TLDR AI 2024-05-09

Microsoft is developing a new AI model named MAI-1, which reportedly boasts about 500 billion parameters, aiming to surpass other major AI models by Google and OpenAI.

Apple has announced its next-generation Apple Silicon chip, the M4, the company's first chip designed specifically for AI. Featuring a 3-nanometer architecture and debuting in the 2024 iPad Pro, the new chip boasts a 10-core CPU that is 50% faster than the M2 and a new Neural Engine capable of processing 38 trillion operations per second. It enhances applications that leverage AI, although specific AI features for iPadOS haven't been detailed yet.

Google DeepMind and Isomorphic Labs have developed the 3rd generation of AlphaFold, a powerful protein folding prediction model. It is launching AlphaFold Server, which is a free way to interact with the model. AlphaFold 3 is 50% more accurate than previous generations. It correctly predicted the folded structure of the spike protein on Coronavirus OC43.

A novel labeling method that leveraged two passes of VLMs resulted in an extremely detailed pairing of images and text. The captions have far greater detail than any previous dataset and can aid in training strong multimodal models.

A document released by OpenAI outlining how it wants its models to behave in various scenarios.

A new diffusion model introduces a unidirectional block attention mechanism to overcome memory limitations in generating ultra-high-resolution images.

DocRes is a new model that simplifies document image restoration by handling five tasks: dewarping, deshadowing, appearance enhancement, deblurring, and binarization within a single system.

QoQ, a novel quantization algorithm, accelerates large language model inference by using 4-bit weights, 8-bit activations, and 4-bit KV cache.

ChemFlow is a new framework designed to enhance molecular science by efficiently navigating chemical space using deep generative models.

Teenage users on Character.AI are turning to AI chatbots for companionship and emotional support, which they find helpful but potentially addictive. There are concerns about the appropriateness of AI in mental health, especially when bots give problematic advice or increase reliance on virtual interactions.

GitHub Copilot Workspace, now in a limited technical preview, offers a powerful programming tool that lets users code in plain English, from planning to implementation, within the browser. The AI integrates seamlessly with codebases, suggesting block-by-block code execution and handling complex tasks with less active user input compared to ChatGPT.

Accelerated timelines, competition, and a lack of concern from superiors about real-world effects are themes common across a broad spectrum of the biggest effect companies. Engineers are increasingly saying that a large part of their jobs is focused on satisfying investors and not falling behind in the competition rather than solving actual problems for users. Some have been switched over to AI teams to help support fast-paced rollouts without adequate training, even if they are new to the technology.

The gpt-2-chatbot, which appeared recently in the LMSYS arena, has been confirmed as a new OpenAI model following a 429 rate limit error that revealed details from OpenAI's API.

OpenAI is developing a search engine for ChatGPT to allow the AI to search the web for answers.

The data curation platform adds to its $11m seed round from September, aiming to expand its team and continue building for enterprise.