What Ai Can't Do

Just finished an insightful little piece from our friends @ ElementAi, "What Ai Can't Do?"

Yoshua Bengio and Daniel Gross, provide a compelling, pragmatic and honest assessment of the Ai landscape, both the immediate potentials and limitations. So what can Ai do? And what can't it do?

Well, it's great at "perception", especially in a a the 2-d, theoretical world. Is this a cat? Is this a dog? Find the picture of the dog. Ai is exceptional at these tasks. But it is challenged with things such as spatial recognition (eg. 2 items cannot hold same space at same time), because, according to Yoshua Bengio:

“In order to make sense of language, you need to understand our world.”

To date, Ai has worked with very deep, but very limited datasets. limiting its understanding of the world beyond the task in front of it. Algorithms require excessive domain-specific-data, but real intelligence, real knowledge, comes from the varied experiences and exposures to the world humans have as we grow. Knowledge, innovation and creativity come from those experiences in one area of life that are transferred to another. They come from not just data but the context of the data, and correlations with independent data sets around other topics and experiences.

Humans are unique in our ability to draw these correlations. Think of the poet and musician who uses the loss of a lover to write a song? What of the painter who uses spray paint and chalk to invent a new art style, after watching a street artist do so? Think about how the Swiffer, a 2 billion dollar industry, was invented by a residential electrician who didn't work for P&G, but still suggested static electricity was a great way to capture dirt.

Ai can also understand and generate language but in a very simplistic format and far from human communication abilities. Human communications are a complex web of verbal, non- verbal, visual and sensory experiences. Ai is capable of binary language exchange, a stimulus such as a question to Alexa, and a response, based on a large data set that is associated to the topic. But the creation of new thoughts and independent thinking is far from Ai.

Finally, as Daniel Gross suggests, "an intelligent agent makes good decisions." Ai is capable of making algorithmic generated predictions based on its data set, and level of learning and neural net sophistication, but the final decision, with the needed cross experiences for knowledge, is still a very human function. To be an intelligent agent, a level of inherent “consciousness” is required, that ability to reason, to adapt thinking and learning as the environment changes and grows. The ability to shift the outcome by changing inputs, and backward engineer such insights, is still a very human role.

“Ai doesn’t have ability to think past what it is doing... the reality of Ai is it has nothing to do with what you see in movies.”

As Geoff Hinton discussed at the 2017 Rotman Machine Learning event: "we are far from either the utopian or apocalyptic type events we are debating." And in his, Yoshua's and Daniel's comments lies a stark warning, there is a danger in misunderstanding the potential and threats of Ai. Hype, for either the potential or threats, can limit the development of this technology by dividing our efforts, ironically creating the opportunities for malicious use of the Ai.

Human to human relations will still require well... humans. But Ai will be brilliant at augmenting human interactions with insights to things like sentimental and emotional understanding, helping people with empathy and targeted experience designs. What to do with those insights and how may be guided by Ai, but the choice and true understanding is very human.

For now.