fnbr

Yes, they do. It’s typically slightly less than what you get initially, although that can vary depending on your performance.

I am a DL researcher at a top industry lab.

I'm completely unsurprised by this. Regularly, at lunch, I'll ask my coworkers if they know of any DL applications that are making O($billions), and no one knows any outside of FAANG.

FAANG is making an insane amount of money due to DL. Outside of them though, I don't know who's making money here. When I was interviewing for jobs, there were a ton of startups that were trying to do things with DL that would have been better done with a few if statements and a random forest, and that had a total market size in the millions.

I think that, eventually, there'll be a market for this stuff, but I'm not convinced that it's anywhere near being widespread.

I was also a consultant before my current role. The vast majority of non-tech firms don't have their data in well organized + cleaned databases. Just moving from a mess of Excel sheets to Python scripts + SQL databases would have made a HUGE difference to the vast majority of clients I worked with, but even that was too big of a transformation.

Basically, everyone with the sophistication to take advantage of DL/ML already has the in-house expertise to do it. There's almost no one in the intersection of "Could make $$$ doing DL" && "Has the technical infrastructure to integrate DL".

They don't allow Kindle purchases in the Amazon app on iOS, maybe this is why?

See: https://www.amazon.ca/gp/help/customer/display.html?nodeId=G...

Unfortunately, this isn’t true in Canada. MDY is very common, especially in the English parts. See:

https://en.m.wikipedia.org/wiki/Date_format_by_country

Also, you typically save a few hundred bucks on your insurance by having monitored alarms, so it pays for itself.

For me, I use 20g of coffee, ground to a 20 on a Baratza Virtuoso, with a 275mL of water in a French press, brewed with boiling water for 4 minutes.

If you change any of these variables- grind size, water temperature/volume, amount of coffee, etc, you drastically change the caffeine content.

They’re really useful (and arguably state of the art) in situations with small amounts of data. Deep learning is really hard to productionize, so classical techniques like SVMs and random forests are widely used in production. Deep learning is too, but not as much as you’d think.

Yes, exactly. I've learned much more in the last two years working at a big tech co than I did at a small software company in the midwest, making me a much stronger engineer.

> I'd guess that the people worth $600,000 Facebook salaries are already living there and working for Facebook.

Yes! I think this is true. However, the more interesting development to me is what happens 10 years from now. The engineers who are making $600k in SF are typically 10 years into their career (if not more). This is the biggest knock against expanding into places like Omaha- there's plenty of entry-level talent, but little, if any, mid-level to senior talent.

In Edmonton (Canada), where I'm from, there's an awesome university that graduates tons of CS grads who immediately head to the Bay Area, because they can make 4x more there. As such, most of the work in town isn't sophisticated- there's lots of web development shops, but not many jobs doing sophisticated engineering at the level of, say, Facebook or Google.

As a result, you're right, there's not many people who could get a $600k offer from Facebook. But! If these companies start supporting remote work, we could see tech ecosystems developing, and we could see new grads staying in these places, and building ecosystems around them.

To answer your question about:

"why, if there were such big savings on the table, these companies hadn't already moved in this direction."

I think the answer is as follows:

1. New grads need more experienced talent to help mentor them and ensure that they're seen as productive.

2. All of the companies' offices have historically been in certain locations, so we either need to a) have the new grads move to somewhere we have experienced talent, or b) need to have experienced people move to where the new grads are.

We see both of these- companies regularly will start new offices once they can get sufficiently experienced people to move there, but only when they can get that. This is the main reason (aiui) why many of the large tech cos in the Bay Area are concentrated in the Bay- there are no senior level people who are interested/willing to move.

Good on you for doing this- I think the apology is great and shows TripleByte listens to feedback. I also think that taking on LinkedIn could be amazing for the broader ecosystem- LinkedIn is terrible, and anything competing against them would be awesome, so I wish you luck.

It’s mostly an issue that A2C isn’t designed for adversarial environments. It also doesn’t have any notion of hidden information, while other algorithms (eg CFR) explicitly handle this. There’s a well-known phenomena of cycling, where agent A will beat agent B which beats agent C which beats agent A; A2C can exhibit this. Think of rock/paper/scissors- AlwaysRock beats AlwaysScissors which beats AlwaysPaper. To avoid this, you typically need to do some sort of averaging.

