#harness

New post: Wikipedia and friends: learning through code https://lu.is/2026/08/wikipedia-and-friends/ #llm #wikipedia #wikidata #opensource #harness #openknowledge #commons #glam

The second (Not final) iteration of the Harness WU 'operational language' visualistion... ...looks better, needs a little more tweaks so its more comprehensible... BTW: I love the oldy worlde steampunk theme, I may apply it to the whole harness as an option :) #Harness #Visualisation #Vibecode

Kanga on the fishing pier, Quarton Lake, Birmingham, Michigan. April 2014 . 🌐42°33'4.98" N 83°13'36.786" W . ◉ smc PENTAX-A 50mm F1.4 📷 Pentax Pentax K-5 🎞️ISO 80 ⏱️1/180 sec f/4.5 (𝑓=50 mm) . . #dog #animal #collar #companion #dog #harness #metal railing #nature #outdoor #pentax_fa5014 #pentaxk5 #pet #pier #pitbull #posing #waterfront #pitbeagle #pittie #dogs #dogsofinstagram #dogsofmastodon

Dont be like oak. Be like reed in the storm. I have this problem in my Ai #Harness. I take a complex problem, decompose it, and then varioius engines to bruteforce it... The decomp is into a WU (Work Unit) chunk and sometimes they fail. I had one roll through 400 iterations (I stepped away from the keyboard for a few days /eyeroll) My solution is, to run a decomp again on that failing WU. Rather than the whole (15 page smoketest) prompt... ...and again... ...and again... ...and again... until something that doesnt die pops out... I can do this because I am rotating throug various models. Using the stochiastic nature of #LLM as a strenght... now a weekness 😀 #Vibecoding

I have also added accounting page to the #Harness Shows Cost, Energy use, Water use (Open and Closed loop system) and CO2 emission (assuming you are running the datacenter on the dirties coal/methane) So far, in the few weeks I was developing the harness... the total multi-model use equals running 14 water boiling kettles for an hour. #Datacenters #Vibecoding

The spinny 3D Ai Harness multi graph... ...why? Because we can. Engines that hit their compute limits are red. Some of the small wireframe engines have not had their "Power" evaluated yet, so the Harness router has them on the end of fall-through hierarchy. The skulll is a local guardrail free - model. Oh, and I bought some API time on togetherAi. Because its cheap and gives me access to GLS and Kimi models which are comparable with #Anthropic Opus Using those two models, I now can have a viable alternative to Opus/Sonnet with adequate tokens. #Vibecoding #Harness

...aaand Im out of compute. Implemented a visual dashboard for the engine router because it was an algorithmic black box. Its still computing the new revised WOOFER V2 values (more angry judge so we dont get 100%). So the values are still fluid. But you get an idea. There are three classes of decomposed WUs (Working Units), the engine router assigns the engine to the WU based on WU difficulty. There is a manual slider now so I can tweak difficulty floor. #Vibecoding #Harness #Harnessengineering