Browse Comments — Clean (de-noised)
Close reading of the corpus at each pipeline stage: raw → clean → relevant → coded.
140
comments matched
· page 6 of 7
The #1 most used model on OpenRouter right now costs $0.07 per million tokens (Hy3 preview). The #2 costs $5.00 (Claude Opus 4.7). Wild gap. check out for more upto date information
So something like Qwen, but using significantly more specialists. Is it useable for software dev tasks today?
Mike Pappas ensemble architectures may indeed outperform brute-force scaling, but don't you think that orchestration complexity, latency, and consistency across specialized models will be the real test at production scale?
Is it on hugging face ?
Is Deepaseek available in the US?
See Michael Lentz - told ya!
The voice ≠ transcript point is the strongest part of this ; curious how the orchestrator handles cases where the wrong sub model gets picked .. does aggregation recover, or do errors compound? :/
Interesting
Is there a simple way to try it out?
I keep getting this advertising for some reason. I looked into it and their 1% is not OpenAI's 100%. It's not like the general purpose LLM's that we know.
What stands out here is less the headline cost comparison and more the architectural philosophy shift: specialized systems collaborating dynamically rather than one increasingly massive generalist model attempting to do everything at once. If the efficiency and performance claims continue holding up at scale, ensemble-based approaches could end up influencing far more than voice AI alone.
This is an interesting shift in thinking, especially moving from single large models to coordinated ensembles. If the efficiency gains are real, it could meaningfully change how people approach compute heavy AI systems. The emphasis on voice nuance and context is a real gap in current systems, so that focus makes sense.
Fascinating. I'm looking forward to testing it!
Huge one, congrats! VCs have been underwriting this thesis for months - keeps surfacing across the newsletters we track at Byblos. Excited 🙌
Def going to read up on this, if this works at scale, you may be on to something 👏
Oh wow I didn't know anybody was doing this already! This is the way forward.
So, it's a factory of a lot of SLM's?
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?