Model

Alibaba has presented its largest AI model, with a billion parameters. The question is whether at this point that means something

by

The Chinese giant Alibaba has announced a new language model, the largest they have announced to date. It is called Qwen-3-Max and presumes that it has more than 1 billion parameters. The biggest. It is the last model within the series Qwen3 which was launched in May of this year and, as its name ‘Max’ indicates, it is the largest to date. Its size is given by the parameters, 1 billion to be exact, while the previous models of its series reached a maximum of 235,000 million. According to South China Morning Post (Which owner Alibaba), his model stands out in understanding of language, reasoning and text generation. Benchmarks. The results of the benchmarks place QWen3-Max ahead of competitors such as Claude Opus 4, Deepseek v3.1 and Kimi K2. If Gemini 2.5 Pro or GPT-5 does not appear, it is because they are models of reasoning and have only compared rapid response models. As they point out in Dev.toboth Gemini 2.5 Pro and GPT-5 obtain higher scores in mathematics and code, so reasoning models continue to have advantage in those areas. Qwen3-max-preview can already be tested free of charge. Benchmarks shared by Alibaba. Parameters. The parameters are all the internal variables that a model learns during training. In other words, it is the knowledge that the model has obtained from the data with which it has trained and allows it to interpret our requests and generate their answers. In theory, the more parameters, the model will have more and better capabilities. It also implies that it needs more computational power both to train and to execute the model. More does not mean better. The speech of the parameters remembers that of the megapixels with the first cameras. A 100 megapixel sensor will take larger photos than a 10 sensor, but there are other crucial factors that affect image quality such as sensor size or lens luminosity. Quality data. More parameters can be translated into more learning capacity and more resolution of complex tasks, as long as quality training data has been used. It is obvious: a language model that has been trained with redundant, incorrect or biased data will learn and continue to reproduce those errors in their operation. There are more. In 2022, the laboratory Deepmind from Google, discovered that many models were oversized in parameters but underlined in data. To demonstrate it they created the Chinchilla model with “only” 70,000 million parameters, but four times more data. The result was that it beat Gopher, a model with four times more parameters. Architecture. The architecture of the model is another decisive factor in order to achieve an efficient model; A standard architecture is not the same that forces the model to use its entire neuronal network, than one like Mixture of experts which consists of many smaller networks. It would be something like having an expert committee each with a specialty. In this way, the model can choose your expert for each query and not have to use the entire network. For example, with this technique, Mistral manages to use only a fraction of his parameters And so it is faster and cheap to execute. Image | Markus Winkler, via Pexels In Xataka | The ASML-Mistral alliance reveals the European plan B: if we cannot manufacture chips, we will at least control how they are manufactured

The world is waiting for Depseek’s new great model to compete with GPT-5, but Depseek has other plans: the agricultural AI

by

At the beginning of the year, the Chinese startup Deepseek put the world of AI up with Deepseek R1a free and open source model that was placed at the height of GPT-4 or Claude. After the coup on the table, in Depseek they have been quite quiet, but now we know what its next objective is: the agriculture. Before the end of the year. A few days ago Bloomberg reported that Deepseek is working on an advanced and very ambitious agent. He will be able to perform multiple tasks with minimal user intervention and will learn as he works. According to sources close to the company, the founder of the company Lian Wenfeng is pressing his team so that the new agentic model is ready before the end of the year. The company has already taken a step in this direction with the Deepseek v3.1 presentation Just two weeks ago. As detailed by the company in A post in Wechatits new model improves performance in reasoning tasks and agricultural abilities. A step back. Deepseek R2, the expected successor of the successful model with which Deepseek revolutionized the industry making begging. Instead they gave us Deepseek v3.1 and now the rumors suggest that their next great launch will be an AI agent. What is happening? There are voices, such as This Chinese journalistthat they see this turn to the agricultural AI as a way of taking a step back and getting away from the expensive and competitive career of the foundational language models. That The generative AI is reaching its roof It is something that is being talked about Since last year. GPT-5 is the test more recent than The big jumps are a thing of the past. If we add to this that China has a more conservative way of proceeding, with more long -term strategiesDeepseek’s turn towards an agriculture instead of launching Depseek R2 makes sense. Restrictions Although we have seen The most ingenious forms to make fun of themUnited States restrictions on chip export to China are also impacting the plans of many Chinese and Deepseek companies do not get rid. This also involves extra pressure that forces new routes with which to market their products. In fact, there is something striking in Deepseek v3.1 and it is that the model has been specially designed for Chinese chipswith the objective of Avoid dependence on foreign chips. Generate income. The agricultural AI opens another way for Deepseek, one in which you can get benefits more easily. Large language models have a problem: They cost a money and monetize them is not being a simple task. Given this, IA agents rise like a Most reasonable business model. Deepseek R1 has already given a whole lesson in Resource efficiencyIt makes sense that the company wants to opt for the fastest path to the benefits. A more conservative position. Although He has trimmed positionsChina lags in AI in terms of investments and access to the most advanced chips. Despite this, his approach in this AI race is being different. We see it in your Bet on the Open-Source wave “Personified“But perhaps the biggest difference is that, while their competitors in the United States continue to squander billions, in China they are choosing to be more conservative and not waste. This turn to the agents is in that conservative line to achieve a more sustainable industry. Image | Matheus Bertelli, via Pexels In Xataka | There is a city in China that is measured face to face with Silicon Valley: welcome to Hangzhou, the house of the ‘Six Little Dragons’

