. . AI: A growing-up day for AI (12.7.23)

Author

Marshall Kirkpatrick
December 07, 2023

Friends, here are the stories the AI community is discussing most today. Particularly notable is the fun research from Anthropic that finds improved performance if you end your AI prompt with "Here is the most relevant sentence in the context:”

There's something about these stories today that feels like an air of maturation: cybersecurity and bias resources from Meta and Anthropic, an edge model from Stability, knighthood for Yann LeCun!

I hope these stories are useful and interesting to you. As always, this is a data-informed view of what’s hot among AI specialists today. Thanks to everyone who’s been sharing this growing newsletter with friends and co-workers!

First impacted: The French AI ecosystem
Time to impact: Medium

Meta AI leader Yann LeCun has been made a Chevalier de la Légion d'Honneur (Knight of the Legion of Honor) by French President Macron at the Élysée Palace. The French Legion of Honor was a meritocratic system of recognition established by Napoleon Boneparte in 1802. French AI is a pretty big deal. [via @ylecun] Share by email

First impacted: AI developers, Cybersecurity professionals
Time to impact: Medium to long

Meta AI has initiated a project called Purple Llama, aimed at creating a fair platform for the creation of secure and ethical generative AI applications. The project offers licensed tools, assessments, and models for cybersecurity risk monitoring and filtering LLM input and output. [Announcing Purple Llama: Towards open trust and safety in the new world of generative AI] Share by email

First impacted: AI developers, Edge device users
Time to impact: Medium

Stability AI has introduced StableLM Zephyr 3B, a chat model that they claim is 60% smaller than 7B models, making it suitable for edge devices. According to benchmarks on platforms such as MT Bench and AlpacaEval, the model reportedly excels in generating contextually relevant text and can efficiently manage multiple linguistic tasks. [Introducing StableLM Zephyr 3B: A New Addition to StableLM, Bringing Powerful LLM Assistants to Edge Devices — Stability AI] Share by email

First impacted: AI researchers, AI model users
Time to impact: Medium

Anthropic's latest AI model, Claude 2.1, is said to have shown a 30% drop in wrong answers and fewer false document support claims, according to a blog post from the company. Additionally, if the following is added to the end of a long prompt, Anthropic says, it makes a huge difference in the model's willingness to answer a question based on a single sentence: “Here is the most relevant sentence in the context:” Prompting: it's all about narrowing the focus. [Long context prompting for Claude 2.1] Share by email

First impacted: AI developers, Bias auditors in AI industry
Time to impact: Short

Anthropic has introduced a new dataset to evaluate bias in language models across 70 potential uses, including loan applications, visa approvals, and security clearances. The company reports that it has utilized this dataset to detect discrimination patterns and devise strategies to reduce them. The study reportedly found both positive and negative bias in the Claude 2.0 model under certain conditions, and demonstrated methods to significantly reduce both forms of bias. [Evaluating and Mitigating Discrimination in Language Model Decisions] Share by email

First impacted: Information seekers
Time to impact: Short

AI firm Perplexity.ai today marked its first year serving users. Founder Aravind Srinivas credits the company's success to their focus on providing answers, not just links. While I too like using Perplexity, it's not like extracting answers from web pages and displaying them without ads is something Google simply hadn't thought of yet. [via @AravSrinivas] Share by email

"Language Model Development" Emerges as New AI Subfield

First impacted: AI researchers, AI developers
Time to impact: Long

AI researcher Colin Raffel suggests that "language model development" has become a new AI subfield, with its own assumptions and challenges separate from deep learning. In a widely-shared blog post Raffel outlines the main issues in this subfield, such as managing the model's actions, enhancing its efficiency, and incorporating it into other systems. [Language Model Development as a Subfield?] Share by email

That’s it! More AI news tomorrow.