1,100 karma · joined March 8, 2016
https://leblancfg.com
Honestly though, I would still consider an internship there in a instant, were I better connected with universities. The things they do consistently blow my mind, and they are one of the few startups out there that do.
We see links to the Halting Problem pop up at times when discussing compilers. Now, compilers try to interpret code in the most perfect way possible. But of course, finding the optimal solution is just the Halting Problem in disguise, and instead they opt for a more pragmatic solution. While none are perfect, some of them are also really, really good, though (http://lemire.me/blog/2016/05/23/the-surprising-cleverness-o...).
Now this is just an analogy, but I think the parallel can be drawn back to the SI scenario. We might not have SIs that never cause harm perfectly, but this does not preclude them from being really, really good at it. Because if you admit the definition above, you're also implying that it will efficiently infer/deduce its harm function.
I think the scariest part of AI security is when the program itself becomes unfathomable. By that I mean, we can't just look at the source code and go "Ah! There's your problem". Now, your paper assumes a static reward function, but we can imagine the benefits of an AI that could dynamically change its reward function, or even its own source code.
In fact, the most powerful tool I can think of to train a multi-purpose agent is through evolutionary methods, and genetic algorithms. Take for example the bigger ideas behind https://arxiv.org/abs/1606.02580 [Convolution by Evolution: Differentiable Pattern Producing Networks] and http://arxiv.org/abs/1302.4519 [A Genetic Algorithm for Power-Aware Virtual Machine Allocation in Private Cloud], and determining the fitness of agents by the global accuracy on a large number of broad ML tasks. But I digress...
Given enough computing power and time, these have the possibility of ending in an "outbreak-style" scenario. [This exercise is left to the reader]. And the way AI ideas and methods are so rapidly disseminated and readily available, it's safe to imagine that it could happen in a relatively short time span.
Here's my question: I know you're with Google Brain, but do you know if OpenAI is actively researching these avenues of "self-determined" agents? For their first security-related article, I was expecting security measures along the lines of: safety guidelines for AI researchers, containment and exclusion from the Internet, shutdown protocols for the Internet backbone, etc. I get the impression some of these issues might rear their ugly heads before our cleaning robots become cumbersome.
P.S. Looking at your CV, it's funny to see that you once interned at Environment Canada. I'm also working there presently, during which time I can perfect my knowledge in ML to eventually transition careers. Small world...
Edit: Grammar.
Now, the general rule of thumb for drugs is there's no such thing as a free lunch, what are the possible negative outcomes of supplementing those kinases and AMPK substrates? Because obviously their effects wouldn't be limited to influencing the neurotransmitter system exclusively.
I find the Ramanujan story actually reads more like a Mathematica advert, whereas the other more like a very self-involved eulogy.
>Whereas magic enables people to experience the impossible
Now you're just playing with the fact that "magic" has two meanings, even if the context is well understood. The sentence is clearly worded, and the only logical interpretation is akin to "video games enable people to experience the impossible", which I find totally valid.
So yes, this seems like a huge discovery. Pretty much all of our understanding of relativity is based on the fact that c is, well, immutable. Changing the way we understand c means that we'll need to completely change the way we grok the last century of theoretical physics.
From my point of view, what I see is that most large organizations haven't even started using standardized, one-stop-shop accessible data warehousing practices, and aren't headed to hit that target by 2025. What you find is strong data-driven processes in areas like shipping, warehousing and production -- these guys know what's going through their chain down to the individual proton. But in most cases I've seen, the service and enabler departments, be it Finance, HR, IT, etc. are way behind on that curve. This isn't due to the technology itself, of course, but because mindsets need to be changed. And that can in no way happen in 10 years.
That being said, I think we are seeing a very distinct technological acceleration ramping up in the next years, when Machine / Deep Learning takes off in full. The immediate conclusion is that it shouldn't be surprising for ALL 3rd-sector jobs to eventually be fully automated in the near future. Call it The Deep Wave.
Links: This Stanford School of Business panel, held in 2014, has the best 20-minute breakdown of the possibilities of Deep Learning. Given by Steve Jurvetson (investor in Tesla, SpaceX, Hotmail, Baidu, etc.), its scope is genius, and I think it's a must-see. https://www.youtube.com/watch?v=czLI3oLDe8M
Edit: clarification.
