dojomouse

The main reasons they don't do this are that it's a fairly known quantity from an ML perspective (going from sequences of images to representational features), so wouldn't be proving that much to be able to do (c.f. the various Atari benchmarks which adequately learned actions to achieve rewards working with pixel inputs)... but at the same time would consume a huge fraction of the computer resource they really want to be targeting at the core timing/tactics/strategy problems... which is where they're really going beyond what's been demonstrated elsewhere with RL.

I agree it'll be even cooler when it all justworkstm end to end, but in terms of incremental 'holyshiticantbelievethatworked' this is at least as big a step as it will be when they add in direct visual input.

This seems basically the same as an ion engine without the need to BYO ions, relying on the existing ion/electron mix in plasma being sufficient and suitable. Which makes it seem fairly credible. Potential for much higher efficiency than ion engines too if a plentiful supply meant you could sacrifice thrust/ion.

Cathode Ray tube comparison below isn't entirely valid, as the goal of a CRT isn't to generate net thrust.

Finish working through 'AI - A modern approach'

Finish reading 'From AI to Zombies'

Get to a point with Rust that it's my go-to language for personal app development and general scripting. Super happy to see Rust feature so often on people's lists btw.

Get to a point with general Machine Learning understanding and proficiency that I feel I can usefully contribute to OpenAI work.

Build a retaining wall under my house, and have it still be retaining the things it's intended to retain come the end of the year...

++ this. The branches of maths that are relevant to ML are all pretty extensive, and you can do a huge amount of applied ML and understand the underlying theory while understanding only a fairly small subset of (for e.g.) probability theory.

Starting by learning the maths will mean you learn a lot of stuff which isn't directly relevant. Not the worst thing that could happen, but you'll be a hell of a lot more directed (even if you want to learn the theory - and I would recommend learning at least some) if you pick a decent ML course and learn the maths you bump into as you go.

http://cs231n.github.io/ is one of the best general hands-on introductions I've found. The TF tutorials are pretty good too if you just want to try some things out, but I predict that once you've worked your way through the TF tuts you'll still not really understand what's going on and will feel a bit like you just learned the magic words that made the black box dance some particular dances.

As for notation - https://en.wikipedia.org/wiki/List_of_mathematical_symbols is really good for when you stumble on something unfamiliar.

Good luck.

AI simply working to satisfy the 'benign' goals given to it by human programmers, without our having adequately considered all possible consequences (which, guaranteed, we won't), can be incredibly dangerous.

You should read some of Elizier Yudkowsky's writing on the topic, or Nick Bostrom's 'Superintelligence'. I think they articulate quite well the concerns that prompt many of the statements from Musk, Hawking, etc.

Can't tell if you excluded startups where the founder had a previous successful startup (solo or otherwise) but it doesn't look like it?

I think this would massively distort your results. Founders with a history of success are vastly more likely to succeed in a new venture for numerous reasons, but most people who are interested in these stats (both founders and investors) are interested in the pre-big-win founder case.

Also, I suspect far more 'startups' have a single founder, purely because it's so incredibly easy to 'start' something alone. Having a cofounder implies a certain level of organization and commitment - it doesn't really work with that pet project you dabble with occasionally. If solo dabblers far outnumber 'team' dabblers then you'd expect them to be over represented in successes even if their success rate was much lower.

| Graham is arguing against an enormous straw man.

Word. I don't know anyone who understands enough to actually have a discussion (e.g. isn't some back to nature hippy) and seriously thinks we should attempt to minimise the opportunities to create new wealth. Everyone I know who argues for reduced economic inequality favours increased minimum wages and/or redistributive mechanisms (income/wealth tax). I don't know if Graham has encountered people who seriously argue against permitting opportunities to create wealth, or just thinks those people exist, or has some sort of anti-redistribution agenda (which would function to reduce the potential gains that could accrue to startup founders... though not significantly).

I was particularly struck by this suggestion:

"And if there are people getting rich by tricking consumers or lobbying the government for anti-competitive regulations or tax loopholes, then let's stop them. Not because it's causing economic inequality, but because it's stealing."

