Unconventional AI's Vision - Unconventional AI aims to revolutionize computing by drawing inspiration from the brain's efficiency, targeting AI ubiquity [1, 40, 41] - The company is focusing on analog computing to achieve greater efficiency compared to digital systems, especially for AI workloads [4, 9, 15] - The company's goal is to find a paradigm analogous to intelligence within five years and build a scalable solution for manufacturing [34, 35] Technological Approach - Unconventional AI is exploring energy-based models, diffusion models, and flow models due to their inherent dynamics [26] - The company is building a mixed-signal chip, potentially one of the largest analog chips ever built, for its first prototype [48, 50] - The company plans to release open-source resources to encourage experimentation and collaboration [27] Industry Perspective - The increasing energy consumption of data centers, currently using 4% of the US energy grid, is a major concern, potentially rising to 8%-10% [16] - The industry faces a potential shortfall of 400 gigawatts of additional capacity over the next 10 years to meet AI demand [17] - TSMC is considered a key partner, while collaboration with Nvidia and Google remains a possibility [36, 37, 38] Company Strategy - Unconventional AI is building a practical research lab environment, encouraging exploration and innovation without premature manufacturing constraints [56, 57] - The company seeks individuals skilled in mapping algorithms to physical substrates, energy-based models, dynamical systems, and analog/digital circuit design [47, 48] - The company emphasizes agency and empowerment for its team members, fostering a culture of ownership and learning from both successes and failures [59, 60]

Understanding neural networks through sparse circuits:OpenAI (@OpenAI):We’ve developed a new way to train small AI models with internal mechanisms that are easier for humans to understand.Language models like the ones behind ChatGPT have complex, sometimes surprising structures, and we don’t yet fully understand how they work.This approach ...

Model Training Optimization - The industry has been training neural networks for 9 years [1] - The industry actively uses 16 ways to optimize model training [1]

Tesla's Dojo and Nvidia Dependence - Tesla is currently dependent on Nvidia's A100 and H100 GPUs for training its video-based neural networks [1] - Nvidia's GPUs are general-purpose chips optimized for broader markets like LLMs and gaming [1] - Tesla's Dojo is designed to break free from this reliance on Nvidia [1]