Apr 30, 2024

• GPT-4.5 or GPT-5 being tested on LMSYS?

• gpt2-chatbot, exhibiting capabilities beyond known GPT-2 models, was likely running on a server operated by OpenAI, as inferred from API error messages and its sudden unavailability on chat.lmsys.org.

• Deep Dive into XZ Utils Backdoor

• The XZ Utils backdoor was a sophisticated cyber attack targeting Linux systems, involving the insertion of malicious code into the XZ compression library, which is widely used across various Linux distributions.

• Memary: Open-Source Longterm Memory for Autonomous Agents

• memary enhances autonomous agents by enabling long-term memory storage in knowledge graphs, facilitating in-depth user knowledge inference and relevant information retrieval for more meaningful interactions.

• Alice's adventures in a differentiable wonderland

• Simone Scardapane's book, Alice’s Adventures in a differentiable wonderland, serves as an intuitive primer on neural networks, emphasizing differentiable programming and the basics of function optimization through automatic differentiation. Differentiable programming

• Malleable Software in the Age of LLMs (2023)

• Large Language Models (LLMs) are poised to revolutionize end-user programming, enabling even non-programmers to create and modify software tools through natural language interactions, thus democratizing the ability to harness the full power of computing.

• [D] Do Leads in an AI/DS/ML team always have PhDs, is it a requirement?

• Leads in AI/DS/ML teams often hold PhDs, as observed in examples from both a company specializing in evolutionary algorithms and a multi-billion dollar online marketplace.

• TL;DR AI are better community-written abstracts for research papers

• TL;DR AI offers a solution to keep up with AI research without succumbing to the fear of missing out (FOMO).

• Debugging Tech Journalism

• Tech journalism often succumbs to sensationalism and lacks depth, particularly in its coverage of the tech industry, exemplified by the misinterpretation of OpenAI's Q* project as a potential threat to humanity without adequate context or understanding of its capabilities.

• Replacing Judges with Juries: Evaluating LLM Generations with a Panel of Models

• Using a Panel of LLm evaluators (PoLL) composed of diverse, smaller models outperforms a single large model like GPT4 in evaluating the quality of outputs from other LLMs, demonstrating less intra-model bias and greater cost efficiency. Read more

• Tour of CLIPS (2022)

• CLIPS, a rule-based programming language developed by NASA in the mid-'80s, facilitates focusing on domain business logic over software execution mechanics, leveraging the Rete algorithm for efficient rule execution order and caching common calculations.

• Level of Gaussians: Real-Time View Synthesis for Millions of Square Meters

• The Level of Gaussians (LoG) approach significantly enhances the real-time rendering of large-scale scenes by efficiently managing the detail levels of Gaussians based on the viewer's perspective.

• [R] NExT: Teaching Large Language Models to Reason about Code Execution

• NExT introduces a method to enhance Large Language Models' (LLMs) ability to understand and reason about code execution by teaching them to inspect execution traces and reason with chain-of-thought rationales, addressing the gap in semantic understanding of program run-time behavior. Read the paper

• [D] How can attention mechanisms retrieve meaningful information over long distances when using RoPE or ALiBi?

• RoPE and ALiBi mechanisms inherently penalize distant token relationships, potentially obscuring valuable information located far within the token sequence.

• [R] CRISPR-GPT: An LLM Agent for Automated Design of Gene-Editing Experiments

• CRISPR-GPT is an AI-powered tool designed to streamline the CRISPR-based gene editing experiment design, integrating LLMs with domain-specific knowledge and external tools for comprehensive support.

• Replacing Judges with Juries: Evaluating LLM Generations with a Panel of Diverse Models

• Using a Panel of LLm evaluators (PoLL) composed of diverse, smaller models outperforms a single large model like GPT4 in evaluating the quality of outputs from other LLMs, addressing the challenge of intramodel bias and the difficulty in accurately evaluating LLMs.