It is not necessary to be a specialized engineer in artificial intelligence or a head of thought from Silicon Valley to develop AI tools. In this way, the prayer ready to smile and one could say that the AI has already surpassed its masters. It is not yet. But it is quite rare, behind notebooklm, the service of the IA developed by Google to help the collection of documents and data, is a passionate writer of technology.
Steven Johnson definitely does not have the profile that we expect and that his meeting with Google teams would be unlikely. In the summer of 2022, this author of a good dozen books was approached by Clay Bavor, then head of Google Labs, and Josh Woodward. They had impressed an article written in the New York Times magazine about “the potential of language models as a significant change for software.” The two men offer him to come to work in part “to develop a new AI -based research tool.”
“I received a Clay email saying: ‘You don’t know me, but I would really like to chat with you. We have a small team, some engineers, a designer and laboratories dedicated to creating prototypes.’ It seemed to be a great idea,” Steven Johnson laughs at Tech & Co.
A tool that understands what is working
From his inexperience as an engineer in the midst of engineers, he finally attracts a force that brings a lot to the project: to make a research tool capable of supporting the user and especially to understand it. “We not only wanted the user to discuss with an AI on the basis of the general knowledge of you. We wanted to be able to say: ‘Here are the documents on which I work. Here are my research project, my business plan and an overview of my competitors,” he summarizes. “And the model would respond or begin to interact according to shared information, and not just their type of knowledge.”
Observing that the brightest models of currently have enormous knowledge, but do not understand the context of demand or its participation when they have to administer monumental amounts of information, documents of hundreds of pages, images, interviews to listen, give life to Notebooklm with the mission of supporting the user.
For this, the tool must be able to summarize the investigation, support the analysis and verify the elements, answer questions too. And this, whatever the type of formats (PDF, audio, Internet links, YouTube, etc.). Hours of work won and a “remarkable improvement in research, writing and creative process,” Steven Johnson progresses.
Truly launched in May 2024, Notebooklm has an operation that adapts to researchers, journalists, writers, students or academics. Anything based on information search and requires establishing links between data, structuring or thinking notes to draw chronologies, guides, articles, presentations, etc.
A tool that creates more real thorough podcasts than life
But where the tool turns out to be even more impressive, it is its ability to create audio summaries from sometimes gigantic documents. Thanks to the arrival of Gemini 1.5 and today Gemini 2.0, Notebook You can transform the sources integrated into the podcast type conversation to synthesize everything to listen to it anywhere.
“It is a powerful way to learn and remember the information listening to two people who argue the subject,” Steven Johnson excites. And all this in just a few minutes of design, whatever the original language of the documents. But where Notebooklm is highly efficient, it is the ability to interrupt the two “virtual interlocutors” to ask questions through the menu and request additional information. Then they adapt their discussion.
Until now, representation in the Podcast format has not been accessible in English. As of April 29, Notebooklm offers audio in French. On the other hand, it will not be possible to interrupt the podcast to evolve.
The Frenchman took to reach “veracity.” “It works very well in English and is credible, because it is a real conversational audio model, not separate voices,” says Steven Johnson. Because the AI model was drawn on the basis of more than 200 hours of study recordings, with two sites that argued, to understand intonations, reactions, the way of speaking too.
“We needed a conversational French to obtain a real French version,” he explains. “Each language is interrupted differently. In each one, the way of informing their agreement or disagreement in the conversation is made of different sounds. If we had hurried to have the podcast in English, that would not have had the magic of a fluid and natural conversation that we wanted to obtain.”
Source: BFM TV
