pinouchon

Neurosym labs | Zurich (maybe NY/Boston, maybe SF/Bay area) | ONSITE, INTERNS, VISA | Full-time

I am considering starting an AI/ML research company in the field of neurosymbolic programming (which sits at the intersection of deep learning, program synthesis and probabilistic programming). The goal is to be 100% focused on producing research that is made publicly available, guided by intellectual curiosity and a desire to move the field forward.

I am looking for co-founders or early employees. Although there is an existing company & office that will be ready by the end of summer, not all details are figured out yet. The location is 80% likely to be Zurich (keeping the current office), but could be in the US. The funding is privately secured by other ML-related ventures I was involved with in the past. The name is also a placeholder name and will change. I can give more details privately.

Roles and compensation:

  - Full-time employees: base of around 100-250k USD/annum + equity with standard vesting (negotiable).
  - Interns or residency program: base of around 80-180k USD/annum

(Other research interests: self-play, world models, common sense reasoning, program induction, causality, bayesian inference)

To apply, send me an email to b.crouzier at gmail and including "Neurosym labs application" in the title. Please reach out also if you just have pointers or other things you want to discuss, as I'm still exploring all options available.

This institute is founded by some of the best people in their respective fields. I think probabilistic programming is one of the most underrated fields in AI. Have a look at this: https://www.youtube.com/watch?v=GYQrNfSmQ0M&t=1000s or this: https://www.youtube.com/watch?v=TFXcVlKqPlM for an intro. So I highly recommend getting in touch with them if you have the skills/interest.

Look for Vitalik, balaji srinivasan, naval ravikant podcasts, also chris dixon

Interestingly, the poet with the highest level understanding can be clueless about UTF-8 or what a pixel is

Yeah. Have a look at this for a good summary of the ideas: https://www.youtube.com/watch?v=lhYGXYeMq_E

I'm using the qgmlwy layout for a few years now, pretty happy with it. I remapped caps & added a new modifier to have a "special chars" layer and it works pretty well. Overall it's more trouble than it's worth, if I had to redo the whole journey, I would simply learn to touch-type with querty (maybe colemak), but I'm still happy with the end result.

Before that, I had a strong interest in AI & ML, read the Deep Learning book and did the cs231n mooc. So I could train basic neural networks with pytorch but that was it.

We have experience trading since march 2020, so no proper experience in a bear market. We deployed our first algos with nontrivial amounts (> 100usd) the night before the march crash. And yes we did get rekt that night :) Trading strategies work. That it still works is a huge mystery to me. We did beat BTC, we have no long/short bias, and are doing the best during volatile periods (up or down)

Posted a comment on this thread, https://news.ycombinator.com/item?id=26321793, then had people contact me and had very good conversations as a result.

Also, I met my co-founder on reddit/r/algotrading and it was a total game-changer for me (reddit is not HN, but similar)

Deep Learning (by Aaron Courville, Ian Goodfellow, and Yoshua Bengio) Read cover to cover, then applied these techniques to Crypto Markets (turns out, crypto markets are pretty inneficient)

Do you think air pollution can be a cause of headaches and do you have good sources for reading about this? Recently I have been having a constant headache after moving to Budapest. Went to the doctors, they found nothing wrong with me. I went back to Germany and France for two weeks, felt better there. Back to Budapest, felt worse again. I also tried to move to Bielsko-Biala in Poland and the headache has gotten even worse here (2 days since I'm in Poland) It seems far fetched that pollution is the cause of my headache because most people do not have headaches there, so I want to read more about it.

Then it's a matter of investing that money wisely for a few years

https://arxiv.org/pdf/2006.08381.pdf A take on Josh Tenenbaum's hard problem of learning

I have very similar issues. Two things worked for me:

- have a daily tracking system where you have clear simple goals (such as going to bed at a certain hour, the first thing you do in the morning should be productive...) and try to have "winning streaks". I use an google sheet for this and color red/green tiles based on success/failure.

- Find a buddy to work with. Since I started working with a cofounder next to me, my productivity shot up simply because I feel bad doing crap while he is working.

or trading == investing (and changing your mind very often)

For the last year or so I have been working on a ML-based trading system in the domain of crypto with two friends. I made more in 2 months than I used to in a year. This is after thousands of "full positions swings" and millions of trades (short and long). We are now experimenting with different classes of trading strategies to reduce risk.

We would like to find 1 or 2 more people to work on this project, we need people who can tolerate risk and skilled at data engineering: data pipelines, psql, pandas, numpy, data visualisation, setting up servers. Ideally also skilled at machine learning / deep learning and who has tried his hand at trading systems. If interested, my email is in my about info.

It depends how data hungry those models get. It might be the case that to train GPT-10 you actually need more data than there is available in the universe, so you have no chance of training it.