Stealing is an extremely imprecise and emotive word, and he makes no effort to clarify. Which is usually what you do when you want a lot of people to feel like they agree with your arguments even when they might not if you went into detail. But the weirder part of the statement (ignoring the fairly broad 'tricking consumers', which would include most people who market their products) is that he specifically identifies tax avoidance (which I'd argue ISN'T stealing) as something we should eliminate... "not because it's causing economic inequality, but because it's stealing", even though the main reason we have tax AT ALL (and hence laws around when you do/don't need to pay it) is to reduce economic inequality!

So... all our existing tax laws, and their economic-inequality-reducing redistributive effects, are wise and good and infractions should be rigorously prosecuted because they'd amount to stealing. But any NEW tax laws that might increase the redistribution of wealth to increase economic-equality would be bad and misguided? Because where we are, right now, is the perfect optimum, despite the widespread evidence of increased concentration of wealth and its negative consequences? Riiiight.

Finally, last I checked there are various areas which have lower tax than the USA. Yet somehow plenty of startups are still formed in the USA. Which casts doubt on the 'the startups will move elsewhere if you change anything to reduce their direct rewards' argument.

I love it, but it let this go completely uncensored:

"This transformative change in paradigm allows us to move beyond conventional thinking and kickstart the data driven era."

Gonna go ahead and flag that as a bug :-)

Uber's brand doesn't seem very strong to me ("Oh yeah! Uber! They're the ones who deny any responsibility for accidents and attempted kidnappings of users of their service, right?"), and in any case it's only going to be significant for the tiny fraction of the travelling population who currently use Uber. The vast majority (for whom the human chauffeur model is too expensive) will have no brand loyalty at all, and are there for the taking. Also, for Google, getting the necessary visibility is basically effortless. They have Maps (which Uber integrated their hail system into for a reason), and they have incredible robot cars that will be featured everywhere when they appear anywhere.

As for the trip data:

a) I don't think this is either worth as much or is as hard to get as Uber would like their investors to think. You only really need average bulk commuter flows in order to get a near optimal fleet distribution, and this is surprisingly easy to get surprisingly quickly, and especially easy to manage when you don't need to worry about the vehicle distribution constraints human drivers create.

b) Google have this data in spades through Android device tracking. It doesn't matter that it's anonymized and aggregated, because individual trips don't matter; only having your cars in roughly the right place to respond to specific demands.

I think Uber have grossly misrepresented the extent to which this market (personal transport using self driving vehicles) will be 'winner take most', and equally grossly misrepresented the extent to which their current position creates any sort of moat guaranteeing them the 'win'.

So much criticism of Google! My question: Why on earth would Google want to cede any significant chunk of the personal transport value stream to Uber - or acquire them at a $40Bil+ valuation - when the main (only?) semi-uniquely valuable thing Uber have (a large network of drivers) is worth approximately zero dollars in a self-driving vehicle scenario. Uber are valuable and effective in a human-driven vehicle model because they have achieved critical mass in the pool of human drivers on their network, and are in a position to grow that pool. In a self driving vehicle model critical mass of vehicles/drivers is available to anyone with a decent line of credit.

Exactly not quite what they sound like.

Interesting implications (which I'm sure people who've been thinking about it for longer than me have already considered) for the development and evolution of Friendly Superintelligent AI.

The way this is presented, personhood (which involves behaviours that we'd like a friendly AI to display) is an adaptation based on mutual benefit. We're not nice to other people because kindness and empathy is an unavoidable consequence of intelligence - rather we're nice because we're better off in a society in which we're known to be nice. And that's because we've got fairly limited personal capabilities so are heavily dependent on efficient cooperation with other people.

An intelligent entity without this constraint (e.g. a superintelligence that had access to advanced nanotech and could do pretty much whatever it wanted) w.r.t interactions with humans might not (probably wouldn't?) have any evolutionary pressure towards achieving and maintaining personhood. Which would mean we shouldn't take the approach of "Let's just make an advanced intelligence, because its intelligence will surely mean it behaves with empathy towards us".

