A 28.9 million parameter language model is running on an $8 ESP32-S3 microcontroller.
The model operates entirely on the chip, generating text at about 9 tokens per second and displaying it on a small screen. This achievement is possible because most of the model's parameters are stored in flash memory, using Google's Per-Layer Embeddings technique.
A recent exploration into optimizing Postgres LISTEN/NOTIFY-backed data streams has demonstrated scalability, achieving 60,000 writes per second on a single Postgres server with millisecond-scale latency.
LISTEN/NOTIFY allows for efficient notification-based streaming, avoiding the resource-intensive polling method. The initial implementation faced throughput limitations due to a global lock during NOTIFY, capping performance at 2.
FLUX 3 x Mimic is a new video-action model developed in collaboration with Mimic Robotics and deployed at Audi.
FLUX 3 is a multimodal foundation model, trained jointly across images, video, and audio. Over 95% of the compute cost is dedicated to video prediction, requiring the model to learn motion, contact, and cause and effect to generate realistic videos.
Anthropic has introduced Claude Opus 5, a new model that enhances the Opus tier for long-running agents and improves coding and professional work.
Claude Opus 5 is positioned as a cost-effective alternative to Claude Fable 5, offering comparable intelligence at half the price. It sets new benchmarks in coding and knowledge work evaluations such as Frontier-Bench and GDPval-AA, although it trails Mythos 5 in cybersecurity tasks.
AI companies are reportedly concealing a massive amount of debt, as they invest billions in large data centers for complex AI models.
A recent investigation by Nikkei Asia reveals that five major US tech companies — Alphabet, Microsoft, Amazon, Meta, and Oracle — have an estimated $1.65 trillion in off-balance-sheet debt. This figure surpasses the $1.35 trillion in debt they officially reported for the last quarter. Meta alone accounts for about $420 billion of this hidden debt, drawing parallels to Enron's notorious financial collapse in 2001.
Copilot now charges for usage based on the listed API rates.
This change prompts a comparison between accessing models directly and utilizing Copilot’s integrated coding workflow. The shift involves evaluating the coding processes, policies, and harness work surrounding Copilot. Andrea, a Senior Developer Advocate at GitHub, highlights these aspects, bringing her extensive experience in developer tools to the discussion.
GigaToken supports a wide range of CPU hardware and is compatible with most commonly used tokenizers. It can be used with its own API for maximum performance or in compatibility mode with HuggingFace Tokenizers and Tiktoken.
An experiment was conducted to investigate if AI labs are optimizing their models specifically for Simon Willison's pelican-riding-a-bicycle benchmark.
Over 1,000 SVGs were generated using seven advanced models and evaluated by an LLM judge. The analysis utilized Claude Fable 5 and included a grid of 48 prompts, combining 8 animals and 6 vehicles. The pelican-on-a-bicycle prompt is part of this grid and has become a well-known informal benchmark in AI discussions, especially on platforms like Hacker News.
GitHub is restructuring its bug bounty program to enhance the experience for researchers collaborating with its team.
The changes aim to refine the program's focus, though specific details on the new structure or incentives have not been disclosed. This marks a shift in GitHub's approach towards engaging with security researchers. The program's evolution reflects GitHub's commitment to improving interactions and outcomes in its bug bounty efforts.
OpenAI introduced OpenAI Presence, an enterprise AI agent platform for deploying voice and chat agents.
This platform supports organizations in implementing AI-driven agents for both customer service and internal processes. It is designed to integrate seamlessly with existing workflows, ensuring reliable operation. OpenAI Presence is specifically tailored to meet enterprise needs, providing a robust framework for AI agent deployment.