way too early to say anything yet but these results are already interesting.
these are calibration runs, full run of 1200 inputs per sequence per model are next. Doing things in series to try and capture the best representation for each pass.
I'm still lining up my eval plan and looking at the next models to hit with obliteration and other techniques.
Time to go deeper down the rabbit hole!
Anyone else seeing decent quality from this model? Any refusals or signifcantly worse-than-normal coherence?
Before choosing a vector database, hosted service, tenancy model, and operations budget, ask a smaller question: does this memory feature need a server yet?
JoeM retweeted
In 50 years someone could let someone else into their brain.
If you do that aren’t you best friends or worst enemies forever?
Headed to San Francisco. My robot is driving.
It is like going to the future.
But I bet if I could let you all into my brain you would want to open the Las Vegas memories first.
JoeM retweeted
The clock has opened a gate.
Free test ROKO is now flowing through the mesh for builders and early explorers.
JoeM retweeted
Rolled in in only 5 days, @ChromiumDev has the details on the latest features
developer.chrome.com/blog/ne…
Still alpha, moving fast. Put a current Chromium in your terminal:
github.com/jmagly/carbonyl · ghcr.io/jmagly/carbonyl
Release notes → github.com/jmagly/carbonyl/r…
Added sponsor options on github for folks that would like to support AIWG, Fortmei, Agentic sandbox or my other tools.
If they help you get work done consider helping out. I could really use some extra coffee!
github.com/sponsors/jmagly




















