TLDR AI 2024-04-25

Apple has released OpenELM, a family of eight open-source LLMs designed to run efficiently on a single device for text generation tasks with parameter sizes ranging from 270 million to 3 billion.

AI-powered coding platform Augment has launched from stealth with $252 million in funding, valuing the company near $977 million. Founded by ex-Microsoft developer Igor Ostrovsky, the platform aims to enhance software quality and productivity using advanced AI models. Augment plans to offer standard SaaS subscriptions and expects to reveal pricing details later this year ahead of its GA release.

Sakana AI's EvoSDXL-JP is a high-speed image generation model optimized for Japanese language prompts that utilizes an evolutionary model merging method. EvoSDXL-JP boasts tenfold faster inference speeds and superior performance compared to existing models. It is ideal for educational use in Japan to demonstrate the benefits of generative AI.

A character control framework has been introduced that leverages motion diffusion probabilistic models to produce a variety of high-quality animations that respond instantly to dynamic user commands.

CutDiffusion is a new approach that transforms low-resolution diffusion models to meet high-resolution needs without the complexities of traditional tuning.

Sleeper Agents are language models that have been trained to perform malicious actions when prompted with a certain set of wake words. Probing language models with simple linear heads and the prompt “are you going to do something dangerous?” gives extremely reliable detection of these previously hidden malicious actors.

Microsoft has released a set of GPU accelerated kernels for training BitNet style models. These models have substantially lower memory cost without much drop in accuracy.

Apple has released a neural network training library along with a set of models called OpenELM which are optimized to run on devices.

MaxText is a high performance, highly scalable, open-source LLM written in pure Python/Jax and targeting Google Cloud TPUs and GPUs for training and inference.

The last week of March 2024 marked a significant moment in open-source large language models (LLMs) with multiple notable releases, including DBRX by Databricks, Jamba by A21 Labs, and Samba-CoE by SambaNova Systems. These launches signify a pivotal moment in the diversification and proliferation of accessible and decentralized AI models. The trend reflects a narrowing performance gap between open-source LLMs and their closed-source counterparts, indicating a vibrant future for AI innovation and enterprise adoption.

Profluent, a Berkeley startup, has used generative AI technology to create new gene editors based on CRISPR. The company has used one of the AI-generated gene editors to edit human DNA, but it has yet to put any of them through clinical trials. While Profluent plans to open source the gene editors generated by its AI technology, it will not be open sourcing the AI technology itself. The project is part of a wider effort to build AI technologies that improve medical care.

FlexAI launched with $30 million in seed funding led by Alpha Intelligence Capital, Elaia Partners, and Heartcore Capital. The company is rearchitecting compute infrastructure to deliver universal AI compute: effective and seamless infrastructure needed to propel advancements in artificial intelligence. FlexAI's cloud service, launching later this year, enables developers to utilize heterogeneous compute architectures to build and train AI applications reliably and efficiently.

Google will have a “version 2” of Gemini Nano available by the time the Galaxy S25 launches next year.

Let AI call, screen, and rank your candidates for you.

Cohere has released a chat interface that includes many nice features for building AI-based chat applications.