Gödel, Escher, Bach, and AI (2023)

Douglas Hofstadter, author of Gödel, Escher, Bach, warns that despite the impressive capabilities of large language models like ChatGPT, they cannot replace the authentic and reflective voice of a thinking, living human being. He illustrates this point by sharing an experience where a state-of-the-art language model, GPT-4, was asked to compose an essay about why he wrote Gödel, Escher, Bach, resulting in a piece that, although well-written, lacked the depth and nuance of a human author.

Read more

LearnLM: An experimental task-specific model for learning science principles (2024)

LearnLM is an experimental task-specific model trained to align with learning science principles, capable of inspiring active learning, managing cognitive load, adapting to the learner, stimulating curiosity, and deepening metacognition. The model can be used with specific system instructions to achieve various learning goals, such as test preparation, teaching new concepts, rewriting text for easier reading, and guiding students through learning activities, all while incorporating key learning science principles like adaptivity, active learning, and stimulating curiosity.

Read more

AI Scientists Are Told to Remove 'Ideological Bias' from Powerful Models

The National Institute of Standards and Technology has issued new instructions to AI scientists that remove mention of "AI safety," "responsible AI," and "AI fairness" and instead prioritize reducing "ideological bias" to promote "human flourishing and economic competitiveness." The change has raised concerns among researchers that it may lead to the development of discriminatory AI models that harm marginalized groups, as the new directive appears to downplay the importance of addressing biases and misinformation in AI systems.

Read more

Inside Elon Musk's 'Digital Coup'

Elon Musk's loyalists at DOGE have infiltrated dozens of federal agencies, pushing out tens of thousands of workers and siphoning millions of people's sensitive data, as part of a "digital coup" to overhaul the US government. With access to vast amounts of data, including investigative files, tax records, and medical histories, Musk's operatives are now poised to unleash AI on the government's systems, furthering their goal of radically transforming the federal bureaucracy.

Read more

How Pickle Files Backdoor AI Models

Python's pickle module has a major security flaw as it can execute arbitrary code when loading data, making it risky to handle untrusted files. The pickle module operates like a mini interpreter, executing its own opcodes from a binary stream, and its ability to invoke arbitrary functions makes it inherently unsafe, allowing malicious pickle files to be used to exploit vulnerabilities.

Read more

Operationalizing Machine Learning: An Interview Study

Organizations rely on machine learning engineers to deploy and maintain ML pipelines in production through a process known as MLOps, which involves a continuous loop of data collection, experimentation, evaluation, and monitoring. Researchers conducted interviews with 18 machine learning engineers to identify key challenges and best practices, revealing three crucial variables for success: Velocity, Validation, and Versioning, and highlighting areas for improvement in tool design.

ArXiv

AI and the value of privacy-preserving tools to distinguish who is real online

The increasing capabilities of AI have made it easier for malicious actors to conduct online fraud and deception, highlighting the need for a balance between anonymity and trustworthiness online. Personhood credentials, digital credentials that verify a user is a real person without disclosing personal information, offer a potential solution to this challenge, providing a way for individuals to demonstrate their trustworthiness and for service providers to reduce misuse by bad actors.

ArXiv

Machine Learning Operations (MLOps): Overview, Definition, and Architecture

The goal of industrial machine learning projects is to develop and deploy ML products quickly, but this is often hindered by the challenges of automating and operationalizing them, which is where Machine Learning Operations (MLOps) comes in. Through research and expert interviews, a clearer understanding of MLOps is established, including its principles, components, and roles, providing guidance for ML researchers and practitioners to successfully automate and operate their ML products.

ArXiv

AutoHete: An Automatic and Efficient Heterogeneous Training System for LLMs

Transformer-based large language models have shown impressive capabilities, but their training is limited by GPU memory constraints, and existing heterogeneous training methods introduce significant overheads. The proposed AutoHete system addresses these issues by dynamically adjusting training parameters and using a priority-based scheduling mechanism, resulting in a 1.32x-1.91x throughput improvement over state-of-the-art systems.

ArXiv

A Proof of the Collatz Conjecture

A new fixed point theorem in metric spaces is presented, and its application to the Collatz conjecture is explored. The theorem is utilized to examine the Collatz conjecture, potentially shedding new light on this longstanding mathematical problem.

ArXiv

OmniAI: A unified Ruby API for integrating with AI providers

OmniAI is a unified Ruby API for integrating with multiple AI providers, including Anthropic, DeepSeek, Google, Mistral, and OpenAI, offering a consistent interface for features like chat, text-to-speech, and embeddings. It allows for effortless switching between providers, making integrations more flexible and reliable, and provides examples and documentation for various use cases, including chat, text-to-speech, speech-to-text, and embeddings.

Repo

As an experiment I "vibe coded" this CLI tool to generate better context for AI

mkctx is a CLI tool that prepares code for AI interactions by creating a formatted directory tree and extracting file contents in a clean, structured format. The tool, developed using "vibe coding" with AI language models, allows users to generate context from their codebase, filter files, and provide custom instructions for large language models (LLMs) like Claude.

Repo

Show HN: Basic Memory – Build a knowledge graph from Claude conversations

Basic Memory is a tool that allows users to build persistent knowledge through natural conversations with Large Language Models (LLMs) like Claude, storing information in simple Markdown files on their local computer. It enables bi-directional communication between humans and LLMs, allowing them to read and write to the same files, and creates a traversable knowledge graph that can be navigated semantically.

Repo

Twitter's Navi: High-Performance Machine Learning Serving Server in Rust (2023)

Twitter's Recommendation Algorithm is a set of services and jobs that serve feeds of Tweets and other content across various product surfaces, including the For You Timeline and Recommended Notifications. The algorithm uses a combination of data, models, and software frameworks, such as SimClusters, TwHIN, and GraphJet, to rank and filter content, and is open-sourced to allow the community to contribute and suggest improvements.

Repo

AI challenge: StructEnv, when donet meets JSON

The StructEnv format is a proposed configuration format that aims to balance simplicity and structure, providing a human-readable and machine-parseable syntax for structured configuration data. The format uses a KEY=VALUE pair syntax, supports nesting, arrays, and objects, and includes features such as type inference, key escaping, and multiline strings, with the goal of being compatible with common environment variable formats.

Repo