joshvm

33 ms·

361.
by joshvm 4y ago
To add to this, companies at Google-scale tend to have a huge variety of ML related jobs, ranging from low level things like optimising libraries for different hardware, to the more general research positions where people are working on the
362.
by joshvm 4y ago
> The ultrasonic cleaner and distilled water are more expensive and have greater ongoing costs Distilled water should be pretty cheap. You can sometimes buy it as "ironing" water. I use it for my watercooled PCs (plus coolant c
363.
by joshvm 4y ago
Lifespan is not dictated by hard disk space, it's dictated by consumables like liquid helium, other cryosystems and fuel for stationkeeping thrusters. Most big NASA projects are specced to just promise enough to get funded (in terms
364.
by joshvm 4y ago
This is discussed in the article, as the sibling (Mike) comments, JWST uses the Deep Space Network (DSN) for comms. The issue isn't so much blackout (the segments are 120 degrees separated^), but that contact time is expensive and you
365.
by joshvm 4y ago
You can use yolov5 - here's a repo I made for an Ultralytics competition: https://github.com/jveitchmichaelis/edgetpu-yolo Note: typical constraint is RAM and changes to the EdgeTPU compiler which now fails to co
366.
by joshvm 4y ago
In principle you can directly replace function calls with their equivalents between frameworks, this works fine for common layers. I've done this for models that were trained in PyTorch that we needed to run on an EdgeTPU. Re-writing i
367.
by joshvm 4y ago
To add to the sibling comment. The challenge is not converting the weights as such. Pre-trained model weights are just arrays of numbers with labels that identify which layer/operation they correspond to in the source model. The challe
368.
by joshvm 4y ago
There isn't any reason you can't run a neural net on a CPU. It's still just a bunch of big matrix operations. The advantage of the GPU is it's a lot faster, but "a lot" might be 1 second versus 10 seconds, and
369.
by joshvm 4y ago
You could just run this on a desktop CPU, there's nothing stopping you in principle, you just need enough RAM. A big memory (256GB) machine is definitely doable at home. It's going to cost 1-2k on the DIMMs alone, less if you use
370.
by joshvm 4y ago
Journals are increasingly requesting that authors provide mandatory data availability statements. It doesn't necessarily mean that the data should be open, just that the authors have to declare if it is or not. Trouble is it's ver
371.
by joshvm 4y ago
Yes, I'm also curious what will happen there. The Express is still faster (20 vs 30 mins), but costs twice as much. And you have to get it from Paddington. It sits in a weird niche where the alternative rail routes weren't that
372.
by joshvm 4y ago
> I don't think its about best practices, its about good design and communication. I would argue these should be included in best practices for software engineering. > So should a mechanical engineer PhD be designing and making a
373.
by joshvm 4y ago
See my sibling comment on this. I think this is an ideal case for an RSE: if you have a shared codebase that ends up being contributed to by multiple members. That avoids legacy problems where someone contributes, leaves, someone else modif
374.
by joshvm 4y ago
The challenge is that usually PhDs are not writing software that's designed for production, is very specific (for a single conference or journal paper) and often the utility of the code is nebulous until the end of the project. So what
375.
by joshvm 4y ago
Yeah let me be clear. PhD students absolutely should get guidance from experienced engineers (so I was a bit over-zealous with "assist" in my parent post). But this should be more like understanding best practices, and they shou
376.
by joshvm 4y ago
RSEs should not be working on, or assisting, research at the PhD level. By that I mean news/skunkworks stuff. They should be taking the output from a PhD and turning it into core software for the group. That doesn't solve the coll
377.
by joshvm 4y ago
P2P lending is a thing, and was quite popular in the UK until most of the big players (e.g. Zopa) went private. You can invest a small amount, usually as low as $100 and it goes either to personal or business loans. There are companies that
378.
by joshvm 4y ago
This was the subject of a Silicon Valley skit: https://www.youtube.com/watch?v=HcXu4_K1tMQ This approach is used commercially. Our office is in a co-working space and we have a smart fridge in the kitchen. I should point ou
379.
by joshvm 4y ago
The duct fan + filter combo is how lots of cylindrical consumer units work (e.g. Levoit, Philips, etc). They just use crappy fans which don't really pull any air through, or they're very expensive for what they are. In terms of av
380.
by joshvm 4y ago
Perhaps a stupid question. Why go to the trouble of bolting together four filters, instead of attaching the duct fan to a cylindrical HEPA filter? Then you just need two plates: one to adapt the duct inlet and one to seal the base of the fi
381.
by joshvm 4y ago
I would imagine this is mostly the domain of on-call engineers/SREs (wasn't this historically a problem that was fixed by a reboot?) and you have to assume that you're going to be called at 4AM on a Saturday morning. So even
382.
by joshvm 4y ago
TF Research Cloud. Especially if multiple people could use it with eg model checkpoints: > 5 on-demand Cloud TPU v3 devices, 5 on-demand Cloud TPU v2 devices, and 100 preemptible Cloud TPU v2 devices for free for 30 days So up to 7k hour
383.
by joshvm 4y ago
They don't offer direct search, but isn't this what "key moments" is in search results? Try eg how to change a lightbulb . I believe SeekToAction works even if the uploader didn't put chapters in. This was a relati
384.
by joshvm 4y ago
For hobbyists in the US, Sparkfun or Adafruit are a good start. Otherwise go to the big suppliers: Digikey, Farnell and Mouser. China (Aliexpress) supplies pretty much everything else. Many dropship on eBay and Amazon too, if you want conve
385.
by joshvm 4y ago
Yes exactly, when I saw $100k I thought that sounded pretty good for an academic salary basically anywhere else except Switzerland. A better comparison is somewhere like Oxford/Cambridge which both have extremely high rent/income
386.
by joshvm 4y ago
It can happen both ways. My PhD ended up being a computer vision project and while I could program better than most new grad students I didn't have any significant experience with image processing, visual geometry or CV (or ML for that
387.
by joshvm 4y ago
If you're comparing to expensive places to live, Switzerland pays academics well. Postdoc salaries are around 85-100k (CHF, which is about the same as USD these days). Many universities don't have much in the way of a pay ladder t
388.
by joshvm 4y ago
There are plenty of variations on this, I'm surprised no one has mentioned Falsehoods programmers believe about names yet: https://www.kalzumeus.com/2010/06/17/falsehoods-programmers-... (as someone wi
389.
by joshvm 4y ago
Yes, it is. One of the biggest providers is called GoGo who in turn use satellites from SES. My comment about satellite was that I've also worked in very remote places using much poorer links and HN still works , amazingly.
390.
by joshvm 4y ago
My experience with in-flight message-only WiFi is that they're just really slow and the ping times are long. Some services are actively blocked, e.g. Skype wouldn't work at all even for text messages, but browsing the internet is

More ›