queuebert

Pipe to command-line LLM: https://justine.lol/oneliners/

BTW, the "curse of dimensionality" technically refers to the relative sparsity of high-dimensional space and the need for geometrically increasing data to fill it. It has nothing to do with storage. And typically in vector databases the data are compressed/projected into a lower dimensionality space before storage, which actually improves the situation.

Especially in conversation, but calculations are typically done in SI units to avoid mistakes.

https://en.wikipedia.org/wiki/Foe_(unit)

https://en.wikipedia.org/wiki/Barn_(unit)

Scientists don't use either. If using SI units, they would say 1.7 x 10^7 m, or 1.7e7 in code, which is even shorter.

Pareto is a power-law distribution.

GOOG, which has a reputation of being slightly bloated, made $138 billion in revenue across 140,000 employees. That's $2.3 million per employee.

Especially this given how selective they are in hiring. There are likely hundreds of people salivating at the chance to replace you.

> ... he had contracted pneumonia in April and suffered a stroke following an MSRA infection.

That should be "MRSA", for Methicillin-resistant Staphylococcus aureus. Terrible fact checking.

(But they did helpfully provide a hyperlink to Facebook-related stories.)

I wonder who buys these. Crypto miners? My institution would make it nearly impossible to buy a secondhand supercomputer.

Does anyone know if desktop scaling on Wayland is fixed now?

On 22.04 scaled desktops caused certain apps, such as Obsidian, to be blurry.

That's with ESM, which is only free for up to 5 machines, IIRC.

DuckDuckGo has generative search now too. Both in LLM chat form and in a beta results box at the top of the results called DuckAssist.

As a person with reading comprehension skills, I'm still not sure what a use case is for LLMs for me. Everything I try feels like outsourcing a homework problem to a middle schooler.

There is so much more you can pick up by reading a text yourself, even quickly, than comes through when an LLM summarizes a text or answers a complicated question. The way the text is written, the language chosen, the punctuation, the sources chosen, etc.

Have you tried Ubuntu Kylin [1]? If so, what did you not like about it?

1. https://www.ubuntukylin.com/

I want to see an LLM that generates answers without the letter 'e', like the novel Gadsby by Ernest Vincent Wright.

Is that an entire book or any part of any book? Does it count audiobooks? I know a lot of people who claim to "read" a lot of books, but they actually are listening to audiobooks.

Surprised Amazon hasn't implemented this.

> the possibility for ethical, sustainable business models is still very real.

Those exist everywhere in the trades. Every major city has dozens of small businesses doing plumbing, electrical, roofs, etc., with the goal of sustainable, loyal customers based on doing a good job at a fair price.

Please, please, please call the final product Halfwit.

Seriously though, this is a very interesting biologically inspired idea, since not all neuronal pathways fire all the time.

It seems to follow that, if you can predict which weights you won't need, then you should be able to compress the model architecture permanently, at least for certain use cases.

As an old Slashdot user with a low user ID, I disagree. Not RTFA was definitely scorned as it was on the Usenet. But you had limited ability to vote people down, so perhaps it appeared more socially acceptable.

Depends on whether you measure workloads as "jobs" or "flops". If "flops", I would hazard that the bulk of computing on the planet right now is happening on GPUs.

Aaronson does quantum computing. He might be slightly biased.

Or cuBLAS. In practice, if I'm going through the trouble to rewrite math in C++, I'd rather just make GPU kernels.

Talk about the downtrodden.

That will happen if the other currency is also inflating at the same time.

This appears to be working because the data are highly correlated, i.e. similar magnitudes across the vector. I wonder how well it would do with completely random input, not just uniform floats in the [0, 1] interval, but floats across all exponent combinations.

Did we just figure out a DoS attack for AGI training? How large can a robots.txt file be?

There are three kinds of people in my experience:

1. bzip2 -1

2. bzip2 -9

3. What's bzip2?

A huge amount of time is spent optimizing for #3. Maybe instead we should offer descriptive commands that convey the goals. Say, "squash", "speedup", and "deflate", or some such.

It's one of those things you saturate your brain with for a few days, then put it down, and two weeks later in the shower you figure it out.

> Alas the tantalizing prospect of real time raytracing seems to be backburnered under the AI hype onslaught.

Can you elaborate on this? Is that because ray tracing needs only simple vector units, while AI is driving tensor-type architectures?