kahnjw

I could have been more clear. My point is that aggressively reopening might still be dangerous, but it'll take a while for the public to convince themselves it's safe enough to start going out, even though it isn't. I'd imagine this would be driven by both consumers not wanting to get sick _and_ businesses trying to avoid liability for employees and customers.

This perfectly illustrates the "opening up" paradox. If you, as a Government, aggressively open but the public still sees risk, not much will change. People won't go outside much because they perceive risk. We'll have to wait much longer to actually know whether opening aggressively is a mistake or not. From my perspective the risk of opening too early and starting a second wave much later in the year, is worth taking very seriously.

I'd say anyone drawing conclusions from a a few weeks data has a 50% chance of getting burned at this point.

Becoming a massively profitable megacorp isn't the only winning formula.

Were Linkedin, Instagram, Beats by Dre, WhatsApp, Tableau, Skype, GitHub, MuleSoft all failures because they were acquired for billions, making lucrative paydays for their founders and investors?

Let's leave it at this then: if capital is completely miss-allocated and a bubble has been inflating over the last 5-10 years as you claim, then we'll hear the proverbial pop in the next three to six months as a rapid pullback in consumer spending unwinds nearly all VC backed growth stage companies.

Ask that to someone trying to find housing off Sand Hill Road in Palo Alto.

Are we talking about investment strategy or cherry picking data for the sake of arguing?

1. There are plenty of companies on the path to IPO that didn't take 1B+ in VC money 2. The "sharing" platforms are expensive investments because there are so many players fighting for market share.

We're talking about a strategy of fast growth vs slow and steady. All the companies we've mentioned so far invested in fast growth early on, whether from VC or reinvestment.

I'm not sure what you're getting at. Here are the facts: Companies following these accelerated growth trajectories now make up a total of 4 trillion in market capitalization depending on how you count it. That's really just the FAANGs, not the smaller companies that are profitable or on the road to profitability [1]. If you count everything you can safely say the number is closer to 8 trillion.

Every year, VC in the US _as a whole_ invests roughly 100B [2]. If you cut out non-growth and non-tech sectors I'd guess that number total goes to around 40B, and roughly 100B (very rough number) globally.

So yeah, some money gets "wasted" but it creates huge market capitalizations that are around two full orders of magnitude larger than a single years investment, and growing strong year over year.

[1] https://www.investopedia.com/terms/f/faang-stocks.asp [2] https://www.prnewswire.com/news-releases/us-venture-capital-...

Except for dozens of counter examples that make up for literally multiple trillions of dollars in market capitalization.

Correction: Dominos not Pizza Hut.

I'd counter that Pizza hut has a key logistical advantage of vertical integration. A delivery service is a very different business than a restaurant chain that does its own delivery.

I agree that GrubHub, Doordash, and to some extent Uber seem bloated when considering the sum total of the markets they play in. That doesn't mean these business models aren't sustainable, though. Some companies allocate resources to a few areas that turn into profit centers, some don't. The ones that don't will be sold off or parted out. And the cycle will continue. I'd wager that one of these companies will survive and turn out to be a profitable, healthy business in the next few years. The rest will probably be sold off or slowly downsized.

More broadly, to your criticism of SV's investment strategy, resource allocation is a hard problem. If you want to direct large sums of capital at certain business verticals, do you want to grow slowly and steadily over a 20+ year period only to find that the economics don't work, or do you want to fail fast with some extra waste in the middle? Failing fast has some upside to it, though I understand why I consistently hear this criticism on this site. It feels like the last decade has seen the pendulum swing towards fast money and back a little. I don't think were as far off from a healthy middle ground as some might argue.

Some private companies offer RSU grants. I've been offered them.

There are other factors at play. Most employers offering options don't pay as much in total as those employers offering straight up RSUs.

I got an RSU offer from Pinterest a couple years back when they were private, though I turned it down.

Words matter, concretely define "open ended"? Did you just add that phrase to preemptively nullify evidence to the contrary?

Deep learning has surpassed human level performance on many tasks [1][2]... (could add more you get the point).

