Salesforce Inc. today announced the open-source release of its in-house family of  “large action models,” called xLAM, that it says offer lower cost and higher accuracy than much bigger artificial intelligence large language models on the market today.

The company also announced xGen-Sales, a proprietary model trained to handle complex autonomous sales tasks and enhance Agentforce, a Salesforce platform that allows users to design AI agents capable of interacting with customers.

The Salesforce AI research division created the xLAM family of AI models to simplify the creation of AI agents that perform actions instead of simply create content, which allowed the team to reduce the overall complexity of the models. The resulting xLAM models, according to Salesforce, are smaller, are streamlined for tool use and more performative than their much larger counterparts, which must perform a different set of conversational, summarization and generative capabilities.

“The difference between a large language model and a large action model is a LAM is a fine-tuned LLM that’s been optimized for function calling — so when you, the user, passes in an ask, it produces a command like, ‘call this app,’ or ‘call this call Python program,’” Shelby Heinecke, senior AI research manager at Salesforce, told SiliconANGLE in an interview. “It produces an action that needs to be taken in order to answer that question. I think that’s what at the heart of what a LAM is.”

Through fine-tuning xGen-Sales to specific tasks, such as sales-oriented operations, it is capable of delivering more precise and rapid responses. It can generate customer insights, enrich contact lists, summarize calls and track the sales pipeline. Salesforce said that the xGen-Sales model is a step towards the next generation of the large action model AI and it has already eclipsed other much larger models in internal testing.

The xLAM model family is anchored by the ultra-small xLAM-1B model, which the research team has given the nickname “the Tiny Giant.” According to the team it has outperformed significantly larger models in tool use and reasoning tasks including OpenAI’s GPT-3.5 and Anthropic PBC’s Claude. That’s despite that fact that it’s built with only 1 billion parameters.

The compact size of xLAM-1B also enables it to run on mobile devices such as smartphones and tablets. There, it could be used to automate commands for a weather app such as querying data for display from local stations through function calls, running through the correct actions to condition it for the user to understand it and then present it on screen.

Salesforce has released the xLAM-1B version open-source alongside three others on Hugging Face open source for developers and enterprise users to experiment with. This includes xLAM-7B, a small model for academic exploration for limited GPU resources; xLAM-8x7B, a medium size mixture-of-experts model for industrial applications; and xLAM-8x22B, a large mixture-of-experts model that allows robust high-performance application building but requires major computational resources.

Heinecke said that when developing these new action-oriented models, one of the biggest challenges was the data needed to train them. The xLAM-1B model, in particular, needed to be significantly smaller than its LLM counterparts so it needed to be fine-tuned with very specific data to get it cut down to size.

“This is constantly the bottleneck in AI,” Heinecke said. “This function-calling direction for AI is very cutting-edge. There are only a few publicly available open-source datasets for it.”

Given the lack of maturity in action-oriented, tool-using AI models, the team needed synthetic data generation to fill in the gaps to generate enough data to train and fine-tune the model to get it compact enough while still operational and performant enough to work.

At the same time, Heinecke noted, there has been this trend around generative AI models to build them ever larger. Given how xLAM-1B outperformed models many times its size, with specialized models it doesn’t need to be this way.

“If you look up some of these models, for example Google PaLM and GPT-3, they’re hundreds of billions of parameters,” Heinecke noted. “But now, as we’re faced with deploying them, we’re seeing how managing models of that size can be troublesome given costs and latency Now we’re thinking: Can we achieve the performance of these large models in a smaller package?”

Image: SiliconANGLE/Microsoft Designer