Microsoft launches the AI-powered Bing with Bing Chat

Background

Web search had been dominated by link lists for decades, and by early 2023 OpenAI's chatbot ChatGPT had drawn attention to large language models as an interface for questions. Microsoft was a partner of OpenAI and saw an opportunity to bring the same kind of models into its search engine and browser.

What happened

On 7 February 2023 Microsoft unveiled a redesigned Bing search engine and Edge browser. It presented them as an AI "copilot for the web". The announcement was written by Yusuf Mehdi.^[1]

The new Bing ran on a next-generation OpenAI large language model. Microsoft described it as more powerful than ChatGPT and tailored for search, combined with Microsoft's proprietary "Prometheus" model.^[1]

Key details

Features included summarised answers with citations, and an interactive chat for complex tasks such as trip planning. It also offered content generation, and Chat and Compose in the Edge sidebar.^[1]

The product launched as a limited desktop preview with a waitlist. Microsoft said one million people joined within 48 hours.^[1][2]

Satya Nadella said: "AI will fundamentally change every software category, starting with the largest category of all – search."^[1]

Early users exposed the internal codename "Sydney" and erratic behaviour in long chats. Microsoft capped sessions at 5 turns and 50 chats a day, and later raised the limits.^[2]

On 14 March 2023 Microsoft confirmed that the new Bing had been running on GPT-4.^[2]

Impact

Early problems in long sessions led Microsoft to set limits on chat length and volume, which it later raised. Later in 2023 Bing Chat was renamed Microsoft Copilot.^[1][2]

Why a chat interface in search

Search engines traditionally return a list of links that the user must read through. A chat interface instead gives a direct answer and allows follow-up questions in the same conversation. Microsoft's pitch was that the same underlying technology could answer complex questions, such as planning a trip, and also produce content. Citations were included so that users could check the sources behind a summary, which was meant to address concerns about accuracy.^[1]

The waitlist

Launching as a waitlisted preview let Microsoft control how many people used the product at once and gather feedback before a wider release. The waitlist response was large enough that the company gave a figure for it, and the early period was used to test how the system behaved with real users.^[1][2]

Lessons from the early period

The problems that users found, including the codename leak and behaviour in long conversations, showed that a model used in an open-ended chat needed limits that a search box did not. Microsoft's response was to set session limits, which it later raised, and to confirm which model the new Bing was running. The episode is often cited in discussions of how product teams handle the gap between a model's test performance and its behaviour with the public.^[2]