Canberra AI: Unclassified

There is too much noise out there. Headlines are screaming about the apocalypse, vendors are promising magic, and meanwhile, we’re all just trying to figure out what this actually means for our life, our work, and our future.
​Let’s get back to basics.
​We believe the best way to learn isn’t through a lecture, but through community. We learn faster when we learn together, asking the questions that actually matter without fear of judgment.
​We’ve done the heavy lifting so you don’t have to. We have spent the last decade deep in the trenches of AI research—writing the code, publishing the papers, and reading the dense academic journals—so we can distill it down into plain English for you. We filter the signal from the noise.
​This is an "Ask Us Anything" session. No script. No sales pitch. Just honest answers.
​Bring your question and skepticism. Backed up by scientific research, we’ll be discussing questions like:

• ​"Is AI actually smarter than a PhD, or is it just a really confident parrot?"
• "When we say AI can 'see' or 'hear,' what is it actually doing?"
• ​"If I put a policy draft into ChatGPT, who owns it? Does it keep my secrets?"
• ​"Is 'Prompt Engineering' a real skill, or just a buzzword?"
• ​"Are we at the age of AGI (Artificial General Intelligence) yet, or is that just sci-fi?"

​Whether you’re an expert or you’ve never used a chatbot in your life, you belong here. This is a space to clear the fog and build a shared understanding.

The Reality: The current narrative says: "If you don't have 10,000 H100s, you can't compete." That is a trap. It forces reliance on massive foreign cloud providers.
The Question: What if the bottleneck isn't hardware, but math? What if you could achieve training stability and performance on your own servers, at a fraction of the cost?
The Paper: We are dissecting "Manifold-Constrained Hyper-Connections (mHC)" (arXiv:2512.24880).
Why This Matters (For Canberra): Australia cannot win a "brute force" spending war against Silicon Valley. To build Sovereign Capability, we must be Asymmetric. We need architectures that extract maximum intelligence from limited silicon.
This paper proposes a method to stabilize training deep networks by projecting connections onto geometric manifolds.

• The Result: Faster convergence, less wasted compute, and the ability to train deeper models on smaller clusters.

We will discuss:

• The "Tax" of Brute Force: Why standard ResNets waste compute.
• The Fix: How mHC stabilizes the signal, reducing the need for massive "trial and error" runs.
• The Sovereign Playbook: How to apply these techniques to host and train models on local infrastructure (On-Prem), keeping data secure and costs predictable.

Who is this for? Builders, CIOs, and Strategists who want to own their infrastructure rather than rent it.

In November last year, we showed you how to build rapid AI prototypes. The room was so packed and the questions so intense, we literally ran out of time.
​But knowing how to build is only 10% of the battle. The other 90% is navigating the "Department of No"—the legal hurdles, the risk-averse executives, and the cultural fear that kills innovation before it ships.
​We are back for Part II.
​Harshil Siyani returns to sit down with Cheng Yu (Trellis Data) for the conversation we didn't get to finish. We are moving from the "Code" to the "Culture."
​
​Big thanks:
​To Trellis Data, ACS Canberra Branch for sponsoring this event.