TLDR AI 2024-02-01

Building production-ready AI applications is hard and expensive. Anyscale empowers builders to iterate fast with a simple user interface. Fine-tuning and deployment are 10x faster with automatic scaling and proactive resource management. You get the best cost-performance benefits using serverless or fully hosted on Ray.

Ray is trusted by OpenAI, Uber, AWS, Spotify and other AI leaders. Now, you can build your GenAI, conversational AI, or computer vision app easily — here’s how:

→ Fast-track GenAI with Anyscale Endpoints: Connect, fine-tune, and serve popular open-source LLMs in minutes — privately, securely, and cost-effectively.

→ Deploy at scale with Anyscale Platform: Go big with the industry's best-performing Ray distributed framework.

⚒️ Start building ⚒️

The AI community is abuzz with a leaked large language model called "miqu-1-70b" that was shared on HuggingFace and 4chan. The model shows performance rivaling OpenAI's GPT-4 and is linked to Mistral's Mixtral 8x7b. This incident, acknowledged by Mistral's CEO as a leak by an enthusiastic customer, highlights the growing capabilities of open source AI.

Dipendra Misra presented Layer-Selective Rank Reduction (LASER), a technique that improves large language model (LLM) accuracy by replacing a weight matrix with a smaller approximation, at the January Microsoft Research Forum.

The newest version of Llava, a vision language model, has improved OCR, reasoning, and world knowledge. It matches Gemini on some tasks. The Llava team plans to release data, code, and the model.

Weaver is a series of models specifically trained to tell stories. The largest model (34B params) outperforms GPT-4 on a storytelling benchmark.

This paper introduces a benchmark and two datasets for scene and handwritten texts. Featuring original, corrupted, and assistant images, the novel Global Structure-guided Diffusion Model (GSDM) leverages text structure to efficiently recover clean texts. It shows significant improvements in recognition accuracy and image quality.

Monte Carlo Tree Search (MCTS) is a state-of-the-art search algorithm. It was used in AlphaGo and researchers are now working to integrate it with language models. MCTS is notoriously tricky to implement.

MF-MOS is a new model for better detecting moving objects in autonomous driving using LiDAR technology. It uniquely separates motion and semantic information, using residual maps for motion capture and range images for semantic guidance.

Norton is a new approach to improving long-term video understanding by addressing the multi-granularity noisy correspondence issue in video-language studies.

FireLlava is a new open vision model trained on data that makes it possible to use for commercial tasks. It matches the performance of the original Llava, but not yet to the level of Llava 1.5.

ServiceNow, a $150b market cap company, last week announced that its gen AI products drove the largest net new ACV contribution for its first full quarter of any of its new product family releases ever, including its original Pro SKU. We’re already seeing enterprise-level AI applications drive meaningful revenue growth – that’s cool.

AI has gotten good enough to accomplish multiple tasks in under a minute, which means organizations and employees are going to have to emphasize using AI for good, not bad.

Fly to Hawaii for $149, Greece for $289, or anywhere else. Get 93% off lifetime access to Dollar Flight Club for $129 (was $1,690) before midnight. Comes with 3-day money-back guarantee*

Automate resume screening with AI.

GPAvatar introduces a new method for creating 3D head avatars from images that use a dynamic point-based expression field and a Multi Tri-planes Attention module.

A library that allows for creating autonomous AI agents in Python.