Godmother of AI Warns Against Regulation, Apple Intelligence System Prompts, Faster and Cheaper AI, and More
Machine learning resources and updates 8/12/2024
Here are the most important machine learning resources and updates from the past week. Follow me on X and Threads for more frequent posts and updates.
Society's Backend is reader-supported. Support the Society's Backend community for just $1/mo to get the full list each week. Thanks to all paying subscribers! 😊
Last week’s updates:
‘Godmother of AI’ warns California bill SB 1047 could stifle innovation and harm academia
‘You are a helpful mail assistant,’ and other Apple Intelligence instructions - The Verge
Understanding the best practices and ideas for LLM-enabled RAG systems[RAG]
Structured generation hurts LLM reasoning performance (Paper Explainer)
Interesting Content in AI, Software, Business, and Tech- 08/07/2024 [Updates]
Hollywood’s Divide on Artificial Intelligence Is Only Growing
Interviewing Ross Taylor on LLM reasoning, Llama fine-tuning, Galactica, agents
Google is set to supercharge Google Home with Gemini intelligence - The Verge
A new generation of African talent brings cutting-edge AI to scientific challenges
Palantir jumps 11% on Microsoft partnership to sell AI to U.S. defense, intel agencies
Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters
‘Godmother of AI’ warns California bill SB 1047 could stifle innovation and harm academia
Dr. Fei-Fei Li, the "Godmother of AI," warns that California's AI bill SB 1047 could harm innovation and academic research. She argues that the bill's requirements are overly restrictive and could stifle open-source development and collaboration. Li encourages policymakers to adopt more balanced AI regulations that promote growth and address real issues.
‘You are a helpful mail assistant,’ and other Apple Intelligence instructions - The Verge
Apple's latest developer betas include AI features with pre-prompt instructions stored on Macs. These instructions guide AI tools to avoid negative themes and hallucinations. Users discovered these backend prompts in the macOS 15.1 beta files, revealing how Apple’s AI functions operate.
🥇Top ML Papers of the Week
This week's top ML papers cover advancements in various domains, including real-time object segmentation, structured generation's impact on LLM reasoning, and the use of LLM-based agents in software engineering. Tools like Transformer Explainer and frameworks like RAGFoundry enhance understanding and performance of LLMs. Novel approaches such as Self-Taught Evaluators and Conversational Prompt Engineering improve model accuracy and personalization.
How I Use "AI"
The author uses AI language models to help with programming tasks, saving time and effort. They find these models especially useful for generating initial code and handling simple scripts. Despite being a skilled programmer, the author prefers the convenience of AI for quick solutions.
What Makes Machine Learning so Hard for Software Engineers
Machine learning is hard for software engineers because it requires a different mindset, embracing uncertainty and statistical thinking. Unlike traditional programming, ML involves extensive experimentation and lacks clear success criteria and time frames. Software engineers must balance their existing skills with new ML concepts, focusing on one area at a time to avoid feeling overwhelmed.
Understanding the best practices and ideas for LLM-enabled RAG systems[RAG]
Retrieval Augmented Generation (RAG) uses AI to search a knowledge base, retrieve relevant information, and feed it to a language model to answer queries. RAG improves information extraction by speeding up searches and reducing retraining costs. The article explores best practices for building effective RAG systems and will break down complex topics in a series of posts.
Gemini 1.5 Flash price drop, tuning rollout complete and improvements to Gemini API and Google AI Studio
Google has significantly reduced the costs for using Gemini 1.5 Flash and expanded its API to support 100+ languages. Google Workspace users now have easier access to AI Studio, and developers can tune the Gemini 1.5 Flash model for better performance. Additionally, the Gemini API and AI Studio now support PDF understanding, and documentation has been improved for a better user experience.
Achieving Human Level Competitive Robot Table Tennis
Researchers from Google DeepMind have developed a robot capable of playing table tennis at a human competitive level. The project involved teams working collaboratively to achieve this feat. They have shared detailed results, including match videos, to highlight their progress.
A recipe for frontier model post-training
The traditional methods of Reinforcement Learning from Human Feedback (RLHF) are becoming outdated, with newer techniques emphasizing synthetic data and extensive data curation. The latest approaches prioritize multiple training rounds, data filtering, and leveraging model-generated instructions to enhance performance. Industry leaders like Meta and OpenAI are focusing on these new methods to improve model quality significantly.
Introducing Structured Outputs in the API
OpenAI's new Structured Outputs feature ensures API-generated outputs strictly match developer-supplied JSON Schemas. This improves reliability by constraining models to valid schema tokens and training them on complex schemas. Safety policies remain intact, with Structured Outputs available for models supporting function calling.
Keep reading with a 7-day free trial
Subscribe to Society's Backend: Machine Learning for Software Engineers to keep reading this post and get 7 days of free access to the full post archives.