TLDR AI 2024-08-21

Enjoy great food, drinks, and networking while learning from the successful AI journeys of top companies. Hear about the tools used, lessons learned, and wins on their path to production.

🗓️ When:Wednesday, September 4

🕺 Who's attending: Industry experts from DataStax, Priceline, PayPal, NVIDIA, IBM, WorkSpan, and Unstructured

🥂 What's on: Hackathon, demos, cocktails, networking — all the good stuff

🗽 Where: NYC | Pier Sixty at Chelsea Pier

Attendance is limited, so RSVP today to be considered!

OpenAI has entered a multi-year licensing agreement with Condé Nast, allowing it to display content from the publisher's brands, such as Vogue and The New Yorker, within its products, including ChatGPT and SearchGPT.

OpenAI has launched fine-tuning for GPT-4o, allowing developers to customize the model for specific use cases with their own datasets. It is offering 1 million free training tokens per day through September 23.

Meta FAIR researchers have introduced the Self-Taught Evaluator, a method to train evaluative LLMs without human annotations, potentially enhancing the efficiency and scalability of LLM assessment. Using the LLM-as-a-Judge concept, it iteratively generates and refines responses to create a training dataset, demonstrating improved performance on benchmarks like RewardBench. This technique could enable enterprises to leverage unlabeled data for LLM tuning while acknowledging the importance of a well-aligned seed model and the limitations of benchmarks.

This is a great paper that discusses how brittle modern agent systems are and what the path forward could be to design learned systems. Its authors use programming languages as a test bed where agents can be designed and run unsupervised.

Automating and optimizing chip design with AI can reduce costs and meet demand for stronger chips. This method was tested on many circuits, generating compact and efficient designs with a success rate of 84.6% using an Alpha Zero based approach.

A fast, cheap, and state-of-the-art 3D mesh generation model that can be conditioned on images or text. Specifically, it uses signed distance function supervision and a cascade of steps such as a normal map generator that offloads different tasks to various submodels.

Join OctoAI for an expert panel discussion on balancing innovation with data security and compliance to help businesses deploy GenAI applications securely and efficiently.

Tune in on August 27th and learn to:

• Ensure data security and regulatory compliance
• Navigate deploying Generative AI in private environments
• Implement quality evaluations at an enterprise scale

Sign up today for this free virtual roundtable or to receive the recording.

A new framework was developed for generating X-ray medical reports that enhances performance while reducing computational complexity.

TraDiffusion is a new method for guiding text-to-image generation using mouse trajectories instead of traditional box or mask controls.

Extremely fast and efficient optical flow code that can be run on low power devices such as phones and some security camera systems.

AI companies have overestimated the readiness of generative AI for widespread commercial use, leading to costly missteps in product development and market integration. As they try to correct course, they must overcome five key challenges: ensuring cost-effectiveness, improving reliability, safeguarding privacy, enhancing safety and security, and refining user interfaces. These obstacles highlight the gap between AI's potential capabilities and the practical realities of deploying AI systems that meet consumer expectations and integrate seamlessly into existing workflows. The path to widespread adoption will likely unfold over a decade or more, rather than the rapid timeline some have predicted.

Current language models predict one word at a time, which gives very little room for planning and reasoning. There are ways around this such as Chain of Thought prompting. However, potentially using diffusion models, which can spend more diffusion steps per token, could be a way to improve model reasoning.

Unitree Robotics has optimized its $16,000 G1 humanoid robot for mass production, showing off its improved agility with abilities such as leaping and climbing stairs.

A fun utility that animates a curve rider sled going down loss functions, showing when they converge and when they're too spiky.

SkyScript-100M is an extensive dataset designed to enhance the generation of high-quality shooting scripts for short dramas.