reinforcement learning - Tech News from Simmons Systems

Tech News from Simmons Systems

reinforcement learning

A collection of 42 posts

Comparing Supervised and Reinforcement Learning for Fine-Tuning Reasoning LLMs

A recent discussion on the Hugging Face forums raises an important question for AI practitioners: whether to use supervised learning or reinforcement learning when fine-tuning a reasoning-focused large language model. The conversation centers on which approach produces better results for models designed to handle complex reasoning tasks. Supervised fine-tuning involves

NVIDIA Launches Nemotron 3 Open Models to Power Agentic AI Systems

NVIDIA introduces the Nemotron 3 family of open models featuring Nano, Super, and Ultra sizes to help developers build efficient, transparent multi-agent AI applications. The new models utilize a breakthrough hybrid architecture to deliver leading accuracy and unmatched inference speeds.

OpenAI Introduces Robotic Testing Environments for Multi-Goal AI Research

A new technical report unveils a suite of challenging robotics tasks for OpenAI Gym that focus on multi-goal reinforcement learning. The environments feature sparse rewards and aim to drive future algorithmic improvements.

Tencent Unveils Hunyuan-T1, the First Mamba-Powered Ultra-Large Reasoning Model

Tencent releases the official version of its Hunyuan-T1, a cutting-edge reasoning model built on the world's first ultra-large Hybrid-Transformer-Mamba MoE architecture. The model leverages massive reinforcement learning to deliver top-tier performance with twice the decoding speed.

OpenAI Introduces CriticGPT to Catch AI Hallucinations

OpenAI unveils CriticGPT, a new model designed to review and find errors in code generated by ChatGPT. This AI critic aims to solve the growing problem of hidden hallucinations as language models become more advanced.

Cognition Labs Unveils Devin, an Autonomous AI Software Engineer

Cognition Labs introduces Devin, an AI agent that independently plans, codes, and deploys software applications. The tool significantly outperforms previous models on benchmark tests but faces GPU limitations before enterprise adoption.

matrix multiplication

AI System Breaks 50-Year-Old Matrix Multiplication Speed Record

DeepMind's AlphaTensor uses reinforcement learning to discover faster matrix multiplication algorithms, surpassing human intuition. The AI breaks a 50-year-old speed record previously held by Strassen's algorithm.

DeepMind's DeepNash AI Conquers Complex Board Game Stratego

DeepMind develops DeepNash, a model-free reinforcement learning agent that achieves a 97% win rate against AI bots in the complex board game Stratego. The system uses advanced game theory and neural networks to master imperfect information scenarios.

DeepMind AI Achieves Breakthrough in Nuclear Fusion Plasma Control

DeepMind and the Swiss Plasma Center successfully use deep reinforcement learning to control and sculpt superheated plasma inside a tokamak reactor. This breakthrough opens new avenues for advancing nuclear fusion research.

Smaller AI Models Outperform GPT-3 Through Human Feedback Training

OpenAI researchers introduce InstructGPT, a language model trained with human feedback to better follow user instructions. Despite having 100 times fewer parameters, this aligned model consistently outperforms the massive GPT-3 in human evaluations.

quantum computing

Nvidia and Quantum Machines Leverage AI to Tackle Quantum Calibration

Quantum Machines and Nvidia are using reinforcement learning on Nvidia's DGX platform to keep quantum processors calibrated in near real time. This partnership brings the industry a step closer to achieving reliable, error-corrected quantum computers.

ETH Zurich Teaches Four-Legged Robot to Climb Ladders Using AI

Researchers at ETH Zurich successfully train a quadruped robot to climb standard ladders using reinforcement learning and specialized grippers. The new system achieves a 90% success rate and climbs 232 times faster than previous state-of-the-art methods.

Nvidia Announces Project GR00T Foundation Model to Power Humanoid Robots

Nvidia introduces Project GR00T, a new foundation model designed to help humanoid robots understand language, process environments, and mimic human actions. The company also unveils the Jetson Thor computing platform and updates to its Isaac robotics tools to support this advanced embodied AI.

NVIDIA Secures Top NeurIPS Awards for Generative AI and Generalist Agents

NVIDIA earns prestigious NeurIPS 2022 awards for innovative research in diffusion-based generative models and generalist AI agents. The company presents over 60 projects at the conference, focusing heavily on synthetic data generation.

MIT Teaches Mini Cheetah Robots to Play Soccer Goalie Using AI

Researchers utilize a reinforcement learning framework to train MIT's Mini Cheetah robots to block soccer balls in under a second. The same underlying technology shows promise for future offensive robotic soccer skills.