π€ AI Funding - Supporting DeepIndaba Attendees
This article discusses the funding disparity in the AI field and highlights the ML Collective's efforts to support DeepIndaba attendees. The ML Collective is raising funds for the third year in a row to help attendees who may not have access to sufficient resources.
Key Points:
β’ Addresses the funding gap in AI, focusing on underrepresented researchers.
β’ Supports attendees of the DeepIndaba conference.
β’ Third consecutive year of fundraising efforts.
π Resources:
β’ ML Collective β - AI community supporting researchers
β’ DeepIndaba β - African AI conference
β’ Nahid Alam β - Contributor to the initiative
β’ SavvyRL β - Contributor to the initiative
Image
π€ LLM Bounding Box Output - API Efficiency
This article describes the challenges of obtaining LLM bounding box output using APIs and highlights a solution that simplifies the process. It focuses on overcoming difficulties with argument order, formatting, range, and normalization.
Key Points:
β’ Simplifies the process of obtaining LLM bounding box output.
β’ Automates the handling of complex API parameters.
β’ Saves significant time and effort in development.
Image
π€ AI Talent Acquisition - Meta's High-Profile Hire
This article reports on Meta's recruitment of Ruoming Pang, formerly of Apple's AI models team, for a compensation exceeding $200 million. It discusses the implications of this large compensation package in the context of current industry salaries.
Key Points:
β’ Meta hired Ruoming Pang with a significant compensation package.
β’ Apple did not match the offer.
β’ Highlights current salary trends in the AI field.
π€ High-Performance Computing - QuACK Kernel Library
This article introduces QuACK, a new memory-bound kernel library for SOL written in Python using CuTe-DSL. It emphasizes QuACK's performance compared to existing libraries like PyTorch's torch.compile and Liger.
Key Points:
β’ Achieves 33%-50% faster performance than optimized libraries.
β’ Implemented entirely in Python using CuTe-DSL.
β’ Operates on H100 with 3TB/s memory bandwidth.
π Resources:
β’ Wentao Guo β - Principal developer of QuACK
β’ Ted Zadouri β - Contributor to QuACK
β’ Tri Dao β - Contributor to QuACK
Image
π€ Deep Learning Education - NVIDIA DLI Course at Flatiron Institute
This article reports on NVIDIA's Deep Learning Institute (DLI) delivering a two-day course on deep learning fundamentals and data parallelism in PyTorch to researchers and interns at the Flatiron Institute.
Key Points:
β’ Provided training on deep learning fundamentals and PyTorch.
β’ Focused on data parallelism techniques.
β’ Addressed researchers and summer interns at the Flatiron Institute.
π Resources:
β’ NVIDIA AI Dev β - Provider of the DLI course
Image
Image
π‘ Book Recommendation - The Three-Body Problem Series
This article is a personal recommendation of the "Three-Body Problem" novel series by Liu Cixin. The author shares their enjoyment of the series and encourages others to read it without spoilers.
Key Points:
β’ Highly engaging science fiction series.
β’ Author recommends reading without spoilers.
β’ Enjoyed during a period of injury recovery.
π Edge AI - Orange Pi 5 Plus Object Detection
This article discusses using an Orange Pi 5 Plus for AI object detection at the edge. It details how to leverage Metis M.2 and Voyager SDK to achieve this.
Key Points:
β’ Enables standalone edge AI object detection.
β’ Uses Metis M.2 and Voyager SDK.
β’ Runs on Orange Pi 5 Plus.
π Resources:
β’ Axelera AI β - Provider of Metis M.2 and Voyager SDK
β’ Guide on how to turn your OPi5 into a standalone edge AI device β
Image
π Space Launch - Rocket Lab's Production Facility Tour
This article describes a tour of Rocket Lab's production line for Electron and HASTE launch vehicles, hosted for General Stephen Whiting of U.S. Space Command.
Key Points:
β’ Hosted a tour for U.S. Space Command.
β’ Showcased Electron and HASTE launch vehicles.
β’ Supports both national security and commercial missions.
π Resources:
β’ Rocket Lab β - Space launch company
Image
Image
Image
π€ LLM Evaluation - Detecting "Thinking" in LLMs
This article discusses a new ICML paper exploring methods to distinguish between LLMs merely reciting memorized text and exhibiting genuine "thinking."
Key Points:
β’ Explores methods for evaluating LLM "thinking."
β’ Differentiates between memorized text and genuine understanding.
β’ Identifies when a model performs interesting tasks.
π Resources:
β’ Florian Gallwitz β - Author of the ICML paper
β’ Robert Csordas β - Co-author of the ICML paper
β’ JΓΌrgen Schmidhuber β - Co-author of the ICML paper
β’ iDivinci β - Related research
Image
π€ LLM Output - Addressing Unexpected Output Format
This article describes a user's experience with unexpected LLM output formatting, specifically a mismatch between expected and actual output grid size.
Key Points:
β’ Encountered unexpected 3x3 output format.
β’ Issue resolved by understanding output grid size selection.
β’ Highlights the importance of task parameters in LLM usage.
Image
βοΈ Support
If you liked reading this report, please star βοΈ this repository and follow me on Github β, π (previously known as Twitter) β to help others discover these resources and regular updates.
