Search the archive

A BENCHMARK FOR EVALUATING CODE REPOSITORY EDITS

RES-Q is a new benchmark for assessing the ability of large language models to edit code repositories based on natural language instructions.

PROMPT ENGINEERING TOOLKIT

A great prompting toolkit almost entirely made from Sonnet 3.5 that helps test different prompts' performance.

THE FUTURE OF LLM-BASED AGENTS: MAKING THE BOXES BIGGER

This post shares two critical approaches that help take Agents out of the playground and into the real world: long-term planning and system-level robustness. These introduce capabilities to develop higher-order plans for Agents while allowing mid-episode adaptation as well as systems approaches to intelligent orchestration of the models that provide higher accuracy and performance.

HOW AI TRANSFORMS THE TRAVEL VALUE CHAIN

Online travel has yet to crack the end-to-end trip experience. AI offers fresh strategies for rebundling travel services, potentially shifting profits. Decision support and coordination enabled by AI will be crucial for companies to assert control over the travel rebundling process. The travel industry's future may see incumbents leveraging AI for competitive advantage, while traditional travel agents face market erosion due to AI efficiencies.

AI-POWERED CONVERSION FROM ENZYME TO REACT TESTING LIBRARY AT SLACK

Slack engineers utilized a hybrid approach, combining Abstract Syntax Tree transformations and AI Large Language Models, to automate the conversion of over 15,000 unit tests from Enzyme to React Testing Library. The team achieved an 80% success rate in automated conversions by integrating Anthropic's Claude 2.1 AI model and capturing the DOM tree for React components. This innovative project underscores Slack's commitment to enhancing developer experience and productivity through the application of AI. It is part of ongoing efforts to stay ahead in the ever-evolving frontend landscape.

R2R

R2R is an RAG engine with a RESTful API and prod features.

CHATGPT FOR MAC DESKTOP (PRODUCT)

The ChatGPT desktop app for macOS is now available for all users.

MEDICAL MODEL ON PAR WITH GPT 4 (HUGGING FACE HUB)

This 7b model, trained on medical data, tied with GPT-4 on human preference evaluation when measured by medical doctors.

ETCHED RAISES $120M SERIES A TO BUILD AI CHIPS FOR TRANSFORMERS

Etched has announced a $120M round led by Primary Venture Partners and Positive Sum Ventures, as well as angel investors including Peter Thiel, Stanley Druckenmiller, and David Siegel. It is working directly with TSMC's Emerging Business Group to produce chips on their state-of-the-art, 4 nanometer node. Its chips can achieve 500K tokens/second. One 8xSohu can replace 160 H100s.

STABILITY AI NEW FUNDING

In addition to a new CEO, Stability AI was able to secure a “significant infusion of capital” from existing and new investors.

APPLE INTELLIGENCE AND AI MAXIMALISM

Apple is segmenting its AI approach, differentiating between 'context' and 'world' models, with the latter, such as ChatGPT, being used for open-ended queries but without personal context. OpenAI's position is not highly defensible, given the lack of moats in large language models (LLMs) and the rise of competing models. Apple's strategy suggests a commoditization trend for LLMs that leverages on-device capabilities and private cloud to provide AI features while viewing services like ChatGPT as interchangeable components.

RAG FINE-TUNING

Together AI and Morph Labs have put together a great blog post on tuning models for retrieval augmented generation. They showcase some uses of synthetic data as well.

IMBUE TRAINS 70B THAT MATCHES GPT-4

Imbue has trained and released an extremely powerful 70B language model. It uses Imbue's custom optimizer and some great data filtering techniques. The model was trained with zero loss spikes.

EVALALIGN FOR TEXT-TO-IMAGE EVALUATION

EvalAlign is a new metric designed to improve the evaluation of text-to-image generative models. Unlike existing metrics, EvalAlign offers fine-grained accuracy and stability. It focuses on image faithfulness and text-image alignment.

you deliver to users. Create a free account

FINE-TUNE FLORENCE 2

Microsoft's new many-to-many vision model can be tuned for specific downstream tasks. It isn't quite as powerful as PaliGemma, but is easy to run in PyTorch.

2:4 SPARSE TRAINING IN PYTORCH

The PyTorch team has written custom kernels to make use of sparse cores, which are usually reserved for inference only.

CONTROL VIDEO GENERATION TRAJECTORIES WITHOUT TRAINING

FreeTraj is a tuning-free method to control the motion trajectories in video generation using diffusion models. It modifies noise sampling and attention mechanisms to guide the generated content.

MOLECULAR MODELING WITH TRANSFORMER

GeoMFormer is a new Transformer-based model designed to improve molecular modeling by learning both invariant and equivariant features.

MAKING MY LOCAL LLM VOICE ASSISTANT FASTER AND MORE SCALABLE WITH RAG

Researchers were able to make an LLM voice assistant more efficient and scalable by categorizing data, precomputing embeddings, and dynamically generating examples.