The alphastar paper and blog post do a good job discussing these issues as they had similar problems. I’d say that’s a great starting point (and then following their references).

Blog post:

https://deepmind.com/blog/article/alphastar-mastering-real-t...

An additional problem with this is that they use A3C here for trading. A3C is known to not be suitable for adversarial environments (e.g. board games, like Chess).

I wrote a paper that demonstrated that A3C is as exploitable as a uniform random strategy in board games (specifically, some poker variants): https://arxiv.org/abs/2004.09677

(Exploitable is a technical term that is defined in the paper; basically, it's "how much can someone who knows everything about your strategy beat you by?")

So I would be very surprised if this survives contact with other traders.

Yeah, absolutely. Now, the high end is fantastic, but the average meal was _not good_.

I agree completely. Doing novel things with ML is very, very, very hard, and (afaict) hasn’t been particularly successful at the startup level. But operatonalizing existing techniques is _very_ promising, and where I think a lot of the wins will come from.

I largely agree with what you're saying. The only point I disagree with is 3. That's the exact same logic I used to make a large investment on March 6th. If I had time-averaged the investing, I would have saved a ton of money.

The point is that it's near impossible to predict market bottoms, so time-averaging helps with that. If your point is that you should increase investing now, I totally agree, but you should invest the money over a period of weeks/months, not in one lump sum (which I don't think is what you're advocating).

This is exactly the right answer. Broad based index funds, invested over time.

You could also consider a 70/20/10 portfolio of S&P 500, international equities, and bonds.

Can you link to the tweet stream?

In general, no. The highest performance software only runs on Nvidia GPUs (or internal hardware, like Google’s TPUs).

Why isn’t this view shared by Alphabet?

That's high, but not crazy high. A Staff Engineer at a FAANG can make ~$500k. See:

https://www.levels.fyi/# (L6 at Google, for instance.)

I know engineers in Canada who are making similar amounts. There's not a lot but it's not nonexistent.

I worked with a team which used raw pointers instead of unique_ptr as they use more memory (twice as much, I think). The core data structure was a tree, and the tree was several hundred GB in size. So they were doing whatever they could to shave off memory.

They’re not paying them that much- $160 000 over four years, including tuition. So it’s $40 000 a year, which is a great deal imo (for both parties).

I assume that, with sufficient screening, you can get someone contributing relatively quickly, and start getting value. I think the key is the screening- they’re not giving this offer to every CS student. At the places I’ve worked, we’ve definitely got value out of interns comparable to a junior developer.

In my experience, most pure python code will be 10 to 100x faster if it’s rewritten in C++. So I just profile it as usual using cProfile, try to make algorithmic improvements (eg caching) and then, if I need another order of magnitude, rewrite it in C++.

So basically anything where the hot path is in pure Python, rather than a standard library method.

In my experience, it's all cars in Canada- I've never seen one that doesn't have it (other than imports). I guess you could probably get away without it on the west coast.

Most cars do start though. It’s really a maintenance issue, where it’s bad for your car to start it at -30 (as the linked article says). In my experience, a newer car in decent shape will start consistently.

Yes!

Yeah, but that's an embarassingly parallel task, so you can recreate it with sufficient resources. If there's a sufficiently lucrative application, businesses will happily pay that.

Yeah, but it's still a gamble. Unless you become a VP/Director it's tough to get to a sufficiently high level to make it worth it. It's rare, for instance, for someone from a start-up to come in as a Staff/Senior Staff engineer (L6/7).

For Google, that's entry level (L3). L4 (between fresh graduate and senior) makes $250k TC.

L4 is ~2 years out of school.

Two examples:

- Kenya has a dominant electronic currency, M-Pesa

- South Korea rapidly developed into a globally competitive developed country, despite being poor in the middle of the 20th century (same applies to Japan, just slightly less so).

Presumably they have a supervised model that they've trained on all their labeled data so far (possibly pooled across clients). They'd use this to estimate labels for their data, and then have humans correct it. They're basically doing the standard supervised data training loop.

If I had to guess, the long term plan probably is to move up the stack and sell the models to their clients.