Unrelated to AI:
See my latest Around the Web newsletter; watch my latest video essay at episteme Engine.
[AI/ML + Bio] 🤖 🦠 🧬
[I]: 🧬 GenePT: A simple but hard-to-beat foundation model for genes and cells built from ChatGPT
Summary: The paper introduces GenePT, a foundation model for genes and cells, utilizing ChatGPT embeddings based on literature summaries. GenePT leverages text descriptions of individual genes from the NCBI database and GPT-3.5 to create gene embeddings, which are then used to generate single-cell embeddings. This model shows comparable or better performance in various biological tasks against other single-cell foundation models, without the need for extensive data curation and pretraining, highlighting a simpler and efficient alternative for gene and cell modeling.
Technical details: GenePT is a novel approach that simplifies gene and cell modeling by utilizing GPT-3.5 to process text summaries of genes from the NCBI database. This results in 1536-dimensional gene embeddings. Cell embeddings are generated through two methods: GenePT-w, which averages gene embeddings based on gene expression levels, and GenePT-s, which uses sentence embeddings of genes ordered by expression levels. Unlike traditional models that rely on extensive gene expression datasets and intensive training, GenePT leverages existing literature, making it a simpler and more resource-efficient alternative. Its performance in various biological tasks, such as gene property classification and cell type annotation, is comparable or even superior to more complex foundation models, demonstrating its effectiveness and efficiency in gene and cell modeling. GenePT represents a significant step forward in efficiently utilizing literature-based information for biological modeling.
[II]: 📊What Should Data Science Education Do with Large Language Models?
This paper examines the transformative impact of LLMs on data science education, emphasizing the need for a pedagogical shift to develop skills like creativity and critical thinking alongside AI-guided programming.
[III]: 📄Can large language models provide useful feedback on research papers? A large-scale empirical analysis.
Summary: This study evaluates the effectiveness of LLMs like GPT-4 in generating useful feedback for scientific papers. Through extensive empirical analysis involving comparison with human peer reviews and a survey of researchers, it reveals that LLM-generated feedback can be a valuable resource, especially in contexts where expert human feedback is inaccessible.
Technical details: The study employed GPT-4 to create an automated pipeline for generating feedback on scientific papers. The system parses the full PDFs of papers and constructs a specific prompt combining the paper’s title, abstract, figure, and table captions with main text. This prompt is fed into GPT-4, which generates feedback in a single pass. The effectiveness of this feedback was tested through two approaches: a retrospective evaluation comparing LLM feedback with human peer reviews from 15 Nature family journals and the ICLR conference, and a prospective user study involving 308 researchers. In the retrospective analysis, an extractive text summarization and semantic text matching process was used to evaluate the overlap between LLM and human feedback, showing an average overlap of 30.85% for Nature journals and 39.23% for ICLR. The user study further demonstrated that LLM feedback was deemed helpful by the majority of participants and often provided novel insights not covered by human reviews. However, limitations include the LLM's tendency to focus on certain aspects of feedback and its difficulty in providing in-depth critiques of method design.
[AI X Industry + Products] 🤖🖥👨🏿💻
[I]: 🤖Gemini 1.0
I am not sure there is much of any bigger news now in terms of AI models than Google’s new multi-modal AI model, Gemini. It’s some mind-blowing stuff.
Here is the Gemini announcement video.
Gemini can turn images into code, understand unusual emojis, make sense of natural environment, explain reasoning in math and physics, excels at competitive programming (Alpha Code 2), understand outfits, processing and understanding raw audio, guess movies from scenes, unlocking insights in scientific literature.
[II]: 📹Another AI Video Creator
Pika Arts. I wonder how they compares to runway ML. I have been using runway ML for some time now, actually I opened an account a year ago, and their products, while not the best yet, has been improving. I am currently on the waitlist of Pika Arts, we will see how it goes.
[III]: ℹ Introducing PPLX Online LLMs
Perplexity AI unveils ‘online’ LLMs that could dethrone Google Search? Not sure about that.
[IV]: 👩⚖Write Patents with AI.
Solve Intelligence, a legal tech startup, helps attorneys draft patents for IP analysis and generation.
[V]: 🌄 New Amazon’s AutoML Patent
Automated machine learning pretrained model selector: The patent describes a system and method for automated generation of machine learning models using pretrained models. These pretrained models serve as a starting point, which are then augmented and retrained to meet specific client requirements.
[AI + Commentary] 📝🤖📰
[I]: 💻 What would an LLM OS look like?
Cam Pedersen's essay explores the potential of integrating LLM into operating systems. Pedersen argues that while current LLMs, like ChatGPT, are limited by prompt-based interfaces, the future is likely to be 'agentic,' allowing for more autonomous and decision-making capabilities. The concept of an LLM OS is envisioned as a modular architecture for agentic behavior, with applications far beyond simple chat interfaces, potentially making computers vastly more productive tools for human interaction with information. Pedersen also discusses the technical feasibility of running such models locally ("on the edge") rather than in the cloud, suggesting that devices could house more autonomous, semi-intelligent agents, effectively transforming computers into much more powerful tools.
[II]: 🏢Rebuilding Organization with AI
Prof Mollick’s One Useful Thing Substack publication as established itself as a useful blog in insightful commentaries on the ongoing LLM revolution. This article discusses the evolution of organizational structures in response to technological advances, particularly AI. It highlights historical shifts in organizational design, from the New York and Erie Railroad's 1855 organizational chart to modern-day ongoing AI integration. The article emphasizes the need for contemporary organizations to adapt and rebuild their processes to effectively incorporate AI, drawing parallels to past technological impacts on organizational change.
[III]: 💻Optimizing LLMs for Real World Applications.
This article was a high-level summary of a recent Generative London event. The write up discusses key aspects of deploying LLMs effectively. It covers four main areas: 1) Quantization techniques, focusing on model compression without accuracy loss. 2) Fine-tuning versus prompting, exploring strategies for specialized tasks and the future of these approaches. 3) Total cost of ownership optimization, including model pruning and runtime tweaks for cost reduction. 4) Adapting to the evolving model landscape, emphasizing the need for real-world use case evaluation. I suppose their discussion aims to guide developers and founders in practically implementing generative AI while considering the latest developments.