It’s almost 11 p.m. in Man Group‘s offices in London. The humans have all gone home.

But the computers are just waking up.

Man runs about $43 billion in assets through quantitative trading. Algorithms do most of the work, with people writing the code to build them and monitor for any anomalies after the fact. The machines are trading about 21.5 hours per day, from the open of the Asian markets to the close in the U.S.

It’s a strategy the firm has been utilizing for 30 years. And now, it’s seeking to be a pioneer in the next phase of quant: Machine learning.

The difference between machine learning and traditional quant is that with the newer technology, the computers do not need to be told what and how to trade — they find patterns themselves and put their own buy and sell orders on accordingly.

“We view it very much as the future; it’s where we’re spending more money and research than any other place today,” said Sandy Rattray, the chief investment officer of Man Group in an interview from the firm’s London offices. “We will see it steadily increase in all areas and environments where we are less dictating what the model should do and more letting the model learn from the data on its own.”

He said machine learning is already helping to deliver returns better than that of traditional quant. Rattray declined to specify the exact prevalence of machine learning in their funds today, but said he would not be surprised if in five years, half of their quant trading will be supported by machine learning.

It’s not just limited to quant. Man is applying machine learning to help discretionary managers interpret data better. One example is instructing computers to listen to multiple earnings calls at a time, whereas the individual portfolio manager can only listen to one.

“Can they interpret the earnings calls as well as a human can? No,” Rattray said of the machine-learning-enabled computers. “But they can do a lot more. You get a lot more breadth out of the computers than you do out of the humans.”

Man has partnered with the University of Oxford to study quantitative finance, and in the last year, carved out a new specialization to research machine learning and data science.

Of course, the potential for machine learning in global finance is tremendous, but it’s also started an engineering arms race, where some of the largest firms will have an advantage. A recent survey by Barclays found that 62 percent of hedge fund managers were employing some sort of machine learning capacity for the investment process.

While machine learning has garnered significant excitement among many investors in hedge funds, it’s not without its challenge. One of the key issues is that the computers cannot yet communicate why they make the trades they do. If a fund starts posting big losses as a result of what the machine has been trading, that will be a difficult conversation to have with investors, who want to know exactly why returns were down.

Another big hurdle is acquiring talent — coders to build out these very sophisticated programs. Historically, recruitment wars centered around compensation — a firm could often just increasing a prospect’s salary to encourage them to come. In the coding world, that’s quite different. Compensation doesn’t matter because they are seeking people who also think they could start their own company that becomes the next Google or Facebook and they’ll be a billionaire.

Ellis’s pitch is the speed and dynamism of finance relative to Silicon Valley.

“The fantastic thing about financial markets is that they move terribly quickly and they’re always different, so intellectually it’s a very interesting problem,” said Luke Ellis, the CEO of Man Group, which runs $96 billion in total assets across varying strategies. “We’re able to give them responsibility and let them run money into an algorithm very quickly compared to you know, you go work at Google and it’s years or decades before you’re making a big impact on the mothership.”

That pace also requires Ellis to constantly be focused on staying ahead of the competition.

“Being passive in terms of investment approach doesn’t work at all, people work out what you’re doing, they see your trades, they get ahead,” Ellis said. “So we have to do a lot of continuous work to improve algorithms to find alpha but also improve the trading we’re doing every day in order to make sure other quants can’t find our footprint in the market.