We're building a similar service at Teleport, and recently ran a few numbers on the benefits of this sort of trip sharing assuming large scale and self driving EV's. Really can change the entire transport experience in cities - cleaner, faster, safer, cheaper. http://teleprt.com/50-cities/

I looked at it in some detail but they were a pain to assess. Wouldn't give me any details of the API or feature set unless I created an account... for a non-refundable $200 fee. I pointed out how ridiculous this was, and they said that if I went through their signup process and committed to paying the $200, they would let me look at the documentation without actually paying the fee. It was all so vague and ad-hoc I found it pretty hard to trust whatever might be behind the curtain.

(edit: Looks like fee is now $99, and there's a little more detail on their developers page.)

PX pricing is actually pretty good if you've got high transaction volume. But it's terrible for small biz. Most people I know default to Paypal.

Please Stripe, please come to NZ fast. You've been talking about it for more than a year. I know we must look like smaller-poorer Australia, but I promise there are people here who'd use you!

It depends heavily on the level of refurbishment required between flights; agreed. I don't think there's enough examples to say "typically". The shuttle didn't come close to targets, but SpaceX seem to be progressing pretty well.

At the very least, a new orbital launch venture which is existentially dependent on SpaceX failing to achieve something that they're very well on track to achieving seems pretty commercially risky... and that's in the already highly commercially risky domain of space launch in general :-)

I hope the roadmap for these guys includes controlled descent and landing for the rockets - otherwise they haven't got a snowballs chance against SpaceX. The aerospike dependence on aerodynamics implies it will be much less effective when running in reverse for the landing phase - better have a way around this. When you look at the breakdown of fuel cost vs vehicle cost, it's obvious that reusable rockets will win so resoundingly that everything else will probably fall by the wayside.

Oh, of course. Nanomachines.

So basically we should celebrate this contemporary not-really-an-innovation because some future speculative innovation might make it useful... and ignore the fact that it's much more likely that the future speculative innovation would make it entirely redundant. I love it that your future sees us pulling CO2 directly from the atmosphere with nanomachines, but still driving around in cars that run on blocks of aluminium taken from aluminium oxide through smelting.

If I use your logic then I can just say "There's no issue with everyone running their cars on diesel". After all, one day we'll have nanomachines which can pull the soot, sulphur dioxide, CO2 and water vapour from the atmosphere and convert it all back into diesel again for a squeaky clean closed loop system running off the clean energy technology I have also implemented at massive scale.

I don't know why we're even bothering with clean diesel. Or any pollution controls at all for that matter. Nanomachines will sort that right out!

How're they coming along by the way? Oh, they're not ready? In the meantime then we should probably stick with assessing merit based on near term available technology, rather than saying "this thing is great, because I will solve all the things you don't like about it with my nanotech magic wand which I'm sure I'll have access to on a suitable timeframe".

Not at all. I'm pointing out that whatever process you assume to get carbon out of the atmosphere, you should plug the inputs and outputs of that process into your calculation.

Speaking of circular processes though... does this one look familiar?

a) Angrily and dismissively state flawed views

b) Have flaws in views pointed out

c) goto a)

Of course it's not purely circular as with every iteration there's more anger and less substance. But whatever.

So, they respond to numerous fairly well founded accusations that their 'idea' is laughably lacking in solid analysis with... a cherry picked selection of accusations and rebuttals which are also laughably lacking solid analysis.

My favorite part is where they used, as a defence to accusations that their idea won't fly commercially: "We are still in R & D, and we haven't even calculated the cost for our prototype". Riiiiight. They haven't even calculated the cost of their prototype, but you can totally trust their claims that the product will pay for itself. I wonder what assumptions they've made about future energy prices. My guess - none at all.

Uhh... I disagree. Firstly, I don't see the circle in the following:

- SolutionX puts lots of CO2 into the atmosphere compared to alternatives, and would be extremely uneconomic if we included in the cost/unitSolutionX the cost required to extract that CO2 from the atmosphere again.

