mlsu

32 ms·

511.
by mlsu 2y ago
I mean, this is true of any low-level language like C++ or C. Manual memory management. I think it does drive a lot of people nuts who have only ever worked with GC languages. I think your reaction to Rust depends on where you approach it f
512.
by mlsu 2y ago
I don't believe that quantization comes for free. Someone made the observation that llama3 models quantize "worse" than llama2 models do -- they suffer in quality from quantization far more. My intuition is that a model which
513.
by mlsu 2y ago
Your causation is reversed. Financial speculation happens because people believe (reasonably) that the price of a home will increase. Why do they hold this belief? Because the demand has increased while the supply is kept artificially cons
514.
by mlsu 2y ago
For the civilians like me in the room: - LRIP: low-rate of initial production, the early phase of weapons development where kinks are being ironed out before mass-production begins - C-130, C-17: large, US-made cargo aircraft that we have a
515.
by mlsu 2y ago
I'm not really sure your self-driving analogy is apt here. Waymo has cars on the road right now that are totally autonomous, and just expanded its footprint. It has been longer and more difficult than we all thought, and those early te
516.
by mlsu 2y ago
If I had to guess: It's gendered: women are using LLM's for roleplaying/text chat, and men are using diffusion models for generating images.
517.
by mlsu 2y ago
The voice modality plays a huge role in how impressive it seems. When GPT-2/3/3.5/4 came out, it was fairly easy to see the progression from reading model outputs that it was just getting better and better at text. Which was
518.
by mlsu 2y ago
this is one of those things that makes me feel like I'm insane. - if housing is to be an investment - it must appreciate faster than inflation - if it appreciates faster than inflation - it must become unaffordable there is literally N
519.
by mlsu 2y ago
Don't think about the money. Money flows in one direction, but corporate decisions, organization of resources, manufacturing activity -- real physical flows, energy flows, mass flows -- move opposite. What "finance" really is
520.
by mlsu 2y ago
I wanted to upgrade my M1 because it was going to swap a lot with only 8 gigs of RAM and because I wanted a machine that could run big LLMs locally. Ended up going 8G macbook air M1 -> 64G macbook pro M1. My other reasoning was that it w
521.
by mlsu 2y ago
They know that anyone who has bought an M3 is good on computers for a long while. They're targeting people who have m2 or older macs. People who own an m3 are basically going to buy anything that comes down the pipe, because who needs
522.
by mlsu 2y ago
>(You could argue that it's better in some sense for Mr Zuckerberg to control these resources than for the average investor to have that control. I'm sympathetic to that argument, especially if you carefully state what sense yo
523.
by mlsu 2y ago
No, I don't have a problem in general with returning capital to shareholders. I'm not making an anti-capitalist argument here. I'm saying. For humanity, net good. Facebook could do a stock buyback with all of that H100 money.
524.
by mlsu 2y ago
I would challenge that a little bit -- I think it is a good thing, given that they have open-sourced the results of the project. It seems to me to be far better than a company which simply gives 0.1% of US GDP back to shareholders in the fo
525.
by mlsu 2y ago
Yeah. Amazing, isn't it? I will say: I don't necessarily like to say "lucky," or even "privileged". Luck incites tricky emotion because there's an implication with luck that you didn't deserve it. A g
526.
by mlsu 2y ago
These conversations where highly paid software people complain about insanely minor things (the code linter is the worst part of your job??) are actually kind of nice to read, in a funny way. The privilege of having pixels be the most str
527.
by mlsu 2y ago
And, on top of that, the first few fixes will actually work! They will look good to PMs and to the business! Look at the velocity ! The velocity is good enough to scale back the size of the team, it's good enough to mandate a pace tha
528.
by mlsu 2y ago
Interesting! That is more than I thought. Honored to have caused a nerdsnipe. In the grand scheme of things, though, most of these are quite small -- 7b range. A 7b model is nothing to sneeze at but it's not megacorp resources either.
529.
by mlsu 2y ago
The fact that you have enough reading comprehension to summarize my comment, deconstruct its arguments, and formulate a counterargument puts you in an extremely small minority. You and I spend time around people who read, who have reading c
530.
by mlsu 2y ago
Mid-level managers at a lot of companies still have no clue what LLMs are or how they work. These companies (like databricks) want to have their salespeople upsell such companies on "business AI." They have the base model in the
531.
by mlsu 2y ago
Far fewer than 600,000 of those are pretrained. Most are finetuned which is much easier. You can finetune a 7B model on gamer cards. There is basically the big guys that everyone's heard of (google, meta, microsoft/openAI, and ant
532.
by mlsu 2y ago
Books are a vestigial organ. Nobody uses books as their primary mode of accessing information about the world anymore. What is interesting is that the cultural mores around this vestigial organ have persisted to such a great degree. Harry P
533.
by mlsu 3y ago
The CRV was stopped with hazards and lights off, at night, in the center lane of a 10 lane freeway. Certainly a human driver could have easily made this mistake.
534.
by mlsu 3y ago
Certainly an interesting looking chip. It looks like it's for recommendation workloads. Are those workloads very specific, or is there a possibility to run more general inference (image, language, etc) on this accelerator? And, they me
535.
by mlsu 3y ago
8x22b. If this is as good as Mixtral 8x7b we are in for a wonderful time.
536.
by mlsu 3y ago
Feedback Systems is probably the book you are looking for: https://fbswiki.org/wiki/index.php/Feedback_Systems:_An_Intr...
537.
by mlsu 3y ago
It's a clever bit. It is intellectually amusing to read, and I'm sure to write, a character that the reader hates. But the detachment becomes exhausting quick. It's a dead end; you spent all of this effort being clever, with
538.
by mlsu 3y ago
I wonder if there is a clever/more efficient shortcut that could come from before the sample is taken on each token. We have the logits after all.
539.
by mlsu 3y ago
imho, and you can call me an atomic gardener if you wish, nuclear tech is a bad example. If we had actually gone "all in" on nuclear to produce electricity, our and the world's clean energy footprint might result in warming o
540.
by mlsu 3y ago
Disagree. The input/output structure (tokens) is the interface for both inference and for training. There is an emergent intellect embedded in the weights of the model. However, it is only accessible through the autoregressive toke

More ›