A.I. startup Cohere lands a top YouTube exec in another sign of NLP’s acceleration

One overwhelming trend that was clear from the conference was the way in which advances in natural language processing are rapidly transforming business. The technology is enabling unstructured data to be pulled directly from written documents and analyzed and summarized. It’s also powering the advent of more useful chatbots and the creation and editing of images from text prompts for both entertainment and marketing. These NLP-based technologies are going to shake up the way a lot of established companies do stuff. It is also going to create some amazing opportunities for startups to build potentially big businesses on the back of these capabilities, fine-tuned on smaller data sets for various industry verticals.

If we needed another proof point, this week brings news that one of the new breed of A.I. startups has scored a major coup by hiring away a top executive from one of the FANGs to lead its operations. Cohere AI, which itself was created by alumni of Alphabet’s Google Brain research division, has hired Martin Kon, currently the chief financial officer at YouTube, to be its new president and chief operating officer. (Kon’s official title at YouTube is business finance officer, and his role is more expansive than might be the case for a traditional CFO, even at a company with $28 billion in annual revenues, as Alphabet’s YouTube had in 2021. Kon has responsibility for YouTube’s strategy, finance, business operations, and commercial data analytics and reports directly to both YouTube CEO Susan Wojcicki and Alphabet CFO Ruth Porat.)

In an exclusive interview with Eye on A.I., Kon tells me he sees natural language processing as “the next disruption and transformation in how we, as humans, inform and entertain ourselves.” He points out that 90% of enterprise data is unstructured, and that NLP is the key to unlocking it. The problem, he says, is that building a large language model of your own requires very expensive supercomputing infrastructure. And while Google has that kind of infrastructure, much of it is focused on using A.I. to generate improvements in Google’s existing product suite, not in rolling out products aimed at enterprise customers to help them transform their own businesses, Kon says. Cohere, on the other hand, is all about bringing “Google-quality A.I. to the masses,” he says. The company has a close partnership with Google that has enabled it to build its NLP capabilities using Google’s data centers, including computer chips, called tensor processing units (or TPUs) that Google built specifically for A.I.

Aidan Gomez, Cohere’s co-founder and CEO, says Kon is the right person to help Cohere bring these products to market. “The major project for Cohere right now is turning in a go-to-market direction,” Gomez says. And he says Kon has experience from YouTube in how to bring new products to market and create a successful community (creators in the case of YouTube, developers in the case of Cohere) around those products.

This week Cohere debuted a large language model that can understand text in more than 100 languages—which is more powerful than other NLP models that tend to be trained only in a single language, usually English. According to Gomez, the company is now working on rolling out products in three areas: one that can create coherent blog posts and other kinds of writing, one that can summarize large amounts of text, and one that can do search and information retrieval. (For more on what this new NLP revolution in search might mean, see the Brain Food section of today’s newsletter.)

And with that, here’s the rest of this week’s A.I. news.

Jeremy Kahn
@jeremyakahn
jeremy.kahn@fortune.com

A.I. IN THE NEWS

Apple scales back its self-driving car plans and delays launch date. That’s according to a Bloomberg News report that says the company will no longer try to produce a fully self-driving electric car as its debut in the automotive sector. It has also delayed the launch date for the vehicle by at least a year to sometime in 2026, Bloomberg said, citing sources familiar with the project, which it said has been dubbed “Titan” within the Cupertino, Calif.-based tech giant. The company had originally hoped to launch a car without a steering wheel or peddles, but now realizes the technology to do so isn’t yet developed enough.

Stack Overflow forced to restrict ChatGPT-generated answers. The coding question-answering site Stack Overflow was forced to ban users from uploading answers provided by OpenAI’s remarkable new chatbot interface ChatGPT. “The primary problem is that while the answers which ChatGPT produces have a high rate of being incorrect, they typically look like they might be good and the answers are very easy to produce,” wrote Stack Overflow’s content moderators in a statement quoted by tech publication The Verge. The moderators said they simply could not vet the volume of plausible-but-wrong answers users were creating with ChatGPT and then posting to the site. On Twitter, some users said Stack Overflow was the first victim in what would be a coming wave of A.I.-generated misinformation that might obliterate trust in the written word completely.

Large language models and generative A.I. are making rapid inroads into drug discovery. A fascinating report in the Wall Street Journal looks at how similar A.I. systems to those that underpin chatGPT and Stable Diffusion are being used by researchers to create A.I. systems that can generate new proteins or possibly even new drugs based on small molecules, directly from natural language descriptions of their desired function. “Technologies like these are going to start addressing areas of biology that have been ‘undruggable,’” Sean McClain, founder and CEO of Absci Corp., a drug discovery company in Vancouver, Washington, told the Journal. 