- Lots of CO2 in the atmosphere is bad

-> SolutionX is either bad because it's uneconomic, or bad because it leads to lots of CO2 in the atmosphere.

-> We shouldn't use SolutionX until some factors change such that SolutionX is better than alternatives, rather than worse.

That's linear reasoning. If you want to change one of the underlying assumptions, namely that it's expensive to capture CO2 (and convert it back to carbon in this case... where at least the CO2 is coming from a concentrated point which reduces the cost of the process) then it's still linear reasoning with a slightly different result.

For it to become circular there would have to be some step where we said "If we had SolutionX it would be cheap and easy to get CO2 from the atmosphere and turn it back into carbon, but we don't have SolutionX because the methods of getting CO2 from the atmosphere without SolutionX are expensive so SolutionX is dirty".

That, interestingly, was used and is used as an argument against solar cells (e.g. solar cells made on power grids consume energy that may have a large CO2 loading). But it's easily revealed as bullshit through the use of EROEI (Energy returned on energy invested) accounting, and it's easy to show how 'SolutionX' (solar in this case) makes itself clean over time.

How is that the case with this aluminium fuel cell concept? Any advances in energy generation and CO2 capture are probably going to happen independently of aluminium fuel cell production. There's certainly plenty of incentive for both already, the aluminium fuel cells won't create a new market.

In conclusion, you're right that an aluminium based energy storage cycle would be more attractive if we had vast quantities of cheap and clean energy and low emission method of capturing CO2 and turning it into carbon. But.... 1. We have neither of those things 2. So for now an aluminium energy storage cycle is probably unattractive for most applications 3. If we did have both of those things it would totally change the economics of all sorts of other solutions, so you'd need to recalculate those as well before doing any comparisons between aluminium fuel cells and (e.g.) synthetic gasoline.

It's more like saying 'if you're looking at a new solution you better account for every step of the process, including costs and losses, and compare to viable alternatives, before deciding it's a good thing.

The logic isn't circular. Rather, you're conflating multiple independent processes. Even if we did have a commercially viable way to capture atmospheric co2 and turn it back into carbon that still doesn't mean we should then burn the carbon to make batteries. We could also bury it, and use normal rechargeable batteries.

It's highly recyclable. Aluminium oxide (the result of the reaction in the battery) is one of the feedstocks for industrial aluminium smelting. Recyclable isn't necessarily the same as 'efficient' or 'a good idea' though. It's energy intensive (haven't checked the efficiency of 'recharging'), and additionally has, as a byproduct, massive amounts of CO2 even before considering CO2 emitted in generating the electricity required for the smelting process.

I can't see how that a single use battery like this (which also needs swap infrastructure) has any net advantage over the rechargable swappable packs + fast chargers as Tesla use.

Maybe it could have applications in aerospace where existing lithium ion cells don't have sufficient energy density to even come close to a viable solution. But you'd want to compare it to other 'single use' options like synfuel.

Really enjoyed the read, but I don't understand the line about going from 500km/h - 1000km/h in 5 seconds... treetop level or otherwise. Is that really possible? I think my math is off if so.

Mass with a 900kg fuel load looks to be around 12,000kg Max thrust is 162kN for both with afterburner.

500km/h = 139m/s 1000km/h = 278m/s Going from one to the other in 5s means an average acceleration of 27.8m/s^2... or almost 3 g, which is monstrous!

But how does this work? Solving F=ma for a gives a=F/m = 162000/12000 = 13.5m/s^2

And that's before accounting for the not insignificant thrust required just to overcome the aerodynamic drag.

Is it just an understandable overestimation of the acceleration due to the direct experience of it? Or am I missing something?

Either way, awesome.

