Are you talking about github.com/joshuadavidt...? I've been testing it too. It seems pretty good about injecting `bd prime` into the context, but it certainly would be nice if the todo list on the right of the opencode tui was synchronized with the subset of beads being worked on.

Install Taskfile.dev and make a global task for it. That way, you've got a 66.6% chance of getting right instead of just 50%.

Sadly, it's not local-first, so data still goes to a cloud. "Architecture: Smart glasses connect to user's phone via BLE; phone connects to backend; backend connects to third-party app servers running the MentraOS SDK" - the project's own instructions for coding agents.

"I really wish I could steer with my coffee mug"

They got a great deal on introductory offer. www.gsa.gov/about-us/new...

It looks beautiful and intuitive, and I would absolutely be trying to make the logo balanced instead of focusing on just the color palette. Maybe it would help if you could rotate the background without affecting the selected colors?

We have gotten worse at illusions over the years. Sad. en.wikipedia.org/wiki/Mechani...

Temperature regulation in space is like peak HVAC tech. NASA has a great article from 2014 that talks about how they do it on the ISS and how they solved an equipment failure. appel.nasa.gov/2014/03/04/t...

It doesn't use GPS, and the documentation says it won't work without user consent to automatically update their location in Teams when they connect to specific devices or wifi networks. learn.microsoft.com/en-us/micros...

... and there's a downdetectorsdowndetectorsdowndetector.com ? Where does this rabbit hole end??

That's something they think a lot about at Cloudflare. They discuss some strategies here: blog.cloudflare.com/private-rate...

The radiation induces current in circuits, which can flip bits and fry chips, but that risk can be mitigated somewhat with things like shielding and older (larger) IC fabrication technology. NASA is a good resource on the subject. s3vi.ndc.nasa.gov/ssri-kb/topi...

Oh, absolutely! And I'm just saying that re-rolling can produce results that are just as good. It's a gamble as to whether the results are better one way or the other.

Reinforcement learning does not happen during inference. The probability space shifts with each token during inference (regardless of the source), but it does not have a signal indicating whether results are better or worse. It only mimicks learning because that's what it "saw" in training.

It's still randomly selecting tokens. Explaining mistakes can bias it towards making more because that's likely when interacting with someone who makes a lot of mistakes, but re-rolling can result in it randomly selecting tokens that produce acceptable results. Much like gambling, YMMV.

That's a pretty apt description, but it happens during model training, not inference. Training does not happen in real time while you interact with the model.