Pete Warden

+1. My favorite essay on the topic: https://petewarden.com/2023/09/10/why-nvidias-ai-supremacy-is-only-tem…

You want https://usefulsensors.com/, specifically https://www.crowdsupply.com/useful-sensors/ai-in-a-box This is from Pete Warden: https://petewarden.com/

Don’t Shave That Yak! - https://seths.blog/2005/03/dont_shave_that/ (this is the original with the borrowed EZPass) 2020 https://news.ycombinator.com/item?id=21930441 (81 comments) Don’t shave that yak – God loves lazy programmers https://petewarden.com/2010/06/26/dont-shave-that-yak-god-loves-lazy-p… 2012 https://news.ycombinator.com/item?id=4034034 (3 comments)

Hey, DVC maintainer here. For those who interested in this topic, I like this one about the same problem - https://petewarden.com/2018/03/19/the-machine-learning-reproducibility… (industry focused) and an excellent talk from Patrick Ball - https://www.youtube.com/watch?v=ZSunU9GQdcI&t=1s how they structure data projects.

OK. I was going by this - https://petewarden.com/2016/05/03/how-to-quantize-neural-networks-with… I haven't seen 8bit training implemented in any (public) frameworks yet - that's not to say it's not possible. If it works then that's great, especially for specialised hardware.

Believe Pete Warden ( http://petewarden.com/ ) works on TensorFlow now.

Yes, your comment was a paragon of unemotional wisdom. You have no interest here other than cool exercise of rational thought. You aren't upset at all. It's those other people who have let their feelings run away with them. Just keep telling yourself that. Because as we know, feelings like empathy and concern have no place in a discussion about how humans should treat other humans. Funnily, I agree with you that nerds are behaving badly now that they have power. [1] But I disagree with your conclusion. What happened to Sierra is a great example of nerds misusing their sudden social power. Now we need to start taking responsibility for the broader effects of our works and our previously insular culture. [1] For others, this is a good analysis: http://petewarden.com/2014/10/05/why-nerd-culture-must-die/

I think this phenomenon happens all over. Cities -- and regions within cities -- can have their own peculiar culture, while the culture of rural areas varies on much longer distance scales. Because of this, I suggest that the idea that North American is best thought of in terms of regions might be an unhelpful one. OTOH, the evidence from Facebook apparently suggests otherwise: http://petewarden.com/2010/02/06/how-to-split-up-the-us/

Why all the negativity, everyone? This is an impressive little framework - if you need to have the source code, it's not for you, but as-is, it is already useful for - at the very least - personal image recognition projects. I love his excitement in his blog post and video about it where he trains a program to recognize his cat: http://petewarden.com/2014/04/08/how-to-add-a-brain-to-your-smart-phon…

For deep learning or chips for deep learning? With an EE/Chem/Math background you should be set to go :) For deep learning, I highly recommend the cs231n course materials (available for free online) and the Deep Learning Book by Goodfellow et al. For chip design for deep learning specifically, it's a fairly new field, and a lot of commercial interests, hence why a lot of it isn't available in the form of an "Intro to Deep Learning Chip Design" course. But, the basic point is that deep learning can both train and perform inference in astonishingly low precision. For training: https://arxiv.org/abs/1502.02551 For inference, there are so many papers confirming this fact that there really isn't one key paper to point to. This article is a good introduction however: https://petewarden.com/2015/05/23/why-are-eight-bits-enough-for-deep-n… Other than that, stripping out the cache hierarchy, HBM memory, etc. are the obvious steps to take in targeting deep learning. Note that it's my opinion that deep learning chip startups (such as ours) need another "secret sauce" beyond simply lowering precision, we're not like Intel which can just say "okay, here's $100mil, make me an 8-bit GPU, Go."