Microsoft had a saved secret. His new AI model for Copilot is the clearest statement against Openai’s domain

by

Since the fever broke out by generative AI, Microsoft has opted for OpenAi models to give life to key functions in some of its most important products. It is not strange if we remember that those of Redmond They invested more than 10,000 million dollars in the startup directed by Sam Altman. However, this alliance of convenience It has been showing fissures for months And, as time goes by, The rivalry between both parties becomes more evident. It was Microsoft itself that, a year ago, included OpenAI in his list of competitorstogether with Amazon, Apple, Google and Meta. And it is OpenAi who, According to The Informationinsists on not wanting to share its avant -garde technology when it arrives, if it arrives, the AGI. Even if you try to make up, the link is no longer as solid as in its first days. And now there is another chapter underway. Microsoft AI begins to show their own letters Mustafa SuleymanCEO of Microsoft AI, who assumed the position when the association with Openai had been consolidated for years, It has just presented two internally developed models. They are advanced proposals that users can already prove and reflect the company’s ambition: “Create AI applied as a platform for products.” One is a real novelty and another we already knew. Let’s look at the details. Mai-1-Preview. It is the great novelty. It is a Mixure-OF-Experts model, in the style of GPT-4O or GPT-5, designed to offer great capabilities in resolution of instructions and useful responses for daily consultations. According to Suleyman, it is the first model trained from beginning to end in Microsoft AI’s own laboratories. The most striking thing is that it can already be tested. Just enter LMARENAselect Direct Chat and choose Mai-1-Preview. In the coming weeks it will also arrive in Copilot, although only “for certain cases of text use.” It is paradoxical: the text functions of the Microsoft Chatbot work thanks to OpenAi, but now it will begin to live with its own technology. Developers can also request anticipated access to API. MAI-VOICE-1. It is a voice generation model that stands out for its expressiveness and naturalness. For some time now it promotes functions such as Copilot Daily (news summaries) and Copilot Podcasts, although only in English. It is also available in COPILOT LABSwhere you can try different voices, styles and narration tones. One of its strengths is efficiency: it can generate a complete minute of audio in less than a second using a single GPU. With this, it is one of the fastest and most effective voice systems that exist today. Microsoft defends that the voice will be the interface of future assistants of AI. And he wants to advance with a high fidelity solution capable of responding in different scenarios. Could I have resorted to Openai and your GPT-4o in audio version? Yes. Do you want to do it? Everything indicates not. “Much more to come. We have great ambitions for what comes next: advances in the models, an exciting roadmap in computation capacity and the opportunity to reach billions of people through Microsoft products. We are building an AI for all,” said Suleyman in X. It remains to see what course the relationship between Microsoft and OpenAi will take. What is clear to users is that there will be more variety and more tools to experiment. Images | OpenAI In Xataka | Goal wants us to use AI when we don’t know what to say at WhatsApp: This is how your new option for writing assistance works

