Physical Address
304 North Cardinal St.
Dorchester Center, MA 02124
Jaap Arriens | Nurphoto through Getty Images
Large language models such as those developed by MicrosoftThe OPENAI supported firm will be comfortable this year in the midst of the rapid advances towards next -generation artificial intelligence agents and more agile and open source rivals, according to the main technology executives.
Last week, the Chinese firm of the Deepseek launched R1, a Open source reasoning model that claims to rival OPENAI’s O1 model both in cost and in performance. The open source refers to the software whose source code is available for free on the open web for possible modification and redistribution.
This week, growing awareness of the new Deepseek model led to a severe fall in actions of Nvidia and other technological giants, such as investors feared a possible reduction in reduction in powerful graphics processing units necessary to train and execute advanced workloads of AI.
Nvidia Lost about $ 600 billion in market capitalization Monday: the largest one day fall for any company in the history of the United States.
Meanwhile, executives and scientists from the main laboratories of AI are talking about a change of large language models to the so -called “AI agents” who can carry out actions in their name.
The LLM are the fundamental technology behind today’s generative applications. However, experts believe that an impulse towards AI systems of Agentic, which incorporate LLM technology, this year will erode the value of these models.
LLMS mercantilization
The LLMs are expected to become a merchandise in the near future as technology becomes increasingly advanced and the costs involved in their training and execution continue to decrease.
Thomas Wolf, co -founder and scientific director of Hugging Face, suggested that the LLMs will be integrated more into smart systems linked to the company’s own databases.
“I think people are moving from this madness around the model, understanding that, thanks to the open source … Many of these models will be free and freely available,” Wolf told CNBC. Hugging Face is a popular code repository for open source AI projects.
“In the Internet Revolution, we are moving from the creation of websites such as the main business to build native Internet companies, therefore, the Airbnb of AI, the AI strip,” he added. “They are not the model. It is the system and how the model is useful for tasks.”
The arrival of the DEPEEEK R1 model last week is seen by some technology CEO as to Other llm of indicators are becoming more and more commercial.
R1 incorporates some key aspects that differentiate it from other models in the market. For example, the model incorporates a “mixed precision” frame that uses a 32 -bit 32 -bit floating numbers combination (FP32) and floating point numbers of 8 bits of low precision (FP8).
The latter is faster to process, but it can be less precise. However, instead of trusting one or the other, Depseek USA FP8 for most calculations and changes to FP32 for certain tasks where a greater degree of precision is required.
“As AI becomes more efficient and accessible, we will see its Skyroket use, making it a merchandise from which we cannot have enough,” Microsoft CEO Satya Nadella wrote On the social media platform X, in an apparent reference to Depseek.
Meanwhile, Matt Calkins, CEO of the American Software firm Appian, told CNBC that Deepseek’s success simply shows that AI models will become a merchandise in the future.
“In my opinion, we will see a commercialization of the AI. Many companies will achieve a competitive AI, and the lack of differentiation will be bad for the first in the first high expense movements,” Calkins said by email.
What are AI agents?
In the last two years, much of the focus has been in products with Food from LLM, from generative chatbots such as Opensei chatgpt to tools that can produce images and video content.
LLM -based generative applications such as chatgpt depend on a certain amount of users to continually request the results, while, in contrast, AI agents are designed to be more action oriented.
If you want to reserve an appointment with the doctor, for example, you can ask ChatgPT with information about its closest GP practices, similar to how the website would consult using Google search engine.
In the future, technology executives predict that the AI of the next generation will involve the agents who can find out where their closest header is, provide times available for their doctor and reserve an appointment, all directly within a single platform . This is expected to transform the way we interact with the web.
“Throughout the decades since the website was invented, the standard interface of the web pages and the menus really has not changed much,” said Bobby Blumof, director of Technology at Akamai, to CNBC previously.
“We are living in this tyranny of being tied to a screen and being tied to web pages and menus that you should remember how to sail.”
The change of agents models
Last week, Arthur Mensch, the Mistral CEO, said he believes that the “approach should change to systems” that integrate both language models and contextual commercial data.
“That’s where this is changing,” Mesch told Arjun Kharpal of CNBC at the Annual World Economic Forum in Davos, Switzerland.
“It also means that the industry that is adopting it will distill its experience in those systems, and that is where that will really promote the value of AI in those industries.”
Mistral, which is supported by MicrosoftIt has quickly become one of the most forceful companies in Europe. Founded only two years ago, it was Last valued by investors in around $ 6 billion in June 2024.
The trend of agents has already won considerable traction. Last week, Operai launched his own agent, called operator. Operai describes it as “an agent who can go to the web to perform tasks for you”, adding that he is trained to interact with “the buttons, menus and text fields that people use daily” on the web.
Microsoft, who has invested billions of dollars in Openai, Rolumed the agents of AI itself In November.
Anthropic, a Amazon-The start of the support founded by former Openai research executives, last year introduced “computer use” Which is a capacity that allows their AI agents to use computers as humans to complete complex tasks.
Dario Amodei, CEO of Anthrope, told Andrew Ross Sorkin of CNBC in Davos last week that the use of the computer should be considered as a precursor to a “broader vision” that integrates these agents with common use applications.
“We are planning this year in 2025 to build something that we call the ‘virtual collaborator’,” said Amodei, describing him as “an agent who operates on his computer.”
“You can give you tasks. You can say, write this function for this application and write some code, compile the code, verify the code, talk to your co -workers in Slack, or in Google Docs, or on some other platform,” explained the Chief of Anthrope.
Anthrope plans to gradually build pieces of this vision during the next year, while ensuring that it is “safe and controllable,” Amodei added.
Paul O’Sullivan, director of technology at Salesforce for the United Kingdom and Ireland, predicts that verbal communication will be more important in the new world of AI agents. Salesforce was possibly early for AI Agent game, as it launched its own agentforce system in September before Microsoft, OpenAi or Anthrope.
“Converse more,” he told CNBC in a recent interview. “The screens will become smaller, our eyes will get worse. But they will probably be replaced by mobile devices that are more powerful than laptops anyway.”
