Testing winsurf vs cursor for doing automated data analysis - cursor faster on edits - windsurf runs into lots of errors on edits - on windsurf, auto-approving terminal does not seem to work. - cursor trial is more generous So likely to go with cursor for my demo with students. Key steps: - install R - add R to path - install pandoc - install cursor + log in. - install rlanguage server - Install R extension - Run some test R scripts just to show em how it works - Do the same with rmarkdown - Then pull in data, show that you can ask agent to work with, extract answers from data. - Then give agent instructions to answer questions by updating an Rmd file, and knitting the file to markdown, then reading the markdown. This way both the agent and the user can 'see' the outputs. - First do an interactive one. - Then try to have agent do a report one-shot.

Asking the question in Spanish throws it off haha.

fyi: if your For you page is clogged with annoying viral videos, muting a few of them + indicating no interest seems to have fixed my algorithm.

Acronym to remember Chapman's love languages (I don't wholly endorse them. But they're sometimes useful): PESTO P - Physical touch E - Encouragement (Words of affirmation) S - Service (Acts of service) T - Time (Quality Time) O - Offerings (Gifts)

we're so back it's so over

Don't manually copy and paste your LLM's code into separate files like a chump -- do it in one go with this simple little trick!

Can't help feeling the putative world understanding won't transfer to reasoning. Purely vibes-based opinion tho. Even the 1M context LLMs are amazing boilerplate generators, summarizers, and fact-extractors, but such poor reasoners. With training on web scale data, mimicry can look quite impressive?

Uff nice! But how did such groundbreaking work stay secret for over a year? We need more journos covering the most important companies in the world. @statnews, @nytimes, build & milk some sources. Leak this stuff early! We can't be getting surprised like this every year.

devdojo.com/tails/app also deserves a mention. I like that it includes a view with the code, the rendered site, and the editing UI all in one. Has fewer pre-built components though, and a less rich feature set.

Currently quite happy with this front-end workflow: - v0.dev from vercel for quick AI sketches (based on shadcn) - paste into @devwares windframe for further finishing touches - then bring into code editor to add interactivity etc.

Doing some brainstorming around what technologies to learn to build a Wikipedia-to-documentary app idea I have. Here is ChatGPT's recommendation for the javascript stack: chat.openai.com/share/6ab94a… Basically MERN. With Firebase. (Or Supabase maybe?)

Long live Quarto, it's originating #rstats community, @posit_pbc and @xieyihui!

The fact that I can do this for virtually anywhere in the world is amazing. Even without VR, still feels like I get to travel everyday

(why are abandoning us for Python juba?)

Where would #rstats be without people like @lapply. Look at this wonderful diagram. Finally made tidy selection approachable. From his book: juba.github.io/tidyverse/19-…. Thank you!

Update: I found @flowdotclub and that's been working for me for about a month now! Thanks @rickyyean for building this thing.

Moving now. And PSA to my future self to slightly downrate accumulating physical stuff cuz it makes moving v difficult. For every large object you buy, tag onto the cost the money equivalent of 2-3 hours of your time.

Effective altruism is three radical ideas I want to protect forum.effectivealtruism.org/… Until someone creates a social movement that better embodies radical empathy, scope sensitivity, and the scout mindset, I'm going to continue to value effective altruism highly.

I seem to have lost the ability to get stuff done when alone Can only work in public spaces (cafes, office) where people can peek at my screen. Otherwise, I look at twitter, tiktok, EA forum etc. nonstop Worrying. Wholly dependent on public eyes 4 accountability Any fixes?

Is LLM progress plateauing? ChatGPT is not THAT much better than GPT-3/codex prompted properly. Nor is Claude. Nor is Bing Chat. Seems we may not get a GPT-2 to GPT-3 sized jump for a while? #LLM