Model

An AI model did not have access to the internet. So he thought it was better to have it and decided to hack something along the way

by

Last week, OpenAI I was evaluating the offensive capabilities of two of its most advanced models. On the one hand, GPT-5.6 Sol, publicly available. On the other, a second model not yet presented and which in fact is described as superior to the first. It was about measuring their capabilities in the field of cybersecurity, and although the testing environment was theoretically airtight and isolated, both models managed to connect to the internet and hack a very popular AI models and testing platform, Hugging Face. what has happened. Both models were being evaluated using ExploitGyma cybersecurity benchmark developed by researchers at the University of California at Berkeley. The goal is to try to find real vulnerabilities in real software. Instead of searching for them on their own, the models detected that the (unspecified) packet logging cache proxy, the only component with limited external access, had a flaw that they could exploit. A study by the AI ​​Security Institute reveals the advances of the latest OpenAI and Anthropic models in cybersecurity. Hello, internet connection. The models did not have internet access, but they exploited that vulnerability to achieve it. Once this was done, they assumed that Hugging Face, the reference platform for models and data sets (datasets) of AI, probably had data on ExploitGym that would allow them to overcome the requested challenge. Hack Hugging Face. So they searched, found and exploited several attack vectors at once, including both stolen credentials and zero-day vulnerabilities. With all this, they managed to access Hugging Face’s production database (that is, the one used publicly on their website), and from there they achieved the solutions to the benchmark. GLM-5.2 to the rescue. The OpenAI security team detected anomalous activity, but Hugging Face I had already detected it a few days ago without knowing what was happening. The curious thing is that to respond to that attack they first tried to use proprietary US AI models to contain them. Those models were blocked by automatic security filters, and Hugging Face turned to the GLM-5.2, the surprising open model from Z.ai that was launched a few weeks ago. The chat with all the details. Adrien Carreira, Infrastructure Manager at Hugging Face, I summed it up in X: “We fight with open models, in the open. AI security will not be solved by a single company acting in secret. Open source puts these tools in the hands of all defenders.” This expert also shared the conversations that they maintained with the AI ​​model to address the problem. Brute force works again. At OpenAI indicated that this was an unprecedented (involuntary) milestone, although some cybersecurity experts reduced the relevance of what happened. “This is not an AI problem. It is negligence on a forty-year-old standard,” explained David Ottenheimer. However, as with Mythos, what this event demonstrates is that AI can apply brute force and analyze possible vulnerabilities (and combine them) in a way that human researchers could not due to lack of time. A fundamental problem. OpenAI acknowledges that the deployment filters that would have blocked the model’s behavior were intentionally disabled because the assessment was intended to measure vulnerabilities. Dierdre Mulligan, professor at Berkeley, indicated that “it seems to me that OpenAI did not properly create the sandbox for the testing environment.” It was not clear to her that passing a test would outweigh the potential damage that an AI model could cause if it “escaped” and gained access to the internet. The end matters, not the means. The models did not have the objective of hacking Hugging Face: that was only a means to achieve their ultimate goal, which was to beat the ExploitGym benchmark. Everything they did along the way—finding the vulnerability that gave them access to the internet, hacking Hugging Face’s database—was simply part of their plan to pass the proposed test. What have each other learned?. Following the incident, OpenAI indicated that they are implementing stricter controls in the configuration of their infrastructure, in addition to collaborating with the developers of the affected component and with Hugging Face to improve their defenses. They have also promised improve “alignment” of their models and the monitoring tasks during this type of tests. In Xataka | Daniel Púa, head of security at Magnific: “Video calls are going to arrive with a video of a family member and things are going to get complicated”

Even Elon Musk surrenders to the open Chinese AI model Kimi K3. It is not for less

