AI Research Insights — Dreamforce

Kate Strachnyi
3 min readSep 20, 2024

--

Caiming Xiong, VP of AI Research & Applied AI at Salesforce

At Dreamforce 2024, I sat down with Caiming Xiong, Vice President of AI Research and Applied AI at Salesforce, to discuss the work his team is doing with AI models. Caiming also shared interesting details about their progress on large action models for Salesforce, and the importance of data quality.

I’m so honored that I had the opportunity to partner with @Salesforce for Dreamforce 2024!

I spent most of my time at Dreamforce learning about Agentforce, building my own agent, and talking to Salesforce customers and partners; so it was very helpful to chat with Caiming to get an understanding of the technical side of the work Salesforce is doing.

Large Action Models

Caiming talked about Salesforce’s latest AI development: xLAM (Large Action Models); these models go beyond just generating text; they can perform actual API requests.

“xLAM is for action,” said Caiming. “It knows what API to call, the correct format, and ensures it’s robust enough to avoid formatting issues.” This level of precision is critical to the development of AI agents that seamlessly interface with other systems.

While these models are flexible and useful for many tasks, Caiming noted the difficulty of applying pre-trained models across different industries, as each has its own specific needs. “Simple things, like getting the weather forecast, are easy, but handling more complex tasks, such as managing databases, requires further investigation and refinement,” he explained. The ultimate goal is to train these models in domain-specific contexts so they can accurately handle both simple and complex tasks.

Prioritizing Data Quality

Caiming emphasized the importance of data quality when constructing meaningful AI models. He said, “You can put a ton of data in Snowflake, but if the data is bad, you won’t get good results.” This highlights the old phrase, garbage in, garbage out, and illustrates a fundamental challenge for AI teams working with real-world data.

While foundational models need diverse data to understand the different ways people communicate, action models require well-processed, accurate, and structured data.

One challenge, according to Caiming, is that action model data is scarce and not as readily available online compared to datasets for natural language processing (NLP) models. Salesforce addresses this by spending time synthesizing data that is both realistic enough to train models and accurate enough for action-driven tasks, such as taking a customer off a product.

Salesforce AI Models — What’s Next

Caiming clarified that the goal of these models is not to build Artificial General Intelligence (AGI) just yet. Instead, Salesforce is focused on practical applications — leveraging large action models and language models to solve real-world issues and deliver value for businesses and consumers.

One area where these models are being applied is in DevOps, where they aim to improve the efficiency of developers and operations teams. For example, AI models can assist developers in creating implementation plans and integrating new functionality — making the process easier than browsing documentation or reviewing code. “We want these action models to assist developers in being more efficient, allowing them to focus on higher-level tasks while the model handles routine work,” Caiming said.

Ensuring Safety and Trust in AI

We also discussed the potential risks of AI-generated code, especially in terms of security. Caiming introduced a framework called INDICT (Internal Dialogues of Critiques for Code Generation), which ensures that AI-generated code undergoes multiple rounds of checks to catch any issues. “Much like a human developer reviewing their code multiple times to find and fix security vulnerabilities, this framework applies the same process to AI-generated code,” he explained. This approach helps make AI-generated code safer and more reliable.

Salesforce is serious about refining AI models and pushing them into real-world applications beyond text generation. Their innovations prioritize high-quality data, real-world adoption, and trustworthiness, helping ensure that AI becomes a productive tool for businesses.

As Caiming said, AI will not replace humans but will help companies scale more easily by automating repetitive tasks, allowing employees to focus on strategic work. Salesforce is paving the way for a future where AI enhances efficiency, improves security, and drives business growth.

Go to Salesforce.com to learn more and make sure you explore Agentforce.

#DF24 #SalesforcePartner

--

--

Kate Strachnyi
Kate Strachnyi

Written by Kate Strachnyi

Founder of DATAcated | Author | Ultra-Runner | Mom of 2

No responses yet