pheug

What do you mean low margin? They're making 45% gross / 30% net, that's pretty respectable I would think.

Most of them just rent servers / VPS boxes, so ultimately your security depends also on whoever operates the servers themselves. NordVPN got recently powned through remote management software vulnerabilities in one of the datacenters they were renting from - https://techcrunch.com/2019/10/21/nordvpn-confirms-it-was-ha...

> Is it the log & clickthrough data?

It's a huge advantage and critical ingredient for ranking in all major websearch engines as far as I know. Google gets billions of queries and clicks every day. Your startup? Pretty much none of that. How are you going to train your ranking algorithm with no data?

> Imagine the privacy advocate conniptions if that were open-sourced. AOL got in huge trouble when they open-sourced their click-logs circa 2002.

Right, logs with such level of detail as individual user sessions would never get public for this reason, too much legal risk. But even simple aggregated datasets of the form "how many clicks did this query-url pair get" would be very useful to bootstrap a competitor and can be anonymized much more effectively.

Why exactly should he have known this, especially as an SWE? Google is a gigantic 110k+ people company by now. It's totally possibly to spend years head deep in code specializing on some obscure parts of codebase with near zero understanding how everything else works outside of your narrow area of focus.

I don't recommend ProtonVPN - tried it and if you're torrenting they throttle you to 20Mbps. Shady past connection to NordVPN / Tesonet also a concern.

Fastmail makes it very easy (unlike many others) to send emails from any *@yourdomain.com email as a sender. But if you only need to receive email, then almost any paid email service I know supports such catch-all addresses, even gmail/gsuite, there's no lock in.

Protonmail probably the most hyped, but personally I switched to Tutanota. I like their commitment to open source - you can build their webmail client yourself. They support full text search (unlike Protonmail) and also 4-5x cheaper than Protonmail/Fastmail, if you're fine with their cheapest 1GB storage plan.

On company issued hardware it's a fair assumption to make that they log everything you do and MITM your HTTPS requests.

The greed as well? Normal jobs aren't paying anything close to 200-300k+ that you'd be pulling at FAANG/HFT.

On a shared memory system there's little need to do that - there's much more parallelism to be had from accelerating fine grained operations, like matrix multiplications to compute each layer's output.

On a distributed system, splitting up layers between machines to do distributed training is pretty much what Google initially designed Tensorflow for. Generally it scales less well due to the need to communicate massive amounts of data between nodes and much lower network throughput than what GPU/TPU memory provides.

Actually it parallelizes extremely well, so that large companies are able to create monster models like mentioned in the article in the first place by just throwing money at the problem with TPUs and similar highly parallelized accelerators. It just doesn't lend itself well to distributed computing due to e.g. throughput requirements.

It seems to be working as designed:

"experiences that were clicked but not booked (which we treated as negative labels)"

You seem to be talking about dimensionality reduction, that's not what I was meant. Distillation is training a different model with a cheaper architecture (CNN, LSTM) on the outputs of an expensive teacher model like BERT. This has nothing to do with dimensions.

Distillation is usually used today to tame its resource problems at scale - you run BERT to squeeze out maximum signal from your training data and then distill the model e.g. into cheap CNN for inference.

Monthly can actually make perfect sense for this kind of service as a signal that the user's still alive. But yeah, $6/month is way too high.

Don't know about upper management, that's a whole different game at these giant companies. But low and middle managers get hired all the time without a personal connection, especially if coming from a different FAANG. Industry experience in a similar role is still a prerequisite of course, i.e you're unlikely to get hired as a manager if you weren't managing or a team lead before.

What you describe mostly applies to smaller companies or niche domains.

FAANG hiring is much more formalized and your connections hardly matter beyond getting invited for an interview.