A Busy AI Day: ChatGPT Power Tip, DARPA, UN, Hybrid AI (8-10-23)

Hello friends, including the many new subscribers who joined this morning after we started talking about this newsletter publicly, like on LinkedIn.

You’ve arrived at a good time! Today has been a big day in AI. We have 9 significant stories today, ranging from the useful to the political to the deeply technical. These stories have risen to the top of our weighted analysis of AI community engagement.

Enjoy!

-Marshall Kirkpatrick, Editor

ChatGPT Mega-instruction Enhances Accuracy of Responses, Says Jeremy Howard

Fast.ai co-founder Jeremy Howard shared a long set of text to add to ChatGPT’s “custom instructions” section, now available to most users, that he says has improved ChatGPT's GPT-4 accuracy, at least for himself. He’s a very well-informed guy and his instructions have been viewed more than 1.2 million times since he shared them last night. The instructions (link below) tell GPT-4 to clarify context before answering queries, and avoid mentioning limitations or ethical concerns that advanced users don’t need to be reminded of. If you use the ChatGPT front end, it could be good to check out. [via @jeremyphoward]

AI Engineer Summit Announces Speakers

The AI Engineer Summit launched a new website and revealed initial speakers for its October event in San Francisco. Aiming to be the biggest community of engineers building with AI, founder @swyx says the event already has more than 10,000 people signed up for the livestream and 600% more than the venue's capacity have applied to attend in person. [AI Engineer Summit]

DARPA Initiates AI Cyber Challenge

DARPA has launched the AI Cyber Challenge (AIxCC), a contest to find and fix software vulnerabilities using AI. So you know this stuff is serious. The contest is in partnership with top AI firms including Anthropic, Google, Microsoft, and OpenAI. The AIxCC has two tracks: Funded and Open, with funding for up to seven small businesses. The aim is to create AI systems to address cybersecurity problems, particularly in critical infrastructure, with final events at DEF CON and Black Hat USA in 2024 and 2025. [DARPA AI Cyber Challenge Aims to Secure Nation’s Most Critical Software]

UN Plans High-level Advisory Body on Artificial Intelligence and International Governance of AI

Does AI need international governance? The UN is forming a high-level AI committee and accepting nominations until August 31st. The committee aims to propose international AI governance recommendations. Nominations from diverse experts, including government, industry, civil society, and academia are encouraged. Several AI community members have highlighted the need for social scientists and civil society experts. [Nominations_vF]

Marcus, Lenat Propose Hybrid AI Approach; Bridging Gap Between LLMs, Classical AI for Trustworthiness

Gary Marcus and Doug Lenat's recent paper discusses the shortcomings of large language models (LLMs) in AI, such as poor reasoning and unpredictability. They propose an alternative AI method involving curated knowledge and rules of thumb to allow logical deductions. They argue that a dependable general AI needs a blend of the LLM and this formal approach. [Getting from Generative AI to Trustworthy AI: What LLMs might learn from Cyc]

Misguided Hiring of ML Experts Criticized

Chris Albon, Director of Machine Learning at the Wikimedia Foundation, criticizes firms for hiring costly machine learning experts without proper support infrastructure, leading to role overload and resignations. Albon calls it "A mistake I see WAY too many times." Peers piled on in agreement. [via @chrisalbon]

Machine Learning Python Cookbook Second Edition Released

The updated "Machine Learning Python Cookbook" by Chris Albon (see previous story) and Kyle Gallatin is now available. It features 200+ standalone recipes for machine learning, designed for those versed in Python, pandas, and scikit-learn. [Machine Learning with Python Cookbook 2, Gallatin, Kyle, Albon, Chris, eBook - Amazon.com]

Amazon Web Services Previews Generative AI Agents in Amazon Bedrock

"Agents for Amazon Bedrock - now in preview - extend foundation models to understand user requests, break down complex tasks, carry on conversations & take actions to fulfill requests." The technology can automate tasks like managing retail orders or processing insurance claims. The initial introduction of Amazon Bedrock and Amazon Titan models was announced in April. [Preview – Enable Foundation Models to Complete Tasks With Agents for Amazon Bedrock | Amazon Web Services]

Torch.compile() Boosts Vision Model Speeds; Up to 30% Faster Inference Across Various Tasks

Merve Noyan has shared benchmarks that she says show that torch.compile() can boost transformers vision models' speed by up to 30% during inference, depending on the model and GPU. The tests were done on different vision models, tasks, hardware, and batch sizes using torch 2.0.1. The Hugging Face guide also provides benchmarking code for tasks like image classification, object detection, and segmentation. [Optimize inference using torch.compile()]

And that’s it for today! Watch this space for more (but probably fewer, 9 is a lot) tomorrow. And please tell a friend that you think could benefit from either general or specific knowledge of these stories.

thanks,

Marshall