DeepSeek, Operators and agents: latest news on AI
The year 2025 begins with a bang for the advancement of generative artificial intelligence, with major innovations that are redefining the field and opening up new scenarios. From China comes DeepSeek, a model that has surprised with its efficiency and open source approach, while OpenAI tries to surprise (but fails completely) with the launch of Operator, an agent capable of interacting directly with the web. In addition, OpenAI CEO Sam Altman announced the imminent arrival of the O3-mini model , and there is increasing talk of 2025 as the year of AI agents, destined to interact with the real world in ways never seen before. Let’s find out all the details together.
DeepSeek: the Chinese AI that is changing the rules of the game
DeepSeek has caught the world’s attention thanks to a revolutionary approach to its training. Unlike large Western models that use huge amounts of hardware resources and billions of dollars of investment, DeepSeek was trained with much smaller resources.
DeepSeek is a China-based startup whose controlling shareholder is Liang Wenfeng, an investment fund co-founder and former trader. He had started exploiting artificial intelligence to make investments in the stock market, and that is where his story comes from. The initial intention was to acquire hardware so that he could create an efficient system to train a model for use in stock markets. The idea was not a success, and so the company changed target, going on to hire dozens of bright young minds (cheap compared to the Western market) and create something innovative.
The features of DeepSeek: lightweight, yet efficient training
So it first released the V3 model with 671 billion parameters and trained only 2 months at the negligible cost (compared to big tech) of just over $5 million. After a few days it released the R1 model, which challenges directly in the rankings of the best language models because of its “reasoning” engine. Basically, just as OpenAI’s O1 model does, it takes time to “think” before providing the answer. It should be pointed out that it does not really think in the traditional way, but merely generates text before the answer (visible in plain text), in which we see the model arguing internally with itself to try to formulate hypotheses.
In addition, Deepseek has also distinguished itself in the quality of its work by the efficiency of its memory use. Unlike its more popular competitors, it uses only small portions of its brain at a time to optimize response times that are very fast indeed. This shows that extreme computational power is not always needed to achieve amazing results.
This news had significant market repercussions: major stocks related to AI hardware, such as those of GPU manufacturers, plummeted in the stock market. This is because DeepSeek represents a clear signal that the industry may no longer depend so much on expensive hardware to train advanced models.
But what about quality?
It is talked about a lot precisely because it can currently compete with the most popular models at extremely low cost.
Looking at LMArena’s ranking of the best LLMs, one can see that the two models mentioned already occupy two positions in the top 10.
I have made comparisons and the quality on some aspects is almost on par with ChatGPT. I have noticed shortcomings on the software development part, but on text generation it produces high quality outputs.
For those who have already built software using OpenAI through the use of its API will be able to reduce costs significantly by simply changing the API key, since DeepSeek uses virtually the same syntax.
Open source and accessible to all
Another important element of DeepSeek is its open source philosophy. The model has been released under an MIT license , one of the freest ever, which allows anyone to use and modify it without special constraints. In addition to the model’s weights-which represent the numerical data that the AI uses to run and which allow the model to be run locally-a detailed scientific paper has also been published . This paper describes the training process step by step, making DeepSeek replicable by any interested developer. Here in our team, we were able to launch a scaled-down version of DeepSeek R1 with a common performance computer.
The doubts
They are not just mine but those of many.
Let’s talk first of all about privacy. If we use the model hosted by them we are certainly faced with privacy regulations that are certainly less reassuring than those in the European Union. One has to understand whether to trust putting sensitive data in the hands of the Chinese. However, it must be said that by using the Open Source version hosted locally the problem would not exist since the data would never leave our computer.
Another point to discuss is Chinese censorship.
As can be seen in this image, the team becomes very protective of the Chinese government.
Those who know the history will be well aware of the Tiananmen Square transcripts, which are available in any Western historical source, but in China it is still well-secreted from the public. OpenAI also operates bans on what ChatGPT can and cannot say, but more than censorship we talk about making the model reluctant to give out dangerous or adult-oriented information. In this case with DeepSeek it sets an important precedent to think about. I don’t think it is fair that a company releasing such a product can influence with its bias what is right or wrong to show the user.