How to edit your free photos on Google Gemini, and what can you do with the Nano Banana model

by

Let’s explain How to edit your free photos using Geminiusing the new “Banana” model that allows you to modify or add or remove elements. This is the new great evolutionary leap of the AI, which after Chatgpt allowed you to create images from photos now comes the internal edition in Gemini. This is something that you will be able to Make in the free version of Geminiyou don’t need payment. You just upload a photo and ask to make the changes you want, you don’t have to choose any specific function or model, because it is already something integrated within Gemini. Here, you should know that the great advance of this function is that the changes are made directly on the original photo. What you say, but Everything else is going to be kept in the photonothing will be changed. How can you edit your photos with To edit your photos, you first have to upload it and then Ask Gemini to make any change. Sometimes it will be enough to directly ask you to put it in black and white, for example. Other times, you can start the prompt with “In the original photo, XXX” And then write the change you want me to do. Thus, Gemini will know that you want to keep the content of the photo as it is. First, Gemini now allows you Make touch -ups in your photoshow to adjust the color, change black and white, and alter any adjustment you want. You don’t have to use an app for something so simple, you can ask Gemini directly. You can also delete objects from a photo. This is something that until recently also needed third -party apps, but both Chatgpt and now Gemini already allow you to do it. They eliminate the object you want, and cover the bottom of the photo so that it seems that there was nothing. When you upload a photo of yours or another person, you can also change elements in it. You can change colors, hairstyles, people, and even clothes that is being put in the photo. Do you want to know how platinum hair would look like? Easy. Do you want to know how you look in a suit? Easy. Everything else in the photo remains the same. You can also change some elements of one photo for othersalthough not always with realistic results. Do you want glasses? Do you want to? Do you want your cat to become a dinosaur? Everything else in the photo remains the same. In addition to this, you can also Upload two photos and make the content of one appear in the other. In the example, I have uploaded a photo of mine teaching a drawing, and I have asked to change it which appears in a second photo. Just be sure to specify in the photo you want to make the changes and what they are. And finally, you can also change the back of a photo. Thus, if in the photo you are in a place you can make it look like you are in another, or add elements that change the environment where you are. In Xataka Basics | How to use Gemini to summarize YouTube videos or ask questions about its Android content

The English Court changed the credit card model. And along the way it has become a Fintech

by

In just two years, El Corte Inglés has financed 1,372 million euros in purchases outside its commercial universe, according to reports Expansion. All for allowing you to use your card anywhere in the world. A silent transformation that has made these department stores an unexpected rival for traditional banks and Fintech emerging. Why is it important. With 11.7 million cards in circulation, the chain has converted its loyalty program of the sixties into a powerful financial weapon. Has moved more money than many Fintech that aspire to revolutionize the sector. The transformation. The change arrived in 2022 Thanks to an agreement with Mastercard. For the first time in its history, the card was the surroundings of the department stores to compete with the banks in neutral territory. The strategy is being a success. They maintain traditional advantages (interestless payments up to 60 days in ECI, 30 in other shops) but add functionalities Fintech: mobile application, notifications Pushcontrol of limits and geolocated promotions. The numbers. The total volume of the card reached 3,778 million euros in 2023, adding internal and external purchases. It is far from the 2018 record (4,895 million), but the new modality is growing. 1,372 million financed outside ECI in two years. Only 191 million in modality Revolving (Interest of 19.56% TAE). Free: without issuance or maintenance expenses. 4% bonus in Repsol for future purchases. The play. The English Court already had the largest loyalty clients in Spain. I just needed to open the doors. A much more interesting and advantageous play than the Fintech forced to find users from scratch. Yes, but. His proposal remains conservative. The revolving model barely represents 5% of the business, well below the standards Fintech. They prefer the safe margin of the interestless interest that monetize via cross consumption. The threat. For traditional banks, the English Court represents something worse than a Fintech Disruptive: a competitor with financial muscle, a huge customer base and brand credibility built for decades. Financiera El Corte Inglés, 51% controlled by Santander since 2013is the giant of consumption financing in Spain. In perspective. The chain has achieved what many financial startups only dream of: monetizing an existing user base without burning millions of euros in marketing. Its competitive advantage is not so much the technology and the confidence of 11.7 million Spaniards who already carry their card in the portfolio. In Xataka | Mercadona earns more and more money selling money, no food: the effect of interest rates on their results Outstanding image | Financiera El Corte Inglés

