but this is great. I hope this actually becomes a format that wraps the weights and transformer module (maybe this can also be NAS-optimized too?). Maybe it would even work for video?
It's like calling gzip but instead of compression level you choose kolmogorov complexity level
> There are some minor kinks that need to be worked out, such as the fact that each image takes around a day to generate on mobile, but this is more than acceptable in certain domains. Website visitors, for example, are well-accustomed to such loading times, and would barely notice any difference.
While clearly satirical, it's definitely quite thought-provoking from various angles including the basis of information, representation of data, and even copyright. It's like watching a movie, writing a book based on it, and then making another movie based on that book.
Another potential weakness that isn't immediately clear from this experiment is if the experiment was run much longer (disregarding cost) then perhaps then the agent's memory could be susceptible to more long term memory compaction corruption and thus made more compliant?
If the threat model was weighted by the stakes, then I wonder how the author would reassess their comfort level. Put to the extreme, the experiment could be whether the AI assistant could be trusted to keep a dangerous AI in a box a la https://rationalwiki.org/wiki/AI-box_experiment where the stakes are assumed much higher
In my interpretation, 躺平 (tang ping / "lay flat") and 摆烂 (bai lan / ..."let it rot" I guess) overlap in meanings in the "do nothing" aspect, but in the tang ping / lay flat case, it is out of learned helplessness / nihilism, while bai lan / let it rot case, it is out of purposeful negligence / noncompliance, which suits geohot's vibe
I recently sifted through a bunch of tagged entries in a text file, where each entry had a json array of image names, but the images resided on a remote server. I basically wanted a program that could detect if the cursor was on an image name, and display it on the right.
There's a bajillion ways to do this, some might even involve writing an html file and launching a remote server and tunneling and using a browser and what not. But no! ChatGPT wrote 20 lines of elisp code. I add a tramp basepath, open the text file, and got exactly what I needed. Need any behavior changes, callbacks, transformations? Modify, eval expression, new behavior!
I asked ChatGPT what other system I can use to achieve the same effect. The best answer it gave was neovim. No, neovim can't do that with the same degree of ease.
Disappointing and amazing at the same time.
The drawback of the emacs approach in my case is the tramp latency. To speed things up we'd want to add prefetch and that's gonna be much more than 20 lines and C-x e
I work a lot in the terminal and so vim/nvim is what I open the most, but the rise of LLMs has actually increased my use of Emacs recently. It fits in a niche of rapid interactive text processing. For example, I just found myself having to sort through a bunch of markdown files in a directory. While a tool like ranger could give dynamic preview, I wanted the preview to work with a todo list so I could mark which files still remain.
In the past, coding up this interaction would be a chore, but Gemini makes quick work of the elisp. After a minute, it's doing exactly what I need: cursor on markdown link, open to right pane, C-c C-c toggles the todo item and closes the pane; C-u C-c C-o opens in obsidian. This functionality can be achieved in any number of ways but Emacs is fantastically well suited for a problem like this.
I'm a little sad that VSO languages didn't start the computing revolution because if they did, we would all be writing LISP. But instead, we are relegated to Python and JavaScript!
I think another part of it is culture. Outside of Western tech circles there's far less a culture (it feels) to invent new languages. To my knowledge Ruby's Matz is the only notable exception, and he's highly unusual as a Japanese. Then, using non-ascii-friendly character set is an even greater challenge.
Non-ascii encodings are harder to program in due to the need to switch in and out of input methods.
That said, some languages like Arabic and Japanese (and possibly Korean and Hindi) lend naturally to VSO token ordering, which maps directly to LISP syntax, so it's unfortunate that there isn't a lot of interest in this. It would be lots of fun. Maybe agents will make this possible!
Here are some interesting examples.
- https://github.com/nasser/--- (Arabic)
- https://honoka.nukenin.jp/Introduction/Loop.html (Japanese)
- https://github.com/wenyan-lang/wenyan (Chinese, which is SVO like English)
My colleagues and I wrote a programming language (really just a python wrapper on unicode characters) in cuneiform. The novelty was that we tried as much as possible to use actual mathematical concepts from 3rd (and early 2nd) millennium BCE Mesopotamia. We published it in Sigbovik 2024 (https://www.sigbovik.org/2024/proceedings.pdf) as well as a full neural net implementation using the language.
Circa GPT-3.5 to GPT-4o I was involved in some research in figuring out how to make LLMs funny. We tried a bunch of different things, from giving it rules on homonym jokes [1], double-entendre jokes, fine tuning on comedian transcripts, to fine tuning on publicly rated joke boards.
We could not make it funny. Also interesting was that when CoT research was getting a lot of attention, we tried a joke version of CoT, asking GPT4 to explain why a joke was funny in order to produce training set data. Most of the explanations were completely off base.
After this work, I became a lot less worried about the GAI-taking-over narrative.
Funny is very, very hard.
[1] without a dictionary, which at first seems inefficient, but this work demonstrated that GPT could perfectly reconstruct the dictionary anyway
There are very good, less known, models that produce funny and highly creative outputs when nudged in a good way. Premier models are just plain meh in this space.
main drawback is that it's not lossless ;-)
but this is great. I hope this actually becomes a format that wraps the weights and transformer module (maybe this can also be NAS-optimized too?). Maybe it would even work for video?
It's like calling gzip but instead of compression level you choose kolmogorov complexity level
reply