Neuralink faces federal probe over animal deaths. The U.S. Department of Agriculture and federal prosecutors are probing Neuralink, the Elon Musk-founded company that hopes to use brain implants (and quite a bit of machine learning) to allow people to communicate with computers just by their thoughts, according to Reuters. The investigation follows whistleblowing reports from company employees that, under pressure from Musk to perfect a surgical robot needed to implant Neuralink’s sophisticated brain-computer interface, the company needlessly caused suffering to and ultimately killed hundreds of lab animals. On Nov. 30, a week before Reuters reported on the federal investigation, Neuralink posted a blog outlining its animal husbandry practices and what it said was a rigorous framework for assessing animal welfare.

EYE ON A.I. RESEARCH

Amazon says reinforcement learning has enabled it to hold 12% less inventory. A team of researchers from the company published a paper on a system they call “Deep Inventory Management,” which they trained to try to better balance the amount of inventory the retail giant needed to hold in its customer fulfillment centers and warehouses in order to meet shifting customer demand. Hold too much inventory and the company loses money due to increased working capital costs. Hold too little and the company loses out on potential sales.

The Amazon team created a new algorithm to help balance these demands and then trained it on a dataset of 80,000 products, looking at their fluctuating demand and inventory levels over a two-year period from August 2017 to August 2019 (it’s notable that this was pre-pandemic.) They said they then tested the system in real life on a portfolio of products during a 26-week trial. The result was that the system was able to reduce the amount of stock held by 12% with no reduction at all in revenue.

The research is significant because it shows how companies can increasingly use more sophisticated A.I. techniques, such as reinforcement learning—but it does require that they have enough good data to build a reliable simulation of their operations. You can read the Amazon paper here on the non-peer-reviewed research repository arxiv.org.

FORTUNE ON A.I.

ChatGPT gained 1 million users in under a week. Here’s why the AI chatbot is primed to disrupt search as we know it—by Steve Mollman

The CEO of Amazon Web Services likes to hire people who are ‘restless and dissatisfied.’ Here’s why—by Geoff Colvin

Elon Musk’s history with OpenAI—the maker of A.I. chatbot ChatGPT—as told by ChatGPT itself—by Steve Mollan

Robots are coming—and it doesn’t look pretty for workers. Get ready for long hours, less pay, and fewer jobs—by Will Daniel

From silicon chips to tortilla chips, the A.I. ecosystem takes shape—by Nick Rockwell

BRAIN FOOD

Is NLP and chat a Google killer? That’s what a lot of people, including Fortune’s own Jacob Carpenter, have been asking following the debut of OpenAI’s ChatGPT.

I also thought it was notable that among the products that Cohere’s Gomez told me that the company is planning to roll out is one that will use new NLP capabilities to power search. And it turns out that Cohere — and although they haven’t explicitly said it, probably OpenAI — aren’t the only ones who think search is ripe for disruption and that NLP is going to be the tool to do it. I also caught up this week with Richard Socher, the former head of A.I. research at Salesforce, who is now the founder and CEO of You.com, a startup that hopes to challenge Google’s search dominance. (Socher admits that this goal is “somewhere between coolly ambitious and stupid,” but says that “someone has got to try it.”)

You.com’s product is a simple to use search platform that is open to developers who want to build their own niche search apps on top of You.com’s underlying tech. So, for instance, there’s an app called YouCode that allows software programmers to search in natural language for a particular piece of software code that will perform a certain function. Another, called YouWrite, has a text generation function that can create a blog post on any topic. There’s also an image search app that allows a person, if they can’t find exactly what they want, to use the open source text-to-image generation A.I. Stable Diffusion to create a novel image of exactly what they want directly from the You.com search interface.

Socher says he sees companies being able to build on top of You.com to create interfaces that will provide better answers from corporate knowledge bases and databases. This will enable people to find Slack messages, emails, and other documents, including policies, contracts, and sales material, without having to rely on keyword searches that are cumbersome and often fail to retrieve the information people actually want to know.

Of course, the biggest problem with using NLP interfaces such as ChatGPT for search is that there is, right now, no easy way to tell if the information being presented to a user is accurate—or completely made up. It’s also difficult to constantly retrain these models to be sure they return up-to-date information. And right now there’s no easy way to get the large language models to cite the sources of any factual information they produce.

But both Gomez and Socher told me they think that these problems are solvable. “I don’t think it is technically an impossible task,” Socher says, noting that You.com has the right mix of people—experts in both search and A.I.—to potentially do it. For his part, Gomez says Cohere is doing some research on this area, which is called retrieval augmented generation.

So while Google’s dominance may not be in jeopardy yet—cracks in the foundation of its power are starting to appear.