HOW TO SOLVE LLM HALLUCINATIONS

Lamini has developed Memory Tuning to reduce hallucinations by 95%, embedding specific data into models more efficiently without compromising general knowledge.

OPENGLASS

Turn any glasses into AI-powered smart glasses.

OPENAI PUSHES BACK VOICE MODE

OpenAI's advanced Voice Mode, demonstrated during its Spring Update, will launch in alpha to select ChatGPT Plus users in late July.

LLM RADAR (PRODUCT)

LLM Radar is a free tool that scans for new LLMs and delivers them straight to your inbox so you can stay ahead in the AI game without any effort.

OPENAI WALKS BACK CONTROVERSIAL STOCK SALE POLICIES

OpenAI has reversed its previous restrictive policies, now allowing current and former employees to participate equally in annual tender offers for stock sales. This change addresses concerns about liquidity access, especially given OpenAI's high valuation, and eliminates provisions that allowed forced share repurchases.

AMAZON DEVELOPING AI CHATBOT TO COMPETE WITH CHATGPT

Amazon is reportedly developing an AI chatbot called "Metis" to compete with OpenAI's ChatGPT, expected to be unveiled later this year. The chatbot, which will be accessible via web browser, will utilize a new foundational model and share infrastructure with "Remarkable Alexa," Amazon's upcoming paid-tier Alexa version.

OPENAI BUYS MULTI

Continuing the acquisition spree, OpenAI buys more foundational technology for using the desktop.

TOP 40 CVPR POSTERS

A list of curated computer vision papers from the 2024 CVPR conference.

TOUCAN TTS IN 7000 LANGUAGES

Toucan recently released a set of new text-to-speech models that it has extended to work with every ISO-639-3 standard language.

IMPROVED PASSAGE RANKING

PE-Rank is a new approach to passage ranking that improves efficiency by using single passage embeddings for context compression.

1300 TOKENS PER SECOND ON MAC

Implementing batch parallel KV cache in MLX leads to dramatic inference time speed up for synthetic data generation and model completions.

MIXTURE OF ATTENTION IN LLMS

The Mixture of Attention (MoA) approach optimizes sparse attention in large language models by tailoring unique sparse attention configurations for different heads and layers.

TRAIN VISION MODELS IN TRL

TRL is a Hugging Face library for training transformers with reinforcement learning. This example allows you to also perform the same process for vision-based language models like LLaVA.

CAREER OPPORTUNITIES IN AI

Popular developer Shawn "swyx" Wang talks about hidden opportunities for traditional software developers looking to transition into AI. Specifically, looking at building on top of current tools, using AI for summarization, and more.

REMOTE SENSING CHANGE DETECTION

This project introduces CDMask and CDMaskFormer, two new models for remote sensing change detection.

APPLE INTELLIGENCE

Apple has integrated generative AI into its core applications, enhancing features like Safari summaries, Mail categorization, and Siri's functionality, rather than creating standalone AI products, showcasing the company's focus on privacy and user control.

LLAMA TTF

This post shows how to run a small Llama language model in a font file.

ABILITY AI RAISES $1.1M PRE-SEED ROUND

Ability AI, a marketing startup, has raised a round to build its marketing agent technologies.

JUNE

June is a local voice chatbot that combines the power of Ollama, Hugging Face Transformers, and the Coqui TTS Toolkit.

OPENAI ACQUIRES ROCKSET

Rockset is a powerful database with indexing and querying capabilities. OpenAI has acquired the company to improve its retrieval infrastructure.

SNAPCHAT AI TURNS PROMPTS INTO NEW LENS

Snapchat has introduced an on-device AI model that enables users to create custom AR lenses from text prompts. A rollout is expected in the coming months for lenses and by year's end for creators. The latest Lens Studio update adds AI tools for generating 3D assets, face masks, and textures from prompts or images. Snapchat's push into AI has also produced AI-generated snaps for subscribers and a ChatGPT-powered chatbot for all users.

HEYGEN RAISES $60M SERIES A TO SCALE VISUAL STORYTELLING FOR BUSINESSES

HeyGen, an AI video generation platform, has secured $60M in Series A funding to enhance its tools for creating and localizing studio-quality videos rapidly and cost-effectively. Recently profitable with $35M in ARR, HeyGen aims to democratize visual storytelling for businesses of all sizes.

EVIDENCE OF LOG-SCALING IN PERSUASION

Super persuasion is the fear that as models grow larger, they will get substantially more persuasive. There is weak evidence to suggest that larger models aren't substantially more persuasive than smaller models. However, they may be able to be tuned to be more persuasive.

SMARTER CRYPTO TRADING

MacroHFT is a new approach to high-frequency trading (HFT) in cryptocurrency markets that leverages reinforcement learning to improve decision-making and profitability.

BETTER EXPLORATION IN MULTI-AGENT MODELS

Researchers have improved QMIX, a popular method for multi-agent reinforcement learning, by adding a local Q-value learning method within a maximum entropy framework.

use case. Register here