i expect one day models will essentially all exist in a system like this, with sensors like real time vision, audio, and a variety of other signals being passed in. urgent decisions (reflexes) will be on a bypass route, and everything else goes through some amount of iteration and confirmation. we can use several base trains or loras or temperatures to improve variety and creativity.
only it'll be with z stacked angstrom chips (maybe with adiabatic logic) and optical everywhere it make sense, so it uses a fraction of the electricity and returns almost instantly.
@emmysteuer It IS unregulated. I was also shocked to find that private enterprise is equally inefficiency and it wasn't just the military. But after 20 years of work I've found the number of truly efficient processes and industries are very few.
My fable/opus usage as orchestrator/front end review (respectively) ran out several days faster than Sol max doing all research, build, testing, and backend reviews.
I like Fable but it's only a hair over Sol max and almost not worth the headache of fallbacks. With Luna pricing and the stronger overall model lineup I'm tempted to just cancel Claude and pick up a 2nd GPT sub.
So I guess orchestration is the way. You could prob pick up a $20 and get thru a week using only Luna xhigh for build and testing, just set a subagent hook in CC to call Luna or have it open in a terminal and send directly to the other session.
@trashh_dev Inuyashiki, Made In Abyss, Daily Life of the Immortal King (not obscure but very entertaining), Summer Time Rendering. Haven't updated MAL in a long time so most of these are older
@Ananth7e It's been pretty useless since 4.6 IMO. Sonnet was cheap, strong, and fast. Unfortunately it seems they've struggled to get a good distill for either model since then.
With Luna being significantly cheaper to use, I'm trying to think of how to maximize value with it. Has anyone used Luna xhigh as a primary subagent for research and building? Is the quality as strong as Sol only?
I've been quite happy with Sol max quality of work, I think it'd make a strong orchestrator. If reset lands I'll run a 3 way test tonight to see how Luna does under instruction vs Sol handling all roles.
Yup that's my take as well. It's pretty widely accepted at this point that harness and scaffolding is just as important as the model. Not allowing a model to work as intended just means the bench no longer has value, because it's no longer representative of the model's capabilities or the experience people will have with it.
I dunno if people just haven't tried it or what, but I have a sci-fi three.js first person explorer cooking with Sol max and it's excellent. And it will probably end up being like 15% of my weekly