Last week, Sergey Brin of Google gave a talk at the World Economic Forum describing how surprised he was by the great breakthroughs in deep learning and machine intelligence just in the past few years: Davos 2017 - Sergey Brin: The Future of AI and Google
Brin is probably much closer to objective reality and advanced technology than anyone else at his level of economic power, but to know what the real risks and opportunities are coming up there is no substitute for digging deep into the front lines – front lines like the major symposium last month, where many of the “Deep Mind” Google people whom Brin talks about got to speak about the next big wave:
If you go to the web page of that symposium (where I saw 2,000-3,000 people in the audience), you will see names and links to those technology leaders at Google and other major players – and you will also see my own name on top, because I originated a lot of this technology, and led the NSF actions which actually caused the great rediscovery (new to the computer science world) which happened just a few years ago.
The main theme of this symposium was that the deep learning which the whole world knows about is just now starting to hit a second great wave. A reporter asked me how I assess that wave, and the new startup of Juergen Schmidhuber who organized that symposium; the overall summary I sent him was:
Please forgive my delay; this is a pivotal moment for a huge technology, and I have often caused more harm than good by accidentally putting the wrong spin on this kind of thing.
The kind of deep learning which has received wide application so far -- starting from a successful grant to Ng and LeCun which I funded at NSF (see the 13 attached slides of my talk at the symposium) -- is just the first step in a huge stream of technology with risks and opportunities much greater than people imagine even in the IT field. In 1988, at one of the two big neural network conferences of that year, I remember Jasper Lupo of DARPA saying "this is bigger than the nuclear bomb"; I did not realize at the time just how precise his words were, but as I look at the next wave of what we can do (and what a few people have already done in industrial and military settings ) -- I do hope we will be careful as we let this huge new genie out of the bottle.
The kinds of neural networks -- robust recurrent networks -- which Schmidhuber is implementing are far more powerful than the earlier generation, and they are an entry to even more powerful systems, systems which do indeed have the potential to outsmart human brains. How safe is it to disseminate this powerful type of technology more widely? It is very important that we think hard about this question, in a way which society has not done right in the case of brain-computer interface (BCI), another important new technology.
With BCI and longevity technology, I often think we would have been better off just to let sleeping dogs lie. But here, the latest issue of Scientific American argues that we need to develop true artificial intelligence, fast, before we get locked into a new pattern of IT-based top-down control by corrupt and/or confused human political opportunists. Schmidhuber's company is firmly addressing the next big steps needed to get to such true artificial intelligence, but I worry what could happen if the wrong people use this technology in the wrong way. Still, we are really at a crossroads, as the Scientific American article argues, and to find our way to a safer path, we need to understand what we are doing; I hope that the attached slides, as well as Schmidhuber's company, can play a pivotal role in helping us all understand what we are doing better.
An industry guy at a later discussion asked for something on "what ARE those next big steps even after NNAISENSE?" To some extent, the slides and the links in the slides give some answer to that – above all, the link on the last slide, to a paper written for a NATO workshop, for futurists and decision-makers interested in the big picture. The most futuristic part of that paper, proposing some useful first steps towards a quantum level of intelligence, was initially written before certain leaks appeared on cybersecurity, and I worried what I could safely talk about; however, since I did write about that aspect, please note that my previous blog post on backwards time communication summarizes the outcome of some further work on “step two” of a path towards useful backwards time communication. Yes, the well-mapped road ahead goes very, very far now.