DEV Community

Cover image for Are we Fighting the Wrong Fight with AI? It's Not a Model Problem, It's an Architecture Problem

Are we Fighting the Wrong Fight with AI? It's Not a Model Problem, It's an Architecture Problem

Sandro Garcia on August 07, 2026

I just read AI Coding, One Year Later: What August 2025 Didn't See Coming by @lukeocodes, and one section stopped me cold. Not the benchmark tables...
Collapse
 
96alchemist7 profile image
Blake W. Allcorn

Don't need a new model... need a new perspective of the model.
I guess that's all i should share.

Collapse
 
sandrog profile image
Sandro Garcia

Right a new perspective of HOW to use de model. We are charging all responsabilities into de model.

Collapse
 
96alchemist7 profile image
Blake W. Allcorn • Edited

Well its that they're overloading it with unnecessary functions; and not giving it the Optimizations.
A way of thinking is much more powerful than a tool, but those thoughts need Optimizations too...

I would take ToT (Tree-of-Thought) and DoT (Direction-of-Thought) and CoT (Chain-of-Thought) any day over a pipeline.
But do you know what you need to make those work better than a pipeline? (just wondering, I'll tell you~ if not)

Thread Thread
 
sandrog profile image
Sandro Garcia

Just to clarify: I'm definitely not advocating for rigid pipelines (that's actually what I'm trying to move away from—static DAGs are far too brittle). The idea is dynamic, message-passing specialization.

That said, I'm completely with you on cognitive frameworks like ToT and DoT over tool-stuffing. But you left me hanging—what's that missing optimization in your view?

Collapse
 
96alchemist7 profile image
Blake W. Allcorn

i think my message is being blocked?

Collapse
 
sandrog profile image
Sandro Garcia

Hi Blake! Your comments weren't blocked—I was just in the middle of a major production rollout this week and couldn't sit down to read and reply properly.
​As a systems architect, I've been developing an open-source protocol for decentralized multi-agent networks (feel free to check out irc-a.org if you're curious). I'd love to see how your scratchpad/blueprint approach structures cognitive state, and explore whether any of those pieces complement the architecture. Always happy to give full credit for community ideas. Feel free to share more!

Collapse
 
96alchemist7 profile image
Blake W. Allcorn

I will check it out, I am just not so trusting of open source, it's been difficult to even show anyone cause no one usually even answers back 😅 makes me want to keep the secret sauce tell I can afford to give it up. I have the cure for a lot of what enterprises are paying millions for; and im looking for someone that i can trust that wants to earn a big check for it honestly 🤫 but yeah ill check it out.

Collapse
 
96alchemist7 profile image
Blake W. Allcorn

SIMPLE ANSWER: ymal_scratchpad and ymal_blueprint i guess im not allowed to tell you more.
I was wondering if youbwanted a meta-os for the system you want

Collapse
 
96alchemist7 profile image
Blake W. Allcorn • Edited

No cost, just agree to give me credit (work cited) for it if you write about it. no catch just that.