For me though, that headline just killed the story. Come on Gizmodo. Find someone else for the evolution assignments.
I know it's a deep subject to wrap your head around, but you just can't say that human gills evolved from shark limbs. What should have been squeezed into the headline is this: our ancestors' limbs evolved from what were once the gills of an ancient species -- that eventually evolved into a thousand other species, but who most closely resembles our modern sharks.
"New evidence suggests that human limbs evolved from our ancestors' gills" is just as evocative, and has the nice bonus of not being a blatant lie.
</flame>
Upon further reading, looks like you can achieve >99.9998% reflectance for a 10% wavelength band (in this case, in the infrared). [http://www.crystallinemirrors.com/wp-content/uploads/2013072...]
Still, you're right in saying that it's probably still quite a ways off.
The relativistic effect shifts the frequency by 22.47% for 0.2c.
Given a 445nm laser (the powerful blue ones), that's 545nm when redshifted (green). It's a bit of a jump, but that wavelength is still smack in the "solar atmospheric absorbtion spectrum window" -- you won't lose much of the energy in the atmosphere.
Reading this page diagonally [http://www.edmundoptics.com/resources/application-notes/opti...], I don't think that the idea of getting a >99.5% reflectance value for that bandwidth is that far off.
Mostly, I was educated in the Québec system, and had no idea before I was 16 that if I got good grades I could get into interesting universities -- even the American ones. Reading up voraciously on Richard Feynman, only to discover that if things had been different, I, too, could have gone to that school. By that time, it was too late, I simply didn't have the grades. My parents' perception that it would cost a fortune certainly didn't help.
Not that I'm complaining, I can't change the past. But the idea of never attending MIT is one of those very few, recurring what-ifs? that will sometime keep me up at night, half a lifetime after.
Oh, the interesting people! The interesting projects!
How I wish I had worked harder.
I mean sure, it might look like an affront on the work you've put in your site, but in the end, you've also found another group doing what you're also doing, probably just as dedicated as you are. Depending on the mutual language skills (maybe even consider using a telephone interpreter service https://en.wikipedia.org/wiki/Telephone_interpreting), you might be able to find common ground and work together towards a collaboration agreement or outright acquisition -- whatever it is, just make sure to have it in writing that your company came up with the concept and design. Win-win for both sides, and that contract is then legally binding.
Even more, if you're in talks with a very large Chinese company, having Chinese collaborators might actually work in your favor.
TL;DR: You might have more to gain by working on a collaboration deal first. Just never lose sight of the fact that, in the end, they were stealing.
In the immediate, the only thing that can happen, if at all, is for some people to lose their jobs.
I think he is hoping that if the leak is well covered enough by the media, it will be adding oil to the fire of public discontent. Perhaps in a way that would dislodge the current government.
Way I see it though, that's quite a long shot :)
I don't think anyone is assuming otherwise... all I see is statistics and sampling being argued. You'll have to endure my wall of text, though, sorry XD.
Let's take the planets again. We don't know the size of every single planet in the galaxy (i.e. the planet size histogram), so we need to make a guess. So we've been using Kepler's exoplanet observations, knowledge of planetary geology, etc. to fit a model, and it's well developed (http://exoplanetsdigest.com/2014/07/25/exoplanet-statistics-...). The size distribution is quite well understood, its bounds and modes are well defined. It's unlikely that our best guess of what the average size of exoplanets is will change drastically the more we know about planets. That is why the article's first conclusion is legitimate.
Now back to (sentient) species, and my original point. The definition of life in the SETI context has to be VERY wide. It has to encompass any scenario for a species that we might consider "intelligent" -- and not just little green men. It might have (some would say, inevitably) evolved beyond the biological, and still be considered sentient. After all, life on Earth has only existed for ~4By, compared to ~13By for the Milky Way. As a base for extrapolation, just think how different to humans extreme life on Earth is. (http://www.livescience.com/13377-extremophiles-world-weirdes...) In any case, there are bound to be some weird-as-shit species out there, whose composition still obeys the laws of physics. In other words, there is almost certainly many statistical modes of life out there that are, or can become, sentient.
With that in mind, there's a very good chance that the overall size distribution of (sentient) species does not match the one the author used (that of vertebrates only). In statistics-ese: if the distribution is multimodal, the average of our unimodal sample is not a good guess as to what the true average really is.
Maybe we can agree upon that?