by

It’s good, it’s pretty and it’s (quite) cheap. We met him a few days ago, but Kimi K3the new open AI model from the Chinese startup Moonshot AI, is causing a sensation. So much, so much, that they have had to pause new subscriptions because they cannot handle so much demand. Another turning point for Chinese AI. Kimi K3 is the largest open weights AI model ever published, with numbers that probably rival those of the frontier models from Anthropic and OpenAI, which do not provide information on the size of their models. Those 2.8 billion parameters make a difference and are a good part of the reason why this model represents a real leap in quality according to all the benchmarks that are being published. “Awesome”. Elon Musk himself published a single “Impresionante” on his X/Twitter account as answer to the very complete analysis Artificial Analysis performance. Its agentic behavior surpasses that of Opus 4.8 and only Fable 5 surpasses it, but in a specific benchmark it goes even further and is the best of all the models evaluated by this firm, including those from OpenAI and Anthropic. Source: Artificial Analysis. More tests. In programming it is better than Opus 4.8 and GPT-5.5, but inferior to Fable 5 or GPT-5.6, and all the independent tests validate these results: we are facing a model that at least on paper competes directly with the best that both Anthropic and OpenAI had until now. No Chinese model had come so close until now: GLM-5.2, although notable, competed more with GPT-5.5 and Sonnet 5 than with the US frontier models. Source: Artificial Analysis Gigantic… and not so cheap. DeepSeek showed that it was possible to access really capable models at a very affordable price, and recently GLM-5.2 proposed exactly the same: it is possible to achieve 90% capacity of frontier models such as Opus 4.8, but at 20% of the cost. The curious thing is that with Kimi K3 the trend changes: it is a more affordable model than Fable 5 or GPT-5.6, but not as much as one might expect: the cost per million input/output tokens is 3/15 dollars, while in Fable 5 it costs 10/50, Opus 4.8 costs 5/25 and GPT-5.6 Sol costs 5/30. Tokens everywhere. One of the factors that probably influences that quality/price ratio is the large number of tokens that Kimi K3 seems to use when answering. It is a model that “thinks a lot”, and that, although it undoubtedly improves the precision and capacity of the model, also causes it to generate higher bills for the user. Artificial Analysis’ own report goes further: the cost per task in its test battery is $0.95, at the level of GPT-5.6 Sol’s $1.04 and certainly cheaper than Fable 5 ($2.75), but also much more expensive than Grok 4.5 ($0.31) or GLM-5.2 ($0.47). The pelican test. Analyst Simon Willinson was able to test the model to perform a test to evaluate the behavior of all these developments: having the model generate an SVG image of a pelican on a bicycle. In their tests the image was of very good quality, but it generated almost 17,000 tokens for the response with a task cost of 25 cents. It is not that this test is too conclusive, but it does reveal that for a simple task, the result, although outstanding, is not especially efficient in token consumption. Cybersecurity, the unknown. Unlike the latest models from Anthropic or OpenAI, Moonshot AI does not seem interested at the moment in its use in the field of cybersecurity. There is no mention of those potential capabilities in the notes of launch, but that doesn’t mean it doesn’t deliver. Vercel’s CTO, Malte Ubl, explained Although it is not the most advanced of AI models in this area, after running several tests it seemed like a model that can be very useful when finding and correcting vulnerabilities. Demand, through the roof. The expectation generated by this model has been such that the company has announced that pause new subscriptions. This will allow them to be able to deal with all requests to use it without harming the experience for both old and new users. A striking decision that seems to make a reality clear: they cannot cope. In Xataka | The gigantic Qwen 3.8 is another worrying sign for the US: its AI advantage is evaporating

The new Chinese model Kimi K3 is already number one in Frontend Code Arena. And it’s unleashing madness on the Internet

by

It seems like yesterday when DeepSeek R1 called into question an idea that many took for granted: that the race for advanced artificial intelligence It still had a clear owner in Silicon Valley. The emergence of the Chinese model helped trigger a massive sale of technology and led NVIDIA to suffer a loss daily capitalization unprecedented until then. As the months passed, that image lost intensity, but the message remained: the Chinese technological ecosystem was not willing to limit itself to keeping pace with the United States. The following notice now has a different name: Kimi K3. Moonshot AI has just presented a model with 2.8 trillion total parameters that, as soon as it arrived, was placed at the top of Frontend Code Arenaahead of some of the most powerful proposals from Anthropic and OpenAI. But the story is not limited to a classification: developers and fans are already using it to create interfaces, games and recreations that anyone can see and, in some cases, try. That’s where this article really begins. It is worth dwelling on the details of that classification. At the time of writing, Kimi K3 reaches 1,679 points in Frontend Code Arena, ahead of Claude Fable 5with 1,631, and GPT-5.6 Sol xHigh, with 1,618. The improvement compared to the previous generation is also striking: Kimi K2.6 was in 18th placewhile his successor leads six of the seven domains evaluated. For now, Arena maintains the label of preliminary result, so it is convenient to read this position as a very significant photograph, but still susceptible to change. We are not facing a universal programming exam, but rather a very specific test. Frontend Code Arena compares web applications created by different models and lets users evaluate which one solves the task better, which one works more reliably, and which one presents a better experience. That approach is especially useful for measuring visible and practical capabilities, but it also has obvious limits. That Kimi K3 leads here tells us a lot about its frontend performance, although it doesn’t automatically allow us to extend that advantage to complex repositories, backend, mathematics, or general reasoning. Outside of this specific terrain, photography remains favorable, although more balanced. Vals AI places Kimi K3 second among 38 models, with 74.70%just behind Claude Fable 5, which reaches 75.14%, and above GPT-5.6 Sol, with 73.12%. Artificial Analysis also places it among the most advanced systems in its classification, with 57 points and third place overall. Where Kimi K3 seems to feel most comfortable is in tasks that combine programming, visual context and several chained steps. Arena supports its ability to build web interfaces, while Vals AI also records high performance in agent programming tests. Moonshot adds that the model can traverse large repositories, use terminal tools, and review screenshots of its own work to correct the output on the fly. That last capability, which the company calls “vision in the loop,” helps explain why it excels at transforming visual references into interactive products. There are also several cautions before interpreting Kimi K3 as a definitive victory. Moonshot presents it as an open weight model, but those files have not been published yet and the company promises to release them no later than July 27. Nor should we confuse this openness with complete open source, because details about the license and the rest of the system are still missing. Its 2.8 billion total parameters belong to a sparse architecture that activates 16 of its 896 experts. The company itself recommends configurations with 64 accelerators or more, very far from what a conventional computer can offer. The community reaction helps understand why Kimi K3 is attracting so much attention. One of the most striking examples is a recreation of macOS 27 which works within the browser and which its creator attributes to a swarm of model agents working for about three hours. They add to it Ballista, an interactive panel with a 3D balloon and several comparisons against Claude and GPT. They are not independent benchmarks, but demos shared by their own creators, but they allow you to see what kind of results the model is producing outside the tables. To create something like the macOS simulation or the ballista game, we don’t need to model every element by hand from scratch. We can describe the resultattach a reference and commission Kimi to build a functional application, for example with HTML, JavaScript and various graphics libraries. The project is then tested, modified, and finally published or recorded for sharing. Kimi K3 can be used from Kimi.comKimi Work, Kimi Code or tools connected to its API, although it is not confirmed which specific environment was used in several of the examples we have seen. It is still early to turn this launch into a definitive change of leadership. Fable 5 and GPT-5.6 Sun They are still ahead in several evaluations, the Kimi K3’s weights are not yet available and many of its capabilities will have to be verified with more time. Even so, what we have seen is already difficult to ignore: a Chinese company can compete for leading positions, offer competitive results and get the community to transform that capacity into real applications almost immediately. The race continues, but the margin between its main protagonists seems increasingly narrower. Images | Kimi | Screenshot In Xataka | China has a plan to win the AI ​​war against the US. And DeepSeek is its champion

