tantony

Claude Opus 4.5 has been a big step up for me personally, and I used to think Sonnet 3.5 was good. It is an amazing deal at $20.

Just yesterday, it helped me parse out and understand a research paper - complete with step-by-step examples (this one: https://research.nvidia.com/sites/default/files/pubs/2016-03...). I will now go ahead and implement it myself, possibly relegating some of the more grunt-work type tasks to Claude code.

Without it, I would have been struggling through the paper for days, wading through WGSL shader code and there would be a high chance that I just give up on it since this is for a side project and not my $job.

It has been a major force multiplier just for learning things. I have had the $20 subscription for about a year now. I bump it up to the $100 plan if I happen to be working on some project that eats through the $20 allocation. This happens to be one such month. I will probably go back to the $20 plan after this month. I continue to get a lot of value out of it.

My name is on the blue plaque. I (and many others) spent countless hours on these two vehicles. Many of us are now working 12-hour shifts performing initial operations tasks.

The majority of us didn't know about the plaques until after the spacecraft were packed up and ready to be shipped to the launch-site. It was a nice surprise when we learned about it. Feels good to know that we got to sign our name on our work.

I was able to get pretty far using Clojure and this package instead of Scheme: https://github.com/sicmutils/sicmutils

I was able to make "CloJupyter" notebooks with the examples from the book. You can see some of my notes here (only goes till Chapter 2): https://www.thomasantony.com/projects/sicm-workbook/

I have personally experienced this. I went to a concert in NY while living in the midwest. I was a pretty broke grad-student at the time and hence opted not to get a hotel for the night before the return flight. The concert ended late and we reached the airport by around 3:30 AM for a 7AM flight. I think I took some naps at the gate before the flight and got very little sleep on the flights.

I barely remember anything from that night. The concert itself is mostly blank except for one or two moments. I remember some moments of driving late at night afterwards (getting lost at one point) and having a very late night dinner at an IHOP. I barely remember getting on my connecting flight.

Overall a very surreal experience.

Now I make it a point to get proper sleep on such trips. What is the point of doing these things anyway if we don't get to keep the memories?

Before buying the Framework, I used a MacBook Pro for ~11 years. I am perfectly comfortable with the command-line and ssh into a Linux VM every day at work (windows laptop). I use a tmux+neovim setup that mostly works without any problems (except for random LSP issues at times).

I can just get all of the advantages of a *nix shell but with a better UI, using macOS. I have zero regrets from switching back to a Mac after my month-long experiment. My only regret is that I didn't order the M1 Max as soon as it was out and couldn't get the amount of RAM I wanted (it had a long wait by the time I decided to buy).

I bought a Framework 13 when it first came out, with exactly this hope. I gave up after about one month of madness and ordered an M1 Max Macbook Pro. I also installed Windows on the Framework with WSL. Now it gets used very rarely (like for when I want to do Windows development, or for my parents to use when they visit) and mostly sits around gathering dust.

Desktop linux has burned me one too many times for me to trust it again.

There's already plenty of it in space.

Magnetic field drops off by an inverse-cubed law (as opposed to inverse-squared for RF). So it would be a lot more impractical to cover large volumes.

> tech-illiterate

The guy wrote two books on Lisp. Lol. What are you talking about?

And why do you need HTTPS to access a static HTML blog which doesn't even have a single form? What are you hoping to hide here?

> Physics for Entertainment by Yakov Perelman

I read the same in Malayalam! That book was one of my earliest introductions to practical physics. He deconstructs pretty much all the early "perpetual machine" designs in the book.

I also particularly remember the chapter with the experiments with soap bubbles as being very interesting as a kid.

Wow! I used to fly out of the Ames airport and earned my PPL there. It turns out that it wasn't always at the same spot.

I looked it up and apparently my house was on land that used to be the airport back in the day! My old house is actually on the map shown there.

[1] http://www.airfields-freeman.com/IA/Airfields_IA_W.htm#ames

[1] http://www.airfields-freeman.com/IA/Airfields_IA_W.htm#ames

Based on the paper, results, video etc., this is either legit, or a massive fraud. I guess we will know soon enough.

If it is real though, this is a massive step up for humanity. The stuff of science fiction.

Someone is probably going to port it to llama.cpp soon.

The speed at which this stuff is developing makes it very hard to keep up and contribute in a meaningful way. Especially when you have a separate full time job.

Would love to see your code for this btw.

Vicune

This is what happens when running inference on a neural network:

Input (list of numbers) -> (Bunch of math operations) with (other numbers) -> Output (also a list of numbers)

This applies whether you are talking about image classification, image generation, text generation etc.

The model defines what the "(Bunch of math operations)" part is. As in, do these multiplications, then add, then a tanh operation etc.

The weights define what the "(other numbers)" are. Training is the process of figuring out these weights using various methods - some of which involve example inputs/outputs (supervised learning), others don't require examples (unsupervised or self-supervised learning).

It uses "ggml", a tensor library written in C to do the math. The models weights are converted to a format that can be used by GGML. The higher level structure of the model is created in Rust be leveraging ggml data structures and the weights are passed into the C library.

PyTorch is only needed for reading the original weights while converting them.

Stable Diffusion runs pretty fast on Apple Silicon. Not sure if that uses the GPU though.

I think one reason in this particular case may be the 4-bit quantization.

I am working on https://github.com/ggerganov/llama.cpp/issues/77 to "librarify" the code a bit. I eventually did want to Rewrite It In Rust™, but OP just beat me to it.

This is pretty cool. I got it working in about 5 minutes. Inference of the 7B model is super fast. I don't know enough about prompting plain GPT models though.

I have owned a Model 3 in Iowa while living in an apartment with no garage. This was at the beginning of the Covid lockdown where it stayed unused for extended periods of time. The battery was cold-soaked in temps below 20F. Whenever I drove the car, I had almost zero regen-braking, reduced acceleration and the range was decreased by about 30%. For short commutes, that would persist throughout. But whenever I went on longer trips, a one hour drive would get it back to its original performance. The range would stay a bit lower due to higher energy use (~330 Wh/mi instead of 240 Wh/mi in the summer). The navigation would account for this when planning trips and also pre-condition the batteries prior to going to superchargers. On average it added maybe one extra charging stop during road-trips.

XML is old-school, but it works. Any program that processes the XML schema can validate them using XSD files. Repetition can be avoided using extensions like xpath and xinclude.

Where can I find api.py?