How has AI changed over the last five years?
Within the gaming industry
One way to assess the evolution of AI is to trace the history of AI achievement through the ability of AI systems of succeeding complexity and their ability to confront and overcome games of succeeding complexity. Even back in 1947 there was an AI platform that could play noughts and crosses. No-one thinks of that being AI nowadays, but it was impressive at the time.
In 1997, Deep Blue beat Garry Kasparov at chess. But the thing that really put a marker down in data science occurred in 2016 when AlphaGo beat the Go world champion. This is something that data scientists had not expected to happen until the 2030s. The business (DeepMind) behind AlphaGo was immediately acquired by Alphabet for $500m. The rest is history; DeepMind is now one of the twin pillars of Alphabet’s global AI effort.
The past five years have been extraordinary. AlphaGo Zero is capable of not just playing one thing but doing more things; to achieve artificial general intelligence, AI has to be able to do lots of things, rather than just one thing. So, if you can play chess, can you play chequers? The answer is yes. AI can now play games without any instructions at all, i.e. here’s a chess board, but what’s the game? AI has defeated the world’s most successful chess engine.
AI can also successfully play computer games. The story is one of progression, even in hugely complex games.
Regulators have started to engage with AI – albeit much too late. How do you regulate autonomous weapons? China aims to be the pre-eminent AI power by 2030.
In the US we’ve seen Trump sanction a number of AI companies, on the grounds of human rights violations. We’re not convinced that Trump cares about human rights all that much, but he does care about the technological lead that China has over the US. The West has (belatedly) woken up and realised it is way behind in a number of key AI technologies.
Natural language processing (NLP)
Natural language processing (NLP) may sound boring but it allows computers to read, and indeed we use it in our own AI platform. NLP was in its infancy five years ago, but in 2018 OpenAI launched GPT. This had 150m parameters. A year later, Google launched its own NLP platform with more than twice the number of parameters. Baidu then launched its own NLP in 2019. OpenAI then launched GPT-2, with 1.25bn parameters – i.e. 10x more complex than the original GPT platform. Microsoft then developed an NLP with 17 billion parameters. OpenAI then hit back with GPT-3 – with 175 billion parameters. The leading NLP platform now has 530bn parameters.
The leaps forward in parameter density have been extraordinary and with each leap you can do a lot more than previously.
The best NLPs now comprehend language better than the average Stanford undergraduate. This is where we are today.
How do you define AI and what do you think about ‘sentient’ AI?
Everyone has their own definition of AI. You can see ours in our presentation.
Talking about sentience opens a can of worms. In 2018, there was an experiment with two NLPs in Alexa units, and they had a conversation with each other. When left overnight they had invested a new language to converse with each other. To us, this is more interesting than asserting that something is or is not sentiment.
We would regard an AI-enabled platform that uses robots to search collapsed buildings for signs of life as an example of AI. This is a role that would normally be done by a rescue dog. We would argue this is artificial intelligence as the robot can (safely) navigate a very dangerous situation, but others would say it is not AI as the robot is not operating at the same level of cognition or comprehension as a human.
It is important to have your own definition of AI – and to debate what it means. The same applies to sentience. The bottom line is, if something extraordinary happens, we should be prepared to talk about it.
In chess, we see that amongst humans, the average move quality has flatlined over time. Leela AI has enabled humans to understand how AI works, and now humans have become better players.
Learn more about Sanlam Global Artificial Intelligence Fund
Fund Risks
The Fund may invest in shares of companies listed on stock exchanges in the United Kingdom, and outside the United Kingdom, exchange rate fluctuations may cause the value of investments to go down as well as up. Investing in companies based in emerging markets may involve additional risks due to greater political, economic, regulatory risks, among other factors. The Fund may invest in derivatives for the purposes of efficient portfolio management and hedging.