Can AI agents do more than answer questions? In this October 7th episode of GenerAIly Speaking, hosts Jay Busch and Bharathi Srinivasan put that to the test with two demos.
First, Madhu Samhitha Vangara and Tejas Kashinath build a research agent from scratch with Amazon Bedrock AgentCore. It searches the web, arXiv, Hacker News, and GitHub, remembers who you are, and reports back as a noir detective, a sports announcer, or a Greek bard. Afterward, we talk about where research agents could go next, from agents that read research to agents that run experiments on their own, like Andrej Karpathy’s autoresearch project.
Next, Kiowa Jackson and Cole Harrison show how a humanoid robot policy is trained and evaluated on AWS. Using the AWS Physical AI Toolchain, they follow NVIDIA GR00T from fine-tuning in Amazon SageMaker to testing in NVIDIA Isaac Lab Arena, with a success gate it has to pass before it’s registered. Along the way, we discuss robot-use agents, the idea from MIT professor Phillip Isola’s recent essay that a general AI agent could use a robot as a tool.
Get started with Amazon Bedrock AgentCore https://aws.amazon.com/bedrock/agentcore/getting-started/
Try the AgentCore getting started workshop https://catalog.workshops.aws/agentcore-getting-started/en-US
Explore the Arena workflow in the AWS Physical AI Toolchain https://github.com/aws-samples/sample-aws-physical-ai-toolchain/tree/development/isaac-lab-arena-on-aws
Tune in to watch GenerAIly Speaking every other Wednesday on Twitch, or right here on YouTube!
#AWS #AmazonWebServices #GenerativeAI
