@RRunner144
Account based inNorth America
About this account
- Account based in
- North America
- Connected via
- North America Android App
Account-level information from X, not a live location or the device used for a specific post.
Subsea Engineer, Trading Indicator Dev and AI-Memory Layer Specialist | @echo_locate_ai | @rainbowtrends
NFA
Joined May 2019
- Tweets37.8K
- Following1.7K
- Followers7.6K
- Likes47.8K
How it feels getting DeepSeek to run on a ddr3 workstation.
Project Maya v1.1.18 is out 🫡
- DeepSeek: faster long prompts on one GPU, smarter use of PCIe
- Strix Halo on Windows: decode back to full speed
- GLM on Tesla P100 + Responses API
- DeepSeek respects anti-repetition settings; steady speed meter
github.com/mw00/project-maya…
Unreal what this guy is doing. Check for releases at least twice a week.
Strata 0.2.0 is now live in beta! It includes its own in-app harness and terminal-serving interface.
github.com/Niko1221/Strata
A lot of performance improvements and support for low-end devices 👇
Skirt and ribeye steaks are the top two flavor wise. Skirt is my all time favorite, but you have to prep these things to not get a tough cooked steak. Trim the fat, then let these things sit in an elevated wire rack for 30min before you start cooking them. If they aren't room temp when you grill, they will turn into jerky. 2 to 3 min per side then wrap in foil and let it sit for 10min before you serve.
8 Million tokens later and my Z840 has been able to get the @RainbowTrends Runner Bot Algorithm running on @TrendSpider.
Coming up on the 90% complete mark and over 20mil tokens converting my source code over to javascript with a few 3090s and some 5060Tis. Never thought I would ever be able to say this, but ican confidently confirm @RainbowTrends is coming to @TrendSpider. If there is interest in a tutotial, I can share the setup in detail. 2 or 3 agents with the proper setup can convert just about anything. Stuff I failed to convert with Opus.
I had a second machine without an NVME and tried to get Coder running and shockingly it works and outcodes Q6 27B. I repeat, this is running WITHOUT an NVME.
50tok/s w/98K Cache Q8
Windows 11
5600X w/32gb DDR4
EVO 860
2x5060ti 16gb
New update is here!
- Apple Silicon is now supported
- NVFP4 Support, UD-Q4, UD-Q5, UD-Q6, UD-Q8
- Qwen3.6-35B-A3B is also now available
- Another 10%-20% boost
and more: github.com/Niko1221/Strata/r…
Coming up on the 90% complete mark and over 20mil tokens converting my source code over to javascript with a few 3090s and some 5060Tis. Never thought I would ever be able to say this, but ican confidently confirm @RainbowTrends is coming to @TrendSpider. If there is interest in a tutotial, I can share the setup in detail. 2 or 3 agents with the proper setup can convert just about anything. Stuff I failed to convert with Opus.
Watching Qwen 3.8 Flash w/Strata engine on 2× RTX 3090 24GBs build out @RainbowTrends in real time to Javascript on @TrendSpider.
My Z840's numbers trying the IQ3_S, Q4 and Q6 quants on dual 3090s at 250w with the updated Strata engine from @coldniko. Unreal numbers, but sticking with IQ3_S for now, once you average 120+ tok/s its really hard to go back.
New version is now live with another performance boosts and improvements on Strata
github.com/Niko1221/Strata/r…
My daughter always gets a hard time being one of the only girls in robotics clubs. I always tell her about Margaret a few times a year and tell her don't let the boys tell you anything.
We’ve lost an icon. Margaret Hamilton, a computer science pioneer best known for her work at MIT leading development of the flight software for the Apollo missions to the moon, has died at 90. Read about her legacy via MIT News: news.mit.edu/2026/margaret-h…
ALT Margaret Hamilton is reclined inside a NASA spacecraft module, surrounded by control panels and switches.
Well it looks like this week some people are going to find out the hard way that some GM Diesel Fuel Injection Pumps (especially 6.5L) run on optical sensors that don't exactly work when you run the red Kool Aid diesel.
Codex found a way to control the fans on the Z840 which were not meant to run off dynamic fan curves. Now I have it checking both 3090s every 5 seconds, picking the hotter of the two (not always are both cooking) and applying a dynamic fan curve to it. 3090s are getting an extra 50w of power each now.
Update, this took about 5 million of self hosted tokens to get this far. I've been trying to get Rainbow Trends on other platforms and never been able to do it with Claude or Codex. Right now I have an agentic team of Qwen agents chewing through this with unbelievable success. I know people laugh at self hosting your own AI, guys, I flat out COULD NOT get this to work in the slightest with Claude..... Its nothing special, no gimmick 10,000 agent swarm. It's an Orchestrator with two sub agents (one fast, one slow) working in Hermes. Setup is per the below:
Orch- Qwen 3.8 Flash Next w/Strata
Auditor - Qwen 3.8 27B
Speed Worker - Qwen 35B - Tiel Coder
Harness - Hermes Agent - @NousResearch
Orchestrator - Qwen Strata - @coldniko
Slow Worker - Qwen 27B - huggingface.co/unsloth/Qwen3…
Fast Worker - Qwen 35B Tiel Coder
huggingface.co/peculiar-ragd…
