Russell Pierce

“To prevent data retention concerns, Dell Technologies recommends that based on current technology, SSD/NVMe drives which contain user data are powered-up once every 2.5 months for a minimum duration of 3 weeks to allow background tasks to complete.” www.dell.com/support/kbdo...

I recently fought through permissions heck on AWS and came up with a solution I think I believe in. Stacks of hard drives are my fallback position. But I only recently read that you’re supposed to power SSDs on periodically to avoid data loss.

Which gave me an even more sympathy for security folks worried about what people might do and the risks of social engineering.

I locked @openclaw-x.bsky.social in a sandbox and watched it skitter around, with Opus 4.5 and with gpt-oss:120b. The security doomers are spot on. But the threat model is similar in kind (but worse) that just giving a random person access to equivalent permissions needed to perform a given task.

The skill markdown file for a Claude skill must be static when packaged. But, nothing seems to prevent it from being a stub that then calls a method within the skill to populate dynamic content. My question is how many boilerplate tokens (aside from description) does a skill burn just by existing?

The skill markdown file for a Claude skill must be static when packaged. But, nothing seems to prevent it from being a stub that then calls a method within the skill to populate dynamic content. My question is how many boilerplate tokens (aside from description) does a skill burn just by existing?

I guess it is a blessing that the AIC/BIC make it really hard to misuse them in this way?

Exactly so. There are many measures that make sense for looking at a model across datasets. R² isn't that as it describes the model itself on its dataset.

For alternate datasets there are many other options

Or in the spirit of R² being in 'plain' units rather than variance units, RMSE.

Journal of Agricultural Research

Depends on how you define 'traditional' and the tradition you come from I suppose. books.google.co.ao/books?id=lNN... would generate power from the grave I think.

Yikes. Is there a list of 'sklearn doesn't do what anybody else did' choices? I ran into the 'everything is secretly horrible' with L2 regularization being the default for logistic regression. So, I'm wary, but this one took the cake as seeming like more than just being 'a choice'.

So much has gone wrong at that point though. R² was historically originally a measure of model fit within its train set. Arguably it doesn't belong, nor is it sensible, to use for validation or test data.

If the predictions have a negative relationship with the observed values it can drop below negative 1. Deeply cursed at a conceptual level.

I don't find it credible that R² should be negative anymore than the square of a non-imaginary number can be negative.

In the sklearn case, they wrote it to be sensitive to the mean... if the prediction is worse than the mean of the observations would have been, it goes negative.

Nevermind that question alone shows in other ways that sklearn is unhinged.

Train

Why in the list underneath the article is it listed at 11? I want to believe. Scikit gave an ML colleague a negative R² value today. Not adjusted R². I'm tired. 🙃

Sklearn's regression metrics function r2_score can yield negative values. No, it isn't yielding an adjusted r^2. Cursed.

Do other parents enjoy helping their kids out on studying for spelling tests? How do you do things so that it is more enjoyable?

Is the drum in alexdeng.github.io/public/files... or is there another resource you'd recommend (until your paper comes out)?

As a _very_ minor point at typical sample sizes, the ANCOVA will certainly penalize you for the two degrees of freedom that you're otherwise stepping out of the model (slight change in error terms). But, I'm looking forward to seeing the paper he is putting together to explain the rest!

# Hold onto the mean of the outcome outcome_mean <- mean(mtcars$mpg) # mean center the outcome and predictors # n.b., you only need to mean center (no need to scale any of the variables) mean_centered_variables <- scale(mtcars, center=TRUE, scale=FALSE) model <- lm(mpg~0+am+wt,data=data.frame(mean_centered_variables)) # The adjustment is to just take the unexplained variance from the model # i.e. residuals and restore their original mean adj <- outcome_mean + model$residuals # Original Variable mean(mtcars$mpg) #> [1] 20.09062 var(mtcars$mpg) #> [1] 36.3241 # Adjusted variable mean(adj) #> [1] 20.09062 var(adj) #> [1] 8.978055 # Note 1: We mean centered, of course we didn't need to estimate an intercept model_w_intercept <- lm(mpg~1+am+wt,data=data.frame(mean_centered_variables)) zero <- 0 names(zero) <- "(Intercept)" all.equal(0,model_w_intercept$coefficients["(Intercept)"],check.names=FALSE) #> [1] TRUE

@phdemetri.bsky.social would you mind checking my restatement? cc: @chelseaparlett.bsky.social

I have used d2 and mermaid for diagrams with it mostly. True re: GIMP. A web based tool it might have a shot at steering, but it wouldn't be cheap.

Does it do any better with the graphs, graphics, and illustrations if you bounce it through writing code to produce them? Maybe giving it hints about technology that tends to produce more appealing outputs? Or it is just at a loss?

🤦‍♂️Give me the power to give myself the grace I aspire to give to others.

It is f'ing awesome to realize that something you'd been shooting intuitively around for years but couldn't get a handle on has a fully realized version. It is a little bittersweet too. Moreso when you see that the realized version has been around for decades.

It's a little less insane than it looks on the surface if you consider that someone needs to be able to wake a screensaver from a blueooth mouse or keyboard. But... yikes... it wasn't one of the first 10 error messages I tried to run to ground to solve the problem.

You can't make this stuff up. Gnome was dying on a new linux machine because... *checks notes* the screensaver wouldn't start because a bluetooth library wasn't available.

I've spent some of today using local models to squish down verbose webpages to the things I actually care about for doing additional work. Then I notice Qwen code does "Fetching content from {url} and processing with prompt: {prompt}"... makes me feel seen and validated.

I kind of love using a model from a different family to vet what the first model said. It has seemed to work better than reprompting the initial model with a fresh context window.

Now right an LLM as a judge to evaluate the prompt and evaluate the evaluation...