compumetrika

Get a Strix Point or Strix Halo with 128GB DDR5 RAM and you can run gpt-oss 120B at 10-20+ TPS.

Hmmm maybe I should re-read A Fire Upon the Deep.

Currently on the front page there are ~2/30 posts on the front page.

AI hype is everywhere. On HN there is genuinely much higher quality conversation about AI than many other places.

So, no, the question in the title doesn't resonate with me. I don't think there is an obsession, and for the AI discussion happens, I'm happy with the quality of it.

LLMs use pseudo-random numbers. You can set the seed and get exactly the same output with the same model and input.

Great quote, commenting just to 'bookmark' it.

Ollama can pull directly from HF, you just provide the URL and add to the end :Q8_0 (or whatever) to specify your desired quant. Bonus: use the short form url of `hf` instead of `huggingface` to shorten the model name a little in the ollama list table.

Edit: so for example of you want the unsloth "debugged" version of Phi4, you would run:

`$ollama pull hf.co/unsloth/phi-4-GGUF:Q8_0`

(check on the right side of the hf.co/unsloth/phi-4-GGUF page for the available quants)

Not all details are posted yet, but an interesting development.

I haven't tried this yet myself, but have you tried plugging it into GPT or Claude or Perplexity and asking Qs? I've made some progress on thongs this way, much faster than I would have the usual way. Apply the usual precautions about hallucinations etc (and maybe do the "ask multiple AIs the same thing" thing). We don't yet have a perfect tutor in these machines, but on balance I've gained from talking to them about deep topics.

The economist Emily Oster wrote an excellent series of books about pregnancy anf early kid years, where she dives into details about various studies, whether they are causal, etc. It's one of the best practiacl explainations of reading research I've encountered targeted at non-academics, really well done. She has a chapter on peanut exposure allergies and i think inrecall that these early-exposure results are in fact from causal research vs just correlational research (basically there are at least two types of papers out thwre -- correlational and causal. As you might guess causal is harder to get for many reasons). Great books; she may also have published some chapters on her substack (substack came after the book I think).

I realize as I write this that you are probably saying that for your own experience you can't disentangle correlational from causation, to which I would say -- correct!

I've directly experienced this in my own life. Moved during the pandemic to be near family, and coincidentally that is a much more religious and conservative community. Many people I am around daily are much more unplugged from online discussion and much more directly involved in face to face lives. There is a subset who are more "online" and tend to be more stressed/anxious.

Many great recommendations in the comments here about on emergent order in complex systems -- does anyone have recommendations for reading about how such systems de-organize?

Is there a proof or demostration somehwere about maximizing the consent of the governed?

Arrows theorem always has implied to me that the next step should be quantifuing some welfare measure for voters and then exploring which system maximized that welfare measure. "Consent of the governed" sounds like a welfare measure so I an intrigued.

> "Westworld was a warning, not an instruction manual!"

By happenstance I just ran across this by the same author somewhere: https://ymeskhout.substack.com/p/death-of-a-client

Do you have a custom "intro prompt?"

I have one that describes a lot of statistical work I want GPT to help me with.

I got this result the first try:

> Sally has 2 sisters. Since each brother has 2 sisters, and Sally is one of them, there must be one other sister making it two sisters in total. >

I've often found EM's things on LLM usage to be quite helpful, and it seems he's gathered a lot of things on prompting in one place here.

I'm a little surprised no one has pointed out yet that this is incredibly close to cellular automata, which Wolfram has show can get you Turing completeness among other things.[1]

Wolfram's "New Kind of Science" almost certainly has tons of fun things for the author to play around with if so inclined.

[1] https://en.m.wikipedia.org/wiki/A_New_Kind_of_Science#Simple...

I think the data are that we are working a bit less than we used to, it's rather that the norms and mental categories have shifted. For example 40hr work week is now a sort of norm, vs "sunrise to sunset."

The average workday is also much safer and less physically painful than it used to be.

There are two things I have in mind, and if I have time later I'll see if I can find them and update the post -- one chart of total working hours per worker, and one article, about how much added safety has or hasn't slowed down productivity growth (which inevitably has a lot of data in it about workplace injury and death over time).

Children also work much less than they used to.

I think that Keynes wasn't completely wrong in his prediction (...and maybe if AI really kicks off he will be more vindicated, just missed on the timing)

I worked through large chunks of the book a long time ago (perhaps an earlier version). I agree with the other responder-- if you've had zero stats it may be more difficult, but the author does an excellent job of discretizing almost everything, which means that integrals and derivatives are replaced with adding and subtracting. For this book I think the most advanced math you need is multiplication and division. It's a clever trick that relies on the fact that computers are really good at multiplying 10,000 things by 10,000 other things and adding them up (which is what discretization implies here).

Before computers that was much harder to do so you needed clever math tricks to "do it all at once," which gets you to all the analytical methods -- calculus, etc. Still all very useful! Just hard to teach to a broad audience. Downey really leans into the technical advantage to computers provide and thus can cover a lot of material before getting into complicated math.

I'm surprised to see no one has pointed out RMarkdown + RStudio[1] as one simple, low-effort way interface with Pandoc using an IDE you may already use.

I used to write papers and slides in LaTeX (using vim, because who needs render previews), then eventually switched to Pandoc (also vim). I eventually discovered RMarkdown+RStudio. I was looking for a nice way to format a simple table and discovered that rmarkdown had nice extensions of basic markdown (this was many years ago so maybe that is incorporated into vanilla markdown/pandoc).

The RMarkdown page claims:

> R Markdown supports dozens of static and dynamic output formats including HTML, PDF, MS Word, Beamer, HTML5 slides, Tufte-style handouts, books, dashboards, shiny applications, scientific articles, websites, and more.

...which I think is largely due to using pandoc as the core generator.

RStudio shows you the pandoc command it runs to generate your document, which I've used to figure out the pandoc command I want to run when I've switched to using pandoc directly.

This is a bit of a "lazy" way to interact with pandoc. Maybe the "laziest" aspect: when I get a new computer, I can install the entire stack by installing Rstudio, then opening a new rmarkdown document. Rstudio asks whether I'd like to install all the necessary libraries -- click "yes" and that's it. Maybe that sounds silly but it used to be a lot of work to manage your LaTeX install. These days I greatly favor things that save me time, which seems to get more precious every year.

[1] https://rmarkdown.rstudio.com