Rambone

My custom inference engine progress is going well. Already beat my strata numbers on token gen, not yet prefil speeds. Iteration is so slow that I don't think I'll burn a whole $20 claude plan.

Nobody likes a heavy phone. Its actually the worst thing ever once you experience a light phone. Due flopped here

The feeling of having infinite tokens on a frontier model must be how billionaire feel all the time

Im honestly surprised how Nvidia isn't a frontier lab, I guess they are making so much money they don't have the human resources to allocate to it. You would think it would be in their interest to create "frontier" level open models. Nobody cares about the consumer anymore (except you qwen ❤️)

Opus has broken ground on my personal inference engine project. Optimizing specifically for my 2x rtx 3090 setup on qwen flash. The goal is to beat the baseline strata numbers of ~140tps and 3000pp.

Unless they drop 10x compute with astra 6.2 im done with codex. Constant overload

Reach is up people, im getting over 40 views per post 😎 (on average)

I saw someone mention computer use on their phone. In about an hour with my custom harness on termux, my fold 8u now has computer use. The agent now has full control of my phone, can use all apps, browser, etc. Idk why I didnt think of this sooner. Super cool, idk what I'll use it for yet.

I hope the new stealth model is qwen 4 🙏🙏

As soon as astra 6.2 drops im making qwen 4.1 flash

Deepseek, hack my washing machine.

I give it 2 weeks before astra won't allow you to modify open weight llms

When im at work I pretend I have 4 rtx 6000 pros and spawn 10 deepseek subagents to fix my ui

me: make the bg white codex: sure me: why is there a login system

Im surpised less people aren't obsessing over typing speed. I cant dictate at work so my typing speed is a big measure of how productive I can be. Fast typing = more thorough prompts = better work.

Calling an agent big dawg before a task. Does it work better? I ran it and the result was no, it does not make it perform better. X lied. On a serious note I've been really interested in testing the psychology effects speaking to agents in various ways have. More to come.

Someone was going to do it. Its so so much faster then either researching yourself or opening the chatgpt app and getting a long winded response.

Wake me up when they figure out how to make a model that can fix basic things without +200/-15