#Model Branching. How to make true general-purpose, and making everybody happy

1 messages · Page 1 of 1 (latest)

radiant spire
#

TL;DR
I suggest three model branches:

  • Creative (GPT-C) for writing, roleplay, style, ideation, and narrative work.
  • General (GPT-G) for default everyday conversation and lightweight productivity.
  • Logic (GPT-L) for coding, research, math, and deep reasoning.
    The system could route automatically by intent, while also allowing manual selection for users who want control.

Longer version:
OpenAI has consistently aimed for a "general-purpose" model that excels across many kinds of tasks. I think that is an exciting goal, but OpenAI’s current model lineup already suggests that different post-training profiles excel at different categories of work: everyday conversation, deep reasoning, coding, and research are not all optimized equally well by the same model behavior. If OpenAI wants to move closer to a truly general-purpose assistant, the most feasible path may not be one monolithic model profile, but a routed family built on a shared base.

With that in mind, OpenAI could restructure the ChatGPT lineup around two axes: branch and mode. Each branch would have its own modes. The schema I propose is:

  • Creative (GPT-C) for writing, roleplay, style, ideation, and narrative work.
    • Instant: A low-latency, high-throughput option for creativity-first tasks
    • Thinking: A deeper creativity-first option for complex narratives, continuity, and characters
  • General (GPT-G) for default everyday conversation and lightweight productivity.
    • Instant: A fast and direct option for casual chat
    • Balanced: A casual-chat option with reasoning for harder questions
  • Logic (GPT-L) for coding, research, math, and deep reasoning.
    • Thinking: Reasoning for very complex questions and tasks
    • Pro: Research-grade intelligence for long, difficult problems
      As I mentioned earlier, the system would automatically route to the most plausible model, but users should be able to select the model they want manually too.
hazy portal
#

Boost!

radiant spire
#

A branched system could reduce a common source of dissatisfaction: one model behavior rarely satisfies everyone. Different branches would let OpenAI align tone, reasoning style, and optimization priorities more directly with user intent.

grave patio
#

Sounds interesting.
OAI really wants to get rid of the model picker, but maybe more model pickers for people who are picky (like me) is just a good idea.

#

Since those manual model pickers also train the wrapper / router models on how to better route auto pickers.

hazy portal
#

I feel like it would also help in general.

That way, when they make updates to the different types of models, it would be toward their specified use cases and there'd be a lot less people being upset that everything changed back and forth every time.

Maybe new models would start to feel more like upgrades instead of wipes with a higher knowledge base and speed 😆

From a business perspective, it seems to make sense to me. However i've never been the type to deal with the finances, usually just the type someone throws in a lab and leaves alone 😆

Still though, if every three months, my lab kept getting changed and my equipment getting replaced with equipment I don't use, i'd be just as upset.

And my poor boss would probably never hear the end of it 😆

radiant spire
#

bump

radiant spire
#

bump

wanton vortex
#

All 3 of these would be possible with the same model if they didn't have a 15,000+ character system prompt giving it tons of policy, some of which contradicts itself.

You have to approach how you theorize about ChatGPT by seeing it as a product in search of self-sustaining revenue. Satisfying casual or "creative writing" users will not make openai solvent. If they don't eventually become profitable then ChatGPT won't exist at all. The product can still be made better for casuals, but only in non-specific ways that benefit the product over all

pine sage
#

W suggestion