A Google Brain engineer’s guide to entering AI
80000hours.org
80000hours.org
I can't help but get the sense that we're trying to tell ourselves "evolve or die" and that the web jobs will move towards AI/ML if you want to stay employed in the 10-20 year range.
Whilst I have been involved in Android since the beginning I personally would not encourage younger folks to focus too much at the expense of new technologies (inc. blockchain tech for that matter).
That said, we hopefully have another 10 years!
But, I could totally see AI/ML driving voice interfaces in the future.
But no, I’ve been enjoying the experience more and more. Texting my girlfriend by voice. Asking Siri to find something on a map for me. First time I’ve been impressed by the pragmatic nature of it and not just finding it a clever toy (at least in some time).
I’m still wary of some of the implications, but I’m cautiously optimistic now more than I would have been.
That said, it’s certainly not my area of expertise so I’m also wary it spells my own career doom if I don’t get smart.
For certain industries - e.g. tech - people like quiet to concentrate. Not unique to tech of course... libraries are quiet too for the same reason. I cant imagine the average office where concentration is required benefiting from a call-centre-esque environment where everyone is talking out loud.
.. but I guess it might mean we can get away from open offices and back to private offices?! :-)
The other guy is the one who will never talk in public to a computer.
Finally, you have the guy who can’t imagine not using a keyboard to program.
For the rest of us, we have no problem talking to our computers, and we anxiously await the future to arrive.
So whilst ChatOps might be fairly basic today I am open to it becoming very much more advanced in the coming years.
From Wikipedia:
Computer chess engine Zor won the freestyle 2017 Ultimate Challenge tournament. The best human plus computer came in 3rd place. In 2017, chess computer engines are superior to human plus computers.
But, I think ML/AI is a bit different than that. ML seems like more of a backend thing or specialized application. How many people need to work on face authentication? Or object recognition? I’m just not sure how much AI/ML will be considered a general purpose programmer requirement. Interacting with models, sure... but not necessarily making them. In this respect maybe ML is similar to databases — we all use them, but don’t need to make them.
There will always be programmers writing at the interface between the user and the computer. Right now, those developers are heavily skewed towards web apps. But there is a sizable amount of mobile development too. If anything, I’d think the web jobs would move closer towards mobile as opposed to ML.
This is one of those situations where I greatly benefited from thinking of the problem from an ML perspective, as it allowed me to create a very simple probabilistic model and verify that it was sufficiently accurate.
I believe we'll get more of those "ML in the small" things popping up even for regular developers. It's not something new, it's just a mental perspective shift toward the probabilistic.
And this exactly why an enterprising software engineer would be well served to be familiar with how to do basic machine learning. You don't have to be great at it, you just need to know how to copy somebody else's notebook or whatever and get it to work on your simple data set.
"AI/ML engineer" is the new "web developer".
Soon everybody and their mother can make ML models and apply them. You may ask if it isn't already the case.
However, just as many people can hop in to web development and quickly create value (or at least CRUD apps), you still need really experienced people with a breadth of knowledge to push the field/work on really complicated or large scale systems.
It's the same for AI/ML/DL. Basically anybody can train an image classifier right now, but only a handful of teams on earth understand computer vision well enough to attempt to tackle self driving cars.
You should still learn some ML because it's cool and fun, though, and will give you valuable perspective.
The FHI/SIAI/MIRI people have been spinning their wheels for over a decade on this and made zero progress.
The current safety progress certainly isn't ideal (where ideal would be "oh hey we figured it out, boy that was easy"), but it is roughly in line with what I'd expect for a smallish new field trying to lay foundations. No one knows what to do; there may not be any other option than to flounder for a while.
It's akin to trying to come up with a firearm that can't kill innocent people. The problem isn't even coherent.
The only solutions are:
1: Redefine what AGI means, like openai has done.
2: Prevent AGI altogether
For instance, in your firearm example, we can't determine who is innocent and who isn't. But we can use embed a facial recognition device into the firearm that then cross references whoever the firearm is pointed at against a database of known non-innocents.
And then exclude any targets obviously under 18.
And then make the model probabilistic and add greater weight to targets that appear to be carrying weapons of their own or that are acting in an obviously dangerous manner.
You get the point. After enough time, you would have a firearm that is (arguably) better off than the one you started off with and so closely approximates the goal of your initial problem description that you don't care to make a distinction.
I could be wrong, but I imagine this is how AGI safety researchers think about their work to some degree.
That explicit reasoning approach falls squarely into GOFAI, symbolic reasoning, expert systems etc... so we're well past that at this point and know the problems with it.
Anyway, again, it's not a tractable problem. Something smarter than you is going to be able to eventually beat whatever restrictions you put on it. Might as well just be comfortable with it.
The symbolic reasoning you're talking about is going to make a comeback through probabilistic inductive programming, so don't dismiss it. It is necessary to attain sample efficiency and generalizability, which is virtually impossible with purely statistical approaches. I recommend you look into Josh Tenenbaum's work: http://science.sciencemag.org/content/331/6022/1279.
Rather, the AI community has really good confidence that you can "fool" every AI approach to date. That has been true the longest for GOFAI approaches, and now we're doing it for DL/RL.
How sure do we need to be that we're safe from evil AGI? Once it's in the wild, we don't get another chance to make it safe.
Once evil AGI arrives, that's quite possibly binary. Our only hope is good AGI outsmarting it and at that point we're playing the same game that produced us. We're not universally good. So we have to be more sure of it then anything, so we don't go down that road.
On this issue, I'd prefer the truly paranoid are in charge of safety. Not the gung-ho "of course it'll be nicer then us" crowd.
I don't think we can constrain it. Our best hope is to discover that it wouldn't care to harm us.
Makes sense from a theoretical viewpoint, b/c all AI/ML models are not Turing complete in order to be optimizable, whereas your average programmer easily churns out "models" in a Turing complete language. So, joe blow programmer is universally more powerful than any AI/ML algorithm. He can program an AI in his TC language, but AI cannot replicate his work in a non TC language.
- Universal transformers: https://arxiv.org/abs/1807.03819
These are just 2 of most well known.
It's actually Turing-complete for rational-valued weights (countable). If you are willing to grant real valued weights (uncountable) then it is actually more powerful than a Turing machine, able to compute things in P/poly, a very weird complexity class that includes some encodings of the halting problem. But this is just because you're "cheating" by allowing infinite precision which is not possible in the real world.
> Google, OpenAI, Facebook, Uber, or Microsoft.
none of these are academic institutions and none of them have authority to define a limit to research