Sekėjai

Ieškoti šiame dienoraštyje

2025 m. sausio 28 d., antradienis

DeepSeek Story: If you can invest less for a powerful model that has wider adoption because the costs are lower, that's got to be a good thing for the broad-based economy, all the companies that are using AI

 


"For two years, markets' belief that the rise of artificial intelligence would usher in a new era of productivity growth has fueled trillions of dollars in stock-market gains.

Nvidia, the maker of the computer chips at the heart of the AI boom, has been in the vanguard of this advance. Wall Street has perceived the company to have an almost unbreachable defense against competition with its offerings of high-tech chips. The company's rapid growth and windfall profits have helped push other technology firms and the Nasdaq Composite Index to record after record, with giddy investors expecting more of the same down the road.

On Monday, the mood turned sour. DeepSeek, a dark-horse power in artificial intelligence, emerged from China. That rattled big tech stocks, led by a plunge of almost $600 billion in Nvidia, which only last week was the world's most valuable company. Nvidia's fall marked the largest one-day loss in market value for any public company.

DeepSeek released last week an AI model that appeared to perform on par with a cutting-edge counterpart from OpenAI, the U.S. firm at the heart of the AI craze. The twist: Creative engineering tricks meant DeepSeek needed far less computing power. The upshot is that the AI models of the future might not require as many high-end Nvidia chips as investors have counted on.

"This is kind of classic in our industry," Salesforce Chief Executive Marc Benioff said. "The pioneers are not the ones who end up being the victors."

The development turned Wall Street upside down. Nvidia's stock dropped 17% to its lowest level since October. The S&P 500's technology sector lost 5.6%, its worst one-day decline in more than four years. In all, Monday's bloodbath wiped out some $1 trillion from the stock market's value, according to Dow Jones Market Data.

Leon Cooperman, the billionaire stock picker who founded Omega Family Office, is one of many investors who said the euphoria surrounding the sector reached unsustainable heights.

"Every third word out of anyone's mouth was 'AI,'" Cooperman said. "Everybody was bulled up in the market. If you have a contrarian bone in your body, you have to look the other way."

The threat to Nvidia is the largest it has faced since sales of its chips skyrocketed during the budding AI boom two years ago. The chip maker booked more than $63 billion in earnings in its last four quarters, making it one of the most profitable companies of all time, and its shares have surged eightfold since the end of 2022.

For its part, Nvidia praised DeepSeek's advancements and pointed to strong future demand for its products. Deploying AI models "requires significant numbers of Nvidia GPUs and high-performance networking," the company said in a statement.

Many investors had latched onto the notion that AI would unleash a wave of productivity in the economy while powering continued profits in a handful of technology giants. Several said Monday's swoon exposed a deep vulnerability in the market: Many investors had crowded into the exact same AI trade.

"It is difficult to know exactly how to make money on AI," said Mike Ogborne, founder of Ogborne Capital Management, a hedge-fund firm in San Francisco that oversees a position in Nvidia. "This could be the first day of a lot more pain."

DeepSeek is the brainchild of Liang Wenfeng, a Chinese technologist who runs an $8 billion hedge fund called High-Flyer. Wenfeng plunged headlong into the business of advanced AI systems about two years ago when he established DeepSeek and made it his mission to compete with the biggest and most well-funded AI startups in the world.

Until recently, though, DeepSeek went largely below the radar. Executives attended Nvidia's annual conferences in San Jose, Calif., and the company was a big early buyer of Nvidia's chips in China. Even after U.S. export restrictions clamped down on its ability to import Nvidia's most advanced chips, it bought less-advanced chips the company made specifically for the Chinese market.

The big moment for DeepSeek came with the recent release of its "R1" model, which dazzled many users of its app with its ability to reason through tough problems in ways that rivaled -- and some say, surpassed -- OpenAI's capabilities. The company's app quickly rose to become the most popular on Apple's app store.

OpenAI CEO Sam Altman on Monday called R1 "an impressive model, particularly around what they're able to deliver for the price," in a post on X. He said his company would move up some of its product releases.

Dan Cleary, a New York-based founder of PromptHub, a startup that helps users improve their queries to AI systems, said he gave DeepSeek's R1 a multi-step math problem. DeepSeek solved it in about four minutes -- three minutes faster than OpenAI's o1 took. DeepSeek also showed more of the work it needed to get there.

He then asked DeepSeek to produce an image of a pelican riding a bicycle, and to identify an erroneous phrase ("Dan surfs in Portugal") he inserted in the text of F. Scott Fitzgerald's "The Great Gatsby." It did both well.

"It's the first really good reasoning model outside of Open-AI" that has been widely released, he said, as well as the first very good model from China.

Despite the hype, some chip-industry insiders don't believe DeepSeek will supplant AI's incumbents or that its claims of needing small amounts of computing power means Nvidia's business is doomed.

Because DeepSeek made its research and results public, other AI companies can also adopt them, potentially paving the way for other models' improvement rather than posing a direct threat to them.

If DeepSeek indeed delivered its model on the cheap, the disruption to the incumbent AI trade could be profound. But the advance could be good news more broadly, said Joseph Amato, chief investment officer at Neuberger Berman, which manages more than $500 billion.

"If you can invest less for a powerful model that has wider adoption because the costs are lower, that's got to be a good thing for the broad-based economy, all the companies that are using AI," he said.

The run-up in Nvidia and other AI stocks has been marked by intense speculation across markets, where big tech companies have never loomed so large. Since a blockbuster earnings report from Nvidia in early 2023 floored investors, everyday Americans and big institutional investors alike have piled into AI stocks.

