Figma launches new AI features

Stability AI lands a lifeline from Sean Parker, Greycroft

Stability AI secured new funding from notable investors like Sean Parker and Greycroft, amid financial difficulties and leadership changes. The startup behind the image-generating model Stable Diffusion has struggled with cash flow and mounting operational costs. New CEO Prem Akkaraju and the board plan to focus on expanding Stability’s generative AI capabilities while adhering to open-source principles. Despite past mismanagement and ongoing legal challenges, the investment aims to bolster Stability's future in the burgeoning generative media market. (Read more)

Meet Figma AI: Empowering Designers with Intelligent Tools

Figma has unveiled Figma AI, a suite of AI-driven features aimed at enhancing designers' productivity and creativity. Key features include:

- Visual and Asset Search: Enhanced search capabilities using images or text to find relevant designs.

- AI-Powered Content Generation: Tools to create realistic copy and images, background removal, and layer renaming.

- Prototyping Tools: Quick-turn static designs into interactive prototypes.

- Text Prompt Design Generation: Generate UI layouts and components from text descriptions.

These tools are available in a beta program through 2024, and Figma ensures privacy and data protection in its AI training models. (Read More)

From bare metal to a 70B model

A team from Imbue provides a detailed guide on setting up infrastructure to train a 70B parameter model, starting from bare metal hardware. The process includes provisioning machines, setting up a robust InfiniBand network, diagnosing hardware issues, and running health checks. Key learnings emphasize the importance of reproducibility, automated error handling, and the flexibility of swapping hardware. The effort resulted in a stable infrastructure capable of efficient, large-scale model training, and they invite interested individuals to join their team. (Read More)

Meta 3D Gen

Meta 3D Gen (3DGen) is a cutting-edge pipeline for fast text-to-3D asset generation, creating high-fidelity 3D shapes and textures in under a minute with support for physically-based rendering (PBR). It integrates Meta 3D AssetGen and Meta 3D TextureGen for superior text-to-3D and text-to-texture generation, achieving a 68% win rate over single-stage models. 3DGen outperforms industry baselines in prompt fidelity and visual quality, while also being significantly faster. (Read More)

Finding GPT-4’s mistakes with GPT-4

OpenAI developed CriticGPT, based on GPT-4, to catch errors in ChatGPT's code output, finding that it helps users outperform those without assistance 60% of the time. By integrating CriticGPT into their RLHF labeling pipeline, OpenAI aims to enhance AI trainers' ability to evaluate ChatGPT's outputs, especially as models become more accurate and their errors more subtle. CriticGPT's critiques are preferred over unassisted critiques 63% of the time, producing more comprehensive and accurate feedback with fewer hallucinations. (Read More)