What does this mean in 2020?
In a speech Steve Jobs compared the computer to the “bicycle of the mind”. Jobs was referring to a study on the energy efficiency of locomotion of different species in an article in Scientific American. Human walking ranked pretty low in terms of energy efficiency relative to animals. Most efficient was the Condor. Add a bicycle and human efficiency became better than a Condor. Computers, Jobs said were like the “bicycles for the mind”.
That was Steve Jobs’ view of computers in 1980, forty years ago – What has changed and what can we learn?
Today like the 1980s when Jobs made the speech, computers continue to amplify human intelligence, but not with spreadsheets or databases. In 2020 AI amplifies human intellect. AI is now a complementary extension to the human mind. The rate of development of AI is astonishing. From 2015 to 2018, the funding for AI increased by more than 2-fold, about 50% more than other start-up sectors. In the U.S., this increase in AI funding increased by 4.5x in 4 years’ horizon.
These charts show exponential growth in the number of AI start-ups and amount of VC investment over the last decade (Knight, 2018). According to the AI annual report by Stanford, since 2010, the total funding increased from $1.3 billion to over $40 billion in 2018.
Accompanying funding growth, AI usage has surged, a 352% increase in AI-oriented open source solutions, and a 567% increase in downloads. In 2018, 47% of large companies report using AI in their business, in 2019 58%. Behind these trends is the meteoric improvement of AI performance, which has surpassed humans in many aspects. Figure 3 exhibits how AIs beat humans in test sets and validation set accuracies since 2015.
More technical measurements used by linguists, psychologists and other professions validate these improvements in AI across many fields: Image synthesis, semantic segmentation, activity recognition in videos, reasoning over multiple data types, language analysis, and computational capacity. For instance, in the area of language analysis, the Stanford Question Answering Dataset 1.1 (SQuAD 1.1) which contains over 100,000 reading comprehension question pairs, surpassed human performances in 2018, while the newly developed SQuAD 2.0 which included over 50,000 unanswerable questions, used to test that AIs could learn “Know What You Don’t Know”, beat humans in less than a year in 2019 (Figure 4). The General Language Understanding Evaluation (GLUE) benchmark also surpassed humans in the same year (Figure 5). Even the “Google Proof” Winograd challenges thought to be unsolvable by machine have been conquered by AI.
While the accuracy of AI steadily increased, the training time and cost required fell sharply. Figure 6 and 7 illustrate the drops in ImageNet training time and cost respectively. This trend will continue as AIs become increasingly autonomous. At RE5Q we routinely use AIs for labelling and many other AI training tasks e.g. word vectors auto-expansion for custom named entity recognition, judging and evolution, traditionally performed by humans. As S.S.Wilson observed about transportation vehicles, machines will become their own master, meeting new demands that they create themselves. The same will apply to AIs. In an article published in 2018 by the University of Cambridge research centre, it was said that the future of AI mandates that the technologies become independent of human supervision and learn to ‘think’ for themselves. According to Dr. Mateja Jamnik from the Department of Computer Science and Technology, the use of heuristic approaches, or the human-like practical and visual problem-solving approaches, will allow humans to see how an AI is “thinking.” This will allow AIs to explain their “thinking” in terms that the average person can understand. The autonomation of AI will lead us to the phase of Advanced General Intelligence (AGI), which is expected to arrive in the 2020s (UBS, 2020). UBS point to the fact that both Airbus and Boeing have designs for large passenger aircraft with small or no cabin for a human pilot. 2010 saw the first commercial passenger aircraft that could Taxi from the departure gate at Heathrow to the run way, take off and fly to Sydney Australia, land and taxi to the Arrival Gate with no human intervention.
The evolution of AI is reliant on 4 core driving forces:
- More Compute
- Better software and algorithms
- Improved Data and storage (both scale, organisation and availability of large datasets)
- Faster, Cheaper Networking
Of these Data is the least well understood. Data is the foundation for training and validation of AIs, making the quality and quantity of data a pivotal concern. It is forecast by 2025 the amount of data will grow into ten-fold of the 16.1 zettabytes in 2016 (IDC, 2019). The immense amount of data necessitates the enhancement of data rooms for AI: Large training sets can be accessed as needed. Ideally, these data room are fully digitised and structured to support machine learning, with AIs updating and curating content: No human labour required for the maintenance of the datasets. Delivery of improvements in these 4 areas is progressing and so machines continue to evolve.
As AIs evolve independent thinking, this will usher in an era when machines amplify other AIs just as they amplified human minds in the 1980s. In this singularity or AGI phase, machines do not have to work alongside humans. The role of AI will be interesting. Most agree that enhancing human lives would be a worthwhile goal, assuming we like the AI proposed solutions. To extend the metaphor of Steve Jobs, we may say that it will be time when AIs ride their own bicycles. The role of humans in this AI powered bike race is not clear: will we be rivals, passengers or pets?