Here’s a comprehensive AI reading list from this past week. Thanks to all the incredible authors for creating these helpful articles and learning resources.
I put one of these together each week. If reading about AI updates and topics is something you enjoy, make sure to subscribe.
Society’s Backend is reader supported. You can support my work (these reading lists and standalone articles) for 80% off for the first year (just $1/mo). You’ll also get the extended reading list each week.
A huge thanks for all supporters. ?
What’s Happened this Past Week
If you want a good overview of this past week in AI you can check out:
Charlie Guo’s weekly AI Roundup. His newsletter is worth a sub.
’s weekly Roundup at AI Tidbits. Also worth a sub.
The Last Week in AI Podcast by .
The Batch’s weekly newsletter is also a good resource for understanding what’s happening in AI and how it’s impactful.
Also worth knowing:
Papers
ML papers are difficult to keep up with. Here’s the week’s NotebookLM-generated podcast going over important papers you should know:
If you prefer a written overview, check out this The Top ML Papers of the Week by .
Last Week’s Reading List
Reading List
Removing selection bias from evaluation of recommendations
Causal machine learning helps evaluate the effectiveness of Amazon's Fulfillment by Amazon (FBA) recommendations for sellers. To eliminate selection bias, a method called double machine learning is used, which analyzes seller decisions and outcomes simultaneously. This approach allows Amazon to accurately measure how following FBA recommendations impacts seller performance.
Meta releases ‘Self-Taught Evaluator’ AI model to reduce human involvement in AI development
Meta has launched a new AI model called the “Self-Taught Evaluator,” which reduces the need for human input in AI training. This model learns from AI-generated data and can improve itself by analyzing its own mistakes. By moving towards fully autonomous AI, Meta aims to create digital assistants that can perform complex tasks without human help.
A Technical Perspective: Has Google Search Gotten Worse?
The article discusses how Google Search and Google Ads work together, using machine learning to provide users with relevant search results. Many users feel that the quality of Google Search has declined, leading some to switch to alternative search engines. However, the author believes that when properly implemented, Search Ads can enhance the user experience by creating a second, more relevant search feed.
96% Accuracy: Harvard Scientists Unveil Revolutionary ChatGPT-Like AI for Cancer Diagnosis
Harvard scientists have developed an advanced AI model called CHIEF that can accurately diagnose and predict outcomes for various types of cancer. This model outperforms existing AI systems by analyzing tumor images to detect cancer cells, identify genetic profiles, and forecast patient survival. CHIEF's versatility allows it to assist in multiple diagnostic tasks, making it a promising tool for enhancing cancer treatment.
A Selective Survey of Efficient Speculative Decoding Techniques for LLM Inference
By
Speculative decoding improves the efficiency of large language models (LLMs) by allowing multiple tokens to be predicted in a single forward pass, rather than requiring multiple passes. The Medusa architecture enhances this process by adding multiple prediction heads to the base LLM, enabling faster token generation. This approach reduces compute costs and increases throughput by allowing the main model to process several tokens in parallel.
What is AI, and How Do We Govern It?
By
Dean W. Ball and Daniel Kokotajlo emphasize the importance of transparency in AI development and the government's role in creating digital public infrastructure. They argue that instead of just regulating AI, the government should focus on building capabilities that support safety, reliability, and innovation. Ball encourages young people to embrace AI as a tool for creativity and problem-solving, urging them to think about what they can build rather than worrying about job displacement.
Machines of Loving Grace1
AI has the potential to greatly accelerate advancements in biology and medicine, allowing us to achieve decades of progress in just a few years. This could lead to significant improvements in health and quality of life, especially in the developing world. However, there are concerns about inequality and the misuse of AI, which need to be addressed to ensure its benefits are shared broadly.
Building on evaluation quicksand
By
The article discusses challenges in evaluating language models, highlighting issues like contamination and the need for standardized evaluation practices. It emphasizes that many AI labs customize evaluations to fit their needs, making comparisons between open and closed models difficult. Ultimately, establishing common evaluation standards is crucial for transparency and trust in the open-source AI community.
How Uber Manages Petabytes of Real-Time Data
By
Uber's real-time data infrastructure processes vast amounts of data daily, supporting features like customer incentives and fraud detection. It uses technologies like Apache Kafka, Flink, and Pinot to ensure quick and reliable data processing across its global operations. This advanced system allows Uber to make fast decisions and adapt to changes efficiently.
From Features to Performance: Crafting Robust Predictive Models
This guide focuses on transforming raw data into effective predictive models through feature engineering and model training. It covers important techniques for selecting and preparing data, choosing the right algorithms, and evaluating model performance. By mastering these steps, you can improve your data science projects and gain valuable insights from your data.