Is this surprising? I think it's expected, when most published numbers from Google suggest that this is only 2x over DistBelief, which Adam Coates improved upon 6x (http://www.cs.stanford.edu/~acoates/papers/CoatesHuvalWangWuNgCatanzar… in 2013. Coming from a HPC background, DistBelief (and sometimes MR in general) has always felt like a square being forced into a round hole. Sure it "scales" but it's not how you would build a system from scratch to meet the same computational needs. Moving data around is much slower than doing extra math. DNNs provide some interesting opportunities for domain specific optimization with lossy math because empirically it seems that the solutions dont require significant precision. Something like a general Tensor library holds you back from using some of the tricks like synchronization-free hogwild, where it wouldn't be appropriate in other scenarios where more exact computation is necessary. One interesting step towards this is their implementation of lossy compression with a custom 16-bit half float representation. However a lot of people are dialing down the precision even further, including one of the authors of TensorFlow (http://petewarden.com/2015/05/23/why-are-eight-bits-enough-for-deep-ne… Scott Gray, one of the principle engineers at Nervana Systems, has a very wise reply in the github thread - they aren't doing much to optimize for memory locality, and wont have amazing speedups until that happens.

The title is from the original article (https://petewarden.com/2024/10/16/ai-pcs-arent-very-good-at-ai/), the URL was changed by dang: https://news.ycombinator.com/item?id=41863591

Thanks—we changed the URL to that from https://petewarden.com/2024/10/16/ai-pcs-arent-very-good-at-ai/. Readers may way want to look at both, of course!

There is also the much larger confusion arising of whether it's an ML compiler in the sense of being a specialized compiler for ML code (compiles a description of a network to efficient ML-specific low-level IR to enable further optimization and code generation), or whether it's an ML compiler, a compiler (or part of it) that uses ML. The first meaning is extremly common, the top 2 results for "ML Compiler" on google returns[1][2], both using it in the first sense. It's not that ML or ML-like techniques in compiler writing is new but the first sense is definitely at least as popular an interpretation. [1] https://huyenchip.com/2021/09/07/a-friendly-introduction-to-machine-le… [2] https://petewarden.com/2021/12/24/why-are-ml-compilers-so-hard/

>Could a bespoke audio player could enhance low bitrate audio back to sustained listenability? Well, that's what compression is, isn't it? I guess you are asking about how good is ML as a compression mechanism. People are thinking[1] about it. [1]https://petewarden.com/2018/10/16/will-compression-be-machine-learning…

Large percentage of mobile device power consumption is from radio use. Personal tests of phones in airplane mode and without airplane mode show even network keepalive consumes a fair amount of energy. I plan on purchasing the book as reviewed by this blog post: https://petewarden.com/2015/10/08/smartphone-energy-consumption/

Some have written about a "reproducibility crisis" in machine learning research: https://petewarden.com/2018/03/19/the-machine-learning-reproducibility…

Repos, no. But, no coincidence, I have open two relevant tabs right now: "Software 2.0", Andrej Karpathy https://medium.com/@karpathy/software-2-0-a64152b37c35 > It turns out that a large portion of real-world problems have the property that it is significantly easier to collect the data (or more generally, identify a desirable behavior) than to explicitly write the program. "Deep Learning is Eating Software", Pete Warden https://petewarden.com/2017/11/13/deep-learning-is-eating-software/ > The pattern is that there’s an existing software project doing data processing using explicit programming logic, and the team charged with maintaining it find they can replace it with a deep-learning-based solution.

Cool tutorial! It's great to see examples of people building models that are designed to be run in real time on the Raspberry Pi. This post reminds me that I really need to find time to update the RPi TensorFlow Github repo. The TensorFlow team is making progress on official support- check out their nightly binaries here: Model 2/3 B: Python 2.7: http://ci.tensorflow.org/view/Nightly/job/nightly-pi/ Python 3.4: http://ci.tensorflow.org/view/Nightly/job/nightly-pi-python3/ Pi Zero: Python 2.7: http://ci.tensorflow.org/view/Nightly/job/nightly-pi-zero/ Python 3.4: http://ci.tensorflow.org/view/Nightly/job/nightly-pi-zero-python3/ Github discussion: https://github.com/tensorflow/tensorflow/issues/5729 Pete Warden's blog post on cross-compiling: https://petewarden.com/2017/08/20/cross-compiling-tensorflow-for-the-r…

