Rainbird recently exhibited at AI Europe and had the pleasure of mingling with the great and the good in AI research and application.

Google say we’re moving from ‘mobile first’ to ‘AI first’. Microsoft says that Human-AI partnerships can solve society’s problems. IBM says cognitive computing will impact every decision in 5 years.

In 2016 4BN dollars was invested in AI. In all the high profile companies, there is a continuous pressure to innovate. Most large companies are exploring NLP, a few dabbling in voice, and perhaps a spattering in image recognition.

AI is still recognised as being hard as it generally attempts to perform complex tasks in complex environments. However, it is accepted as imperative for large companies to run AI projects and those with AI Labs are seen as having a competitive advantage.

Machine learning is well established with a growing success rate, but training sets for neural networks generally remain large requiring at least 10M + cases. The market recognises the lack of ‘explainability’ of black box solutions as a big problem, especially in compliant industries which has led to growing interest in the integration of machine learning and human cognitive models of reasoning. There is also recognition that open technologies can help bind multiple AI and data solutions together.

In the area of customer service, there is a consensus that enterprise is continuing to on-shore their business processing outsourcing arrangements as labour arbitrage continues to be replaced with labour automation. When a human gets involved in a customer interaction, it is generally accepted that the price goes up by 10x with a large percentage of that conversation being repetitive.

A blend of human / computer pairs, so called Centaurs continues to be a favoured model, blending machine consistency and access insight with human customer service, with the added ability for AI to triage issues towards the best human expert to deal with any specific problem. For business-based automation, “human-in-the-loop” remains the key get-out-of-jail card.

In banking, a number of banks are still seeking to build psychic models, capable of preempting customer needs. This fits into the growing trend of using AI in tailored product design where the smallest viable number is just one.
NLP continue to be hard, especially when dealing with sarcasm. Context remains king. For example, ‘killing’ is bad, ‘killing bacteria’ is good, ‘failing to kill bacteria’ is bad, ‘never failing to kill bacteria’ is good. Understanding the meaning of text continues to require knowledge of who produced it and who it is aimed at. Consider the sentence “Oil prices are going down!”. That might be bad for investors, but good for car owners.

What was the quote of the show? “People remember errors committed by AI, but forget human errors”. An interesting counterpoint to the usual “It’s no longer AI when it works”.