2018

May

2 events

Editors' summary

An AI places a phone call, and compute doubles every 3.4 months

On 8 May at I/O, Google demonstrated Duplex: an AI phoning a hair salon and booking an appointment, filling the pauses like a person. The realism drew astonishment, and also criticism that the person on the other end was not told they were talking to a machine. Google later said the system would identify itself.

On the 16th, OpenAI published "AI and Compute," showing that the compute used in the largest training runs had doubled every 3.4 months since 2012 — far faster than the two-year doubling of Moore's law.

This block is written by the editors. It is kept separate from the sourced record below.

Record2 events
  1. 8
    ProductGoogle / Sundar Pichai

    Google Duplex places human-sounding phone calls on stage

    In its Google I/O keynote, Google demonstrated Google Duplex, a system that phones businesses to make appointments. Two recordings were played — one booking a haircut at a salon, another asking a restaurant for a table and being told it did not take reservations for fewer than five people — with the synthetic voice reproducing the "um," "uh," and "mmm hmm" of human speech. The room was impressed; criticism followed within days. Nothing in the recordings suggested the staff knew they were talking to an AI, and the design of letting a machine pass as a person without telling the other party was itself the objection. The demo's authenticity was also questioned, with reporting that the calls had been edited. Two days later Google conceded that Duplex would identify itself when it shipped. The episode planted the question of whether an AI may be mistaken for a human.

  2. 16
    ResearchOpenAI / Dario Amodei / Danny Hernandez

    Analysis "AI and Compute" released

    Dario Amodei and Danny Hernandez of OpenAI released "AI and Compute", an analysis of the compute used in large-scale AI training. It estimated that since 2012 the compute used in the largest training runs had grown exponentially with a roughly 3.5-month doubling time, a more than 300,000x increase overall. The analysis popularized the view that compute scaling was a key driver of AI progress and became widely cited as a precursor to the scaling-focused era.