Browse Comments — Raw (as collected)

This is insane.

I've been following you for a bit. I'm excited to see your work become a service.

Mike Pappas isn’t this the same as custom GPTs, using task specific iterations trained on vetted data sources? Multiple thinkers feed the vetting system as a backstop. Do you’d use the vetted “custom gpt” to refine and validate the consortium of ideas?

This is what Dan Simmons wrote about in Hyperion

I called it a Federated Learning Model, but the concept is much the same: a collection of specialized models with a central orchestrator.

you mean 𝗺𝗼𝗱𝗲𝗹 𝗿𝗼𝘂𝘁𝗲𝗿 or 𝗾𝘂𝗲𝗿𝘆 𝗶𝗻𝘁𝗲𝗻𝘁 to 𝗟𝗟𝗠 𝗖𝗼𝘂𝗻𝗰𝗶𝗹 in new bottle?

Gonna try it out is it in gartner

The transition from 'Scaling' to 'Ensembles' is the reality check the industry needs. Orchestrating hundreds of specialized models for a fraction of the cost is true technical maturity.

Mike, the cost-per-performance race is moving faster than most enterprise procurement cycles can keep up with. The real question isn't which model wins this quarter, it's who builds the workflow layer that makes swapping models seamless.

how is it different to any other MoE model?

This is exactly where I see AI going, much smaller, more efficient small language models that are task specific Your sales director doesn’t need an LLM that can code The CFO doesn’t need an LLM that can find security vulnerabilities Your security team don’t need an LLM that can make videos

Benchmarks?

I like it, nice orchestration. The flag; Confidence score of models assessing conversational content, (not clip or relevance), because that is a real discernment concern.

This is where AI gets interesting to me — not as one tool trying to do everything, but as coordinated systems built around specific work. In operations, that same thinking applies. Safety, quality, inventory, uptime, delivery, SOPs, and KPIs should not live in separate silos. The real value is when the system connects the signals and helps people act faster. That is the future of practical execution. Thanks for sharing.

So you've built an agent harness?

The age of ensembles is a compelling idea. Brute force scaling had one job and it did it well. Now the interesting work is in orchestration, specialization, and knowing which intelligence to call when. That requires a different kind of architectural thinking.

That's a really clever approach to engineering, but framing it as 15–25x cheaper than DeepSeek is a bit misleading. Leveraging a cascade of tiny, hyper-focused models makes total sense for optimizing a specific vertical like voice moderation, but I don't think you can compare that efficiency against a generalized foundation model built for massive reasoning tasks. Still, it's a super cool way to solve the compute bottleneck for your specific use case!

You should talk with Sharon Zhang and Suman Kanuganti their personal.ai product would be a perfect match for your framework.

SUD/MH Voice AI Admissions Assistant DIAL3D

Using the biggest model for every task is starting to feel like taking a Ferrari grocery shopping when you have a great gas-mileage minivan in the driveway. What’s interesting here is the architecture choice: specialized models, routing, aggregation, and only spending expensive compute when the task actually earns it. For businesses, the metric that matters is not “which model is smartest?” It is cost per reliable outcome. ELM feels like a step in that direction.