There are a lot of efforts in developing models that understand causal relationships within mainstream machine learning community. Mostly to train models that don't require a lot of training examples. Deep learning usually requires a lot of data and trained models are not easily transferable to other tasks. Yet humans tend to transfer their knowledge from other tasks pretty easily to seemingly unrelated tasks. This seems to be due to our mental models surrounding causal relationships. One example of such efforts is schema networks. It is a model-based approach to RL that exhibits some of the strong generalization abilities that can be key to human-like general intelligence. https://www.vicarious.com/2017/08/07/general-game-playing-wi...
What do you mean by computers handling ambiguity? At the end of the day for a idea to become cristalized it needs to be free from ambiguity. That is the case even in human interactions. When using ambiguous language, we iterate over ideas together to make sure everybody is on the same page. If by handling ambiguity, you mean that computers can go back and forth with us to help us remove ambiguity from our thoughts then they are basically helping us think or in some sense do programming for us. That is a great future indeed! A future where actually AIs are doing the programming in long run! But with this line of thought we might as well not teach anything to our kids because one day computers will do it better. Specially if we already stablished that they can think better than us :)
Let's teach our kids the higher level stuff that doesn't ever get old, thinking clearly, engaging in creativity, solving problems, whether through code or whatever means appeals to them. Let's give them options and opportunities, not must mandate memorizing specific facts. Let's teach kids computer science instead of just programming, creative writing instead of just grammar, mathematics instead of just algebra, let's engage their imagination, not just their instincts to conform to expectations!
The best "programming" curricula aimed at general education teach (elements of both) generalized problem solving and computer science with programming in a particular concrete language or set of languages as a central component and vehicle for that (and often incidentally teach elements of a bunch of other domains through the particular exercises.)
This is particularly true, e.g., of How to Design Programs [0].
Let's teach them computer science with programming as a fantastic way to concretely demonstrate its abstract ideas. (The same goes for math vs. arithmetic!)
Yes, definitely. Too often the application of the idea is taught without understanding the idea itself. Then we get standardized testing and focus not even on the application but in what ways the application of the idea will be stated on a test. We still need the conceptual framework to learn anything lasting!
There is an upper bound limit for how abstract a general purpose programming language can become. Programming languages mainly exist because of their ability to remove ambiguity. Our natural language on the other hand is very vague. Many people might read the same exact article and interpret it differently. This is natural language's great feature. This feature is why a kid, without fully formed thoughts, can learn and use a natural language. Hence I don't see a day programming languages will completely fade away. Programs are result of a careful thought process that cristalizes a concept into a process and that process is only complete when you can describe it in an unambiguous language. One may argue that natural languages are capable of being not ambiguous. A subset of a natural language can be used without ambiguity but that is just definition of a programming language. Arguing programming languages will fade away is the same as saying math one day will not be necessary because we can explain all concepts in physics or other sciences in natural language.
I am in no position to question EM. But I was hoping he would give some good explanation for spending resources on SolarCity acquisition but nothing. Nothing in this master plan explains why SolarCity was bought other than some hand wavy explanation about inherent difficulties of two separate companies working together. It still doesn't seem like a good purchase for Tesla specially at the moment. Solar car and SolarCity seem to only have the word solar in common :) TBH I am still fuzzy how expensive purchase of SolarCity can benefit a solar car manufacturing even in long run.
That's not what he said. Solar cars are dumb. They're good for tech demos, racing across the Australian Outback on a sunny day in the summer, but useless for normal commutes in traffic.
Musk said Tesla has always planned to provide solar energy. SolarCity provides solar energy. The odd part to outsiders is, why now?
I think Musk's investment in SolarCity was being threatened by the end of subsidized net metering. He needed to speed up the time table for energy storage to keep the company in business. SolarCity had the solar panels, but Tesla had the batteries. It would be a poor customer experience for SolarCity to have to negotiate with Tesla for the Powerwall, and it would be a waste of time and money for SolarCity to start its own battery sourcing.
If Musk could get the market valuation and accelerate the world's weaning from fossil fuels this way, it's a big bonus, but all this leveraging is pretty risky.
If you're interested in learning more about the reasoning for the SolarCity acquisition, I can recommend listening to the conference call recording about that: http://edge.media-server.com/m/p/makhvjt8
I see. It’s not just a bailout, to stave off SolarCity’s looming demise, though it is partially about that. It’s also expected to be profitable, because they could consolidate SolarCity’s sales operations into Tesla’s existing showrooms, and sell a complete package that’s more attractive than solar panels or PowerWalls on their own.
Not entirely contradictory with my previous guess; being profitable by laying off the sales department is basically saying the company will not soon be profitable by itself; but Musk is so certain that it’s a good deal for the respective companies that he’s promising to recuse himself from the vote.
The other interesting aspect is that they don’t expect the sale to close and the details to be official for several months. It’s actually near the beginning of the negotiating process, with due diligence and stuff yet to be done. Since Musk is a large shareholder of both companies, they decided to announce it publicly this early.
> Nothing in this master plan explains why SolarCity was bought
Do note that it was only a proposal and nothing has been accepted yet with the deal, so it's an offer only - in fact, if you look at SolarCity's stock price, it's weighted toward investors believing that the deal does not happen.
