Search the archive

NOUS HERMES PRO (HUGGING FACE HUB)

The next version of Hermes is available, trained on top of Llama 3 with a newly cleaned dataset. This model excels at function calling and would make a useful agent.

DATABONSAI

Databonsai is a Python library that uses LLMs to perform data cleaning tasks.

GPT-4 CAN EXPLOIT REAL VULNERABILITIES BY READING ADVISORIES

OpenAI's GPT-4 can autonomously exploit certain real-world vulnerabilities with a high success rate when using CVE descriptions, potentially surpassing tools like vulnerability scanners.

The window to invest is closing - learn more here.

META RELEASES EXECUTORCH FRAMEWORK FOR LLM ON EDGE DEVICES

Meta's ExecuTorch Framework is a post-training quantization toolkit that enables running Llama models on various iPhone and Galaxy devices. It is able to get as much as 11 tokens per second on mobile phones running 7B sized language models.

APPLE POACHES AI EXPERTS FROM GOOGLE, CREATES SECRETIVE EUROPEAN AI LAB

Apple has aggressively expanded its artificial intelligence capabilities by hiring numerous AI experts from Google and establishing a secretive AI research lab in Zurich, Switzerland. The lab, known as the "Vision Lab," focuses on developing advanced AI models that integrate text and visual inputs.

TEXT RETRIEVAL WITH FASTER EMBEDDING SEARCHES

Seismic is an innovative approach to organizing inverted indexes that significantly enhances the speed and accuracy of text retrieval using learned sparse embeddings.

EXPLORING MAMBA: A VISUAL FOUNDATION MODEL FOR ADVANCED COMPUTER VISION

The Mamba model is a state-of-the-art approach that excels in handling long sequences without the computational drawbacks of traditional Transformers.

ADAPTING GRAPH NEURAL NETWORKS TO CHANGING STRUCTURES

The Cluster Information Transfer (CIT) mechanism is a new method designed to enhance the adaptability of Graph Neural Networks (GNNs) to different and changing graph structures.

>> June 25, San Fransisco. Register here

CUDA CHECKPOINTING

A new toolkit from Nvidia that allows checkpointing the CUDA state for transfer and restarting. It is useful for distributed training of very large AI models.

FRIENDS DON'T LET FRIENDS MAKE BAD VISUALIZATIONS

Visualizing eval scores for models is key in today's AI research world. However, many charts are bad at conveying the intended information. This repository shows some great charts and comes with dos and don'ts for visualizing results.

META-PROMPTING IN LLMS

Meta-prompting is a technique that turns a single language model into a multi-skilled team. This method breaks down complex tasks into simpler parts, handled by specialized instances of the same model, leading to significantly improved performance across various tasks.

SPEEDING UP LARGE LANGUAGE MODELS WITH QUICK

QUICK is a new set of optimized CUDA kernels that boosts the efficiency of large language models by solving memory issues. This leads to a notable speed increase on NVIDIA GPUs, making AI models faster and more efficient.

DIDDO'S NEW FUNDING WILL BRING ITS SHOPPABLE TV AI API TO STREAMING PLATFORMS

Diddo is a CV/AI API for streaming and media companies that can make their content instantly shoppable without the use of QR codes or second screens, opening up new revenue streams for them. Diddo's API keeps the purchase capabilities on platform and 100% native so it is truly an end-to-end commerce solution for video owners.

COGNITION SEEKS $2 BILLION VALUATION FOR AI CODE-WRITING TOOL

Cognition Labs, which is transitioning from crypto to AI, is seeking a $2 billion valuation for its code-writing AI tool, Devin. Amidst a trend of soaring valuations in AI startups, Cognition's success reflects the sector's growth and the importance of substantial investments in data and computational resources. Key players like Google and Microsoft continue to advance aggressively in AI with large language models.

See the curriculum

APPLE TO UNVEIL AI-ENABLED SAFARI BROWSER

Safari 18 will introduce UI enhancements, an AI-powered Intelligent Search for content summarization, and a 'Web Eraser' for advanced content blocking.

PHOSPHO

A text analytics platform for LLM apps.

FLOWTESTAI

FlowTestAI stands as the world's first GenAI-powered OpenSource Integrated Development Environment (IDE) specifically designed to craft, visualize, and manage API-first workflows.

COPILOT WORKSPACE

GitHub has announced a new project to allow language models to somewhat autonomously finish features and fix bugs.

OPENAI INTRODUCES “MEMORY” FEATURE FOR CHATGPT PLUS USERS

OpenAI has launched a new "Memory" feature for ChatGPT Plus users, allowing the AI to retain information across conversations, which can be managed via settings.

