Felipe Rosa

AI Researcher
Founder & CSO @namastexlabs

[Felipe Rosa @ferosa888]

@faqdev scale

[Felipe Rosa @ferosa888]

@tibo_maker and similarly as you, I found confort in improving the orchestration layer, where solutions like herdr and orca do most of the work, allowing me to focus on the problems actually being solved.

[Felipe Rosa @ferosa888]

@tibo_maker ive been basically hoping to talk about his publicly someday. thank you! feels exactly the same over here, to the point of burning out, losing work life balance entirely. theres probably FOMO involved for most of us.. just for having a max account seats us under .1% worldwide

[Felipe Rosa @ferosa888]

https://t.co/0MPb1mpLE2

[Felipe Rosa @ferosa888]

building autonomous trending research, autonomously. I'll report back as it progresses, including time and costs of this project. some formulas exists for a little while. intellectual contributions to the way we measure are very welcome.

Begin. Day 1 (data extraction still isnt real)

[Felipe Rosa @ferosa888]

I have to say... @OpenAI's harness evolved so much over the past few months that I have felt, for the first time since I began researching GenAI, that some parts are better than my own open-source composition... well done. Let's just hope that we follow a good path for humanity, @sama

[Felipe Rosa @ferosa888]

@raphaelcrosa @deepseek_ai mandei no grupo no mesmo dia, to brincando já kkk

[Felipe Rosa @ferosa888]

@VictorNewsie @birdabo that you should explore open source harnesses and models. being vendor locked to models and their harnesess will get you as far as they want you to go. and during peak hours youll be served with quantized garbage or get silent model routes to smaller ones.

[Felipe Rosa @ferosa888]

@birdabo too busy getting smoked and routing fable requests thru sonnet on max subscriptions.

[Felipe Rosa @ferosa888]

@VictorNewsie @birdabo maybe a sovereignty shot will refresh you in the not enoughness :) cheers

[Felipe Rosa @ferosa888]

same.. whats your stack and model preference today?
me: hermes for personal assistance + custom modules running over a ryzen ai minipc with local models, sol over codex apps for llm judging. everything runs over our labs context framework (https://t.co/t9QqqeJ3zU) + something similar to hedr). fable + sol planning and final reviews, prime-agent + ds40731 for everything else. from 1 month ago, went from maxing 5 max accounts, to barelly using one + less than 5$ in DS costs.. Insane times.

[Felipe Rosa @ferosa888]

@DanDr1s also, its been a potato with gpu capacity exausted (lobotomizing alert). you simply cannot get consistent results.. not to mention hidden model routing, while still charging you full. but if you're happy...

[Felipe Rosa @ferosa888]

@arthurkatcher switch to a open harness and swap models freely, break chains from vendor lock, thats the recommendation here. (eg. pi dev)

[Felipe Rosa @ferosa888]

@thsottiaux codex app server and that custom harnesses are supported. this is essential to me.
improve? faster inference. what happened to the codex-spark project?

[Felipe Rosa @ferosa888]

@pvbuilds @aditya12anand plan and review qwith fable and sol, execute with DS4 over decent harness.

[Felipe Rosa @ferosa888]

@AdityaShipsHQ theres also pi dev and opencode 2 out there... anything is better than rewiring a vendor harness at this point, cheers

[Felipe Rosa @ferosa888]

@AdityaShipsHQ to begin with, USE A PROPER HARNESS.

https://t.co/9ZbIBrAH8G

youre welcome :)

[Felipe Rosa @ferosa888]

Claude code has been routing Fable requests to Sonnet 5, while still charging me for it...

heres the proof I have never even requested sonnet to anything.

whats going on @AnthropicAI ?!