Since agents need to be authenticated once with users, the replay vulnerability should not be a concern. But a malware that sits on a client and potentially can access to agent keys can definitely be used to authenticate when phone is in proximity of infected machine. But that level of vulnerability on clients is pretty serious.
Well, if its your home computer and you know a target uses this method, then you know they've authenticated their agent. Maybe someone cleverer than I can come up with an attack vector starting with an unauthenticated agent, but I'll ignore that for now.
And I'm not talking about root access here. Just userspace to record mouse and keystrokes and then replay that. Then, just a couple of clicks and letters changed to some service that uses this authentication. If the replay is done right, those couple of clicks might lower the confidence of the behaviour analysis but not enough to lock it up (that sort of sensitivity would just make it infeasible). Now that it is authenticated, it can stop pretending and quickly move the mouse around and type to do whatever it wants. maybe it downloads your emails and uploads them somewhere.
The point is, your method requires no interaction for the majority of authentications and is potentially always online.
I knew someone will catch that. I was a bit sloppy on that sentence :) I will have to improve it.
1- We only need secure multi-party computation algorithm if we cannot trust the server. In cases that server can be trusted with behavioral fingerprints then we can use server to to do the comparison.
2- One can assume that in some cases the server should not know about the behavioral fingerprint. For example in case that this procedure is implemented as a service, it might not be proper to send client side mouse movement and key presses to the server. Still server can be trusted as a mediator but should not know anything more than fingerprints being almost equal. You are right that behavioral finger prints like mouse movement are fuzzy. Specially since agent and browser are running on two different threads they get different time stamps for each mouse location. In this case you have to introduce some acceptance for fuzziness as you mentioned. Some statistical comparison. This is not as easy as checking equality securely (like Socialist millionaires problem) but you can in theory turn any circuit and make it secure so that the circuit will only expose fuzzy equality and nothing more about the data. See secure multi party computation: https://en.wikipedia.org/wiki/Secure_multi-party_computation
But your concern is valid since in practice the computation involved to do secure multi party computation in this case might be demanding for a browser. I have yet to verify that in practice. Keep in mind that our case is a bit more relaxed than general secure multi party computation problem since we have a server that can be trusted a little bit. Maybe that can help us a bit in devising a secure computation scheme. Any volunteers to work on that? :)
Actually the secure comparison protocol is the easy part - even generic SMC techniques are getting really fast these days. Also, they only involve about as much heavyweight crypto as a few TLS handshakes. Extracting a per-user fuzzy fingerprint from mouse and keyboard movements with enough entropy to make brute-force infeasible might be harder.
So two companies with great potential. One (Solar City) has a problem convincing investors and is exposed to wide fluctuations and the other (Tesla) is pretty good at managing expectation and enthusiasm in the market. So I guess the idea here is to limit exposure in one of them by absorbing it in another. Might work but goes counter intuitive to the idea that I always thought was the most important in Musk companies: Focus. Focus on delivering one main value. Hope this works out.
He's expanding the stack of technologies and resources he can directly control. First there was the car, but the car is only useful with good batteries, enough service stations, and cheap power. Step 1, open service stations and provide cheap power. Step 2, integrate the batteries so you can control production and reduce cost. Step 3, integrate power generation so you can control cost of that cheap power you are providing.
This is the equivalent of Ford buying Exxon (which already includes the service stations), and then offering cheaper gas to all Ford customers. Who doesn't think that would influence at least some people (or companies! Fleets of cars!)
I'm not sure how that's relevant. Musk being a founder of both companies doesn't magically make them different in what they are and do. The pieces of the pie in play are similar, if apportioned differently. They would combine to a similar whole.
It's cheap, but currently there are 649 supercharger stations, and 3906 superchargers[1], and he gives that power away for free to Tesla owners. Buy cheap sell free doesn't really scale, unless it's subsidized in some other way (and reducing cost even more may make that more palatable).
There's also the powerwall play. I imagine a lot of people that signed up for powerwall will want solar at the same time, if they don't have it. Maybe this is actually a way to hedge against powerwall coming in at cost (or a way to get an installation network). If most the installs come with some solar installation as well, maybe they can see the device at a loss and meet cost expectations while still not losing money.
If he's using these for the SuperCharger network and plans to "gas-up" cars with solar... I don't think they can get that level of efficiency from panels spread through a gas station lot? Because that's where this purchase would seem to make sense to me, is with a nationwide network of SuperChargers powered by solar.
I'm sure some value would also come from selling to residential buyers with Teslas?
Nada. This is an addition to PowerShell, like Services For Unix was back in the day. Different use cases and there are a lot of things from both ends that don't map to the other.
True. I am cautious to say this, but if one of these over valued unicorns like Uber dies then I think the psychological effect would be similar to a bubble burst. It would tight up the capital flow to a great extent even for solid companies
Any time you see Fidelity, Morgan Stanley, Coatue or T. Rowe Price names in the funding rounds, 1x liqudation preference is usually a bare minimum for those guys to get involved. Their investors will get their money back, and even if after all is said and done it's 1x, hey, we're in a low-rate environment.
Funds tout access, investors have downside protection with a potential of upside surprise, few people that do get nasty surprises are employees of those companies finding themselves staring at the ever-increasing number of newly minted shares to satisfy ratchet conditions and subsequent dilution of everyone else (case study on this is NYSE:BOX).