Tesla has launched in China a Model and L. It is the most Chinese (and logic) bet that could do in the short term

by

After months in which he has signed up for a Tesla Model and smallsomething cheaper to boost sales in Europe and press the competition for price, Tesla has surprised us with quite the opposite: a tesla model and longer. Only for China. And it makes all the meaning. The new Tesla Model Ylthat is, the elongated body version of the well -known electric SUV, increases the 150 millimeter wheelbase and its height in another 44 millimeters. A six -seater configuration is adopted (2+2+2) Therefore, a third row of seats is added that has been somewhat controversial. But it is also provided with heating and ventilation to the seats, electrical plates and integrated speakers. They are small improvements to increase the comfort of the passengers that are in tune with the bets of the L versions that the rivals usually offer in China. A category of vehicles that are based on three fundamental pillars: more space and more comfort for passengers. And that, curiously, they are transverse to all types of sizes. Purely Chinese With the presentation of Tesla Model YL, the opinions between criticism and public have been divided as usually happens with Elon Musk’s company. In Hybrids and electric They point out that some media qualify the third row like “going mounted in a convertible”given the little size between the head of the passenger and the roof. They ensure that the space that is free in the legs for a 1.70 -meter person is minimal and that it is an emergency solution for adults. In social networks a whole debate has been mounted following some images in which the same adult is seen sitting in the third row of seats. In one of them You see how it has a lot of free height between the head and the roof and ensures that it measures almost 1.90 meters. In another image it is observed as, with the trunk open, that same adult I would have serious problems traveling comfortably. Personally, I am inclined to think that the second option is the real but that does not mean that Tesla’s bet is bad. Yes, there are rivals that have a third row of more comfortable or practical seats. We have verified it in the Hyundai ioniq 9 or his brother, the KIA EV9. However, they are cars created from scratch with this intention in mind. Tesla is adding a third row of seats with a body L to improve the versatility of the car but, above all, with the intention of improving the comfort of the second row. That third row is an emergency solution and the really interesting is in a second row of a 2+2 configuration car with two good seats with all the comforts desired by a passenger. And in China the bodies L are usual. These cars are versions with longer bodies, with more wheelbase to generate a greater interior space. Europe has tired of selling this type of cars and Audi there, for example, demonstrates that the phenomenon is transverse to the size of the vehicle. The Germans have sold in China the Audi a3 l But also in those cars that, for European standards, already seem huge and do not need more interior space. Example? An Audi A8 L. To get an idea, this last car can exceed 5.19 meters to some endless 5.32 meters or impressive 5.45 meters. The latter refer to Audi A8 l horch. But the most interesting is in its interior seats with resting function or the possibility of adding a fridge in the central compartment, among other possibilities within the company’s customization program. Obviously, we talk here about a luxury car that Tesla does not intend to match but the company’s movement clearly points to proposing an experience in the cabin that is very sought in the Chinese car, transforming the interior into a kind of lounge with more space where feeling as comfortable as possible. Keep in mind that the great Chinese metropolis are huge and that the paths can languish for many minutes before the driver reaches his destination. This has made Chinese cars (and Europeans who want to sell there) have thrown into the arms of the screens and connectivity. But also that very wide rear seats are prioritized. Even sacrificing a good part of the trunk where Chinese cars tend to lose in front of the rivals. This happens because, despite its enormous sizes, in China the car is a Priority urban vehicle. It is rare that you travel with them since the distances are so long and the trains so competitive that either it is traveling in high speed or by plane. That is, cars are spaces with wheels where it is about giving the highest comfort to the client (another good example is the rise of their luxury minorities). And that is why the L -bodies L versions do not seek to have a third row of very comfortable seats. It is about prioritizing central seats in a 2+2+2 configuration and, in the best case, add a third emergency option. That is what Tesla has done. Photo | Tesla In Xataka | Tesla wants to keep certain accidents of accidents with her cars. A judge will decide if they should be made public

Chatgpt now has a “router” that chooses for us what GPT-5 model use. And is choosing the cheapest

