It also helped that 'AI' (in its current zeitgeist meaning use of MLNNs for image processing) has been in use for astronomy - and elsewhere - since well before the current 'deep learning' boom started. Wide availability of hardware to support NN acceleration has made the existing techniques cheaper and faster, but the techniques themselves have not really been new.
[questions]
Nous Research turned heads earlier this month with the release of its permissive, open-source Llama 3.1 variant Hermes 3.Now, the small research team dedicated to making “personalized, unrestricted AI” models has announced another seemingly massive breakthrough: DisTrO (Distributed Training Over-the-Internet), a new optimizer that reduces the amount of information that must be sent between various GPUs (graphics processing units) during each step of training an AI model.But Nous Research, whose whole approach is essentially the opposite — making the most powerful and capable AI it can on the cheap, openly, freely, for anyone to use and customize as they see fit without many guardrails — has found an alternative.Yet, the authors also say that “our preliminary tests indicate that it is possible to get a bandwidth requirements reduction of up to 1000x to 3000x during the pre-training,” phase of LLMs, and “for post-training and fine-tuning, we can achieve up to 10000x without any noticeable degradation in loss.”
They further hypothesize that the research, while initially conducted on LLMs, could be used to train large diffusion models (LDMs) as well: think the Stable Diffusion open source image generation model and popular image generation services derived from it such as Midjourney.To be clear: DisTrO still relies on GPUs — only instead of clustering them all together in the same location, now they can be spread out across the world and communicate over the consumer internet.Specifically, DisTrO was evaluated using 32x H100 GPUs, operating under the Distributed Data Parallelism (DDP) strategy, where each GPU had the entire model loaded in VRAM.By reducing the need for high-speed interconnects DisTrO could enable collaborative model training across decentralized networks, even with participants using consumer-grade internet connections.
The report also explores the implications of DisTrO for various applications, including federated learning and decentralized training. Additionally, DisTrO’s efficiency could help mitigate the environmental impact of AI training by optimizing the use of existing infrastructure and reducing the need for massive data centers.
I remain skeptical (in terms of space or otherwise technical relevance). No matter how cheaply you can train them, LLMs are just too unreliable/unmoored from reality to be used for anything technical, safety-critical, etc. They're great for generating corporate boilerplate or form letters quickly, but..."AI" that would be useful for advanced space probes, ISRU mining on Mars before human crews arrive, etc. etc. won't come from the LLM/LDM type tech at all.
The capability for collaborative task planning originates from the foundation model OpenAI (2023), which directly outputs the collaboration graph in JSON format guided by prompt engineering Mo and Xin (2023).
My extreme skepticism is for... tasks based on interaction with a physical environment.
9/19/24Trajectory Optimization with Reinforcement Learning-driven Control of Multi-Arm Robots in On-Orbit Servicing OperationsCelia Redondo-Verdú & Jorge Pomares (University of Alicante)Intelligent motion planning and control systems are crucial in managing the inherently unpredictable and non-deterministic conditions of space. Trajectory optimization (TO) has been a key method in space robotics for precision control and guidance. However, the effectiveness of TO is limited by its reliance on accurate robotic and environmental models. Reinforcement Learning (RL) has emerged as a promising alternative, enhancing robustness and adaptability in control systems. Unlike TO, RL does not depend on predefined models but learns effective control policies through simulated interactions. This model-free approach has shown potential in handling complex control tasks and adapting to environmental uncertainties and perturbations. This paper proposes an integrated TO and RL approach for enhanced path planning and control of a four-arm robot designed for on-orbit servicing tasks. By combining the predictive power of TO with the adaptive capabilities of RL, this approach aims to optimize the robot’s operational effectiveness. The integration seeks to leverage the detailed pre-planned motion profiles provided by TO, while incorporating the flexibility and resilience of RL-based control strategies to accommodate real-time operational dynamics and uncertainties. The results presented demonstrate how this hybrid approach can significantly improve the robot’s trajectory tracking performance and operational adaptability.
I am going to walk you through some advanced AI magic and bring you up-to-speed on state-of-the-art boundary-pushing when it comes to generative AI. It is a bit of a mystery story too.In this column, I will showcase something that is on the outskirts of everyday generative AI and primarily experimental in advanced AI labs. It is an approach that leverages multiple chain-of-thought processing and incorporates AI-based meta-reasoning. Some believe that this might be an essential ingredient or secret sauce of the new o1 generative AI model, and, by the way, the future for leading-edge generative AI.
The Nobel Prize in Physics has been awarded to two scientists, Geoffrey Hinton and John Hopfield, for their work on machine learning.British-Canadian Professor Hinton is sometimes referred to as the "Godfather of AI" and said he was flabbergasted.He resigned from Google in 2023, and has warned about the dangers of machines that could outsmart humans.
On Monday, Google announced a landmark agreement with nuclear startup Kairos Power to purchase energy produced by seven yet-to-be-built small modular nuclear reactors. The companies claim the deal aims to add upwards of 500 megawatts "of new 24/7 carbon-free power to US electricity grids" — that is, over a decade from now when Kairos promises the reactors will be built.
In the announcement, Google and Kairos claim that the first of the modular nuclear power terminals will be up and running by 2030, with all modules completed by 2035.
The use of small modular reactors to power AI is obviously relevant to spaceflight.QuoteOn Monday, Google announced a landmark agreement with nuclear startup Kairos Power to purchase energy produced by seven yet-to-be-built small modular nuclear reactors. The companies claim the deal aims to add upwards of 500 megawatts "of new 24/7 carbon-free power to US electricity grids" — that is, over a decade from now when Kairos promises the reactors will be built.…Quote In the announcement, Google and Kairos claim that the first of the modular nuclear power terminals will be up and running by 2030, with all modules completed by 2035.https://futurism.com/the-byte/google-nuclear-powerGoogle press release:https://blog.google/outreach-initiatives/sustainability/google-kairos-power-nuclear-energy-agreement/