Published on

Getting to Aha

Presentation
Authors

The fifty-eighth MeetUp of the Machine Learning Singapore Group, was titled : "Build + Voice + Aha!!!". This event focused on Building : While it's super-interesting to observe what's going on in the wider world of AI, nothing actually beats rolling up ones sleeves and actually building a thing!

My Presentation

My talk was titled "The A-ha Moment", and had the following outline:

  • Deepseek + The Aha Moment!
    • Idea: DIY for low $$
  • Yak Shaving with jax.nnx Gemma
    • Actually getting to Aha
    • Runnable Colab to use!
  • Wrap-up & QR-code (the latter to reduce audience distractions)

To set the scene, I lead the audience through some of the significant elements of DeepSeek's R1 release (mainly for those that had not been to my talk in January), then explained the various 'Getting to Aha!' methods. For a GDE sprint during February, I had volunteered to produce a TPU version of this 'Getting to Aha' - so I explained my building journey (which had been given a significant detour due to existing code that simply didn't work as advertised). To round out the event, I showed a (take-home) Colab with suitable 'Aha' signals, but this wasn't TPU-oriented, so my sprint looks like it has become a longer race...

Many thanks to Google for supporting the GCP usage for this project, which was part of their February 2025 #VertexAISprint. My contribution there was titled: "Getting to Aha on TPUs". I also used Colab extensively for testing : Many thanks to the Colab team!

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

In his talk "Building Voice bots both local and not", Sam Witteveen talked through the history of voice-enabled bots, which Red Dragon has been building since 2017. Sam finished with some recent examples, making use both of local models running on his laptop, right through to using the latest Gemini models to create bi-directional conversational experiences.

We were also please to welcome Florian Kowarsch back for a lightning talk "Domesticate the Beast". In his talk, Florian described the two main types of prompt/performance evaluation required with today's LLMs : Evaluating deterministic outputs in classification tasks and handling subjective criteria like creativity and precision.

Acknowledgements

Many thanks to the Google team, who not only allowed us to use Google's Developer Space, but were also kind enough to provide Pizza for the attendees!