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

Salesforce addresses one of the most stable challenges in artificial intelligence of business applications: the gap between raw intelligence of the artificial intelligence system and its ability to constantly perform in unexpected institutions environments – what the company calls “intelligent intelligence”.

In a comprehensive research advertisement today, Salesforce Ai Research revealed many new standards, models and frameworks designed to make artificial intelligence agents in the future smarter, reliable and diverse to use institutions. Innovations aim to improve both the capabilities and consistency of artificial intelligence systems, especially when spreading as independent factors in complex business settings.

“While LLMS may excel in unified tests, plan complex trips, and generate advanced hair, its brilliance often stumbles when it faces the need to carry out reliable and consistent tasks in the environment of unexpected dynamic institutions,” Silvio Savarerez, AI’s chief research scientist, said during the press conference.

The Salesforce batch is towards what Savarese “interprise general Intelligence” (EGI) – AI is specially designed to complicate business rather than the theoretical pursuit of artificial intelligence (AGI).

“We define EGI as artificial intelligence agents designed for this purpose for optimal work not only for the ability, but also on consistency,” Savarerez explained. “Although AGI may conjure up pictures of Superincligent machines that go beyond human intelligence, companies do not wait for this fake distant future. They apply these foundational concepts now to resolve challenges in the real world on a large scale.”

How Salesforce measures and determines the problem of inconsistency in Amnesty International in the Foundation’s settings

The central concentration of research is to measure and treat artificial intelligence inconsistency in performance. Salesforce has provided a simple data collection, a general standard that includes 225 direct direct questions designed to measure the extent of the possibilities of the artificial intelligence system already.

“Today’s artificial intelligence is deserted, so we need to work on this. But how can we work on something without measuring it first? This is exactly what this simple standard is,” explained by Shelby Henick, the first director of research in Salesforce, during the press conference.

For institutions applications, this contradiction is not just an academic concern. One error from Amnesty International’s agent can disrupt operations, erode customer confidence, or causes significant financial damage.

“For companies, artificial intelligence is not an informal hobby; it is an important tool that requires the ability to predict,” Savarerez pointed out his comment.

Inside Crmarena: Salsforce test

Perhaps the most important innovation is Crmarena, a new standard framework designed to simulate real customer relationship management scenarios. A comprehensive test allows artificial intelligence agents in professional contexts to address the gap between academic standards and work requirements in the real world.

“In recognition that current artificial intelligence models are often limited to the complex requirements of institutions environments, we have presented Crmarena: a new framework designed to simulate the real CRM scenarios,” Savarerez said.

The frame evaluates the performance of the agent through three main people: service agents, analysts and managers. Early tests revealed that even with the directed claim, the leading factors have less than 65 % of time calling for jobs for these people.

“CRM ARNA is an internal tool to improve agents,” Savarerez explained. “Stress allows us to test these factors, understand when they fail, then we use these lessons that we learn from these failures to improve our agents.”

New inclusion models that understand the Foundation’s context is better than ever

Among the technical innovations announced, the most prominent Salesforce SFR-Imbedding, a new model for understanding the deeper context that leads to the criterion for the inclusion of the huge text (MTB) through 56 data sets.

“SFR is not just a search. It comes to Cloud Cloud, very soon,” he pointed out.

A specialized version, SFR-Emping-Code, has also been offered to developers, allowing the search for high-quality code and simplification development. According to Salesforce, the 7B parameter release leads the code information retrieval (COIR), while smaller models (400 meters, 2b) provides effective and cost -effective alternatives.

Why the micro -an artificial intelligence models may excel

Salesforce has also announced the Xlam V2 (a large business model), a family of models specifically designed to predict procedures instead of just creating a text. These models start from only 1 billion – part of the size of many leading language models.

“What distinguishes our XLAM models is that if you look at our models sizes, we have a 1B model, we are 70B. This 1B model, for example, is a small part of the size of many large language models today,” he explained. “This small model packs a lot of strength to take the ability to take the next action.”

Unlike standard language models, these procedures models are specifically trained to predict the following steps in the sequence and implementation of tasks, which makes them of special value for independent factors that need to interact with institutions systems.

“The large movement models are LLMS under the cap, and the way we build it is that we take LLM and we bear it on what we call the procedure paths,” he added. He added.

AI Enterprise Safety: How to create a confidence layer in Salesforce handrails to use business

To address institutions’ concerns about the safety and reliability of artificial intelligence, Salesforce SFR-Gard, a family of models trained on both data available to the audience and internal data available CRM. These models enhance the company’s confidence layer, which provides handrails for the behavior of an artificial intelligence agent.

The company stated in its announcement: “The handrails in Agentforce establishes clear boundaries of the agent’s behavior based on the needs of work, policies and standards, and to ensure the work of agents within pre -limited limits.”

The company has also launched contextudgebench, a new standard for evaluating LLM judges in the context-where more than 2000 pairs of difficult husbands were tested for accuracy, accuracy of sincerity, and the appropriate rejection of the answer.

If we look beyond the text, Salesforce unveils Taco, a multimedia model family designed to address multi -step complex problems through thinking and working chains (COTA). This symptom enables artificial intelligence to explain and respond to complex queries that include multiple types of media, with a salesforce claim by up to 20 % on the difficult MMVET standard.

Participation in the work: How the Salsforce’s Enterprise Ai Roadmap is formed

Itai Asseo, the chief manager in the nursery and the brand’s brand strategy, emphasized the importance of the joint agent in developing AI’s ready -made solutions.

“When we talk to customers, one of our main pain points is that when dealing with institutions data, there is very low tolerance to provide already inaccurate answers that are not related,” explained ASO. “We have made a lot of progress, whether it is with thinking engines, with rag techniques and other styles about LLMS.”

ASSEO cited examples of customer custody that results in great improvements in the performance of artificial intelligence: “When we applied the thinking engine in Atlas, including some advanced technologies to increase recovery, along with the methodology and architecture, along with our major competitors.”

The Road to Enterprise General Intelligence: What is the following for Salesforce AI

The Slesforce search boost comes at a critical moment in the adoption of the AI, as companies are increasingly seeking artificial intelligence systems that combine advanced capabilities and reliable performance.

While the entire technology industry follows more effective models with great capabilities, the Slesforce focus on the consistency gap highlights a more accurate approach to developing artificial intelligence-which gives priority to the real world’s work requirements on academic standards.

The technologies that were announced on Thursday will start in the coming months, as SFR-Imbedding is heading to Data Cloud first, while other innovations will run future versions of AgentForce.

Al -Safari also noted at the press conference, “It is not a matter of replacing humans. It is related to responsibility.” In the race to the dominance of Ai Enterprise, Salesforce is betting that consistency and reliability – not only raw intelligence – will eventually determine the winners of the Amnesty International Revolution.