wizzard0

one of the must-have things to have on every windows machine

Um, yes, TimeProvider is cool but works only insofar as ...

> If you provide a fake time provider in all those places

I'm looking for something more akin to https://go.dev/issue/67434 (i.e. testing the code that calls third-party libraries that might just call Thread.Sleep and so on)

Nope, have yet to find a deepseek v2 provider which doesn’t log my prompts

Very cool! Anything similar for .net?

Aider is one tool which works on medium sized repos for me. Make sure to use claude3.5-sonnet or gpt4o, other LLMs are not yet up to the task with the context size required

Impressive visualization! Also kind of keeps the vibe of the internal Google tooling, minimalistic and down-to-earth pragmatic

A pixel takes more than 1 bit to store, too

Authors claim their mechanism is more computationally efficient (both asymptotically and on current hardware) than the Transformer attention, while performing much the same role, yes

Photo copiers replacing digits on scanned financial reports with digits that compress better are a decade old already :)

http://www.dkriesel.com/en/blog/2013/0802_xerox-workcentres_...?

Yes, a GGUF-converted version works fine with llama.cpp for me

“-Chat” means the model has been additionally fine-tuned on “human/assistant” conversation patterns, so the predicted text matches them better and if you feed a question you will get back replies instead of e.g. a plausible continuation of your own phrase

> Our research is motivated by the observation that previous instruction tuning methods force the model to complete a sentence no matter whether the model knows the knowledge or not.

And so... adding the random samples that include "I don't know" to the training datasets brings the accuracy on unknown questions from <10% to over 90%

As much as I love algorithm-heavy tasks, I must admit 90% of software development is working with stakeholders on mapping the business needs to code.

And then the most advanced algorithms are usually just `git clone/cargo add/go get/npm install/your_other_pkg_mgr` away.

I'd say the standard library task is much more about providing powerful composable protocols to integrate all those libraries together (eg see Clojure, Elixir et al)

> many iterations of each prompt

BTW its much faster and cheaper to artive at a good prompt if you sample the model in deterministic mode (ie temperature=0)

By default you have to guess if the difference is due to the prompt change or due to the dice roll, as you’ve noticed, but you don’t need to!

FYI the Asahi Linux installer says “please upgrade that should help” but bumping 13.6 -> Sonoma 14.1 did not help, the SystemRecovery still shows up as 13.5

Or does that mean wait for 14.2?

I don’t think it’s the plants, the fine dust of the pills being _consumed_ all over the world should be enougg

The words “entire athmosphere” and the “plants” might be an exaggregation, but

1) I can readily imagine the _labs_ being contaminated by the reference samples.

2) If the compound is stable, not clumping and not hygroscopic - yes, certainly. “Microscopic” doesn’t do it justice, as there doesn’t seem to be any limit to how small the seed crystals can get.

Since we’re manufacturing literally tons of pharmaceuticals, and carefully manually spreading them over the Earth surface to the pharmacies, so for stable substances it doesn’t seem such an exaggregation to me.

Makes the iPhone uncomfortably hot as well

well i must admit monitoring certificate issuance with LetsEncrypt is quite boring even if you HAVE alerts set up (I do)

so… not surprised. Still cool. What a time to be alive

cuBLAS is cuda, clBLAST is OpenCL

Also, an encoder like these can inject any amount of entropy (eg replace "0" with "100500999-100500999") which gzip can reduce to "100500999-same" but without knowledge of JS semantics has no chance to reduce to either "0" nor the nearby "172913211-172913211")

1) The title is a clickbait, but

2) Thanks for leading with an example of a negative result! That's what any researcher faces every day, unlike what gets published, after all

CPU is still the first-class option, but GGML also supports Metal, cuBLAS and clBLAST for hw acceleration

GGML is the library, GGUF is the new GGML model format.

nvidia might be the king of the hill right now, but the future of AI is reconfigurable analog electronics (~100x more energy efficient already, which will take Moore’s law at least another 10 years for silicon)

Caveat: no backprop :P forward-forward and other algorithms exist though

The points in that post are all valid. Lots of docs are just a waste of letters.

Document the intent.

Why did you decide to spend the time to write this code. What shortcuts were taken (what code doesn’t do!), and why!

Very slick! Are there any similar options that do help with things like masking the user count, and also don't suffer from hash collision when the IDs are short?

Another direction: check digits/pronounce assists?

Would love to be wrong, but to me this means things like "is signature correct" or "does the hash matches" can also be optimized away with --opt-level=3 in 1.72, unless the results are stored in a variable consumed for IO, and not just for branching.

interesting but looks rather chatty;

also completely ignores timing/traffic correlation attacks which are no longer theoretical anymore - and chattiness would likely make them more effective

my tldr: this explains

1) why huge models are important (so the gradient is high-dimensional enough to be monotonic)

2) why attention (aka connections, aka indirections) is trainable at all;

and says nothing about why they might generalize the dataset