TLDR AI 2024-02-09

Google launched a new service that allows you to finally interact with Gemini Ultra 1.0 - it will charge a monthly subscription for access to the model. There is also a companion mobile app.

1X is a robotics company that has made strides in video-to-control models. The company has showcased its robot performing many tasks, all driven by neural networks that emit 10hz control signals from video input.

This paper explores the development of a generalist AI agent capable of understanding and following gameplay instructions, a step towards "read-to-play" abilities. The researchers enhance the agent's multitasking and generalization skills by integrating multimodal game instructions into a decision transformer.

MetaTree is a transformer-based decision tree algorithm. It learns from classical decision tree algorithms for better generalization capabilities.

FunSearch, a new AI-powered method that combines Large Language Models with evaluative algorithms, has made verifiable discoveries in mathematical sciences, including solutions for the longstanding cap set problem and more efficient algorithms for the bin-packing problem. It introduces evolutionary approaches to generating and evaluating code that offers human-interpretable outputs, representing a significant leap in AI-driven scientific discovery.

HuggingFace released a lightweight evaluation library for language model training based on HELM and Eluther AI’s evaluation harness.

You can build a sophisticated and powerful RAG system that runs on your hardware using Ollama, pgvector, and local data.

CogCoM, a new general vision-language model, incorporates a unique Chain of Manipulations mechanism. This allows it to handle multi-turn visual reasoning by actively adjusting input images.

Hear from practitioners solving real-life challenges in productionizing LLMs — including cost optimization, latency requirements, trust of output, and debugging. Join 70+ speakers from companies like Netflix, Google, Databricks, Perplexity AI, and Hugging Face. Attend hands-on workshops for builders. Enjoy musical improv, games, and guided meditation during the breaks. Register now, it’s free!

HelixML was able to better fine-tune Mistral-7B by implementing a suite of qapair prompts that extracted content from all sorts of different perspectives and generating a content-addressed hash for each document.

Google's newly released Gemini Advanced, a GPT-4 class AI model, showcases capabilities equivalent to OpenAI's GPT-4. It shows strengths in explanations and integrating images and search.

This point of view, from an investor, helps to bring a little context to the layers of value within AI right now by breaking them down into the infrastructure layer, like cloud providers and chip makers, modeling and core, like OpenAI and Anthropic, and AI-enhanced products, like all of you that use AI to improve your products.

Singapore-based Brilliant Labs has introduced Frame, lightweight AR glasses featuring a multimodal AI assistant, Noa, that performs visual processing, image generation, and more with integrated AI models like GPT-4 and Stable Diffusion.

Shortwave released a mobile email assistant for iOS & Android that can create instant summaries during its AI Launch Week last week.

A benchmark dataset that contains facts about various anime characters and quotes to assess language model performance.