by

The launch of GPT-5 It has not been up to the expected. Instead of being surprised with the generational jump, too many people have missed GPT-4o warmth. Sam Altman, CEO of OpenAI, anticipated the launch of the New model with the death starbut the Hype It has stayed very far from the moments wow like the Chatgpt images generationthat reached “melter” servers. One of the culprits has a name: “router” or router. What happened. Behind him Model nomenclatures mess of OpenAI (already Remember Xiaom’s mobile messi) and the fact of making the user have to choose which one more, the company chose to launch all (sub) GPT-5 models in a unified way. Now the system would choose the model that estimated the most to resolve the indication. The person in charge of this task is a “router”. This, “quickly decides which model to use based on the type of conversation, complexity, or the need for tools.” The problem is that in these first days, the router usually opts too many times by the basic GPT-5 model, in front of the “Thinking” model or reasoner. What has made the new generation seen as much more silly than Sam Altman came to affirm (he defended it was like have a team of doctors at your disposal). The necessary jump It seemed impossibleand indeed, many users have not felt it. The problem. Openai decides which model is more efficient for the user, but also which model is more efficient for their servers. At the launch of GPT-5, they have not been transparent about what leads them to use one model or another, and users who accuse them of choosing the cheapest model for Save costs. Aidan McLaughlin, employee of Openai, defended in x that prefers GPT-5 without reasoning the majority (65%) of times for how interacts with the user. He assured that they do not use the routing to save costs, but use the one they think is the best model for each consultation, for efficiency. It seemed a successbut it has arrived without complying with what users expect. Altman recognized launch failures. During a question session (AMA) on Reddit held on Friday, Sam Altman read and observed that for many users, GPT-5 did not work as well as 4th. To it He replied That the new model seemed “more silly”, because the router did not work well when they threw it on Thursday. Who warns is not a traitor. On July 19, after announcing great performance in Mathematics OlympiadAltman announced that GPT-5 was just around the corner, in front of images like that of the death star, he was cautious, claiming to want to establish realistic expectations: “This is an experimental model that incorporates new research techniques that we will use in future models” “We believe you will love GPT-5, but we do not plan to launch a model with the gold capacity level of the mathematical Olympiad for many months.” The case remembers Sora’s, who was shown Very green After the Initial announcement promisesor that of O3, which was launched with a performance in benchmarks inferior to the initially promised (Unleashing everything A debate about whether we were already in the AGI). Openai has had to make changes. Before the complaints, from OpenAi they took a great measure: return to the users of the subscription Plus the possibility that the subscribers of the plan (of 20 euros per month) will use 4th again. But, in addition, they have touched the router to better choose the appropriate model to each consultation. Of course, without explaining what leads to use one or the other. The trick (if you pay). When users perceived that the problem of the new chatgpt with GPT-5 was not the model itself, but that the router did not turn to the reasoning model, they launched to look for ways to make GPT-5 Thinking use. The way to do it has been to integrate phrases such as “think about your in -depth response” in the indications, as suggested and demonstrated DOTCSV. In this way, it is achieved that, for Plus users, chatgpt use the Thinking model Without spending the quota of 200 weekly messages Manual mode. This trick does not apply to the free model, which could only activate it without asking for GPT-5 Thinking automatically once a day. Use free reasoning. Another change confirmed by Altman has been to integrate a reasoning button again (hidden) to use the reasoning model in the free version. Trying it, Chatgpt acknowledges using the GPT-5 Thinking Mini model, and we have been able to use it in 10 messages. From there, he used the model without reasoning. Is the successor of O4-miniwhich in April also received for free Reasoning functions With a “Think” button. Result: More use of reasoning. After announcing an increase in reasoning limits, Sam Altman has offered Specific data of the number of users who use reasoning now before the launch, and what indicates is the little that was used before, even among payment users. Free users have gone from less than 1% to 7%, but what the most collides is that those who paid the plus only used 7% reasoning, compared to 24% of now. It surprises in the sense that the reasoning of O1 and O3 supposed one of the great advantages over the free plan, thus the capacities of Deep Research. Image | Xataka In Xataka | Chip restrictions are over to China: now technological war

Hundreds of people attended a funeral. All normal except because the deceased was an AI model that had been removed