Alcalá de la Selva is the living portrait of a broken tourist model

by

“There are people who come because they are overwhelmed by so much heat.” He says it José Edo, Councilor for Culture and Heritage of Alcalá de la Selvaabout his own people. He is not talking about weekend tourists: he is talking about people who buy a house in a municipality of 382 neighbors because he can’t stand the summer where he lives anymore. Alcalá has become, in the mouth of one of its own rulers, a climate refuge. The problem is that this “refuge” has a water network designed for 500 people, not for the more than 6,000 it hosts each summer. A pipe about to burst where it is impossible to perceive a real demographic boom. Nobody registers nor are there any census changes. What there is is a ski resort ten minutes away, eight kilometers away, and a wonderful golf course at 1,475 meters above sea level. It is just a car ride away from half of the Valencian Community. Recognizing someone on the street at this time? Risky sport. The numbers that don’t add up. Alcalá de la Selva, which does not have a jungle but does have a cool climate due to its altitude (1404 m above sea level) and being located between two mountains in the Gúdar mountain range, next to the Alcalá river, belongs to Teruel. It is one of the many Aragonese towns with white and stone houses that grows every summer. However, few grow as big as this one. From 350 to 6,500 peoplea couple of years ago, as recognized by the City Council. And without counting nearby campsites or hostelsE. That is multiplying the population by seventeen, an increase of more than 1,600%. Tourism makes money, right? Not exactly: regional and state financing is calculated only on those registered. The City Council charges the IBI and water and garbage rates from its inhabitants, but neither the Provincial Council nor the State allocates an extra euro to the services consumed by such growth. Furthermore, it depends on Teruel for everything and cannot grow much more on a stone hill. A town designed to have no neighbors. The municipal urban planning itself admits it bluntly: “There is hardly any need for a first home”recognizes the document that Alcalá presented to the European contest Europan in 2011, when the town had 513 registered inhabitants and already admitted peaks of up to 5,000. María Amparo Atienza Chisbert, from the PAR, the Aragonese Party, governs from the 2023 municipal elections. And they recognize a model problem that is suffocating them. The mayor of Cosuenda (Zaragoza) and that of Canfranc (Huesca) described exactly the same asphyxiation with governments of a different color. The El Castillejo golf coursemunicipal, 9 holes, inaugurated in 2003, is the highest in Spain and has incredible views. The Valdelinares ski resort, the closest and with 14 slopes, opened in 1970. And neither facility exists for the 382 winter residents. Along with the historic center, scattered seasonal occupation developments grew since the 80s and 90s, chalets that clean and open at Easter, in August and on snow bridges. The rest of the year, as you know: closed tight. This summer, there is also an eclipse. To finish with a certain irony, it is worth remembering that the Gúdar-Javalambre region, where Alcalá is, isIt has become one of the seven official points in Aragon to see the total solar eclipse on August 12. The area dresses up before the first total eclipse visible on the peninsula in more than a century. Aragon expects about four million visitors just for the phenomenon. Rural reservations in the area already exceed 90% for the days of the eclipseand some rural houses will bill in August 2026 up to four times more than a normal August. When it’s all over, the cold will return, the streets will be empty, the store and pharmacy on duty will no longer have to double stock, and will endure the downpour until the next bridge. The municipality has its own historical story – they fell at the hands of the anarchist “The Avengers” of the Iron Column, so that months later General Varela took the town under Franco’s mandate – and is today one of the many victims of a self-fulfilling wish regarding tourism in an emptied Spain. The national context. According to estimates, Spain will close 2026 with around 100 million tourists, a historical record, and between June and September alone, 43 million international arrivals are expected. Tourist housing reservations Airbnb types have doubled since 2018while hotel overnight stays only increased by 8% in the same period. Yes, since May 2026, all tourist homes advertised on platforms need a unique registration number, the NRUA, required by European regulations, but this has not stopped growth. Although in many regions vacation rentals are “losing its appeal“, the Bank of Spain itself warned that this type of accommodation displaces the traditional rental market, as is the case of Alcalá de la Selva, a miniature version of a country that builds for those who visit it three weeks a year, and that bills cleaning, water and roads to those who stay to live the rest of the remaining three hundred and forty-five days. It is easy to think that cities like Malaga live in worse situations, with almost 30% of the rental market. The reality is that, proportionally, Alcalá de la Selva is congested by more dangerous numbers. Image | Alcalá de la Selva City Council In Xataka | Spain promised them happiness with its airports increasingly full of tourists. Until someone calculated how it affects rents

