[translated by Google Translate]

With Alphafold, Deepmind revolutionized chemistry. The company is now working on competing with Chat-GPT, says the founder in an interview.

by Ruth Fulterer

If you wanted to nominate a polymath today, Demis Hassabis would be on the list of candidates. The 47-year-old is a candidate for the Nobel Prize in Chemistry thanks to his AI Alphafold, which predicts protein folding. And by then he had already had three careers: as a neuroscientist, as an inventor of computer games, and as a chess player. At 12, Demis Hassabis was the second best player of his age in the world.

Hassabis says today that his early chess career prepared him for the pressure he feels today. The head of the company rarely gives interviews. We meet him during the WEF in Davos when he accepts this year's Swiss AI Award. These are particularly difficult times for his company.

Why? His startup Deepmind, which specializes in artificial intelligence (AI), was bought years ago by the Google group Alphabet. Now it has become the focus of the Google group. Because one of the areas that Deepmind is researching is speech AI . And since the company OpenAI triggered a hype with the chatbot Chat-GPT in November, Google has been worried about its position of power. Now it's up to Hassabis and his employees to release Google's competitor product to Chat-GPT.

NZZ: What was your reaction when you tried Chat-GPT, the chatbot of your competitor OpenAI, for the first time?

Demis Hassabis: I was particularly surprised by the reaction of the users. I didn't expect this program to go so viral. I know such systems. Google and Deepmind are also researching speech AI. But we haven't released them for testing. OpenAI already. For the first time, many people are now experiencing directly the status of artificial intelligence.

Google and Deepmind have also not released their voice AI over security concerns. OpenAI dared to advance. And suddenly there's talk of Google search becoming obsolete . Are you now under pressure to find an answer quickly?

A few serious questions remain. Chat GPT sometimes hallucinates, gives incorrect answers. This is typical of speech AI. She can do amazing things but is not reliable. With a chatbot that is intended to entertain, this is not a problem. But what about medical advice? How many mistakes do you tolerate then? The problem with learning systems is that you cannot fully test them. Achieving a zero percent error rate is therefore going to be very difficult.

OpenAI relied on a lot of data for Chat-GPT and on reward systems that should steer the algorithm in the right direction during training. Will your competing product consist of the same components?

My research soul was a bit disappointed at how inelegant the solution to the challenge of voice AI was: simply the brute force of more computing power and data. But this is how you get the best result. So we build on that too.

Hassabis is not yet allowed to speak openly about it, but he does indicate that Google's competing chatbot should draw on the knowledge that Google has collected about information on the Internet. Google has created a huge database that contains sorted knowledge. That could complete the language AI. In addition, the system should be able to quote sources from Deepmind and Google.

When can we try this?

I can't tell you yet. We are currently considering what our product must be able to do in order to go public with it. We can maybe start with a beta version, then hopefully it will be ready soon.

With its program Alphago, Deepmind has defeated the world's best player in the board game Go. With Alphafold you have solved one of the great mysteries of chemistry: how to predict the 3D structure of protein molecules from the sequence of amino acids . That makes you a candidate for the Nobel Prize. Which of your inventions are you most proud of?

It's like choosing your favorite child! But if I have to choose, I say Alphafold. Because it is clearly the invention that brings the most benefit in science. It is already accelerating the research of diseases and medicines. The fact that we can use AI to solve a task that science has been trying to solve for decades shows how much potential it has.

Which areas in which Deepmind is currently researching do you find the most promising?

We recently published in the journals "Nature" and "Science" on the subject of nuclear fusion, more precisely how to control the plasma in fusion reactors - incidentally together with the EPFL here in Switzerland. We helped some great human mathematicians conjecture. We are also working on quantum chemistry. Scientific progress can be accelerated in all areas...

Read the full interview on the NZZ.com website using the link below.