EFFORT ENGINE

Effort enables real-time adjustment of computations during LLM model inference on Apple Silicon chips, striking a balance between speed and quality. The method, which requires no retraining but requires conversion and precomputation, allows models to operate faster by loading fewer weights. The developers are seeking assistance from Swift/Metal engineers to optimize the implementation, which is available for download on GitHub.

OPEN GATO REPLICATION: JAT

GATO by DeepMind was touted to be a generalist agent. A team associated with Hugging Face has trained and evaluated JAT, which is a Jack-of-All-Trades model that performs reasonably on a wide variety of tasks.

FP6-LLM

Reducing floating point precision speeds up training but can be unstable. This paper shows that you can get full tensor core utilization in a new packing format without common instabilities or slowdowns from naive attempts.

STAR-CODER 2 15B INSTRUCT

This model is trained on synthetic and human data. Permissively licensed, it achieves 72.6 on humaneval. Its developers give great insights into how to replicate their data pipeline and apply principles to other problems where synthetic data might be useful.

path. Sign up free and start learning

WHISPER CPP CLI

A fully self-contained speech-to-text system built on top of Whisper by the folks who built Ruff.

MEMARY

Memary allows your agents to store a large corpus of information in knowledge graphs, infer user knowledge through our memory modules, and only retrieve relevant information for meaningful responses.

MISTRAL.RS

A fast LLM inference platform that supports inference on a variety of devices, quantization, and easy-to-use applications with an OpenAI API-compatible HTTP server and Python bindings.

JAX INTERNALS

Autodidax tutorial that shows the inner ideas in the growing Jax framework. It also introduces helpful jargon often used in the community.

AI ISN'T USELESS. BUT IS IT WORTH IT?

This article provides a critical examination of AI and machine learning, arguing that while AI tools can be useful for certain tasks, they often fail to live up to the grand promises made by AI companies. It highlights the significant gap between AI's practical use cases, like coding assistance and language generation, and the inflated expectations set by industry narratives. The article raises concerns about the ethical and environmental costs of AI development, along with the potential for misuse and overreliance on the technology.

BINDING PUBLIC SECTOR AI DIFFUSION

The OMB's new AI executive order policy aimed at the public sector could severely impede AI innovation due to stringent safety provisions and bureaucratic hurdles. Implemented amidst shrinking IT budgets, the regulations may not only slow AI deployment but also potentially halt projects crucial for modernizing government services. Critics worry that these constraints, coupled with funding cuts, could prevent agencies from keeping pace with technological advances in fields like healthcare.

your ticket now to join the community

MOONDREAM LLAMAFILE (HUGGING FACE HUB)

A fully self-contained VLM executable that can run on edge devices based on the Moondream 2 model, which boasts powerful performance for its size.

OPENAI WILL TRAIN ITS MODELS ON FINANCIAL TIMES' CONTENT

The Financial Times has entered into a partnership with OpenAI where OpenAI will use the publication's journalism to train its AI models and develop new AI-driven features for its readers.

01

An attempt to build an open-source ecosystem for AI devices.

Read the full article on the Adevinta blog

APPLE AND OPENAI ARE REPORTEDLY IN TALKS FOR IOS 18 INTEGRATION

Apple is currently exploring partnerships with various AI providers, including OpenAI and Google, to integrate generative AI technologies into the upcoming iOS 18.

COHERE LAUNCHES DEVELOPER TOOLKIT TO ACCELERATE BUILD GEN AI APPS

Cohere launched a new developer toolkit to make enterprise AI application building faster and simpler across any cloud provider so developers can deploy in production at scale. With the Cohere Toolkit, developers have the source code to accelerate their development and quickly set up an application within just a few days.

THE LITTLE SMART HOME PLATFORM THAT COULD

Home Assistant has joined the Open Home Foundation, a non-profit aimed at countering Big Tech's influence by prioritizing privacy, choice, and sustainability in the smart home domain. The transition to a foundation-owned entity allows the platform to scale its consumer appeal while retaining its core open-source principles. Upcoming developments include direct-to-consumer sales on Amazon, certified partner expansions, local voice control hardware, AI integration with Nvidia, and user interface improvements for broader home approval.

LET'S THINK DOT BY DOT

This paper shows that chain-of-thought (CoT) thinking can be obfuscated with ‘...' tokens. This requires training the model but shows that interpreting the CoT steps may not be possible and thinking can be hidden by the model.

VIDEO-LANGUAGE MODELS WITH PLLAVA

Researchers have introduced a new model called PLLaVA, which stands out by adapting image-language AI models for video tasks using a unique pooling strategy.