_yosefk

TFA author. How does the word "dispatching" hint at the order of magnitude of anything? What orders of magnitude do depend on is your competition. Outdoing a CPU at operation X is easier than outdoing a GPU at X [assuming the GPU does X reasonably well] which is easier than outdoing a DSP at X [assuming the DSP does X reasonably well]. If your competition is reasonably optimized programmable accelerators, your opportunities to beat it start shrinking.

Source: my day job

A lot of people including "me" (TFA author, of course "me" = larger team) were doing hardware matrix multiplication for many years before 2012 [let's say before deep learning.] I count "moving data around" as "dispatching costs" [dispatching = figuring out what to do, on what data and where to put results as opposed to actually doing it.]

A group of 100 is 10 groups of 10. Does the autonomy of the group of 10 shrink as predicted by Paul Graham's "virtual person" argument? It might or it might not - and it's certainly not obvious that the constraints imposed on a 10-people company by their relationships with the external world are any better or worse, from any given angle, than the constraints imposed on a 10-people team by their relationships with the company employing the team.

Why do people contribute to projects like Linux or LLVM without getting paid for it? (Of course lots of them get paid to do this, but many are not.) It's not exactly easy calories for a caged lion, the way Graham describes a corporate job taken by recent college grads. Instead, people choose to work in these large groups because they want to make an impact. You can instead contribute to TinyCC or MenuetOS - smaller team, less impact. Are Linux or LLVM uninspiring? Were they only inspiring when they were too small for most of their current practical uses?

If productivity is concentrated in small companies - or companies of size X - why aren't companies of this size drive the other companies out of business? I think that market realities point in the direction of there being economies of scale and diseconomies of scale, without a single one-size-fits-all procedure for determining the optimal firm size for any particular endeavor.

If you work on/for a startup and that startup succeeds, it will more often than not have to grow. If you're unlike Paul Graham who sold his startup and then invested in countless others and made lots of money but did not keep working on the thing he'd built - meaning, if you're what he likes to call maker as opposed to primarily a money-maker - you will want to keep working on the thing you built. And this desire to keep working on that thing is what will prompt you to adapt your views on group size and how people thrive and lions and sugary food. And when you'll see what a hundred people can make out of what was started by ten, it will be very rewarding.

If a collar sits comfortably on a cat's neck, the cat will remove the collar. If a collar sits very tightly on a cat's neck, it's somewhat uncomfortable.

What if you're not home? It is for such occasions that cat flaps were invented. However, today's cat flaps either open for any cat, or they're RFID-based, and for that to work, you implant a chip into your cat, which I find utterly barbaric.

Now what could be a great improvement is a cat flap identifying your cat(s) and letting them in based on computer vision, instead of relying on implants. I'd pay good money for it, provided that it worked in all weather conditions and all cat conditions (a wet and dirty cat needs to get in even more than a dry and clean one but I expect false negatives in these scenarios.)

They don't show a comparison to bfloat16 PEs/FMA. IEEE half precision uses a larger mantissa than bfloat16, and the cost of multiplication is proportionate to the square of the mantissa size. I'd expect much lower gains relatively to bfloat16

What could he learn from the news that would influence his actions or even opinions? If the answer is "nothing", what's the downside for this guy?

Wow - an army guy, "the Father of the Nuclear Navy," who gets things about management that half the tech industry doesn't, instead using supposedly army-like command-and-control methods!

It's certainly the case that neural nets work flawlessly because they're like the brain - ask any journalist. Without doubt, an image of the brain enhances NN performance

(TFA author) If it's a really important patent for an individual inventor as opposed to another one to add to the company's portfolio mainly to grow its size, then of course you want to do prior art search. I wasn't talking about the individual inventor's situation as I'm not familiar with it; perhaps you could elaborate on your reasons why your feelings are mixed after the process. I was talking about the situation of a typical company's employee. And yes, it's an "easy legal defense" for a product company to forego prior art search, but it's not an urban legend, it's real legal advice I heard from real lawyers.

The same applies to the lawyer's experience. If you're an individual inventor, you're gonna look for the lawyer with the best experience and the law firm with the best price (and IMO you're right in that the $1000/hour lawyers aren't necessarily the best for you.) If you work for a company, someone will find a lawyer to deal with your stuff and the lawyer will behave very confidently, in a "business as usual" kind of manner, when you explain them the basics and they keep not getting it, and you might naturally assume that's how things are supposed to work, when instead what you should do is push back against whoever set you up with this lawyer and insist on getting one with relevant experience.

Not sure why this is being downvoted. I admit that I find it hard to see how to control the exfiltration of data once it's been figured out by a badly sandboxed program. But the idea that systematically addressing side channel attacks in a sandbox is really hard seems very valid. I mean, the exploit described in TFA is quite the argument in favor of this point.

What's the difference between France (lost quickly in WWII), Britain (barely evaded invasion), and the US (not invaded, economy grew during the war)? Is it democracy, free markets, or accessibility to German tanks and war planes that is the likely reason for the difference?

8 out of 10 Germans losses were inflicted by the Soviet Army; the USSR didn't rank high on democracy or market freedom but it had no choice but fight Germany, whereas the US could afford to choose the extent of engagement and let the USSR (and Britain) bleed while inflicting losses on Germany. This is not to say that the US policy wasn't sensible for the US, just that WWII is not a clear testament to the latent strengths of democracies and free markets at waging war.

And on the point of democracy specifically, in "The Gathering Storm," Churchill nearly pins the breakout of WWII on the inherent difficulty of conducting sensible foreign policy under democracy (both in the case of the badly designed peace treaties at the end of WWI, and in the case of subsequent appeasement of Nazi Germany.)

I wouldn't want to live where they don't have democracy or a reasonably free market, but neither is a reliable device to either prevent or win a world war.

I think grandparent's point was that the first cause you mention - reducing FB's power over users - does nothing for the users if this power instead goes to other 3rd parties, and perhaps even grows.

GP points out that in a distributed social network, 3rd parties can still mine your data, you still have trouble permanently erasing information, and in fact these problems grow instead of shrink. From a user's POV, what matters is the total amount of 3rd parties over them and the leverage they have against these 3rd parties as a whole. GP's point is that shrinking FB's power by going to a distributed social network might actually increase the total power 3rd parties have over users.

At my end, I wonder where the author ever saw deterministic computer systems. In my experience, you have to work really, really hard to get even a system computing outputs from inputs without processing any external events in-between to produce deterministic results, with things like floating point hardware and software and parallelism APIs, not to mention your own bugs, working overtime against you - to say nothing of faulty DRAMs and other such delightful phenomena.