Sorry, I didn't word that very clearly - yes it's a liability, but it's a liability that's fairly independent of the levitation mechanism. I think it's more clear to say "That's the whole point of evacuated tube transport, to make an artificial...". Whether you use wheels or maglev or air bearings is a separate issue. Hyperloop doesn't turn the atmospheric drag into an asset - it just uses a particular method of overcoming the large liability. Because of how it overcomes it, it might be able to get away with a higher blockage ratio, but to say that with any real confidence you need to specify the vacuum level in the high-vac example you're comparing it to. And then do a bunch of compressible flow math that I've not yet seen for a high-vac ETT. This paper (http://jmt.swjtu.edu.cn/EN/abstract/abstract8587.shtml) looks at the incompressible flow case... but that's kind of missing the point I think.

The atmosphere isn't a liability for Maglevs, you could make hyperloop with maglev tech and it'd still work fine.

The tracks don't need to be powered for most Maglev concepts; superconducting maglev doesn't use powered track. Superconducting maglev is also energetically extremely cheap from an operating perspective - it's arguably energetically expensive during construction because you need a shitload of permanent magnet material (the entire track surface is coated with it). The superconductors are the cheap part from a system perspective. Electromagnetic suspension - like with shanghai maglev - also doesn't need a powered track, the relatively small amount of power required can be supplied by onboard batteries and periodically refreshed. Inductrack can be combined with magnetic propulsion, it's just not a form of magnetic propulsion in itself. There's no obstacle to integrating a separate propulsion stage into the track (as is proposed in the inductrack based container movement system proposed for... LA?)

There are a couple of differences between this and Hyperloop.

1. The hyperloop white paper proposed air bearings as the levitation mechanism, rather than maglev. This could be swapped out without much change to the overall concept (and, in my opinion, should be if it's ever built - maglev gives much more generous tolerances and ease of control).

2. The hyperloop anticipated a low vacuum; around 1/1000 atmospheres. While that sounds like not much air, it's still quite a lot when you want to go around the speed of sound, especially if you want to do so in a tube; pressure builds up in front of the vehicle and causes huge drag. The hyperloop addressed this with an integrated compressor stage in the front of the vehicle; basically sucking in air to avoid the drag. This used quite a bit of power, required cooling, etc... and the resulting drag was still significant to the point that regular medium power 'boosters' would be required along the track to keep the train at a reasonable speed. It also meant the speed was still limited, despite the partial vacuum, to around 1000kph (aircraft speeds). If you want to get away without the compressor you need a much higher vacuum - I don't know exactly where the equation changes but the number usually thrown around is about 1/100,000 atmospheres to 1/1,000,000 atmospheres. This is basically hard vacuum. If you can achieve it then it means no compressor, no significant onboard power source, and no meaningful limit to the top speed (you could go 2000 - 3000 km/h easy if you could make the track straight enough). The bad part is that, all else being equal, it's roughly 100 times harder to make a vacuum of 1/100,000 atm than a vacuum of 1/1000 atm. So... if you're really confident you can build extremely well sealed tubes then you should go for the hard vacuum approach. If you're confident you can't, you should go for the low vacuum compressor approach. If you're not sure, you should build a prototype :-)

The limitations regarding straight lines and the safety issues are roughly identical for the two concepts. The hyperloop is arguably less safe due to having a big spinning compressor right in front of you and a large onboard power source, high pressure steam storage etc.

Yes, so you could write

r + (amount that upward volatility helps) - (amount that downward volatility hurts) > g

No, volatility can increase as well as decrease returns.

I think you've reversed r and g vs ops notation, but yes, agreed. R can exceed g with no trouble, right up to the point where all economic growth is directly absorbed by capital accumulation (or whatever terminology you want to use for piping an annual measure into a cumulative one) at which time g must equal r... a fact of small comfort to the millions of people who, empirically, are getting a tiny slice of pie to live off. At this point volatility may well rear up and change the returns to owners of capital in the form of a revolution (as has happened again and again and again), but that's hardly an ideal form of society. You'd think if we know the mechanism and the result we could implement a fix. On the other hand, if you'd been paying any attention at all to climate change mitigation actions you'd probably be unsurprised that we haven't.

This seems to be what I need. What map data are you using?