TLDR AI 2024-04-18

There have been reports of the open source pioneer Mistral seeking several hundred million dollars of funding to train more models.

Stability AI has made its latest text-to-image AI model, Stable Diffusion 3, available to some developers via API and its new content creation platform called Stable Assistant Beta. The model is still in preview and not yet available to the general public.

Sanctuary AI will deliver its humanoid robot, Phoenix, to a Magna auto manufacturing facility, following Magna's 2021 investment in the company. Phoenix is notable for its walking capability and dexterous hands aimed at enhancing manufacturing efficiency. The specifics of the pilot, including the number of robots and deployment details, remain undisclosed.

The latest completely open 7B language model had an upgrade with a much-improved reasoning score on MMLU. Its development team found that changing the data mixture had a substantial effect on performance. They have provided the exact statistics of new data sources and percentages that led to this improvement.

This paper describes Stability AI's diffusion transformer model for audio synthesis.

The new base Mistral report was released today. Additionally, their own instruction tuned model was released under a similar permissive license. They report a strong MMLU and HumanEval performance along with incredible multilingual English, French, Italian, German, and Spanish performance, function calling, and 64k native context length.

A deep dive interview from industry experts on the challenges and solutions of getting AI models into production and where MLOps differs from traditional engineering. They discuss why so few ML projects make it to production and how to focus as an organization to actually launch.

The Effort library allows for real-time adjustment of the number of calculations performed during inference of an LLM model, resulting in significant speed improvements while retaining most of the quality. Despite some implementation overhead, initial results suggest that the Effort library has the potential to significantly improve LLM inference speed while maintaining quality. The author encourages others to test the 0.0.1B version and provide feedback to further improve the library.

This study revisits using diffusion models for image-to-text generation. It introduces the LaDiC architecture, which enhances diffusion models' performance in image captioning tasks.

Building a successful AI company requires navigating complex short, mid, and long-term dynamics and maintaining elite speed and execution, owning more of the stack, gathering unique data, and leveraging synthetic data generation. As the AI industry matures, companies must adapt to shifting talent dynamics and understand the machine they are building and the axes of competition they are predicated on to create lasting moats and stand out amidst the noise.

Wayve's Lingo-2 is a VLM train in simulation for autonomous driving tasks. It performs driving actions based on video input. Lingo-2 includes descriptions of reasoning for the driving actions it takes based on the scene.

Researchers are developing alternative architectures to address the limitations of Transformers in long-context learning, generation, and inference speed, showing competitive quality at smaller scales but uncertain scalability. The rapid progress in this subfield suggests that the Pareto frontier will continue to expand, enabling longer context modeling and greater inference throughput, ultimately increasing the number of use cases for AI.

Verifying training when sharding across devices can be challenging - this handy snippet verifies that things have been done and updated properly.

Luminal is a deep learning library that uses composable compilers to achieve high performance.

Google Maps will use AI to summarize directions to EV chargers as well as provide reliability and wait times.