Before sharing something that I am visualizing for the end of 2019, some context from the present would be helpful.
This is a link to a 57 minute Democracy Now! program, dated 7-27-2018, with Amy Goodman interviewing Noam Chomsky. https://www.youtube.com/watch?v=fByWYD3PCx4
This is a link to a 9 minute climate change and clean energy video blog about Arctic sea ice by Robert Scribbler (YouTube name of Robert Fanney), also dated 7-27-2018. https://www.youtube.com/watch?v=KyETiESTuyQ
By the end of 2019, we may begin to experience a bump of accelerated global warming as the superior economics and technology of wind and solar energy coupled with advanced batteries and ultracapacitors accelerate the shift from burning fossil fuels to clean, renewable energy. Burning fossil fuels, especially coal, not only puts carbon dioxide and other greenhouse gases into the atmosphere, it also puts particulates and aerosols into the atmosphere which reflect sunlight back into space to partly counteract the greenhouse effect. While CO2 can persist in the atmosphere for thousands of years, particulates and aerosols wash out of the atmosphere in less then two years. Losing that reflective cooling effect in the near term will result in some short term pain for the necessary long term gain of halting the growth in the concentration of greenhouse gases.
That isn’t really the main thing about the end of 2019 that I am visualizing. To put the main thing into context, we can look first at the recent past. Deepmind announced the Wavenet speech synthesis machine learning model in late 2016. One year later, Deepmind announced that they had improved the computational efficiency of that model by a factor of 1,000 so that a second of near-human sounding speech could be synthesized in a fraction of a second. Within a few months of that second announcement, the Wavenet speech synthesis model was incorporated into the latest versions of Google Assistant.
So, it is possible for machine learning software to progress from the first announcement of successful research to commercial deployment within a year and a half.
It is common for seventy to eighty papers describing the latest research in machine learning to be published per day at arxiv-sanity.com. Out of those papers, there are frequently at least four or five per day that are about research in which I am especially interested: continual learning, identity recognition, and multi-agent social choice. A critical piece of almost all of these papers is the ability to create symbolic representations within the model of things from the world such as our speech, our identities, and our political will.
If machine learning models can go from announcements in published research papers to commercial deployment within a year and a half, then some of the significant advances I have read about in the last few months could very well be incorporated into digital assistant software by the end of 2019. We should expect this to happen.
That doesn’t mean that machine intelligence will reach human level by the end of 2019. It doesn’t have to. It only needs to be sufficiently improved. Sufficient at continual learning for speech recognition, language translation, question answering, identity recognition, and global consensus decision making.
By the end of 2019.
July 27, 2018
Updated July 31, 2018
Openai.com released a 3 minute video titled "Learning Dexterity" presenting their success in training a robotic hand to manipulate blocks and other objects. The narrative is a clear and concise description of their methods. It is a remarkable data point on the curve of increasing machine intelligence, and a beautiful piece of engineering. Notice the personal space the roboticists have created for themselves next to their presentation area at the very end of the video.
Now imagine this robotic-hand machine learning model as one module among many which include speech recognition, language translation, question answering, driving, and discovery, and management of all of the other modules so that whatever functionality is needed at any location and moment, is brought forward in the computer's processing and memory. Imagine also that this machine intelligence is distributed with parts of it running on our personal devices as well as in the Internet cloud. It will be learning from each of us.
By the end of 2019, things will just start to get interesting.