage Risks Under Biden-Harris Administration
End users: AI Companies, Policy Makers, General Public
Time to Impact: Medium to long. "The impact of this story depends on the implementation and enforcement of the commitments, as well as the coordination and cooperation among the stakeholders." -Bing
Eight top AI firms, including Adobe, IBM, and Nvidia, have voluntarily committed to managing AI risks under the Biden-Harris administration. They've pledged to focus on product safety, cybersecurity, and robust technical systems. This is part of a wider strategy for responsible AI development, including a forthcoming Executive Order on AI for protecting American citizens. This follows similar pledges from seven other AI companies in July. [FACT SHEET: Biden-Harris Administration Secures Voluntary Commitments from Eight Additional Artificial Intelligence Companies to Manage the Risks Posed by AI]
MIT Technology Review Highlights AI Innovators Under 35
End users: Industry-specific Researchers and Developers, e.g., Pharma, Construction
Time to impact: Medium to long. "These innovations span multiple sectors and are at different stages of maturity. While promising, their broad impact will take time as they navigate technical challenges and market adoption." -GPT-4
A few examples: Connor Coley's ASKCOS AI software helps pharma companies discover and synthesize new molecules. Catherine De Wolf uses AI to reduce emissions and waste in construction by promoting the reuse of old materials. Alhussein Fawzi uses AlphaZero AI to speed up computations, discovering faster matrix multiplication and sorting algorithms. Sharon Li's AI safety feature, out-of-distribution detection, helps AI models identify unknown situations. Sasha Luccioni created a method for tech companies to measure and monitor the carbon footprint of AI language models. [AI Innovators Under 35, 2023]
That's it! More stories tomorrow.