VHRanger

18 ms·

241.
by VHRanger 3y ago
Or the brown M&Ms test from Van Halen: https://www.snopes.com/fact-check/brown-out/
242.
by VHRanger 3y ago
The interesting idea here is that you can have a great contract, but if it's impossible to enforce, it's useless to you. In this case the enforcement issue is that the other party can drag out the fight long enough to make the con
243.
by VHRanger 3y ago
I mean, that only works for a small subset of workloads where the data movement patterns fit, the bandwidth is more important than the latency, etc. The reality is that almost all workloads aren't anywhere near saturating the AVX instr
244.
by VHRanger 3y ago
Compile times for one. Eigen uses C++ templates to do most things, which explodes compile times.
245.
The Acid Test Clause (2017)(stimmel-law.com)
147 points by VHRanger 3y ago|48 comments
246.
by VHRanger 3y ago
AI content generation is only different in scale to content farms. Not in principle. All this means is that relying on an algo to filter things for you will become less useful (easy to game). You'll need to rely more on real humans you
247.
by VHRanger 3y ago
First sentence: sure! Second sentence: nonsense!
248.
by VHRanger 3y ago
Edge is doing computing on the client (eg. browser, phone, laptop, etc.) instead of the server
249.
by VHRanger 3y ago
The killer feature is presumably inference at the edge, but I don't see that being used on desktop much at all right now. Especially since most desktop applications people use are web apps. Of the native apps people use that leverage t
250.
Embezzlers Are Nice People (2017)(stimmel-law.com)
490 points by VHRanger 3y ago|325 comments
251.
by VHRanger 3y ago
More than that, to my knowledge, they can't be used as evidence of crimes. The NTSB runs as a "blameless postmortem" org. Their findings are to be used to change systemic processes.
252.
King Bruno the Chimpanzee(underthebanyan.blog)
1 points by VHRanger 3y ago|0 comments
253.
by VHRanger 3y ago
Also: - Jellyfin. self hosted media streaming AKA "selfhosted Netflix" - Moonlight. Stream games from PC to TV/laptop/phone. AKA "selfhosted stadia" - HomeAssistant
254.
by VHRanger 3y ago
I thought Framework laptop supported linux pretty well?
255.
by VHRanger 3y ago
Yeah, TBH I'm kind of sad about where OpenCL ended up, because it "should have" been what CUDA was used for from 2011-2021. AlexNet, TF, Pytorch, etc. "should have" been written with OpenCL backends. But the driver
256.
by VHRanger 3y ago
Yes, 1.2 is OK. The problem is that the drivers are merely OK. Presumably if you're using OpenCL you care about the performance (otherwise why would you??) and since that's the case, it's the best on no platforms, and there a
257.
by VHRanger 3y ago
Again, you're not getting the point. It's about behaviour. How often have you seen AWS deprecate stuff? The BOTO API for S3 is an ugly mess, but you can still run BOTO2 code from a long time ago. With Google I guarantee you they&#
258.
by VHRanger 3y ago
I didn't say they don't have a deprecation strategy. My problem is that they do have one, and they're very happy to use it. 12 months of notice is laughable for enterprise customers. From planning to getting something rolled
259.
by VHRanger 3y ago
At the end of the day, either you carry around a hidden state, or you have a fixed window for autoregression. You can call hidden states "RNN-like" and autoregressive windows "transformer-like", but apart from those two
260.
by VHRanger 3y ago
People are moving away from OpenCL. The big problem is driver support. The transition from 2 to 3 was a mess, performance is all over the place on some functions between intel/amd/nvidia, and Apple is at the point of just leaving
261.
by VHRanger 3y ago
I don't see any discussion of their deprecation strategy. What's killing GCP is not that they don't have the latest toys, it's that *NO ONE TRUSTS THEM ANYMORE* [1]. Why would you build anything on GCP when there's
262.
by VHRanger 3y ago
Past performance is not a predictor of future performance. While it's possible the gap between GPT5 and 4 is as big as between 4 and 3, it's unlikely. The gap between 2 and 3 was much larger as the one between 3 and 4 (and similar
263.
by VHRanger 3y ago
If you think that, clearly you've never used Oracle SQL!
264.
by VHRanger 3y ago
Yes, like RWKV and Mamba this is a new generation of models that are more like big RNNs than pure transformers we have now
265.
by VHRanger 3y ago
It seems only OK as a model? Looking at the LLM chat leaderboard it's 71st and the 14B version is worse than a lot of 7B models: https://huggingface.co/spaces/lmsys/chatbot-arena-leaderboar... Also, llama.cpp
266.
by VHRanger 3y ago
Yeah, that's just the training stability part to my knowledge
267.
by VHRanger 3y ago
Like RWKV and Mamba, this is mixing some RNN properties to avoid the issues transformers have. However I'm curious about their scaling claims. They have a plot that shows how the model scales in training with the FLOPs you throw at it.
268.
by VHRanger 3y ago
I have a blog post coming in a few days on this topic. But basically, I agree, architecture doesn't really matter. The number of parameters and the training data matter. The tradeoff transformers make is that you saturate the hardware
269.
by VHRanger 3y ago
Yes, Transformers are probably close to the Pareto frontier in terms of efficiency if you want to train something that has a sequence as an input. But they're not inherently *special*. There's a bunch of other model types around t
270.
by VHRanger 3y ago
1. There's nothing particularly special about Transformers, except that it was the first architecture to scale. Transformers scale well because they leverage a GPU well. The drawback to this is the attention mechanism (RAM hungry and f

More ›