Evaluating an LLM’s ability to do a simple thing

People seem to have this idea that large language models (LLM’s) can be relied upon to do complex things. People want to do things like “deep research”, pretending that an LLM can effectively perform the job of a research analyst. People make insulting statements like “GPT-4o…enables PhD-level reasoning”. People seem to be under the impression that LLM’s can be productive on their own 1 2 , that it’s a good idea to let agentic AI loose with access to the ability to submit proposed code to public open-source software, and disparage its maintainer when it is rejected (and I venture to say many are not discouraged by this result).

[%>% more()]

RSS is working

RSS is working

At the advice of Corey Doctorow I’ve gotten an RSS reader for myself and subscribed to a bunch of blogs. It’s refreshingly simple and good-feeling heritage tech, when tech was for the people.

I fixed this website’s parameters to make the RSS work properly (or, at least, as far as I can tell). I’m using a preconfigured theme; digging into the code that forms the index.xml file I saw that the placeholder (is it called a shortcode? a method? I’m not sure) was {{ .Permalink }}, but it was resolving to /; not helpful on the wider web. The solution was to change the baseURL to my domain name explicitly, and create permalink templates in the configuration.

[%>% more()]

Excerpt from G.K. Chesterton’s Orthodoxy

What we suffer from today is humility in the wrong place. Modesty has moved from the organ of ambition and settled upon the organ of conviction, where it was never meant to be. A man was meant to be doubtful about himself, but undoubting about the truth; this has been exactly reversed. We are on the road to producing a race of men too mentally modest to believe in the multiplication table.

[%>% more()]

My Opinion of AI

My Opinion of AI

I was amused by “AI” when it took off in the winter of 2022-2023. My friends and I took turns generating comedic images, probing our imagination for the most incongruous prompts to see what humor could manifest. More than one year later, I’ve developed a distaste for generative technology, but more than that a sense of foreboding. I doubt that I have any new opinions here but I find it worth writing down anyway.

[%>% more()]

The jobs that pay the most money benefit the fewest people

…and the jobs that benefit the most people pay the least.

We’re incentivized to benefit the fewest people and wonder why resources are concentrated into the fewest hands.

2024-09-06: I found this expressed in almost the exact same way in a documentary about burnout.

One of the most pernicious things about our current economic system is that the more your work benefits others in an obvious and immediate sense, the more your work has a clear and undeniable beneficial effect on other human beings, the less you are likely to get paid for it.

[%>% more()]

Acquisition of Valve Software

Acquisition of Valve Software

Apr 2034, Reuters

Sam Bankman-Fried, formerly convicted of various counts of fraud and released from prison early in 2027, has purchased Valve Corp. and lays out his vision for the future. “We’re taking the Steam platform in an entirely new direction”, the 51-year-old entrepreneur said. Bankman-Fried will focus efforts on providing a premiere platform for AI-generated content. “There are so many ideas that AI can help us explore”, said Bankman-Fried during our visit to his 100,000 sq.ft. estate, his voice reverberating across the grand hall.

[%>% more()]

Odds Ratios R Odd

Odds Ratios R Odd

I received a question about interpreting the exponentiated coefficient of a logistic regression (with a logit link). I don’t usually try to understand the coefficients in terms of odds, because I find them unintuitive. And I never really explored why I found them unintuitive until I was asked about them.

[%>% more()]

Updating this website (again)

Updating this website (again)

I’ve done it! Finally got my website all sorted.

TLDR:

  • RTFM. The config section of the docs had the answer the whole time; needed to set config to point to my subdirectory instead of the root.
  • Clear your cache, and the web server’s cache. Band-aid is to open your browser’s network section of the dev tools and disable cache, which apparently also has the effect of requesting fresh content from the web server as well.

I found a php photo gallery at novaGallery. It’s a very pleasant-looking design that uses PHP to serve images from a folder on the server. It handles caching and thumbnails, and can present full-screen lightboxes for the individual images. It’s reasonably quick once the cache is built.

[%>% more()]