Hi. What would be the incentive for someone who has collected an expensive dataset to list in on your marketplace (or dataset-directory, I don't know how you would call it)? I'm just wondering what incentive structure would get a marketplace like this off the ground.

Once you have a big marketplace, then it's a clear value for any engineer who wants easy access to quality datasets (for experimentation, transfer learning...)

This seems like a big deal. Any thoughts?

https://music-grid.surge.sh/#128-32-64-32-8-32-0-32-0-0-0-12... https://music-grid.surge.sh/#320-0-320-1024-128-0-160-0-16-1...

I am using the qgmlwy along with a few custom layers and custom modifier keys. Most of the improvements come from touch typing (not a custom keyboard). But a custom keyboard is still an improvement. The biggest issue for me is not the learning part (even though this took months), but the setup: whenever you want to switch os, reinstall, setting up the keyboard is a huge pain.

I have setup my layout on ubuntu with xkb and this took days of full-time work, because xkb is so convoluted. On osx, I just use a custom C script that transforms each keystroke, the code is ugly but this is easy to work with.

Sometimes I wonder how big a tragedy it is that many well-written questions are simply dismissed because the author found his solution in the process. I'm sure many of those questions are high quality and would get edits, upvotes and answers had the author just pressed the submit button.

Stackoverflow is the best rubber duck because the anticipation of that fear is quite real. And I know that on many occasions I will end up submitting that question

The IDE (I also use PyCharms) is great for editing code (search, refactoring, syntax highlight...) Jupyter notebooks (or JupyterLab in my case) is great for prototyping and data visualisation.

In my work (data engineering and training deep neural nets) I also happen to use both. It's unwieldy, but I can't get rid of either one (pycharms or jupyter lab)

No one suggested standford cs231n: http://cs231n.github.io/. I'd recommend the winter 2016 lectures (by FeiFei Li, Karpathy and Johnson). For getting started with convnets / deeplearning, I think this is one of the best hands on ressources out there.

I'am about to try DaVinci Resolve as an alternative to Premiere (I used Premiere in the past). But I don't want to fork $20 per month because I don't use it that often. A good opensource video editing software (or NLE non-linear editing as they call it) would be great.

Funny how this is in stark contrast with PG advice of staying alive (http://www.paulgraham.com/die.html)

The thing is that Kapla really really don't want the citiblocs videos to be public. I agreed with them do delist those videos because they paid for unexpected plane expenses as part of a project I did with other builders (by that point they had made that request many times). So if I let my brother upload the citiblocs videos (on which he did help but I still did most of the work), that will damage my relationship with kapla a lot. That will probably mean they won't sponsor partnerships I do with other builders (where they pay for travel) or free planks. Kapla know all of that. I cannot fool them saying it's not me who uploaded the videos.

The option that does not damage my relationship with kapla is sharing revenue on exiting kapla videos.

I also proposed to split 50/50 revenue of future videos that we upload to a new common channel. And de-monetizing videos where we work together. But he doesn't like any of those options.

Also the company citiblocs has sold, and their planks are lower quality. My brother doesn't want to play with citiblocs. He wants the existing citiblocs (or kapla) videos to make money off of those videos.

I never thought about that. I actually think this is a great idea. One big issue is a lot of people would be upset at the idea of their reputation getting smaller. Maybe have a new "influence" or "helpfulness" score along with reputation that takes into account the decay you are talking about. Or maybe just have a different scoring mechanism altogether based on how knowledgeable and helpful your are to the SO community.

I use a supervised learning setup (although with a custom loss function).

The kind of data augmentation I do is adding different candles sizes. I validate with 5m candles, but I train with 2,3,4,5,6,7m ones. I also sample more frequently more recent data. I train jointly with ~22 symbols, but in each X with those symbols, I randomly set some to 0, some I invert their price, some I invert time-wise. This helps generalization for some reason. I tried many kinds of noise, but what I described above is what I found to work best in my case.

I have a more ambitious idea to generate synthetic data using self play: have a bunch of agents trading one against another. This create new price data I can train the agents with, and repeat (this self-play training scheme would be similar to what DeepMind did with AlphaGo/AlphaZero). The issue with it is the need to tune the parameters exactly so that the resulting synthetic data is realistic enough that I can tranfer the agents to real data.

For example, during self-play, should you have only trading agents or should you add "retail traders" that buy during bubbles, "normal buyers" that buy only below, sell above certain prices, institutional buyers that randomly move the price a lot in a given direction. This is a lot of parameters to get right, and it's an optitization problem on it own. You could treat this a as two-fold optimization problem such as in this paper: https://arxiv.org/pdf/1810.02513.pdf, but it gets tricky very fast.