U.S. technology mutual and exchange-traded funds attracted $23 billion of net inflows in 2024, the largest annual haul since 2000, according to Morningstar Direct data.

The outsize influence of a few big stocks has led some professionals to argue that the group is more vulnerable than ever before.

"When you see these types of levels of concentration, the megacaps, the biggest companies, tend to have a target on their back," said Michael Reynolds, vice president of investment strategy at Glenmede. "Whether it's a regulatory target, whether it's creative destruction as other companies try to take the mantle."” [1]

People who tried to regulate China, Russia, Poland and Lithuania out of today's economy look now like fools.

 

1. DeepSeek Flips Script on AI --- Chinese dark horse emerges, threatening a market darling and other big tech stocks. Banerji, Gunjan; Fitch, Asa; Langley, Karen.  Wall Street Journal, Eastern edition; New York, N.Y.. 28 Jan 2025: A1.

 

Glimstedt Advertisement: AI Agents – What Are They?


 "Generative AI dominated the AI ​​industry in 2024. One after another, increasingly capable chatbots appeared, such as OpenAI's ChatGPT o1, other AI tools capable of generating images, audio and video, such as AI AI studios from Deepbrain AI, Dream Machine from Luma Labs, etc.

 

While generative AI tools are still looking for mass adoption in various business areas, the AI ​​industry is gradually taking the next step and starting to talk about AI agents or "agentic AI".

 

AI agents are more advanced AI systems that can independently initiate and complete tasks, make decisions, and perform necessary actions with minimal or no human supervision.

 

In other words, generative AI models are designed for someone to create, while AI agents are designed to act.

 

AI agents can be used in a variety of sectors. They can help assess people’s creditworthiness, review and decide on insurance claims, assess and allocate social security benefits, or suggest personalized treatment plans based on real-time health data. These are just a few of the areas where AI agents can be used.

 

OpenAI, which often sets trends in the AI ​​industry, last week unveiled its own AI agent, Operator. This AI agent, according to OpenAI, will be able to fill out forms or order food online on behalf of users.

 

The system is still in the testing phase, and its developers will seek feedback from early adopters to improve it.[1]

 

It is widely known that generative artificial intelligence is associated with many not only technical (e.g., so-called hallucinations), but also legal problems, such as the sometimes borderline legality of using works of authors available on the Internet to train artificial intelligence models.

 

What should you know before starting to use AI agents?

 

In order for AI agents to be able to perform the functions assigned to them, they need considerable autonomy to make decisions and perform appropriate actions to implement these decisions.

 

This essential property of AI agents and other general characteristics of artificial intelligence systems, which are also characteristic of AI agents, raise at least several legal aspects that must be paid attention to before AI agents are "untied".

 

Data and privacy protection. The majority of AI agents currently being developed are based on large language models technology. Therefore, in order to function properly, like large language models, AI agents will have to be trained on appropriate data. In this case, this data will have to be taken from the specific user environment so that the AI ​​agent, having become familiar with this environment, can act independently in it and achieve the user's goals.

 

When collecting this data and transferring it to the AI  ​​agent system, all data protection principles (data minimization, etc.) will have to be ensured. Of course, these principles will have to be ensured even after the AI ​​agent has already started operating, for example, when interacting with the customers of the company that used it and collecting their data.

 

Cybersecurity. AI agents, like any other software, can become a target for hackers. Certain sensitive data can also be leaked to third parties due to errors in the software on which AI agents operate.

 

Liability for damage caused. An AI agent, while independently implementing the goals assigned to it, can cause damage. In such a case, the question would arise as to who should compensate for this damage. Since the AI ​​agent itself is just software, the question of the legal liability of its user or developer would arise.

 

Liability for software defects. In the event that the software constituting the AI ​​agent has quality defects, i.e. does not function properly due to errors left in the development process or arising during software updates, the question of the legal liability of its developers would arise.

 

Compliance with legal requirements. AI agents can be designed to operate in a wide variety of areas. Some of these areas, such as the financial sector, are strictly regulated by specific legal acts applicable to them. Therefore, it is necessary to ensure that an AI agent used in such an area, when independently performing certain actions, does not violate the requirements of applicable legal acts.

 

Conclusion of contracts. Technically enabling an AI agent to perform actions such as the acquisition of goods would essentially mean that the AI ​​agent concludes transactions that should create corresponding rights and obligations for its user.

 

In certain cases, the question could arise whether a certain person or company using the AI ​​agent will properly express the will to enter into a relevant transaction using an AI agent. The other side of the coin could be that the AI ​​agent would assume unwanted legal obligations on behalf of the user. For example, an agent designed to optimize the logistics supply chain could purchase more goods than the company needed at the time. In such a case, it may be difficult to justify that a certain transaction was concluded without the necessary authorization.

 

This list identifies only a few of the main legal risks associated with the use of AI agents. Given how widely AI agents can potentially be applied, there may be more relevant legal risks. Therefore, they should be assessed separately for each individual AI agent.

 

How to ensure that the full potential of AI agents is used?

 

AI agents bring AI technology closer to its main vision - autonomous AI systems that perform the most complex and unpleasant tasks, thereby unleashing the creative potential of people. However, in order for these latest AI tools to truly achieve their intended goals, several factors must be properly considered when developing and using them. It is necessary to properly assess the context in which the respective AI agent will be used, what goals are set for it and who will specifically use it, how complex the technology on which the AI ​​agent is based is or should be, and what are the limitations of this technology, how to ensure the proper use of the AI ​​agent and its smooth further operation. These aspects should be assessed not only by companies developing AI agents, but also by their users, so that they can understand whether AI agents can actually properly perform the tasks for which they are intended."

 

[1] https://openai.com/index/introducing-operator/