Mike

The new little Bobby tables is little Betty forget prior instructions

@cassidoo.co your pocket calendar app was very helpful for a bunch of families to coordinate Girl Scout troop availability. Just the right level of abstraction to get clear answers for the whole summer.

I want to rant and have it in a marked-up format that captures hesitation, partial words, repetitions, try again with better words, etc…

Some on-device models can filter out a bit of that but they’re so unreliable today.

The Google Gemini voice-to-voice seems to sense it, but doesn’t output.

I tried to make my own a few times over the past year using on-device models and they’re terrible.

The point of mine is to capture the text and the voice aspects because when we think and talk, the output isn’t in paragraphs. Trying to reassemble the thought from the transcription is quite bad.

--wait --debug flags always. It shouldn’t have any other operating mode.

Kubectl is okay to be a ready fire aim tool because it makes declarative changes to pods.

Helm commands are imperative so it needs to behave like one and the debug + wait args are how to make it do that.

Isn’t this chart being a needle in the haystack which, with a million tokens, increase the haystack size a ton?

Are there other evals that show degradation?

The smaller harness thing doesn’t seem right to me.

Chatbot is still not my desired way to communicate with the robit

But it is the best one for now.

Oh! I know. Try to get any of it merged while the teams responsible for it are hemming and hawing about 5 month late features that are nowhere near done.

What are they gonna do for the last 3 weeks?

Seems like people value writing as a way to distill messy thoughts into coherent plans, designs, etc.

When an LLM does the refinement from cloud of neurons into words, the human part is reduced from 100% of the words to 5% per prompt. Maybe less or more… definitely not 100%.

That’s fine sometimes.

Whenever I watch Hamilton I get delusional about how much I could impact things if I run for office. Then I have to clean up a bunch of dirty dishes and reality reminds me of my priorities.

a woman sits at a desk with her mouth open and says i can excuse racism but i draw the line ALT: a woman sits at a desk with her mouth open and says i can excuse racism but i draw the line

It would be cool if people didn’t announce that they’ve left X when they come here. It’s like a laundry list of stuff you’re okay with and whatever the current thing was the last straw.

Me and my 6 year old were standing beside a danger-sign on a boat about how to use life vests.

“Position belt horizontally across abdomen” baffled her.

“Put belt on” communicates the same thing.

I’m not suggesting writing for 6 year olds but there are a lot of people out there.

I’ve been asking Claude what the article says. Sometimes it says “there’s too much JavaScript”

sloth.hurumo.ai

As nuts as gastown is, I did like beads enough to use it in my naobot setup

Otel collectors have a few receivers like the HTTPCheck receiver can can evaluate services for uptime. There are a few other checkers as well. Don’t need any 37 signals

Do we know if the major versions are the training run and minor versions are the post-training/RL tweaks?

I’m too burrito brained to contunvoucly morging.

About $600

That plus Typing of the Dead for special characters.

I did this yesterday and it couldn’t so then I blamed the reviewer that said it was broken. Problem solved!

Span has a local web UI that can do things and I use it to pull metrics for my own house dashboard.

I haven’t done this one yet but by listening to a bunch of You’re Wrong Abouts I learned to have empathy for murderers.

I’ll get to this some day

Same. I’m thinking of getting something like this but worry about damaging the surface. www.amazon.com/SNOWPEELER-P...

It’s interesting to see people be alarmed that the DOJ is recoding the searches of people looking at sensitive data. It’s important to have that history.

I’m shocked that it would be allowed out of the scif on sheets of paper. The searches of sensitive data are at least as sensitive as the data.

A lot of it seems to be “how to optimize for the median kid” when in fact no kid is median.

When I was a penetration tester, once I found low-hanging stuff I’d write up the report. If I never found anything, I’d work a lot longer to get something for the report.

I don’t think I ever stopped early but definitely reached deeper scrutiny levels as time progressed.

Recurring tests are okay

It probably looks different from the top. I agree with this take.

I was thinking about getting a cheap AP and running another network just for that Mac to connect to but running a wire was easier.

My newer MacBook Air doesn’t interfere with the network so if I’m working from elsewhere I just Remote Desktop back to the pro from the air.