Published on

Self-Improving Agents

Presentation
Authors

The sixty-second MeetUp of the Machine Learning Singapore Group, was titled : "Self-Improving Agents" - and had talks with a wide variety of ideas related to Agents.

My Presentation

My talk was titled "Self-Improving Agents", and had the following outline:

  • DSPy
  • Darwin Gödel Machines
  • GPU Kernel Scientist
  • Wrap-up & QR-code (the latter to reduce audience distractions)

I had already been working with Evolution methods to create GPU kernels, as part of my involvement with the AMD Developer Challenge (2025). So this talk allowed me the opportunity to talk about the "GPU Kernel Scientist" paper that I put together after the last MeetUp. This paper was Accepted to the ES-FoMo workshop at ICML in Vancouver (in July 2025).

Coincidentally (or perhaps because there's a lot of it in the air at the moment - see AlphaEvolve at last month's MeetUp), the Darwin Gödel Machines paper, with Singaporean first-author Jenny Zhang was also released prior to the MeetUp. That gave me a nice segue into describing what they were doing!

I also couldn't resist a bit of a plug for DSPy, having seen the nice (short) course on Andrew Ng's site.

The slides for my talk, which contain links to all of the reference materials and sources, are here :

Presentation Screenshot

If there are any questions about the presentation please ask below, or contact me using the details given on the slides themselves.

Presentation Content Example

Other Presentations

I was proud to also have two other speakers presenting at the event: Vivek Kalyan, and Leonard Loo.

Vivek's talk "RAG, Agents, RL" described his journey training an Agentic RAG system using Reinforcement Learning for the legal space.

Leonard's talk "Rethinking Superapps: Voice-Driven Multi-Agent Systems with Gradio MCP" showed how LLMs could be integrated with MCP servers provided (ideally) by superapps (such as Grab, Rakutek, RedMart, etc).

Acknowledgements

Many thanks to the Google team, who allowed us to use Google's Developer Space, and provided enough food for about 2/3 of our audience.