Why WPP partnered with OpenAI competitor Anthropic

Entering this week, WPP’s AI brand platform, Open, had integrations with some of the most capable large language models (LLMs) available, including OpenAI’s GPT-4 and Google’s Gemini. But on Wednesday, the advertising giant welcomed a new system into the fold, developed by a lesser-known but equally tenacious company considered by many to be the closest challenger, amongst upstart research labs, to OpenAI. Anthropic, founded just three years ago, will now see its Claude 3 models being used across the marketing process by more than 100,000 WPP employees—the largest implementation of its technology within the advertising industry. The partnership signals how Anthropic is becoming an attractive AI provider for marketers. In March, Accenture struck a deal with Anthropic and Amazon Web Services to bring Claude to its enterprise clients. A week later, Amazon completed a $4 billion investment in Anthropic, further tying its cloud business and customers to the startup’s AI prowess. Check out the latest Ad Age events and award programs Anthropic’s rise to the ranks of Google and OpenAI also sheds light on why more marketers are embracing a multi-model approach. Anthropic’s models are particularly intriguing for WPP because of their strengths in safety, clear value propositions and explainability, said Stephan Pretorius, chief technology officer at WPP. Anthropic has long hitched itself onto the wagon of developing responsible AI, which is proving valuable as OpenAI faces scrutiny from some, including Elon Musk, who believe it abandoned this cause in favor of commercial success. For example, Claude’s content ranks among the least toxic as compared to that of other generative models, and least vulnerable to prompt injection, whereby outputs are manipulated by contradictory prompts, according to a leaderboard developed by Enkrypt AI. Of utmost importance to WPP is Anthropic’s ties to Amazon, whose investment has conferred even more security and tighter guardrails, said Pretorius. Amazon offers Claude 3 through Bedrock, its AI hub connected to AWS. With Amazon at the helm, WPP can scale Claude through Open more easily than if it were accessing the AI like an end consumer, Pretorius said. Claude 3 is composed of three models of varying capability: Haiku, which is the fastest but least powerful; Sonnet; and Opus, the most powerful. These clear and distinct value propositions are an advantage, according to Pretorius, because they can be leveraged for more specific task completion. On account of its speed, Haiku is optimal for quick summarization of a document, he said, whereas Opus is best for achieving more expansive creativity. Pretorius also lauded the models’ ability to explain why they produce certain outputs—an especially important feature when marketers are training an AI with different datasets. Being able to successfully interrogate a model is also key to improving its mistakes. One small but potentially valuable aspect of Claude 3 is the upper limit of text that a user can include in an individual prompt, typically referred to as its context window. All three models offer a context window of 200,000 tokens, or about 150,000 words; in comparison, GPT-4’s context window is 128,000 tokens, or roughly 96,000 words. This means that marketers using Claude can process more files simultaneously—or just bigger ones—when completing tasks. “When you can combine [many documents] into a single prompt, it becomes very powerful,” said Pretorius. Many models While Anthropic and OpenAI are competitors, for their enterprise clients, they may be better understood as complements. “Each model has a different benefit, a different purpose,” said Sahil Agarwal, CEO and founder of Enkrypt AI, the company that ranked Claude low in toxicity on its leaderboard. This means that a one-model approach, despite its enviable simplicity, may not be optimal for marketers looking to get the most out of their AI products. With so many new launches in the AI space—from Google to Meta to OpenAI and so forth—this may come as a relief. But Pretorius warned against compiling different models simply for the sake of collection. Marketers need to experiment with AI tools and apply them to a variety of use cases in order to wield their strengths and weaknesses, he said. Over time, he added, more distinction will emerge between providers as users figure out where their models are best implemented. The influx of small language models (SLMs) is a testament to this growing trend. SLMs require less computational resources to train and operate, and therefore fewer costs, making them an attractive option for small and medium-sized businesses. But as is becoming more common in the enterprise space, the best AI strategy may be a question of “and,” not “or.” “As long as models keep improving, we’ll continue to add them to our platform,” Pretorius said. AI marketing blog How brands are using ChatGPT, DALL-E 2 and other tools Read more here

Why WPP partnered with OpenAI competitor Anthropic

Featured Stories