Browse Comments — Clean (de-noised)

Mel Morris CBE

Unimpressed. Great advertising and self-promotion though. Bawwston.

Peter Signore Yes, it's called the BS model, Bachelor of Science model. It's like siccing 1000 ants on a spider. Who do you think would win? Please don't say Al Capone. Russian.

Exciting leap in AI efficiency

Justine Whitaker

This is a really important shift—and it mirrors what we’re seeing across AI more broadly. We’re hitting diminishing returns on brute-force scaling, so architectures that decompose intelligence instead of centralizing it feel like the natural next step. The ensemble idea makes a lot of sense, especially for modalities like voice where flattening everything into text destroys meaning. If ELMs can truly orchestrate specialized models dynamically at ~1% compute, that’s not just a cost win—it’s a fundamental architectural rethink of how intelligence should be built. Feels less like “bigger models” and more like better systems.

Hmmm..this is a very interesting signal of where AI architecture is heading. As systems move from single monolithic models to ensembles of specialized ones, the challenge shifts from “how smart is the model?” to “how do you reliably orchestrate and supervise many moving parts over time?" It feels like we’re entering the era of AI systems engineering, not just model scaling.....at our company adrondak, we are working on the runtime/execution layer. Our system IGLOO governs what the agent is allowed to do.

Cost of what for what?

Joe Walsh we were just talking about this

Ensemble methods have been around since before the deep learning revolution. The most famous example is the Random Forest, an ensemble of hundreds of specialised decision trees - each decision tree has limited capabilities but adding up these limited capabilities can provably yield higher capabilities. Does your approach build on this idea, your terminology made me think of it?

Jon Hammant really new interesting approach

thanks! Will take a look

Thats amazing. Does this challenge the fundamental inference costs built into the financial models that are powering the "we must send chips to space" or "raise 100B or die" mode the industry seems to be in, OR do you think this has been priced in?

Just FYI: ELM is already snagged by something with a... not great history:

Wenyi Zhu

Joel Harrison

Doesn't pass the sniff test.

Interesting write up Mike Pappas. Orchestrating to the best chip for the workload is what makes I/ONX High Performance Compute so unique. Different types of chips from different vendors (heterogeneous compute) is what we do. I can't help but wonder if there is a way to partner together. We have successfully leveraged much smaller chips to run on far less energy/water, etc that actually run faster. Super interesting either way. Well done!

This direction makes sense. Monolith models don’t scale forever, you specialize, route, and orchestrate. Ensembles aren’t new, but applying them cleanly to perception and decision layers is the interesting part, especially for voice where flattening everything into text throws away signal. The real test will be orchestration cost, latency, and eval discipline.

Wich difference is between other MoE LLMs and Modulate's approach? As i understand, both are about splitting LLM out.