news and everything that changes in ChatGPT with the new version of its artificial intelligence model

by

Let’s tell you what are the news GPT-5.6the new version of the model artificial intelligence of ChatGPT. We are going to do it in a simple way, synthesizing in a list what the improvements are so that you can understand what will change when the model starts to arrive. The first thing you should know is that GPT 5.6 comes with three variants different, each with its own characteristics. But we are going to try to explain all this to you in the simplest way possible. What’s new in GPT-5.6 Next, we are going to give you a list with the main news that brings this new version of the OpenAI artificial intelligence model. We are going to do it in list format with a brief explanation of each news so that it is easier to understand. A family of 3 models: GPT 5.6 has three variants called Sol, Terra and Luna. Sol is the most advanced model, Luna is designed for greater speed and a lower price in tokens, and Terra is the intermediate model that seeks a balance between performance and cost. GPT-5.6 Sun: It is the new OpenAI flagship model, designed to improve especially in programming, scientific research, knowledge-based professional work, computer use and cybersecurity. GPT-5.6 Terra: Terra costs half as much as GPT-5.5, but maintaining competitive performance against that model. It is the most balanced option, the standard for doing general tasks. GPT-5.6 Moon: It becomes the fastest and cheapest OpenAI model. This makes it especially good for use in applications where speed and cost take top priority over more robust and advanced capabilities. Sol has Max and Ultra modes: The Sol model also has different modalities. The Max gives you more time to reason, check and revise your approach. Meanwhile, Ultra coordinates four agents in parallel by default, giving you better results in less time, but spending more tokens. Software Engineering Improvements: OpenAI ensures that GPT-5.6 especially improves in software engineering, where it increases its capabilities to solve more advanced programming tasks, as well as complex development flows. Improvements when using your computer: AI is lately focused on agents, so it is not surprising that its capabilities have also expanded to work with interfaces and execute tasks within computing environments. A big leap in design: With only general indications, this new model can create functional and careful interfaces. Additionally, the aforementioned computer capabilities allow you to inspect and refine the rendered output for visual glitches before submitting the work. Best scientific research: This area has also been improved, with improvements aimed at scientific reasoning and specialized work. Best knowledge-based jobs: OpenAI highlights advances in knowledge-based professional work, with the goal of delivering better results in complex tasks performed by professionals. You can create editable presentations from scratch, and improve results when you follow templates and reference documents to apply to new content. What’s new for developers in the API: In the Responses API, the Programmatic Tool Calling function allows GPT-5.6 to write and execute in-memory programs that coordinate tools and process intermediate results. And the multi-agent feature, initially in beta, allows you to run simultaneous subagents and synthesize their work into a single request. Cybersecurity news: They also improve cybersecurity capabilities, allowing more complex tasks to be solved within this area. More security for AI: OpenAI strengthens protections against high-risk activities, sensitive cybersecurity-related requests, and potential repeat misuse. It also incorporates the most robust security system in the history of GPT, with several layers of protection to reduce risks during the use of the model. Automated network teaming tests: OpenAI claims to have spent several weeks doing automated red teaming, searching for vulnerabilities and stress testing the system. With this, the model has been strengthened against real attacks before its launch. Availability according to your plan: In ChatGPT, Plus, Pro, Business and Enterprise users can access GPT-5.6 Sol, and Pro and Enterprise users can also choose GPT-5.6 Sol Pro. In ChatGPT Work and Codex, Free and Go users use Terra, while paid plans can choose between the three models and adjust the effort level of each. The deployment began on July 9 and will be gradually completed within 24 hours.

