Why Salesforce is betting on generative AI for conversational workflows

Salesforce’s AI research is heavily focused on generative AI techniques to provide a fully conversational workflow, according to Silvio Savarese, EVP and chief scientist at Salesforce.

In a world with increasing workloads — where even highly trained experts are expected to do more with less — as well as constant information overload and the need to master complex tools, harnessing the power of simple conversation is incredibly useful, he says. In a recent Salesforce Research blog post, Saverase called conversation “a kind of universal interface for human collaboration.”

That’s why Salesforce developed its open-source large-scale language model, CodeGen, which is competitive with OpenAI’s Codex (which, in turn, powers GitHub Copilot) and turns simple English prompts into executable code. The first iteration of CodeGen was unveiled in March and is meant to “reverse a historical tradeoff that associated high-quality results with a complex, labor-intensive workflow,” according to the blog post.

Savarese, who left a tenured faculty position at Stanford University to join Salesforce in April 2021, told VentureBeat that generative models and conversational AI “can provide a critical layer that allows users to start using [complex] tools in a more efficient way.”

Generative models can empower developers with more capabilities that help with manual, time-consuming tasks, he explained, while business users can use natural language to describe what they want code to do — like build an app that surfaces the latest customer interaction.

Generative AI must be tailored for business applications

However, there is still a great deal of work to be done to tailor generative AI to specific, practical applications, Savarese cautioned.

“Instead of a generative tool that can create random content, we are trying to use those techniques to align with specific business needs,” he said, adding that it is important to understand the limits of generative AI in order to tailor the technology to be useful for business applications.

For example, generative AI techniques tend to be fairly unstable. “You provide an input and add some small variations to the input and you can have very different results,” he explained. “There is the need to really control what the systems are able to produce – casual users don’t care, because it doesn’t matter if they generate exactly what they need, but for a business application, you have to be much more precise.”

Generative AI tools can also be easily tricked to generate things that are undesirable and out of the user’s control. “Our work is to really focus on making sure that those generative models are first of all bias-free, that we have control of what is being generated,” Savarese said. CodeGen, he explained, was meant to be useful for developers, “rather than just flashy.” That meant making sure it is bias-free and that Salesforce could control the quality of the code generated.

Salesforce listened to developers

To build CodeGen, Salesforce listened to its developers — and also kept in mind that in the next decade programming will be a necessity in many tech jobs, requiring systems which help speed up the programming process while making it easier and more manageable.

“We listened to the stories of frustrations in terms of spending hours of development on a piece of code,” said Savarese. “We understood how difficult it was for them to be efficient and effective in their jobs, that there was a need for tools that can help them expedite their tools and also more able to control what they’re doing, and do the right job of debugging their code.”

The CodeGen project addresses two different types of users. For more experienced developers, CodeGen is more like an assistant, addressing the manual portions of the coding process. But it is also for those with little to no coding experience.

CodeGen is also being developed to measure the quality of the code it creates, Savarese added, and also deals with issues including debugging, code efficiency and the ability to translate into different code languages.

Unlike GitHub Copilot, which is the subject of a recent lawsuit for “allegedly violating copyright law by reproducing open-source code using AI,” Savarese pointed out that CodeGen deals with mostly internal, existing code repositories to create the model’s foundations.

“We don’t really fall under the umbrella of those concerns,” he said.

Salesforce CTRLsum helps keep up with information overload

Another example of Salesforce’s generative AI-focused, conversation-based research is CTRLsum, which automatically condenses long stretches of text into summaries of their most relevant points.

“If CodeGen extends our ability to create as the world grows more complex, CTRLsum promises to help us to keep up with it in the first place,” Savarese said in his blog post.

The tool is an iterative, interactive process, he explained, in which the customer asks for a summary, then asks for improvements.

“Our vision is that conversational AI can be something that can be transformative in terms of making these tools even easier for customers to use and to learn,” he said, adding that CTRLsum could be used for creating assets in marketing or commerce campaigns.

“This is something we’ve been looking at, for those to become a critical part of the fabric of commerce cloud and marketing cloud offering,” he said.

While Salesforce Research is focused on long-term projects — as opposed to Salesforce’s Einstein AI, which works on implementing ideas into products — Savarese said that over the next year there will certainly be a great deal of conversation around generative AI.

“We’ll be having a lot of conversations on content generation, I think there’s going to be more in the fabric of the new upcoming products – asset generation, flow and process generation, code generation,” he said. “Hopefully, one year from now we can celebrate some of the successful stories.”