revalo

This is amazing. I've been looking forward to such a thing for a while now.

It's much easier to work with LISP. It's functional, trees are easy to reason about. We're using an even more simplified language in this work. It's kind of like the MNIST for progsynth research. Scaling this to real and useful programming languages and domains is still non-trivial, and is a major drawback of our work here.

Yeah, this is true! These are more like expressions rather than programs. We were mostly following the language used by previous work, https://arxiv.org/abs/1906.04604

Changing tokens in a program is not differentiable. For me, the key idea is that you can train a neural model to suggest edits to programs by randomly mutating nodes. And when you run this neural model, you get to make edits that are syntactically correct (i.e., a number will only replace a number etc.) according to a context-free grammar.

I wasn't talking just about the US, I think we need very high levels of vaccination everywhere to make it statistically not a threat. We need to control the spread of the virus to levels where it's not mutating at evolutionarily scales, _everywhere_, and the only way to effectively do that is vaccines.

It's not a threat at this very moment, for this very point in time, but since we're allowing the virus to spread at such evolutionarily scales, inside and outside the US, a new variant popping up that evades vaccines is statistically very possible, and pose a significant threat to my own and society's well being.

I'm trying to argue that this is not a personal choice issue, rather a public health one.

Of course, we can never stop all mutations. However, for mutations to be useful for the virus, they do need to happen at an evolutionary scale. Especially because we're lucky that SARS-CoV-2 mutates much slower than Influenza and HIV.

If the rate of transmission reaches the elusive herd immunity, the time-frame for the virus to evolve into evading vaccines grows significantly longer, even with a non-zero mutation rate.

Lastly, antibodies created via the mRNA vaccines have a broader binding affinity to different variants than those from natural immunity [1].

Again, the goal is to not eradicate, but to scale it down to the point where it is statistically not a threat.

[1] https://pubmed.ncbi.nlm.nih.gov/34103407/

Adding to the thoughts before, letting the virus spread and mutate among the unvaccinated to create a new variant sounds like a reset on the progress we've made. With the Delta we got lucky, our vaccines still largely work. I'm not sure how long the luck will hold on with more variants.

I'm all for leaving people alone to their own decisions. But with this, their decisions greatly affect me and society at large.

I've engineered it to be very inexpensive to run. I work on high efficiency video codecs, so one benefit is that I get to sometimes test my codec's and get real user feedback.

I personally don't find virtual stuff fun, the pandemic is really sad, so wanted to make something available for public use. :)

miss you too fam <3

I run a completely free service https://firemeet.io Though I think gather's UX is actually very friendly to non-technical people!

Hey HN! I am a TA for a large undergraduate class at MIT. Teaching virtually definitely comes with its challenges, but the biggest challenge for me was holding office hours.

We used to have a queue system, with hundreds of students trying to get the first spot, with me (along with the other TAs) trying our hardest to keep the queue moving quickly, but algorithmic discussions need time and frequent pauses to think.

I experimented with lots of alternatives, but none of them scaled to hundreds of students (inexpensively) or had strong permission systems. We wanted a clear separation of permissions between the Instructors, TAs, and Students. Having a multiplayer code editor is a nice extra feature that has been useful to help students debug their code.

If you're in a similar position, or just want better breakout rooms, feel free to use the system. The default cap is 100 users to prevent abuse, but I can happily raise it if you reach out.

Would love any feedback!

The difference here is that this library is a one-line drop into your Python back-end to automatically generate the client.

Of course, the route you mentioned is much more powerful and complete, but here I'm trying to offer a way to very quickly build APIs without learning how to integrate any of those tools.

We bought a bunch of ov5647 camera modules from AliExpress. A quick search on the site yields a couple of listings at about $3.26 per camera, here is one of the listings, https://bit.ly/2W0Xui6

We spent the first month of the project on the v2 modules. These are based on the Sony IMX219. Unlike the v1 cameras, the datasheets are publicly available and even then, after much hacking around we couldn't get all the lines to expose at the same time :'/

Thank you! That's super cool :) The fastest strobe we used was a 500ns strobe, which is a 1/(2e6)th of a second exposure. The Pi cameras use the OmniVision OV5647, dynamic range of 67 dB @ 8x gain.

I'm pretty curious as to what you've discovered / worked on so far or have any interesting CV ideas you'd like to see?

My bad, maybe I should take more writing classes.

I guess what's different here is we're not using a $300 DSLR. We're using a $5 raspberry pi camera.

If we did the "let's take a bunch of high speed synchronized pictures with a strobe" on the cheap cameras, we'd see a single row of pixels worth of image, if we're lucky because of rolling shutter.

What we've done here is hack together some software to get a global shutter on these cheap cameras. Now that we can take pretty high speed pictures with $5, we scaled it up to get 16 cameras and took 16 angles of a bullet going though an apple, all at once. The "bullet time" examples in the repo are not computer graphics, they are 16 individual images played one after another.

Yep, that's the exposure time.

LOL, yeah, it looks pretty sketchy. It is a proper bullet trap though, in fact the very same one that Harold Edgerton used for his famous bullet through apple image :)

We reverse engineered the camera module's I2C protocol and figured out how to ask it to to energize all pixel lines at once instead of it going line by line.

The heart of it is just sending a few bytes down the I2C bus, but reverse engineering the register set was the challenging part as the datasheets are given only to OEMs.

Thank you so much!

We use a strobe light instead of a mechanical shutter with a flash duration of 500ns.

To take a picture of a bullet, we take an exposure of 1/2,000,000th of a second. This is super challenging with cheap $5 cameras like the Raspberry Pi camera because of rolling shutter, timing, syncing etc and is traditionally done with more expensive DSLRs.

But yeah, I think you're right the title doesn't seem like it's related, suggestions for a better title?

It's a very generic off-the-shelf CNN + LSTM, with a few bells and whistles.

Username checks out ;)

We had three main intentions with this project, two of them which you've mentioned. This was a hackathon project, and we decided not to go the generic route of what usually wins at such events. We wanted to have fun, and enjoy the process of creation. We ended up insulting our judges, and came last place :)

It's usually racist.

It might be! If a joke makes less sense, it probably is original.

Usually, but sometimes not.

Yes, the subreddit was used.

It's conditioned to output replays way more than make new jokes. But it does make new jokes occasionally.

Hahaha, we're just getting too much traffic. We've added more compute, should be better now!

Yeap, traffic is high. Our GPUs are whirring. Keep trying!