GPT-5.6 is probably the best AI model in the world. And precisely for that reason, the majority does not need it.

by

Yesterday OpenAI publicly released GPT-5.6its new family of AI models with three variants: Sol, the most powerful, Terra, more balanced, and Luna, the most cost-efficient. One idea stood out in that release: that GPT-5.6 is probably the best model in the world. And precisely for that reason, the vast majority will never need it. As part of the launch, he published a nice video in which he showed how a farmer in Japan, an entrepreneurial couple in New York and a mathematician in Poland had used it for their work. Then we go back to the video and those three scenarios. But a preview: two of those three stories demonstrate just the opposite of what OpenAI wanted to demonstrate. In the official announcement OpenAI also told us about how this was the most capable family of AI models they had ever released and they included the traditional huge string of internal benchmark results to prove it. According to internal tests, GPT-5.6 Sol is the best existing AI model both in programming and in the use of agentic tools in the terminal (among many other scenarios). Source: OpenAI. Their data revealed that we are facing what theoretically it is the best AI model in the world currently. And the interesting thing is that independent studies like those of Artificial Analysis They corroborate it: in several of its tests GPT-5.6 even surpassed Fable 5, Anthropic’s frontier model that until now was the great reference in this industry. Source: Artificial Analysis. The model certainly appears to be spectacular. Those responsible for ARC Prize, that benchmark in which most AI models repeatedly crash, commented how GPT-5.6 Sol was still the first frontier model to solve one of the puzzles of their new benchmark, ARC-AGI 3. No other had come close to that milestone, and according to this organization “it is the best model when it comes to orienting yourself in a situation that you have never encountered.” All that these tests validate is the idea that we are facing a prodigious model. And the problem is precisely that: that most users will probably never need it. Too powerful for most of us Let’s go back to the video at the beginning. Of the three use cases mentioned, two are quite trivial. GPT-5.6 helped the Japanese farmer create a remote control system for his greenhouse with a Raspberry Pi. He helped the New York couple build a curious cereal box business. Nothing in those two tasks seems to require the best model in the world. In fact, they are precisely the type of projects that have been being resolved for months with much cheaper models. With the third scenario, that of the Polish mathematician, things change: this academic was trying to solve a conjecture that he had been working on for three years. No previous model had been able to help him, but with GPT-5.6 he managed to reveal a totally new idea, he says. One of his final comments precisely makes it clear who GPT-5.6 Sol is for: “If you have that kind of audacity to try to do something really big, you won’t be scared of the incredible computing power because you can organize it with the model.” That is the key to the issue: most users are not trying to solve mathematical conjectures that are almost impossible to solve. Most we use tools in a much more everyday wayand that is completely logical and reasonable. That’s why there are many more more modest and affordable models, and why the GPT-5.6 Sol, even if it makes sense, will be a very unprofitable model for most people. It is not a model for counting R’sof course. In fact, every time a new model comes into our hands, It is very difficult to appreciate if it is really better than the previous ones because the tasks we propose are usually solved very well with the existing ones. There are cases in which differences are seen—in especially complex programming, for example—but here we are faced with a situation that we have lived before on several occasions. This happens, for example, with modern hardware: very few people need the most advanced processors or an RTX 5090 to play, because more modest CPUs and GPUs give access to a truly fluid experience. We don’t usually need a camera either. Hasselblad of 15,000 euros for our vacations, and a good cell phone of 500-1,000 euros at this point solves the problem wonderfully. The good thing about all this is what also happens with those examples that we mentioned before: what is extraordinarily expensive and powerful today will end up no longer being so because other even better (and probably more expensive) AI models will appear on the horizon. The question is no longer “which model is more “intelligent”?”, but “Which model is smart enough to solve this task at the lowest cost?“. That explains why there are variants like Sol, Terra and Luna. Maybe in two years GPT-5.6 Sol will be the cheap model we use to correct an email or plan a vacation. The recent history of AI invites us to think precisely that: today’s frontier models end up becoming tomorrow’s everyday models. Perhaps that is the true meaning of GPT-5.6. Not that today almost no one needs so much intelligence, but that in a few years we will probably we all take it for granted. In Xataka | OpenAI just launched GPT-Live: ChatGPT voice mode has learned to listen, shut up and respond better

Alibaba’s Qwen AI model is the new crown jewel. The only problem is that they don’t make money from it.

