Salesforce takes aim at ‘jagged intelligence’ in push for more reliable AI

Righteousness Solves one of the most sustainable problems for business applications of artificial intelligence: The space between the ability of the AI ​​system and the ability to perform in the unexpected enterprise environment – the company called the company “jagged intelligence“

Today in a comprehensive research announcement Salesforce AI research Future AI agents have been found in several new criteria, models, models and frames designed to make the smart, reliable and very intelligent and versatility for the enterprise. In addition to both innovations, the AI ​​systems are placed as autonomous agents, especially in complex business, it aims to improve both opportunities and consistency.

“LLMS can be superior to standard tests, create complex trips and complex poetry, reliable and consistent instructions in the environment, Salvio Savarese, Salesio Savarese and AI research, AI research, at the beginning of the AI ​​research

The initiative represents Justforce’s boost toward Savarese’s call “General exploration of enterprise“(EGI) – AI is specially designed for business complexity than how the theoretical of artificial intelligence (AGI).

“We are not only for the ability, but also not only for the ability, but also according to the sequence,” Savarese “explained,” Savarese “can make pictures of super-developing machines exceeding human intelligence.

Salesforce measures and fixes the EU’s inconsistency in enterprise parameters

A central center of the study is to adapt and solve the EU’s activities in accordance. Salesforce presented Simple databaseAi a public benchmark reflecting 225 simple substantiator questions designed to measure how the EU system is actually.

“Today’s AI, we have to work. But we have to work on it. But this is exactly what measurement it? This is exactly what the Benchmark is complete,” Shelby HeaNekke, explained the General Manager in the research media.

This discrepancy for enterprise applications is not just an academic concern. AI agent can break a single error operations, disrupt customer confidence or significant financial damage.

“AI is not a random entertainment for businesses; This is a mission-critical tool that requires free predictions,” said in a statement, Savarese.

Inside CrMarena: Virtual Test Area for Salesforce’s Enterprise AI Agents

Perhaps the most important innovation CrumarenaA novel benchmarking frame designed to simulate realistic customer relationship management scenarios. It provides comprehensive testing of AI agents in professional contexts, eliminating the gap between academic criteria and real world work requirements.

“Current AI models often fell very short to reflect complex requirements of enterprises, crimena provided: a novel benchmarking framework designed to imitate real, professional CRM scenarios,” he said.

The frame evaluates agent performance between three main people: service agents, analysts and managers. Early test, perhaps with the leadership, the use of leading agents in the use of these individuals was less than 65% of the time.

“The CRM arena is essentially a tool to introduce the internal agents,” said Savarese. “This allows us to try these agents, to understand that they failed, and then use these lessons we learn from this failure.”

New EmPedding models that understand the context of enterprise better than before

Was announced between technical innovations, Salesforce was underlined SFR’s placementA new model for deeper contextual understanding that leads to mass text ads (MTEB) in 56 databases.

“SFR writing is not just research. It comes to a lot of information clouds, very soon, Heinecke noted.

A special version, SFR-Lombard-codeHe was also presented for developers, developed high quality code search and development. According to Salesforce, 7B is the parameter Purchase of code data (coir) benchmarkSmall models (400 m 2b) offer effective, efficient alternatives.

Why smaller, action-oriented AI models, may be superior to larger language models for business assignments

Salesforce was also announced Xlam v2 (large action model)Family specially designed to predict movements instead of creating text only. These models start from a total of 1 billion parameters – are part of many leading language models.

“We have a 1B model, we have a way to 70B, we have a way to 70B model, we have a way of large language models. “This small model gives a lot of power in the ability to capture the next movement.”

Unlike the standard language models, these movement models are specifically designed to make the next steps in a task sequence and to make the autonomous agents that need to interact with enterprise systems.

“There are great action models Under the hood, we have a way to build them, and we make it in the direction of action trajectories, we make it in subtle,” heinecke added.

Enterprise AI Security: How to use Salesforce’s printation of the gymniyyril

To apply for enterprises related to AI security and reliability, Salesforce was presented SfroyluFamily family of models in both open information and CRM-specialized internal data. These models strengthen the company’s trust in the company that provides guards for AI agent behavior.

“Agent forces, agents, policies and standard agents based on agents, policies and standards, create open borders to ensure acts within the predefined limits,” company.

The company has also launched In the contextual codeA benchmark of a novel to evaluate LLM-based referee models with more than 2,000 challenging answers, 2000 challenging response pairs to give accuracy, consent, loyalty and response.

Looking beyond the text, the sale was announced TacoA multimodal moving model family designed to solve complex, multi-step problems with chains of thinking and moving chains (Cota). This approach allows you to comment and answer complex requests related to AI’s multi-media type, and requires up to 20% improvement in a difficult MMVET benchmark.

An innovation in the activity: Customer Reviews Salesforce’s Enterprise Forms AI Roadmap

Itai AsseoThe CEO of incubation and brand strategy in the EU research, stressed the importance of customer partners in the prepared AI solution of enterprises.

“When we talk to customers, one of the main pain points we own, there is too low tolerance to provide unkind and non-relevant answers when engaged in enterprise information,” Asseo said. “We have made a lot of progress with other methods of arguing engines and other ways around LLMs.”

Asse quoted samples of customer incubation consisting of an important development in AI performance

Of the general exploration of the enterprise: What is the next for Salesforce AI

Salesforce’s research boosts are income at a time in the adoption phase of the enterprise, because enterprises are looking for AI systems that combine advanced opportunities with their advanced capabilities.

While all the technological industries are following larger models with effective raw materials, the Salezforce’s sequence emphasizes a nuanced approach to the development of AI – one of the prioritizing real work requirements for academic criteria.

Technologies to start broadcasting in the coming months SFR’s placement Initially, the title in the cloud of information will manage the future versions of the Agent force of the Agent.

According to Savarese press conference, “not to replace people. This is the responsibility.” AI Dominantce, Salesforce for the enterprise bets in this sequence and reliability – not only raw intelligence – the event will determine the winners of the AI ​​revolution.