epberry

Are your twenties the “best time of your life” though? It seems to me that you can find happiness and satisfaction at almost any time in your life, if the circumstances are right. I acknowledge that in your twenties you’re more able to stay out all night and have crazy adventures.

At any rate I’ve noticed a funny thing where peers who aren’t doing startups in many cases end up working just as hard. They’re young and out there to prove themselves at FAANG or other places and they make up for their lack of experience but putting in long hours. Same with investment banking, law firms, medicine and many other careers where the path to success is paved with very hard work.

One implication for products, if this work pans out, is that the systems which "adapt to you" (learn your voice, your schedule, what your face looks like) won't have to change the weights of their network. I believe this could lead to much better systems because adjusting the weights of an already-deployed network is dangerous - you could make the system perform poorly. It's also simpler for engineers because you can scrap all that logic and just deploy one RL network.

Sticking with a project in college out of interest, despite pressures from family and friends to ditch it or not work as much on it.

I had to do it: http://www.paulgraham.com/submarine.html

I didn't realize all the pieces of the cloud stack that CNCF was working on and how each piece gives you more flexibility to move between cloud providers. Kubernetes and rkt replace servers and orchestration of course but you also have Open Policy Agent and Spire to replace IAM, Envoy and Jaeger replace X-ray, VPC, Security Groups, Rook replaces S3, etc.

It's kind of interesting because Google and Microsoft have opened a new front against Amazon to counter their strategy of locking you in with services. If software from CNCF is good enough to be used in place of these services on each cloud provider then you can nullify AWS's advantage there and Google can lure you in with ML and Microsoft with their enterprise experience.

But I'm not sure these projects can make progress faster than AWS can release and update their fantastic array of services. Kubernetes being the obvious exception (when is EKS GA?!?!)

Part of me feels like a hidden motivation for doing this was so Google could continue to ask their “implement a BigInt class” interview question :). Cool write up though - I enjoyed it.

Same here. Redirect from http -> https works fine tho.

Interesting, I wasn't familiar with bitonic sorting.

It's late and I may not be thinking clearly or be well informed enough but...

I wonder if this type of thing - implementing algorithms on GPUS that have been the bread and butter of CPUs for generations (parsing regular languages?) - begets a new style of using computer hardware. The sketch shows several clever techniques to get a modest performance gain on the GPU, but what happens when the GPU hardware keeps getting better and the CPU stagnates?

I keep thinking back to that Feynman lecture of computers using the analogy of an office filing system. There the final step was a filing clerk so dumb he could only add one to things but he could do it so fast that he was still better than the filing clerk that could reason out all the operations. Can we extend that analogy to have 8 filing clerks all adding one to different wrong indexes and then choosing the right one in the end? (I'll be honest, my understanding of how GPUs work is rudimentary).

Perhaps the gains we get from doing the wrong things in parallel actually start to outweigh CPUs doing the right things in sequence (ex branch mispredictions).

EDIT: My ramblings aside, kudos to the author for a thought provoking piece.

Same here. This is one of the major opportunities to slow the effects of climate change.

That lecture was exactly what I needed right now - thank you.

I'm all for this. I work with two engineers who are just like this and they do a fantastic job. It doesn't bother me or our CEO at all that they might not buy into the culture or the mission as much as our younger folks.

I’ve experienced some of these same things. But more than that, I applaud you for experimenting!!

Great overview with lots of links. I came across quite a few concepts I barely know - or know them in a different context. I did trip hard over this sentence:

> Distributed systems often have to store a lot more data, than a single node can do so.

It did take me a minute to actually find Parameter Store. There's so much stuff hidden away in the EC2 menus.

Besides the secrets rotation how is this different from EC2 Parameter Store? I’m genuinely curious and will move away from parameter store if this provides some benefits.

Yeah the variable depth-of-field not working well at < 1 ft is a just a property of the hardware sensors. I guess I don't know exactly what they are but if it's time of flight or stereo then they could make it work better at closer ranges but then it wouldn't work as well farther away.

Very cool. I shudder to think of the GPU costs to run these models though. Perhaps they're using TPUs to be as efficient as possible. If you imagine a room occupied in the evening by people for several hours and you have any decent framerate, you're running your pose estimation network on each frame for several hours. And these models are big as far as I have seen. So that pretty much means you have one cloud GPU per camera allocated every evening. I suppose another option is they are running pieces of the network on the device but I think that's unlikely.

Of course these models are getting smaller over time and I'm incredibly impressed that these guys have put together the hardware, computer vision, and cloud setup. I also think they've nailed the MVP - not too easy but not too complicated either assuming they have decent models.

I'm signing up!

The author stressed that this was in NYC in which case I totally understand. $10 per drink actually seems on the cheaper side.

Love seeing this stuff on here. Maybe this reflects poorly on my schooling but I've learned more about literature and philosophy from HN than from anywhere else.

I think the CTOs role here would probably be to make sure a process or team is in place to do this but not actually do it themselves? (For a company the size of oculus)

Agreed. I have never had as good of an experience with IRC or XMPP as I have with slack. Open protocols are still great for open communities, internet discussion boards, etc. but when you're a multi-billion dollar messaging platform and companies are using you for critical communication I don't see any reason to continue supporting them. They've actually gone a step further and put the accessibility features in slack and upgraded their api to support things people used to do on the irc channels.

I used to be surprised when people came across these bugs. I would think: a kernel bug! They must have truly hit an edge case for something like that to skip into production. But then I read Showstopper: The Race to Build Windows NT, specifically the part where that OS shipped with 65,000 known defect (and it was considered one of the most stable software releases of all time), and it became clear to me that these bugs are probably known by the team when the OS ships but just not considered high priority enough to fix. Same with the recent articles about iOS and engineers not getting time to fix P2 bugs. Long story short my entire perspective on kernel bugs (and partly software development) changed and I began to think of them more as known flaws that weren’t fixed because of schedules rather than unknown bugs that were uncovered in the wild.

Ah this is cool too! I assume no neural networks are involved here?

EDIT: why the changing colors in that gif?

Interesting, thanks for the link. I tried a few published methods for doing this a year or so and I found they were quite slow. As in, you'd rather just use a stereo camera to quickly generate a decent pointcloud than use a single view plus a network like these. What did you find your performance (ms/frame) was like?

Interesting - I did not know about their history with gay rights.

Agreed that pytorch feels “pythonic”. I think tensorflow doesn’t because it’s really a C++ API with an extensive set of wrappers and it shows. Pytorch feels like they started with python and added in the C extensions after the fact.

One feature I miss in navigation services is an ability to choose the “easy” route. For example, I live in the Midwest and usually I’d rather take the interstate, even if it’s 5 minutes longer, than a series of country roads. Obviously you can choose this when you know where you’re going but forget it if you’re in an unfamiliar area. A simple heuristic for this would be which route involves the fewest different roads.

> The Beats are boring.

Hear hear. I was always confused about what they offered. Every generation has some pushback against mainstream culture but the best movements are accompanied by something greater. I never got the feeling there was a larger cause with the Beats - maybe that was the whole point.

A sense of community even while living far from the valley. Thanks HN, don’t ever change.