VHRanger

31 ms·

601.
by VHRanger 3y ago
My point isn't when you should be optimizing to the point that the optimization is publishable. My point is that you should be writing fast optimized code by default in ML research because you tend to throw large volumes of data at tha
602.
by VHRanger 3y ago
I mean that's obvious nonsense. Even in ML, which I work in, optimizing a data pipeline or model code to be 3x faster means you can run 3 experiments a day instead of 1. There's a huge chilling effect from slow code in research it
603.
by VHRanger 3y ago
The very next sentence after the famous "premature optimization is the root of all evil" says that "a 12% gain easily achieved should always be done". That sentence was always taken out of context
604.
by VHRanger 4y ago
RIP Hello Internet
605.
by VHRanger 4y ago
How do people come up with variable and function names in a decompilation like this? Do they guess it based on what they end up thinking the function/variable does in the code?
606.
by VHRanger 4y ago
As a CEO you can re-org things (eg. consolidate power) to have impact. Fixing this stuff is the job.
607.
by VHRanger 4y ago
Sources on "they work fine at scale"? Last I checked 4 years ago they didn't at all in terms of costs or pollution to produce/dispose of. Natural stores energy like nuclear or hydro remain best.
608.
by VHRanger 4y ago
Easily, yes. It's easy and safe to dispose of, and doesn't take much space at all. Disposing nuclear waste effectively means putting it under ~3m of water in a place where people won't mess with it. As of 2016, France has pro
609.
by VHRanger 4y ago
Significant obstacles like: - Batteries don't work at the scale of a grid. There are a few 'natural' batteries, like the water level of a hydro dam, but none for renewables. - Sometimes it's cloudy - Sometimes there&#x27
610.
by VHRanger 4y ago
That's a problem for machine learning systems in general. For instance, a recommendation system's output effects what users see, so it effects what they click on. The next training set is statistically dependent on the input of th
611.
by VHRanger 4y ago
How is HN falling for this again, barely 24h after the Bob Lee story showed how this forum is prone to these narratives? The very first link in the article from StatsCan is "Canada's crime rate: Two decades of decline". It sa
612.
by VHRanger 4y ago
This is kind of a short list and very narrow in scope I find? I would have expected more sophistication
613.
by VHRanger 4y ago
At my previous job, we used joblib + a handbuilt file format to persist ML model blobs across versions. The file format was a zipfile with: - a joblib dump of the python class wrapping the model and its data processing pipeline code - a sma
614.
by VHRanger 4y ago
Bitcoin's power consumption isn't a "tradeoff". It's just as centralized as proof of stake systems. Bitcoin's power consumption is just appealing to the religious aspect of the network : if we changed it, it wo
615.
by VHRanger 4y ago
Python does weak typing and doesn't have 1% of the unexpected nonsense PHP has.
616.
by VHRanger 4y ago
Also, banking is dealing with massive legacy codebases. You get to write your own new stuff in crypto, there aren't excuses for fucking up on best practices
617.
by VHRanger 4y ago
Ideally, you'd embed the graph into 2 or 3d first, then visualize it as a scatterplot. Visualizing the edges at scale doesnt yield nice results in general. It's simply too noisy. The way to do it is to reduce the graph to some 300
618.
by VHRanger 4y ago
yes, it's using webGPU to render the whole thing
619.
by VHRanger 4y ago
Not exactly. It's that the model is loading less stuff out of the mmap'ed weights that you would expect. The author of the mmap patch chimes in here: https://news.ycombinator.com/item?id=35393615
620.
by VHRanger 4y ago
They do though. A ton of the neurons end up not activating on any particular layer, leading to a huge waste as you're passing a zero down the layers. I spoke with a PM at TT and he told me an important idea is that we're spending
621.
by VHRanger 4y ago
Notice they require 32gb RAM and encourage 64gb RAM. Presumably the architecture keeps the model in CPU RAM and shuffles it dynamically to the PCIe card network? I'm guessing here. Whatever quirks come out of their hardware, I want to
622.
by VHRanger 4y ago
Last I went it had a tent city along the rail lines, people BBQ'ing metal barrels, the whole thing. It looked really rough.
623.
by VHRanger 4y ago
To be honest even if it fulfills half the expectations itll sell like hotcakes
624.
by VHRanger 4y ago
Moreover: The integers are really indeces into the embedding space. So you'd want to think more that the model maintains a giant matrix (one row = one token; one column is an embedding feature). The array of indices gets the relevant
625.
by VHRanger 4y ago
The idea is that the tenstorrent cards dynamically prune out codepaths that aren't used. Transformers naturally have a big part of the network that is unused on any particular token flowing through the model. You could see that by how
626.
by VHRanger 4y ago
We dont know yet, because they're not available. But Jim Keller has been all-in there for 2 years and they're supposed to be generally programmable.
627.
by VHRanger 4y ago
Cant wait for tenstorrent card to come to public. Grayskull is supposed to be A100 performance for $1000, with some cool features (horizontally scalable by plugging them to each other over ethernet, C++ programmable, sparse computation, etc
628.
by VHRanger 4y ago
The Stalingrad-La Chapelle area, mainly. Yeah other parts of the 18th are really nice. Also to note, I make it a point to avoid, say, inland Baltimore or South Detroit. Whereas I did go to North Paris, which speaks the difference. But compa
629.
by VHRanger 4y ago
As a Canadian, northern parts of Paris, (Seine-St.-Denis and some parts of the 18th) are really sketchy. Maybe not like some parts of American cities (fewer guns to start), but the 18th put me on guard in a way you'd never see in Canad
630.
by VHRanger 4y ago
Why is not having snap packages good? You can just not use it if you don't like it?

More ›