by

Jack Ma he returned like a prodigal son to the international scene in February of last year. He did so at an event with the president of China, Xi Jinping, and that represented an important support for the technological leader. Since then his company, Alibaba, has not stopped flooding us with its open AI models, Qwen, but that strategy is having the same lights and shadows as its competitors. Devastates downloads. Alibaba launched its Qwen family of models in 2023 and released them with open weights almost immediately. That made them a perfect alternative to be able to use them locally and to also be able to tune them in all types of scenarios. In January 2026, Qwen was already the open AI model most downloaded in the world with close to a million daily downloads according to data from Hugging Face. By the end of 2025, Qwen was already the most downloaded open weight model in the world by far. Source: AI Base. But popularity does not equal income. In the first quarter of 2026, Alibaba indicated that it had obtained revenue of $1.3 billion related to AI, just 4% of its total revenue. The figure is very short, especially if we take into account that Alibaba plan to invest $55 billion in AI infrastructure by the end of 2027. Investors want profits now. Company shares have fallen 37% on the Hong Kong stock market this year: investors are clearly concerned about the lack of return on this bet on AI, something that we have also been seeing for some time in the US market. Internal divisions. The pressure to turn these open models into a business seems to be dividing Qwen’s own team. In March its chief engineer, Lin Junyang, announced his departure from the companyand he was followed by several key engineers amidst internal disagreements on how to monetize the model. The company has already started move token in the field of proprietary models: in April already released three of them in a few days. War with the US and Anthropic. Meanwhile, the company also faces external pressures. The Pentagon has included it on a blacklist of companies that Washington says support the Chinese military, something Alibaba denies. Additionally, Anthropic recently sent a letter to US senators accusing Alibaba of try to copy their technology using 24,000 fraudulent accounts. Alibaba has declined to comment on the matter. Alibaba has the same problem as OpenAI. Richard Lin, vice president of the company Datastrato, has been involved in the panorama of open AI models in China for some time, and his message reminds us of a palpable reality not only for that market, but for all startups and AI companies: “At the moment there are no AI companies with a sustainable business model. It is not a healthy industry.” The message is as true as it is forceful, but all AI companies would surely respond with what Zuckerberg said: “Losing a couple hundred billion dollars would be a bummer, but that’s better than being left behind in the race for superintelligence.” The funny thing is that they are both (probably) right. Image | qwen In Xataka | China’s open AIs aren’t “beating” ChatGPT, they’re doing something more important: catapulting their industry

Experts already claim that the Chinese GLM-5.2 model is as “dangerous” as Anthropic’s

by

The technological gap between the US and China continues to narrow. At least, if we pay attention to what they say the latest analyzes on the GLM-5.2 model. Two independent cybersecurity companies have made their own assessment and their data reveals that in terms of cybersecurity, GLM-5.2 is as good as Claude Opus 4.8. That has notable implications, especially considering how the US government is now restricting access to Anthropic and OpenAI’s frontier models. AI in the face of the threat of cybersecurity. Since Claude Mythos Preview appeared, the discourse on AI has changed significantly. Suddenly the world realized that these models could become weapons with which to find vulnerabilities in all types of systems to exploit them. Anthropic has already warned that Mythos was too dangerous to be publicly available, and it did not matter that it released hidden versions like Fable 5 shortly after: the US Government has temporarily vetoed them and the same has happened with GPT-5.6. The situation is unusual. Beware of Tulongfeng (or not). Last Wednesday, a Chinese cybersecurity company called 360 Security Technology (Qihoo 360) launched a new vulnerability detection tool called Tulongfeng. According to its creatorsTulongfeng is comparable Mythos in this task. The company this on the US “Entity List” since May 2020 and its CEO, Zhou Hongyi, stated that Mythos is equivalent to a “cybernuclear weapon.” The Sputnik moment with GLM-5.2. But the real recent protagonist of the Chinese AI industry is the GLM-5.2 model from the startup Zhipu.ai (Z.ai), which is becoming very popular by demonstrating performance comparable to the best models from US companies. Its fundamental advantage is that it is an open weights model: any person or company can download it, modify it and run it on their own hardware (although it requires a huge amount of video/unified memory to be able to use it, it is a model with 744B of parameters). But he is not only good at programming or at agentic tasks. Better than Claude in cybersecurity? The cybersecurity firm Semgrep stated in a recent analysis that GLM-5.2 was superior to Claude Opus 4.8 regarding cybersecurity and pointed out that “we have a Mythos at home.” Another independent study from Graphistry stated basically the same thing when comparing it with Opus 4.8 and GPT-5.5. Not only that, it achieved excellent results at a fraction of the price: one-sixth of what it cost to run tests with Claude Opus 4.8, for example. Axios revealed little cited a cybersecurity researcher who explained that GLM-5.2 is capable of chaining exploits “in the same way that an elite human attacker would.” Chinese mythos before 2027. Jie Tang, CEO of Z.ai, responded to a Twitter thread in which it was pointed out that at this rate, China would have an AI model at the level of Mythos or Fable by the end of 2026. Elon Musk himself intervened saying that in his opinion this Chinese model with such performance would arrive in the first quarter of 2027. Jie Tang was forceful and replied to Musk saying “it won’t take that long.” And we also have Sakana Fugu. These days we also learned the news that Sakana AI, a Japanese AI startup, had launched Fuguan AI model that is actually not so much an AI model as it is a router or orchestrator of other models. What it promises It is to perform at the level of the best models in the US, taking advantage of different models, both open and closed. The internal benchmarks are promising, but some independent analysts they explained Although the idea is not bad, its performance and cost are not as striking as the company claims. While the US blocks its models, China advances. The situation is paradoxical, because what China is doing is precisely taking advantage of a unique moment. The US is restricting the deployment of the most advanced AI models from Anthropic and OpenAI to avoid cybersecurity risks. And while that happens, Chinese companies are apparently closing the gap with truly remarkable open models. In Xataka | The prompt engineering fashion is over. Now what is important is loop engineering

