Google and Anthropic Are Selling Generative AI to Businesses, Even as They Address Its Shortcomings

Anthropic is working on a number of techniques that will reduce hallucinations. REUTERS/Dado Ruvic/Illustration//File Photo (REUTERS)
Anthropic is working on a number of techniques that will reduce hallucinations. REUTERS/Dado Ruvic/Illustration//File Photo (REUTERS)

Summary

‘We’re not in a situation where you can just trust the model output,’ said Eli Collins, vice president of product management at Google DeepMind.

MENLO PARK, CALIF.—Google and Anthropic, two leading makers of generative artificial-intelligence systems, are racing to get ahead of the limitations of their technologies, even as they push forward on efforts to sell them to businesses.

Both companies spoke at The Wall Street Journal CIO Network Summit in Menlo Park, Calif. on Monday evening, acknowledging that their AI systems are capable of hallucinations, where they authoritatively spit out statements that are flat-out wrong. Other challenges, including improving the efficiency of training or teaching the models, as well as removing copyright or sensitive data from training, don’t yet have clear solutions.

The two companies have said they are addressing these limitations, but not all enterprises are ready to put their full faith—and corporate data—into their hands. Corporate technology leaders are under pressure to show that their investments in various AI systems are worth the cost, but that is a difficult sell when the systems aren’t always grounded in reality.

“What tactics can you offer us as we deploy applications with these things, especially in highly regulated or high-risk or highly sensitive areas?" said audience member Lawrence Fitzpatrick, the chief technology officer of financial-services company OneMain Financial.

Jared Kaplan, co-founder and chief science officer of Anthropic, said the AI startup is working on a number of techniques that will reduce hallucinations, including building data sets where the model should respond to questions with, “I don’t know." The idea is that the AI system can be trained to respond only when it has sufficient information, or will provide citations for its answers.

Still, there is a drawback to making an AI model overly cautious. “I think these systems—if you train them to never hallucinate—they will become very, very worried about making mistakes and they will say, ‘I don’t know the context’ to everything. And so a rock doesn’t hallucinate, but it isn’t very useful," Kaplan said.

Google, which last year agreed to increase its investment in Anthropic to up to $2 billion, is betting that customers will want to verify the information AI systems respond with. One solution is to make it easy for users to identify the sources of information that AI systems like its Gemini chatbot send back, said Eli Collins, vice president of product management at Google DeepMind.

“We’re not in a situation where you can just trust the model output," Collins said. “At the end of the day, I’m still going to want to know what the source of the information is so I can go there."

The provenance of model training data remains another unresolved issue. In a lawsuit filed in December, the New York Times said Microsoft and OpenAI exploited its content without permission to create their artificial-intelligence products, including OpenAI’s chatbot ChatGPT.

The tools were trained on millions of pieces of Times content, the suit said, and drew on that material to serve up answers to users’ prompts. But if an AI company were asked to remove certain pieces of content from the training of its model, there is no straightforward way to do that, Kaplan said.

Since the release of AI assistants like Microsoft’s Copilot and Anthropic’s Claude, businesses have sought to retain control over their company data, thereby preventing tech companies from training their models on it, and potentially revealing proprietary information to competitors.

Large language models, once they have been trained on certain data, can’t “delete" that information from what they have already learned, Kaplan said.

Both Google and Anthropic are addressing the biggest barrier in building more powerful models—the availability, capacity and cost of hardware like AI chips that are used for training. “The core thing that you need there is really efficient sources of compute," Kaplan said.

Collins said Google has been working on research developments to address the issue, including its in-house chips, called Tensor Processing Units, or TPUs. “We deploy it in our own data center, so we have fewer constraints," he said.

The largest version of Google’s new Gemini model was already more efficient and cheaper to build than its previous iteration, he said.

Isabelle Bousquette and Steven Rosenbush contributed to this article.

Write to Belle Lin at belle.lin@wsj.com

Catch all the Business News, Market News, Breaking News Events and Latest News Updates on Live Mint. Download The Mint News App to get Daily Market Updates.
more

MINT SPECIALS