Tim Grillmeier

Technologist, theorist, developer, and starter based in Brisbane, Australia. I aim to bring clarity and meaning to a noisy world.

[Tim Grillmeier @timgrillmeier]

With the right company and mindset you begin to notice that problems cause solutions.

[Tim Grillmeier @timgrillmeier]

Exploring carnot line formation in sub-riemannian construction space.

[Tim Grillmeier @timgrillmeier]

I can not begin to overstate how right I was.

[Tim Grillmeier @timgrillmeier]

I’m beginning to lean towards the notion that the memorisation problem in machine learning has its roots in the established definition of an artificial neuron as a unit.

[Tim Grillmeier @timgrillmeier]

If spud is agentic from the ground up, my bet is it will fail to be generally reliable while reliably outperforming agentic harnesses in niche cases.

[Tim Grillmeier @timgrillmeier]

Grammar is a Lie algebra.

[Tim Grillmeier @timgrillmeier]

Applied AI ⊂ Agentic Systems

[Tim Grillmeier @timgrillmeier]

The next major shift in advertising will direct ads toward people’s AI agents, not toward people themselves, if people are targeted at all.

[Tim Grillmeier @timgrillmeier]

Band name #9876

Proxy PR

[Tim Grillmeier @timgrillmeier]

Received my first AI telemarketing call the other day.

Did you know you can use prompt injection techniques verbally to get an AI telemarketer to abandon their script and give you a recipe for cupcakes?

Because you can. And I did.

The cupcakes are delicious by the way 🧁

[Tim Grillmeier @timgrillmeier]

Reworking perceptrons with IG special cases, there might be implications for detecting epistemic cases, like knowing what you know with high/low certainty. This includes hallucination. A neural net with extended perceptrons could show these conditions as signals during inference.

[Tim Grillmeier @timgrillmeier]

Yep, confirmed it. It's not just overfitting, it's a structural deficit. Time to redefine what an artificial neuron should be.

[Tim Grillmeier @timgrillmeier]

I’m beginning to lean towards the notion that the memorisation problem in machine learning has its roots in the established definition of an artificial neuron as a unit.

[Tim Grillmeier @timgrillmeier]

Been experimenting with "Identity Injection" for prompt response cohesion. Pretty insightful results, especially relevant to #aisafety and a unique lens on jailbreaking. Turns out identify robustness in LLMs might be crucial for policy adherence.

[Tim Grillmeier @timgrillmeier]

I trained a single homogeneous transformer on 2 distinct algorithmic tasks simultaneously. No task labels, no routing, no MoE. I explicitly destabilised memorisation midtraining.

The model reformed and stabilised both solutions, achieving joint grokking in 3050 epochs.

#ml #ai

[Tim Grillmeier @timgrillmeier]

9/n) To put this into perspective, the efficiency gains on this mean my 3090 would grok with my approach at the time time it takes a H100 to grok with brute force. I think that's a pretty big deal.

[Tim Grillmeier @timgrillmeier]

1/n) I've been busy on a personal research project lately. A lot of things are falling out of it, including some unique approaches to #machinelearning. I've been able to grok for Fib mod 31 with 100% accuracy from 60k-151M parameter models for effectively memorisation time.

[Tim Grillmeier @timgrillmeier]

8/n) At the start of this thread I said "100% accuracy" what I meant to say was "100% reliability" - no run has failed to grok.