In the fall of 2021, Blake Lemoine, a senior software engineer at Google, was assigned to help the company on a particular AI ethics effort. While working on it, he stumbled upon a tangentially related but separate AI issue. He flagged it to his manager and suggested him to escalate it to the leadership. But his claim was dismissed citing limited evidence.
Mr. Lemoine, turned to Margaret Mitchell, a former co-lead of Ethical AI at Google, for help. With Ms. Mitchell’s support, he gathered evidence and presented it to the leadership. After that, things got rocky for him. In a June 7, 2022 blog titled, ‘May be Fired Soon for Doing AI Ethics Work’, he noted that Google had placed him on “administrative paid leave” in connection with the AI concerns he raised. Two weeks later, the search giant fired Mr. Lemoine for violating the company’s data security policies.
The concerns that had cost Mr. Lemoine his job were centred around the Language Model for Dialogue Applications (LaMDA). Based on his interaction with LaMDA, he claimed the large language model (LLM) to be sentient. Mr. Lemoine was not the only Google AI researcher who raised questions about LaMDA. Other engineers who raised concerns about different aspects of Google AI also lost their jobs. And Google went ahead building its BARD AI service on this language model.
LaMDA, like BERT and GPT-3, is an LLM built on Transformer, a neural network architecture developed by Google and open-sourced in 2017. But Google’s LaMDA was different from the rest of the LLMs available at that time. It was trained on dialogues as input. That made it notice nuances and distinguish open-ended conversations from other forms of language.
Google’s Bard, launched to compete against OpenAI’s ChatGPT, is anything but poetic. In the launch blog, CEO Sundar Pichai did not describe the bot for its poetic or programme-writing skills. Instead, Mr. Pichai claimed Bard to be an outlet for creativity, and a launchpad for curiosity, “helping you to explain new discoveries from NASA’s James Webb Space Telescope (JWST) to a 9-year-old.” To show what Bard can do, he attached a GIF video in the same blog that had three suggested answers to the JWST question.
But the GIF video had inaccurate information. In one of its responses, Bard suggested that JWST was used to take the very first pictures of an exoplanet. However, those were taken by the European Southern Observatory’s Very Large Telescope (VLT) in 2004. NASA has confirmed this information.
This gaffe was reported by Reuters just before Google started to live-stream its presentation from Paris on Wednesday. To make things worse, Prabhakar Raghavan, a senior vice president at Google, in charge of Search, made an underwhelming presentation on what Bard can offer its users. He largely restated what was already known about the conversational AI service. His presentation also lacked any ground-breaking update that could intensify the competition between Bard and Microsoft-backed ChatGPT.
The incorrect answer on the promotional GIF and the underwhelming presentation in Paris had cost Google’s parents dearly. Alphabet Inc.’s stock tumbled nearly 7% on Wednesday, wiping out $100 billion from its market value, while Microsoft’s shares rose 3% on the same day. And just a day earlier, the Windows software maker announced it had released a version of Bing search with ChatGPT functionality.
At the heart of Bard’s unpoetic rendition is the LaMDA. Google’s own Ethical AI researchers have raised questions about it. But a part of that large language model carved out to provide a specific service has failed to enthuse investors and users. The answer to what is Bard’s true capability is somewhere between these two vantage points. And until Google makes the LLM publicly accessible, LaMDA and its carve-out Bard will remain an enigma.