by

The artificial intelligence boom is making us live moments of the most curious. There is people falling in love with an AIothers go to her as to a psychologist or for “Risheat” a loved one. What we needed to see was to celebrate a funeral for a retired model and it is just what happened a few days ago in San Francisco, where more than 200 people gathered to give the last goodbye to Claude 3 Sonnet. The deceased. July 21, Anthropic announced that he withdrew several modelsamong which was Claude 3 Sonnetprotagonist of this funeral picturesque. The model was announced in March 2024 and stood out for its balance between performance and efficiency. Together with him, Anthropic also withdrew Claude 1 and 2, in addition to confirming that Claude 3 Opus would disappear in January 2026. The funeral. It was held last Saturday in a warehouse in the Soma district of San Francisco and more than 200 people attended. They count on This Wired report that the decoration did not go unnoticed: a giant tentacle hanging from the ceiling and several mannequins with an aesthetic and represented different Claude models. In the center of the room he lay the mannequin that represented Claude 3 Sonnet and was surrounded by the most diverse offerings: from the classic flowers to a bottle of ranch sauce. Devotion. It sounds very absurd and clearly there was an aura of irony and humor in the event, but there were also emotional moments and even say in Wired that some attendees cried. Openai has the most popular chatbot, but Claude de Anthropic has a very devout fans community. The reason would be that Anthropic has managed to imbue a warm and friendly personality in his chatbot that has conquered many users. Exists a website where you can see ranking of the users who use Claude most and there are even people doing FAN ART With the company’s logo. Emotional connection Establishing affective links with a machine is not something new. In 1966 it was created Eliza, the first chatbotand many people who interacted with her ended up telling her personal things, as if she were another human, they called him the Eliza Effect. With the AI boom we are seeing how these links become popular and the success of apps such as Replika either Character.AI They are a clear example of this trend that increasingly adds more adherents. Attending the funeral of an AI sounds like a pretty fun plan to spend some time on the weekend, but developing an emotional connection with a chatbot is not always positive. The dark side. There are people whose connection with these chatbots brushes addiction to the point that Online support groups are being createdto alcoholics anonymous. In This study They analyzed how the use of chatbots affects our mental health and, although at the entrance it can have a positive effect, in the long run it is associated with higher levels of social isolation and emotional dependence. The experts alert of the risk of confusing the empathy of a machine with real empathy; A chatbot is always available to speak already It tells us what we want to hearThis causes a false sensation of empathy that can negatively affect our ability to connect with other human beings. Image | Chatgpt In Xataka | A group of experts in AI attended a party in a mansion. The topic of conversation: what will be when AI ends humanity

The best of the model is where it must always have been

by

GPT –4.1, GPT –4.5GPT-4.1-mini, O4-mini, O3, O3-PRO. If these names do not sound at all, you are not alone. Despite its popularity, a large part of the users of Chatgpt He never touched the model selector. Some knew that he was there, but beyond a brief description in the drop -down, few really understood what each one served. And yet they used it daily. Most likely, for months you have lived with the default base model –GPT -4Othe one Openai described as “ideal for most tasks” – without exploring anything else. Someone may ever try, just to discover that the results were not so different or that that was not what he was looking for. In the end, everything remained in the hands of a few advanced users, those who knew how to squeeze those other models to obtain better results. The model selector was a wall for many For a long time, using chatgpt meant making a decision that not everyone was prepared to take. What model choose? Which was better to program? And to write? What if I needed to analyze data? Openai offered a list, a brief and, hopeless description, some generic recommendation. But the only real way of understanding nuances was to experiment. Change model, try, compare results, change again. A process that demanded time and criteria. In an environment where we ask for immediate results, that was an invisible but real barrier. The majority stayed where he was. And the most curious thing is that even paying the subscription, many never used the options that were included. You don’t have to choose anymore. With the arrival of GPT –5, The model selector disappears for free users and loses relevance even among payment. Openai has designed a system that decides by the user: the new model knows when he has to think more and when he is not needed. If the task requires complex reasoning, GPT -5 takes it easy. If it is simple, respond instantly. The old chatgpt model selector, today already out of the scene This changes the rules of the game. Not only because it saves unnecessary time and decisions, but because it finally places intelligence where it should be from the beginning: at the user’s service, without having to know how it works inside. GPT -5 behaves more similar to us. Not because he is aware, much less, but because Start deciding in real time If a task requires thinking or not. As we do. When someone asks us a simple question, we answer quickly. When something demands analysis, we take a few more seconds. That is the logic that begins to replicate this model: if it takes effort, it applies it. If not, prioritize speed. GPT -5 behaves more similar to us In the background, that was what we expected of an advanced artificial intelligence. That understands us without giving technical instructions. That anticipates, That chooses by themselves how to help us better. That the model decides on your own does not mean that we have lost control. In fact, payment users can still manually force the type of response they want. From the chatgpt interface, the mode of reasoning with a button can be activated. But it can also be done with a simple instruction in the message: “Think about it”, “do it fast”, “give me a light answer” … the system interprets these phrases and adjust your behavior. It is a subtle, but important change. Because although GPT – 5 self -managed, It is still possible to intervene and refine experience. Whoever wants depth, can ask for it. Who wants immediacy, too. For those who use chatgpt intensively – at work, in studies, in complex tasks – this change has direct consequences. Before, a user could be paying for a tool that he did not understand at all, without knowing that he was using only part of his potential. Now, even if I don’t have time to experience or learn how everything works, GPT -5 does that job for him. Artificial intelligence finally begins to function as it should: offering the best without asking for explanations. And that means that those who did not know that they were losing productivity … can stop losing it. The difference between those who knew how to combine models and those who not, between those who explored each variant and those who stayed on the surface, have been reduced to minimums. From now on, the quality of experience will depend much more on how we ask for things that we know or not what a model to use. And that is good news, especially for those who never touched the selector. Users who, without changing anything in their way of interacting, They will start getting answers improves, more tight, more useful. GPT-5 equals them with the advanced ones without having to move a finger. GPT -5 does not add a new function. Correct an old incoherence. Until now, the most powerful system was hidden after technical decisions that few knew how to take. It is now in the foreground. Images | Xataka with Gemini 2.5 Flash | Xataka screen capture In Xataka | The AI has so far a nice accessory: with GPT-5, OpenAi wants to place it in the center of all things

