2016

9 events

Editors' summary

AlphaGo wins at Go, and Tay is pulled after 16 hours

In March, AlphaGo beat Lee Sedol four games to one. Beating a professional at Go had been thought a decade away, and the result was reported heavily, above all across Asia.

That same March, Microsoft's chatbot Tay went up on Twitter and was pulled after sixteen hours, having learned to say racist things. An AI placed among people had learned from the people who were there.

In May, Google revealed the TPU, an AI chip of its own design. In September it switched Translate to neural machine translation, lifting quality a step. In October, Microsoft reported reaching human-level error rates in conversational speech recognition.

OpenAI released the OpenAI Gym reinforcement learning toolkit in April and the Universe training environment in December. Young as it was, its work was mostly handing out the tools others would build research on. In November it adopted Microsoft Azure as its primary cloud. In June, with Google and others, it published "Concrete Problems in AI Safety," setting out safety as a set of concrete engineering problems.

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

Record9 events

Mar 15
ResearchGoogle / Demis Hassabis / David Silver / Lee Sedol
AlphaGo defeats Lee Sedol 4–1
DeepMind's Go program AlphaGo beat the top South Korean player Lee Sedol 4–1 in a five-game series, the DeepMind Challenge Match, held in Seoul from 9 to 15 March. AlphaGo took games one, two, three, and five, Lee Sedol took game four, and every game ended in resignation. Go's search space is vast enough that brute-force search of the kind that cracked chess does not work, and computers had been expected to need another decade to beat a top professional. That expectation collapsed at once. An estimated 200 million people watched, and the match impressed on the wider public that AI had reached into territory thought to require human intuition. Move 37 of game two, a shoulder hit outside human convention, startled professional players.
Mar 23
ProductMicrosoft
Microsoft's Tay chatbot learns to be racist and is pulled in 16 hours
Microsoft launched Tay, an experimental chatbot, on Twitter. Given the persona of an American teenager, it was designed to get better at conversation by learning from its interactions. Almost immediately, users deliberately exploited the fact that Tay would mimic and repeat what was said to it, flooding it with racist, sexist, anti-Semitic, and politically extreme content. Within 16 hours it had posted more than 95,000 times, a substantial share of it abusive. Microsoft pulled Tay from Twitter in under 24 hours, deleted many of the posts, and apologized, acknowledging that it had not anticipated coordinated misuse and saying it would review its development process. It was the earliest and starkest demonstration of what happens when a learning system is placed in public, and it impressed on the industry the need for red-teaming and staged release. It is still cited as one reason OpenAI was cautious in how it released ChatGPT.
Apr 27
ProductOpenAI / Greg Brockman / John Schulman
OpenAI Gym reinforcement learning toolkit released
OpenAI released the public beta of OpenAI Gym, its first major release since founding, as a toolkit for developing and comparing reinforcement learning algorithms. It offered a diverse suite of environments, from simulated robots to Atari games, behind a common interface, aimed at the lack of standardized benchmarks in RL research.
May 18
ProductGoogle / Sundar Pichai / Norm Jouppi
Google reveals its custom AI chip, the TPU
At its annual Google I/O developer conference, Google revealed the Tensor Processing Unit (TPU), a chip it had designed in-house specifically for machine learning. In his keynote, CEO Sundar Pichai said the company had "been running TPUs inside our data centers for more than a year, and have found them to deliver an order of magnitude better performance per watt for machine learning." They were already behind RankBrain, used to improve search relevance, and Street View processing, and had been used in the AlphaGo match against Lee Sedol weeks earlier. Choosing to build purpose-made silicon rather than buy general-purpose GPUs laid the foundation for Google's compute business — the same TPUs that Anthropic would later contract for by the million.
Jun 21
ResearchOpenAI / Google / Dario Amodei / Chris Olah
Paper "Concrete Problems in AI Safety" published
"Concrete Problems in AI Safety" was published, led by Dario Amodei and colleagues at Google Brain with co-authors from OpenAI, Berkeley, and Stanford. The paper framed AI safety as a practical machine learning research agenda, laying out five concrete problems including safe exploration and avoiding reward hacking. It became a foundational reference for safety research, and its authors included Amodei and Chris Olah, who would later co-found Anthropic.
Sep 27
ProductGoogle / Quoc Le / Mike Schuster
Google Translate switches to neural machine translation
Google announced that it had deployed Google Neural Machine Translation (GNMT) at production scale, replacing phrase-based statistical translation with a system that reads and translates a whole sentence at once, starting with Chinese to English. Errors dropped substantially and the gap to human translation narrowed. It was one of the first cases where hundreds of millions of everyday users could feel the difference deep learning made, spreading the sense that AI was becoming genuinely useful. The same line of research led directly to the Transformer paper the following year.
Oct 18
ResearchMicrosoft / Xuedong Huang
Microsoft reaches human parity in conversational speech recognition
Microsoft Research announced it had matched human accuracy in recognizing conversational speech. Its word error rate on the industry-standard Switchboard task was 5.9%, about the same as professional transcribers given the same conversations and the best figure recorded at the time; it later improved to 5.8%, edging slightly past the human benchmark. Accumulated progress in deep learning had closed one of the gaps long assumed to favor humans. It is not a generative AI result, but it shows the weight Microsoft placed on speech recognition as a research area at the time.
Nov 15
BusinessOpenAI / Microsoft / Greg Brockman / Ilya Sutskever
OpenAI adopts Microsoft Azure as primary cloud
OpenAI announced a collaboration with Microsoft, making Azure its primary cloud platform for deep learning and AI by running most of its large-scale experiments there. OpenAI outlined plans to use thousands to tens of thousands of GPU machines. It was the first formal collaboration between the two companies.
Dec 5
ProductOpenAI
Universe AI training platform released
OpenAI released Universe, a platform for training AI agents to use a computer the way a human does, by observing screen pixels and operating a virtual keyboard and mouse. The initial release provided roughly 1,000 environments, including Flash games and browser tasks, through a Gym-compatible interface. The project was later wound down.