That's really unlikely. There's no OpenCL for the RPi. GPGPU programming for the Pi is mostly assembler based: https://github.com/nineties/py-videocore https://petewarden.com/2014/08/07/how-to-optimize-raspberry-pi-code-us… (example involves deep learning!) https://www.raspberrypi.org/forums/viewtopic.php?f=29&t=78919 (someone actually tried making an LLVM backend!) I really doubt someone integrated that into TensorFlow…

You can get the Deep Belief stack running on a Pi GPU: https://petewarden.com/2014/06/09/deep-learning-on-the-raspberry-pi/ Aside from that, I don't think the performance will be acceptable. I've been running Spark and Jupyter on a 20-core, 5 node cluster and it's acceptable, but only useful for debugging (because stuff happens slower).

The aforementioned google translate paper suggests so. My hunch is that TPUs work with a mix of 8 bit and 16 bit integer arithmetic on quantized networks with quantized data [1]. As far as I know nobody has managed to make SGD work properly with integer weights and that could be a reason why TPUs are not used for training (yet). [1] https://petewarden.com/2016/05/03/how-to-quantize-neural-networks-with… I could be completely wrong though.

8 bit is enough and I suspect it's what the TPU is using: https://petewarden.com/2016/05/03/how-to-quantize-neural-networks-with…

Based on recent blog posts from some Google folks regarding quantizing neural nets I'm going to guess 8-bit fixed point. For example -> https://petewarden.com/2016/05/03/how-to-quantize-neural-networks-with…

Integration. RPI is nice, but it has a bunch of things that are off spec for the purpose (Pretty decent 3D graphics, but really painful to do GPU work on http://petewarden.com/2014/08/07/how-to-optimize-raspberry-pi-code-usi… and GPU programming is a bit advanced for 11 year olds anyways), and lacking things that allow a diverse book of "labs" to be done in class. Having built in sensors and outputs is critical. If a school had to outfit their computer labs (to the extent they still even have them) with extra HDMI screens for every work station it would really drive up the price of the program, and limit the ability. This isn't supposed to be powerful, it is supposed to be an integrated tool that allows reasonably fun labs and learning sections for a diverse group of kids. You could arguably do the same with shields and arduino, but then you lose the form factor.

A bit of link chasing lead me to http://petewarden.com/2014/08/07/how-to-optimize-raspberry-pi-code-usi… which opens with this extraordinary sentence: "When I was at Apple, I spent five years trying to get source-code access to the Nvidia and ATI graphics drivers" That's at one of the richest development partners in the world. The secrecy is crippling.

ive thought about this too when considering an architecture like the one described ..seems as though storing in S3 is a good first stop for the data as it takes very little care and feeding for a massive amount of data. then, depending on your query and data processing needs you can look to other tools such as elastic search, or elastic map reduce or xyz db. see also : http://petewarden.com/2010/10/01/how-i-ended-up-using-s3-as-my-databas…

This[1] post was on HN about a year ago pointing to the Data Science Toolkit[2] for getting elevation and performing other geographic data lookups. [1]: http://petewarden.com/2013/09/09/why-you-should-stop-pirating-googles-… [2]: http://www.datasciencetoolkit.org/

There was a post on HN a while ago explaining why people should be building / using alternatives to google for geo stuff. The largest reason being that MANY uses are against the TOS of the google geocoding API. Particularly, I think, anything that's not user-facing and free/freely available. I can't seem to dig up a link, but a the case is made here (though not as well): http://wiki.openstreetmap.org/wiki/FAQ#Why_don.27t_you_just_use_Google… EDIT: found it http://petewarden.com/2013/09/09/why-you-should-stop-pirating-googles-…

Pete Warden would like you to use public APIs: http://petewarden.com/2013/09/09/why-you-should-stop-pirating-googles-… . Also, the association of low ground with poverty is old. Toni Morrison riffed on it in Sula: http://en.wikipedia.org/wiki/Sula_(novel)