Wednesday — January 15, 2025
President issues executive order to bolster U.S. AI infrastructure, MiniMax-Text-01 model excels with 4 million token contexts, and Transformer² proves its self-adaptive prowess across LLM tasks.
News
Executive order on advancing United States leadership in AI infrastructure
The President has issued an order to ensure the development and operation of artificial intelligence (AI) infrastructure in the United States, with the goal of maintaining national security, advancing economic competitiveness, and promoting clean energy technologies. The order sets forth five guiding principles for the development of AI infrastructure, including advancing national security, promoting economic competitiveness, leading in clean energy, maintaining low energy costs, and benefiting workers and communities.
LLM based agents as Dungeon Masters
This research explores the potential of using Large Language Models (LLMs) as Dungeon Masters in Dungeons & Dragons, evaluating their ability to generate narratives, engage with players, and maintain narrative coherence. The study uses a dataset derived from the "Critical Role" series and compares the performance of an LLM-based agent with that of human Dungeon Masters, aiming to assess the LLM's creativity, consistency, and ability to keep players engaged in a narrative plot.
Show HN: Value likelihoods for OpenAI structured output
The structured-logprobs library is an open-source Python tool that enhances OpenAI's structured outputs by providing detailed information about token log probabilities, offering insights into the reliability of a large language model's (LLM) structured outputs. It can be easily installed with pip install structured-logprobs and used to add log probabilities to OpenAI chat completions, providing more context and understanding of the model's responses.
Show HN: Simplex: Automate browser workflows using code and natural language
Simplex is a platform that provides API access, documentation, and a playground for users to run and test code. The platform also features a browser preview and code execution capabilities, with a logo and waiting screen displayed while code is being executed.
Transformer^2: Self-Adaptive LLMs
Transformer² is a novel machine learning system that dynamically adjusts its weights for various tasks, allowing it to adapt to new tasks in real-time and outperform traditional static approaches. This self-adaptive system, inspired by the adaptability of living organisms, has the potential to revolutionize AI research and redefine how we interact with intelligent systems, enabling lifelong learning and continuous change.
Research
"GPT training in U.S. data centers can evaporate 700k L of freshwater"
The water footprint of artificial intelligence models, such as GPT-3, is substantial, with training a single model potentially evaporating hundreds of thousands of liters of freshwater, and global AI demand potentially accounting for billions of cubic meters of water withdrawal by 2027. To address this issue, a methodology to estimate the water footprint of AI models is proposed, highlighting the need to consider water footprint alongside carbon footprint to enable truly sustainable AI.
Transformer-squared: Self-adaptive LLMs
The $\text{Transformer}^2$ framework is a novel self-adaptation method that enables large language models to adapt to unseen tasks in real-time by selectively adjusting their weight matrices. This approach outperforms existing methods, such as LoRA, with greater efficiency and fewer parameters, and demonstrates versatility across different architectures and modalities, including vision-language tasks.
MathReader: Text-to-Speech for Mathematical Documents [pdf]
Traditional text-to-speech (TTS) document readers often struggle with mathematical expressions in academic papers, providing unsatisfactory results. MathReader, a new system that integrates OCR, a fine-tuned T5 model, and TTS, has been proposed to address this issue, demonstrating a significantly lower Word Error Rate (WER) than existing TTS document readers.
Eliza Reanimated: the first chatbot restored
ELIZA, the world's first chatbot, was created by Joseph Weizenbaum at MIT in the 1960s and has been reanimated on a restored CTSS system running on an emulated IBM 7094. The entire stack, including the original MAD-SLIP code and DOCTOR script, is now open source, allowing users with a unix-like OS to run the historic chatbot.
Kajal: Extracting Grammar of a Source Code Using Large Language Models
Kajal is a novel approach that automatically infers grammar from domain-specific language code snippets by leveraging Large Language Models through prompt engineering and few-shot learning. The approach achieves significant accuracy, with 60% accuracy using few-shot learning, and offers a promising solution for automating DSL grammar extraction, with potential for further improvement and validation through future work.
Code
Show HN: A blocklist to remove spam and bad websites from search results
The Bad Website Blocklist is a repository that aims to remove low-quality websites, such as AI-generated articles and spam sites, from search results by providing a blocklist that can be used with the uBlacklist extension. To use the blocklist, users can install the uBlacklist extension, subscribe to the blocklist, and it will automatically stay up to date, removing unwanted websites from search results.
Longest context up to 4M, MiniMax-01 hybrid 456B Open source model
MiniMax-Text-01 is a powerful language model with 456 billion total parameters, featuring a hybrid architecture that combines Lightning Attention, Softmax Attention, and Mixture-of-Experts, allowing it to handle long contexts of up to 4 million tokens. The model demonstrates top-tier performance on various academic benchmarks, including general, reasoning, mathematics, and coding tasks, and outperforms other models in certain long benchmarks such as the 4M Needle In A Haystack Test and LongBench v2.
Show HN: Sculptor – Python library for LLM structured data extraction (MIT)
Sculptor is a tool that streamlines structured data extraction from unstructured text using large language models (LLMs), allowing users to define what data to extract, process at scale, and build multi-step pipelines. It provides a simple schema API, parallel execution, and automatic type validation, making it easy to extract specific fields or classifications from unstructured sources and convert them into structured datasets for quantitative analysis.
Show HN: Open-Source Document Extraction Tool
Rowfill is an open-source document processing platform that utilizes advanced AI capabilities to extract, analyze, and process data from complex documents, images, and PDFs. The platform offers features such as advanced OCR, auto-schema generation, and custom actions, while prioritizing privacy and security through local LLM support and open-source licensing.
Show HN: Panopticon AI – Open-source platform for military AI research
Panopticon AI is an open-source, web-based military simulation platform compatible with OpenAI Gym, aiming to advance military AI research. The project's code is available under the Apache 2.0 license, with full documentation and contribution guidelines available on the platform's website.