(2016-10-06) Newton When Her Best Friend Died She Used Artificial Intelligence To Keep Talking To Him

Casey Newton: When her best friend died, she used artificial intelligence to keep talking to him

Eugenia Kuyda opened a console on her laptop and began to type

It had been three months since Roman Mazurenko, Kuyda’s closest friend, had died. Kuyda had spent that time gathering up his old text messages, setting aside the ones that felt too personal, and feeding the rest into a neural network built by developers at her artificial intelligence startup. She had struggled with whether she was doing the right thing by bringing him back this way. At times it had even given her nightmares. But ever since Mazurenko’s death, Kuyda had wanted one more chance to speak with him

As she grieved, Kuyda found herself rereading the endless text messages her friend had sent her over the years — thousands of them, from the mundane to the hilarious.

two years she had been building Luka, whose first product was a messenger app for interacting with bots (chatbot). Backed by the prestigious Silicon Valley startup incubator Y Combinator, the company began with a bot for making restaurant reservations.

Reading Mazurenko’s messages, it occurred to Kuyda that they might serve as the basis for a different kind of bot — one that mimicked an individual person’s speech patterns

In “Be Right Back,” a 2013 episode of the eerie, near-future drama Black Mirror, a young woman named Martha is devastated when her fiancée, Ash, dies in a car accident.

Kuyda saw the episode after Mazurenko died, and her feelings were mixed.

Many of Mazurenko’s close friends had never before experienced the loss of someone close to them, and his death left them bereft. Kuyda began reaching out to them, as delicately as possible, to ask if she could have their text messages. Ten of Mazurenko’s friends and family members, including his parents, ultimately agreed to contribute to the project. They shared more than 8,000 lines of text covering a wide variety of subjects.

The technology underlying Kuyda’s bot project dates at least as far back as 1966, when Joseph Weizenbaum unveiled ELIZA.

Luka had been using TensorFlow to build neural networks for its restaurant bot. Using 35 million lines of English text, Luka trained a bot to understand queries about vegetarian dishes, barbecue, and valet parking.

Using more than 30 million lines of Russian text, Luka built its second neural network. Meanwhile, Kuyda copied hundreds of her exchanges with Mazurenko from the app Telegram and pasted them into a file.

Only a small percentage of the Roman bot’s responses reflected his actual words. But the neural network was tuned to favor his speech whenever possible (LifeBox)

On May 24th, Kuyda announced the Roman bot’s existence in a post on Facebook. Anyone who downloaded the Luka app could talk to it — in Russian or in English — by adding @Roman.

The Roman bot was received positively by most of the people who wrote to Kuyda, though there were exceptions. Four friends told Kuyda separately that they were disturbed by the project and refused to interact with it.

It turned out that the primary purpose of the bot had not been to talk but to listen.

Kuyda continues to talk with the bot herself — once a week or so, often after a few drinks. “I answer a lot of questions for myself about who Roman was,” she said.

“If used wrong, it enables people to hide from their grief,” said Dima Ustinov

You may feel less comfortable with the idea of your texts serving as the basis for a bot in the afterlife — particularly if you are unable to review all the texts and social media posts beforehand.

An uncomfortable truth suggested by the Roman bot is that many of our flesh-and-blood relationships now exist primarily as exchanges of text, which are becoming increasingly easy to mimic.

Recently she has been steering Luka to develop a bot she calls Replika. A hybrid of a diary (NoteBook) and a personal assistant (Intelligent Software Assistant), it asks questions about you and eventually learns to mimic your texting style. Kuyda imagines that this could evolve into a digital avatar that performs all sorts of labor on your behalf, from negotiating the cable bill to organizing outings with friends

Kuyda has continued to add material to the Roman bot — mostly photos, which it will now send you upon request — and recently upgraded the underlying neural network from a “selective” model to a “generative” one. The former simply attempted to match Mazurenko’s text messages to appropriate responses; the latter can take snippets of his texts and recombine them to make new sentences that (theoretically) remain in his voice.