OpenAI launches Operator: the AI agent that uses the web for you
Operator is OpenAI’s new AI agent, designed to automate online tasks by interacting directly with the browser. Unlike classic text generation models, Operator can browse the web, fill out forms, click buttons, and even make purchases on its own. This makes it a perfect tool for simplifying repetitive tasks and improving efficiency.
Currently, Operator is available only in the United States for users of the ChatGPT Pro plan , which costs $200 per month. Although the price is still high for many, this represents a key step toward the practical integration of AI into daily life.
What is Operator and how does it work
The technological heart of Operator is the advanced CUA (Computer-Using Agent) model , which is capable of recognizing the graphical interfaces of websites and interacting with them. This system uses reinforcement learning to improve the accuracy of its actions by asking the user for confirmation before completing critical tasks, such as a purchase or sending an e-mail.
Another strength is customization: users can create workflows tailored to specific sites or recurring tasks, making Operator a highly versatile tool. OpenAI has also provided control mechanisms to ensure that the AI agent remains secure and reliable.
The Potential Risks and Alternatives
Despite its potential, Operator is still in an experimental stage and can make mistakes, as stated by OpenAI itself (and from the testimonies of those who have used it, it still seems to be far from being used in production).
Using such a tool raises questions about the risks of entrusting important tasks to an AI, especially in sensitive contexts such as online shopping or compiling sensitive data. OpenAI claims that the moment we are entering sensitive data its system “closes its eyes” and does not look at what we are typing. But could we trust it?
For those looking for alternatives, similar solutions already exist, such as browser-use systems an open source project available on GitHub and computer-use developed by Anthropic where I have already made a video about it. These tools show how the market is rapidly evolving toward a future in which AI agents will no longer be mere assistants, but true intermediaries between humans and the real world.
Sam Altman’s announcement: the O3-mini is coming.
OpenAI’s O-series models (o1, o1-mini, o3, o3-mini) represent a new generation of artificial intelligence focused on advanced reasoning. Unlike previous GPT models, which were designed primarily for text generation, the “O” models were developed to address complex problems that require in-depth logical analysis and an ability to break down questions into sequential steps.
The o3 model is the latest and most advanced in the series, introduced by OpenAI in December 2024 in the last of the 12 days of pre-Christmas news. One of its major innovations is the use of reinforcement learning, which allows the model to “think” before generating answers. This approach, referred to by OpenAI as “private chain thinking,” allows o3 to plan and reason through a series of intermediate steps to solve complex problems.
The performance of o3 is significantly higher than its predecessor, o1. For example, it achieved a score of 87.7 percent in the GPQA Diamond benchmark, which includes expert-level science questions not publicly available online. Also, in the SWE-bench Verified benchmark, which evaluates the ability to solve real-world problems on GitHub, o3 scored 71.7 percent, compared with 48.9 percent for o1.
OpenAI CEO Sam Altman recently announced on X that o3-mini, a lighter and faster version of o3, will soon be available to the public. In a post on X (formerly Twitter), Altman stated, “Great news: the free ChatGPT level will get o3-mini! (and the Plus level will have extensive use of o3-mini).”
This move aims to make the advanced reasoning capabilities of o3 accessible to a wider audience, giving free users the opportunity to experience the potential of this new model.
2025: the year of AI agents
The year 2025 is predicted to be the year for the rise of artificial intelligence agents, autonomous systems designed to perform complex tasks without the need for constant human supervision. These agents represent an evolution from traditional chatbots, as they can manage complex workflows, make autonomous decisions, and interact with various systems proactively. According to a Financial Times analysis, investors are focusing their attention on AI agent-based enterprise products, which are considered highly lucrative despite their apparent simplicity.
No longer will we speak only of text generators with previously unimaginable capabilities, but of automated software capable of booking trips, making purchases, filling out forms and much more. This evolutionary leap could fundamentally change our relationship with technology, making AI indispensable tools in our daily lives just as smartphones disruptively entered our pockets a few years ago.
Impact on the world of work
The integration of AI agents could bring about changes in the work world. On the one hand, automation of repetitive tasks will free workers from tedious actions, allowing them to focus on more creative and strategic activities. On the other, there is a risk of a reduction of jobs in traditional roles. Several studies warn that automation through AI could increase inequality in the world of work unless the government provides adequate support to small businesses and workers.
To mitigate these effects, we will need to invest in workforce retraining and promote equitable collaboration between humans and machines, but we leave these thoughts to those in authority.
