hikarudo

Very cool!

How does the direct TCP mode work?

Why? The inference server isn't a harness, it's tokens in, tokens out. That's different from a harness.

Sounds super interesting!

What's your approach?

Fun fact: one of Yann LeCun's first 'deep learning' projects was on OCR for postal codes (MNIST).

In Japan, I learned to drink water upside down to stop hiccups. Works for me, whereas drinking normally doesn't.

Thanks for the correction. You're absolutely right.

It's not a market failure, it's just supply and demand. There are many computer components competing for the same resources (fabs, wafers). Demand for GPUs, RAM etc. has increased a lot due to AI, but supply is still the same due to new fabs being huge investments that take years to build. Of course the price goes up.

WhatsApp is end-to-end encrypted. No one at Meta can read your messages.

> Once you reach this stage, the only escape is to first cover everything with tests and then meticulously fix bugs

The exact same approach is recommended in the book "Working effectively with legacy code" by Michael Feathers, with several techniques on how to do it. He describes legacy code as 'code with no tests'.

Also checkout Deepmind's "The Gemma 4 Good Hackathon" on kaggle:

https://www.kaggle.com/competitions/gemma-4-good-hackathon

One trick I use all the time:

You're typing a long command, then before running it you remember you have to do some stuff first. Instead of Ctrl-C to cancel it, you push it to history in a disabled form.

Prepend the line with # to comment it, run the commented line so it gets added to history, do whatever it is you remembered, then up arrow to retrieve the first command.

$ long_command

<Home, #>

$ #long_command

<Enter>

$ stuff_1 $ stuff_2

<Up arrow a few times>

$ #long_command

<home, del>

$ long_command

Why 25%, shouldn't it be 12.5%? 8 ECC bits for every 64 bits.

A gentleman's gentleman!

Here's a great talk by John Hughes, one of the authors of QuickCheck, with real-life examples:

https://www.youtube.com/watch?v=zi0rHwfiX1Q

I used a similar setup for a while: Obsidian for taking notes in markdown, and vscode for coding.

Eventually I moved to using vscode for both. My gigantic notes.md file is always open in tab 1, so I can go to it immediately with Ctrl + 1.

Finding notes in a single file is easier for me than finding them in a bazillion tiny files. And there's less friction whenever I need to make a note (no need to create and name a new file).

Also see "I Pencil" by Leonard Read (1958).

> we've had that idea since at least Hofstadter and his sentient ant nests.

A similar idea is present in Herbert Simon's 'The sciences of the artificial', where he describes a sentient city.

Sincere question: what's wrong with Python f-strings?

Computer vision. C++ and Python.

That was Blaise Pascal.

"Sparks of Artificial General Intelligence: Early experiments with GPT-4"

https://arxiv.org/abs/2303.12712

What an awesome story! Thanks for sharing.

You can split it up into several files, then concatenate the arrays at runtime.

Spot on. You've described Peter Naur's "Programming as Theory Building" paper exactly.

import -window root `date -Is`.png

Greg Egan's "Permutation City" kind of fits those requirements.

"A car is a wheel, and is a wheel, and is a wheel, and is a wheel".

conscious vs unconscious memorization

The opencv_contrib repo does have a module, called "reg", for direct alignment.

Here's a very good visual explanation of double descent:

https://mlu-explain.github.io/double-descent/