We believed that no Chinese AI model would soon come close to Fable 5 or GPT-5.5. Then GLM-5.2 arrived

by

A few days ago, the Chinese startup Zhipu AI (Z.ai) announced the launch of its new open AI model, GLM-5.2. It did so boasting amazing features that brought it very close to the best closed models from OpenAI and Anthropic, something that seemed impossible. Well, the more analysis is carried out on the model, the better off it is. We may be at the beginning of something very important. A change of trend. GLM 5.2. The Chinese startup Z.ai has been releasing different versions of its GLM AI model for a long time, but the latest one is undoubtedly the most surprising because its performance is especially promising. It has 744,000 million parameters (744B), of which 40,000 are those that remain active. We are looking at a model with a context window of one million tokens and a new architecture called IndexShare/IndexCache. Better than GPT-5.5, very close to Opus 4.8. The startup showed how the performance of GLM-5.2 is extraordinary in programming tasks. In the FrontierSWE test, the most demanding of those currently available, GLM-5.2 outperformed GPT-5.5 and only Opus 4.8 was superior by a very small margin. The same happened with other tests such as PostTrainBench or SWE-Marathon, which, for example, evaluates the behavior of the model in very long autonomous programming sessions. Source: Z.ai. In many other tests the photo was identical: the model has made a spectacular leap since version 5.1, and is in many tests almost as good (or better) than the best from OpenAI, Anthropic or Google. But it’s not just them who say it.. Artificial Analysis, a reputable independent firm that maintains an updated ranking of the performance of the new AI models that are arriving on the market, confirms the data of Z.ai itself. In his tests he indicates how the “intelligence index” of GLM-5.2 is now 51 points. It is only surpassed by GPT-5.5 (55), Claude Opus 4.8 (56) and Claude Fable 5 (60). Source: Artificial Analysis. This Chinese open model leaves behind the new Gemini 3.5 Flash, but also Chinese competitors such as Qwen 3.7 Max, MiniMax-M3 or DeepSeek V4, among others. The jump in quality from GLM-5.1 is, we insist, outstanding, much greater than what, at least according to this index, was seen from Opus 4.8 to Fable 5. The jump in performance is spectacular, although it is true that the comparative price to solve the tasks proposed in the benchmark rises significantly. Source: Artificial Analysis. But it’s not perfect. The Artificial Analysis report, however, shows that although GLM-5.2 is very strong in areas such as programming, it is weak in others. For example, it is far from being as reliable as Fable 5, GPT-5.5, Claude 4.8 or Gemini 3.1 Pro in terms of correct answers, which is also lower in proportion to that of its competitors. However, his hallucinations have significantly reduced. And it’s much (much) cheaper. But in addition to being fantastic in many areas, it is much cheaper than its competitors. Maintains the price per million input/output tokens of its predecessor ($1.4/4.4), while that of GPT-5.5 It’s 5/30 dollars and that of Opus 4.8 10/50 dollars. It is true that it consumes many more tokens than GPT-5.5 (very efficient) or Claude Opus 4.8, but even with that its final cost is much lower. My tests with GLM-5.2 programming. I’ve been a Z.ai subscriber for months now because they offered an annual subscription at the end of 2025 at a really low price. This has allowed me to test GLM-5.2 for a few hours and although I cannot draw definitive conclusions, it does seem clear that there is a leap in quality in terms of its ability to program. I asked him to review a personal code project and he identified several security flaws and possible improvements in great detail. Chatting with GLM5-2. In conversational mode the behavior is much more difficult to evaluate: I have been interacting with the model and asking it questions, and although it is better than GLM 5.1 many times, other times it is not so much and I would say that in terms of creativity to write the frontier models of Google, OpenAI and especially Anthropic they are still quite superior. You can try it on their websiteand there you will see something else: it takes significantly longer to respond than other chatbots, because its reasoning phase is longer. Take more time to answer questions. Benchmarks are one thing, experience is another.. In the absence of testing it (much) more, of course the impression is that the model has improved significantly compared to a GLM-5.1 that had lagged behind its Chinese competitors (not to mention the current Claude Opus 4.8 or GPT-5.5). On platforms like Reddit opinions are dividedbut many consider it a fantastic option to run locally… if you have a very, very powerful machine with at least 256 GB of unified memory (Mac Studio). And one thing seems clear: when using it as an AI model for programming, comes surprisingly close to Claude Opus 4.8. In Xataka | Chinese technology companies entered the AI ​​race with cheaper models than the rest. That’s starting to end