[1] https://www.sciencedirect.com/science/article/pii/S2215017X1... [2] https://arxiv.org/pdf/1502.01852v1.pdf

From all appearances this is more or less what China is doing today. Lockdown until the number of new cases in a geographic area is zero, then slowly loosen the grip of suppression and calibrate based on new information. When the second wave starts I'm sure the response will be a little more measured this time.

This is only true if you don't count the cost of massive casualties that are impossible to fully quantify the cost of and only count of the cost of running the healthcare system at existing capacity.

Let's cook up a fantasy scenario where we have unlimited resources and can treat as many patients as needed. Let's also optimistically pretend that only 1% of cases need ICU treatment. Take 1% of 327M and you get 3.2M ICU beds required. Now let's say that at the peak, half the population is sick (likely given the unmitigated exponential explosion scenario), meaning we need 1.6M ICU beds at the peak. We have roughly 100k ICU beds in the country, 1/16th what we need. The cost of those beds would be trillions of dollars. Of course we can't magically materialize ICU beds, so hundreds of thousands will die. I'd choose 6-12 months of the GFC over that _any day_.

Do you not believe these basic facts or just don't have empathy for other people who are at risk?

Sure but what happens if we do nothing for a month and let the virus spread rampant. Then we bankrupt the entire country paying for healthcare assistance packages for those who cannot afford it and are the hardest hit. Or would you recommend we just leave them sick on the streets?

Not sure why this is getting voted down, this is absolutely the case. Either you kill exponential growth via isolation/inoculation or damn near everyone gets sick in an exponential explosion and the healthcare system is overrun. This is simple statistics.

What's the reasoning here?

Connectivity will be the great equalizer in the future.

These words could have come from an article about: the internet, mobile phones, telephones, morse code

With minor modification also: automobiles, the printing press, the steel plow, airplanes, the cotton gin, antibiotics, etc.

None of them were true equalizers, they were incremental steps forward at best.

I see your point though POSIX imposes very few (if any) architecture decisions on application developers. The kind of design choices we’re talking about are very different from those of POSIX-like utilities so I’m not sure if that analogy is a good one.

You could achieve the same with basically any decent concurrency model on a single machine.

Why don't you just checkpoint the model every n steps? NNs fail for a myriad of reasons, you can easily reduce risk by routinely saving state.

Is "graph picture book" the new medium for conspiracy theory videos or something?

This does not answer my question.

The identifying characteristic of capital cost is: does not change over small time intervals, it is fixed up front and amortized over the lifetime of the asset.

This article and OP are talking about price spikes which are a symptom of short term (days, hours, minutes) market dynamics.

The reason that prices spike has nothing to do with capital cost and everything to do with short term demand/supply fluctuations.

As you can read above, I wasn't implying you were. Merely pointing out another, perhaps better, analogy.

I can't make sense of the sequence you've laid out but what do does capital cost have to do with daily (or even momentary) fluctuations in price?

Also, energy prices can be higher on average than elsewhere and dip to zero in some circumstances.

How about UPS/FedEx? They certainly test new products on public roads and have no requirement to make public the resulting data.

Maybe but then we’re no longer making an argument about performance, which is what I was responding to in your initial claim about “everything counts” and numpy shuffle being slow. That’s a straw man argument that has zero bearing on actual engineering decisions.

EDIT: clarification in first sentence

In what circumstance would one measure end to end time budget in training? What would that metric tell you? You don't care about latency, you care about throughput, which can be scaled nearly completely independently of the "wrapper language" for lack of a better term, in this case that's python.

It seems some commenters on this thread have not really thought through the lifecycle of a learned model and the tradeoffs existing frameworks exploit to make things fast _and_ easy to use. In training we care about throughput. Thats great because we can use a high level DSL to construct some graph that trains in a highly concurrent execution mode or on dedicated hardware. Using the high level DSL is what allows us to abstract away these details and still get good training throughput. Tradeoffs still bleed out of the abstraction (think batch size, network size, architecture etc have effect on how efficient certain hardware will be) but that is inevitable when you're moving from CPU to GPU to ASIC.

When you are done training and you want to use the model in a low latency environment you use a c/c++ binary to serve it. Latency matters there, so exploit the fact that you're no longer defining a model (no need for a fancy DSL) and just serve it from a very simple but highly optimized API.