Browse Comments — Clean (de-noised)

Learning an optimal policy over a universe of models, why didn't any one else think of it? 🤔

Fascinating developments

Hey man! We gotta talk. I have built a Distributed AI Infrastructure platform, and one of my platform’s core abilities is the ability to analyze media formats for enterprises to learn from with cross domain learning capabilities. We also do this, while ensuring our clients keep their data sovereign. I have an open-source version of my Distributed AI Infrastructure platform on GitHub (doesn’t come with the above capability, that is the full version - which is already built), but it comes with highly useful capabilities that you and your dev team can easily play around with and integrate into your platform for efficiency gains, intelligence expansion, and cost mitigation on compute/energy within your organization. Excited to talk about AI and orchestration as we are mutually aligned in that category!

Interesting, indeed.

Very cool! The voice use case is very tangible. The idea of using smaller model models + more specialization and structure is certainly appealing as well - will be great to see how ELMs can contribute to this!

How is this different from Mixture of Experts?

So you are claiming that your model beats Deepseek by a 15x on cost? Might want to be clear about this claim. For a listening model that understands emotions, it is okay to charge more because it is a very specialized model

Mike Pappas this is quite interesting! I’m surprised there can be such a large cost reduction - I’d expect the orchestration among “hundreds” of models to eat up any savings. How can I learn more? Also, on a related note: I sent you a connection request to address a UI issue with your site. LinkedIn wanted me to get Premium to message you, so apologies in advance for the public comment.

Mike Pappas this is very fascinating and congratulations on your achievement. As someone that leverages AI heavily in sales and GTM, I can attest to the inefficiency and sometimes lack of quality of current LLM output specifically around call transcripts. I can validate that key human components like sentiment and tone are often lost or mistreated. In the lens of sales and GTM, how do you see the ELM being most impactful?

I would love to see it applied to agentic coding. Is that in your plans?

Mike Pappas attach this to specific agents and orchestrate them into more of generalist interface but with the power of specialists working together. This is really cool.

Literally not even close to a new concept dude this is standard

Coming soon

Mike Pappas the cost angle is what jumps out for me. One of the most common conversations I have with early-stage founders is about running lean, and AI infrastructure costs are increasingly part of that equation. If the efficiency claims hold up at scale, this is the kind of development that changes what's actually accessible to smaller teams. Worth watching closely.

Simon Provencher

AI;DR

The key point for me is that AI performance is not only about model size or raw capability. It is about whether the architecture preserves the context that matters for the task. Voice is a good example. A transcript captures words, but not always meaning. That is where more specialised or ensemble-based approaches could become very important.

How many testing hours and benchmarking has been done with real world scenarios. Let’s just accept today creating something isn’t hard. But validation is the difficult part. After years of working with OpenAi models, I think their models are NOT production ready.

Really compelling direction, moving from monolithic models to orchestrated ensembles feels like a practical path past current scaling limits, especially for richer modalities like voice. Excited to see where this goes 👏

i was very surprised to see my preferred email@ prefix in the doc. i guess i cant talk with you guys sadly. name collisions are serious biz ess :)