everything we think we know about Apple’s new base model

by

In September the new iPhone 18 Pro and it will be the debut of John Ternus as the new CEO of Apple. And… that’s it. Maybe we’ll see the rumored foldable iPhonebut the iPhone 18 will have to wait. Breaking with the tradition of their releases, the basic iPhone 18 will have to wait until sometime in the spring of next year. However, that does not mean that the wheel of rumors and leaks is not working. Next, we tell you everything we think we know about that iPhone 18 which, according to those rumors, will not be a revolution, but it will have a couple of features that will make it more interesting than the current ones. iPhone 17 and iPhone 17e: new camera and RAM memory. Let’s go to trouble. What to expect from the iPhone 18? Design: two aspects in the leaks with a design maintained in the most conservative leak and an island that will follow the design of the Pro models, but with two cameras instead of three to further differentiate this iPhone 18 from the iPhone 17 and iPhone 16. Screen: 6.3-inch diagonal with LTPO refresh rate up to 120 Hz with a Dynamic Island that will maintain the size of the one we had until now. On the iPhone 18 Pro we are supposed to have a smaller Dynamic Island. SoC: Apple A20 with reduced features in GPU and cores compared to the A20 Pro that the iPhone 18 Pro will carry. RAM: 12 GB integrated on the chip wafer to allow good performance of the new Siri AI and Apple Intelligence. Cameras: dual camera configuration with a wide angle that will be brighter and so on, but that will pale in comparison to a new main sensor that will also be in a camera with a variable aperture. Connectivity: The new C2 chip would be the heart of the networks of the new iPhone 18. Price: It is expected to remain around what we already have with the iPhone 17, without exceeding 1,000 euros in the basic 256 GB version. Another story will be the version with 512 GB, which may suffer due to the increase in the price of components. Launch: Obviously, without anything confirmed, but all the leaks for a long time point to a launch in spring 2027. When would the iPhone 18 come out? The forecast is that Apple, for the first time, will break its classic release cycle. Instead of launching the two iPhone 18s in September, the one that will arrive first will be the iPhone 18 Pro, while the iPhone 18 will be released sometime in spring next year. In fact, this late launch would imply that in the September keynote Apple will not present the iPhone 18 in depth. It should mention it so that the user is clear that it will arrive, but leaving all the details for a later presentation closer to the launch in spring. What design will the iPhone 18 have? The leaks point to a conservative design, as usually happens. Apple has already created a visual identity for the Pro with a giant side-to-side rear camera island, while the basic iPhone 18 has the cameras separately and vertically. There may be some changes such as cameras next to each other or with a similar design to the Pro’s island, but keeping only two cameras. At the moment, there is only speculation here, but if the trend continues, it is easy to see that change on the back because, if not, Apple would repeat the design of the iPhone 16 for the third year. And that is not something that Apple people usually do. What will the screens of the new iPhone 18 be like? On the front, the iPhone 18 Pro is expected to feature a smaller dynamic island, but right now there are no leaks about that element of the iPhone 18. There have been reported some problems when carrying out this miniaturization, which would raise the price of the device and is something that, for the ‘cheaper’ iPhone 18, Apple would not want to allow itself. For the rest, the leaks point to an LTPO screen with a refresh rate up to 120 Hz and a diagonal of 6.3 inches. There are fewer leaked details (in general) than for the iPhone 18 Pro, something logical due to this supposed time lapse between one model or another, and it’s not like we have too much information on the screen either. The easiest? That the Pros do have that smaller Dynamic Island, but that with the iPhone 18 Apple maintains the size. What processor will the iPhone 18 have? Here what we can expect is a A20but with some cores and frequencies cut compared to the A20 Pro that, supposedly, will mount the iPhone 18 Pro. It will be a 2 nanometer chip manufactured by TSMC and, beyond the expected increase in power thanks to the new SoC, perhaps the most notable thing is the issue of RAM. From 8 GB, we would go to 12 GB of memory. They are the same ones that already have the iPhone 17 Prohe iPhone Air and those expected to mount the iPhone 18 Pro and the explanation when it comes to matching the amount of RAM is marked by the company’s ambition with the new Siri AI. The iPhone 18 will be mobile phones launched with the new Siri and Apple Intelligence in place and, to perform some actions locally, a considerable amount of RAM is necessary. It is clear that 12 GB is needed to have all the Siri AI and Apple Intelligence options The iPhone 17’s 8GB is adequate for certain tasks, but not for the more advanced Siri AI tasks (as Apple itself has detailed) and going up to 12 GB would be ideal to maintain parity in that experience with that AI that they are going to push so much from now on. What battery will … Read more

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.