His model is not only powerful and free, but it has cost much less

by

Openai yesterday launched its first models with open weights since 2019: GPT-Oss-120B and GPT-Oss-20B. Two reasoning systems that mark a radical change in the company’s strategy after Deepseek R1’s success At the beginning of the year. Why is it important. Openai has gone from defending the safety of closed models to recognize that “it was on the wrong side of history” with open models. The earthquake that caused Deepseek In January he forced them to rethink their entire strategy. The context. Until now, Openai had kept all its models closed since GPT-2justifying it for security reasons. But rivals as a goal with Calls And especially Chinese companies with Depseek, Qwen and Kimi They have gained ground in the open ecosystem. The turning point came when Depseek R1 showed that it was possible to create advanced reasoning models with a fraction of the cost of Americans. Sam Altman then admitted that Openai needed “finding a different open source strategy.” And here it is. In figures: The major model, GPT-Oss-120bworks at a single 80 GB GPU with just 5,100 million active parameters of its 117,000 million total. His little brother, GPT-Oss-20bit can be executed in a laptop with 16 GB of memory (although the speed seems to be very slow with such a configuration, in the absence that we can try it more thoroughly). The current situation. The Benchmarks independent They place GPT-Oss-120b with a score of 58 in the intelligence indexsurpassing O3-mini but being below O3 and O4-mini. In competitive programming tasks it reaches 2622 points in Codeforces, while Depseek R1 achieves less score with a much heavier model. In detail. Architecture uses a system of Mix of experts (Moe, Mixture of experts) that activates only a fraction of the total parameters for each consultation. This allows the 120,000 million model of parameters to activate only 4.4% in each consultation, while the 20,000 million activates 17.2%. Both models maintain chain reasoning ability without direct supervision, following the principle that visible reasoning is necessary to detect undue behaviors. Yes, but. OpenAI has not made public training data, maintaining part of the secret. Nor does it include multimodal capabilities, limited only to text. Hallucinations are significantly higher than in their own models: 49% and 53% respectively compared to 16% of O1. Between the lines. This launch responds to both Chinese competitive pressure and the new Trump administration, which has urged US companies to release more technology to “promote the global adoption of aligned with American values.” The company has worked with partners such as Orange, Snowflake and AI Sweden for real use cases, seeking to validate whether the advantages justify continuing to invest in open models. Deepen. The models are available under Apache 2.0 license on platforms such as Hugging Facewith native support in tools such as Ollama, LM Studio and Services Cloud of the main suppliers. Openai has organized a network Teaming Challenge with $ 500,000 in prizes to identify security risks. This is OpenAi’s response to his Deepseek moment: recognizing that open innovation is not only inevitable, but necessary to stay competitive in a panorama where cost and accessibility matter as much as performance. In Xataka | The GPT-5 paradox: OpenAi needs the greatest jump in the history of AI just when it seems most impossible Outstanding image | Solen FeyissabDima Solomin

Log In

Forgot password?

Forgot password?

Enter your account data and we will send you a link to reset your password.

Your password reset link appears to be invalid or expired.

Log in

Privacy Policy

Add to Collection

No Collections

Here you'll find all collections you've created before.