models

Their models not only perform just as well but are much cheaper

by

Sometimes it is difficult to keep up with the innovations that artificial intelligence is leaving us. However, it is also a good sign that there is voracious competition and that, even if we thought that companies like OpenAI or Google had every chance of completely dominating this sector, not all the fish has been sold yet. And what began as a race to create the most powerful models has become a battle to deliver the best performance at the lowest possible cost. And in this new competition, China leads. Change of third. For years, the conversation in AI focused solely on which model was more capable: who passed more benchmarks, who solved more complex problems, who generated better answers. But that phase is being relegated by another in which price is once again a determining factor in making decisions. This transition marks a turning point, since it is mainly Chinese startups that are demonstrating remarkable ability to produce powerful and extraordinarily economical models. Qwen leads the revolution. As Kai Williams highlights in the ‘Understanding AI’ newsletterAlibaba’s open model ecosystem, known as qwenhas become the most downloaded model family in the world, according to Hugging Face data analyzed by the ATOM Project. “Qwen alone is roughly matching the entire US ecosystem of open models today,” counted Nathan Lambert, researcher at the Allen Institute for Artificial Intelligence, at the PyTorch conference. The Chinese company has achieved something that seemed difficult: create competitive models in practically all sizes, from small to 235,000 million parameters, offering options for any business need. Real business adoption. In addition to the technical figures of their models, the use cases also deserve special mention. In October, Brian Chesky, CEO of Airbnb, caused a bit of a stir stating that his company “relies a lot on Alibaba’s Qwen model” because it is fast, cheap and powerful enough. This statement is interesting because of the context, since we are talking about a top American company preferring to use an open Chinese model, something that is powerful enough to change the perception of the industry. Williams points out in the text that, in addition to Airbnb, there are also other companies that would prefer to adopt Qwen’s models, but for reasons of image or regulatory compliance, they cannot. Mainly because they are models from China. And that could be the great barrier for Qwen and the rest of the Chinese models that would make its adoption difficult. So Chinese startups have a big job ahead of them to try to change that perception in an increasingly complicated geopolitical context. Kimi K2 gave the surprise. If Qwen dominates through volume and versatility, Kimi K2 Thinking It stands out for being possibly the best open model in the world in terms of benchmark test scores. Just like share Williams in the newsletter, Artificial Analysis currently ranks it as the most powerful model not created by OpenAI, Google or Anthropic. DeepSeek and the domino effect. The launch of DeepSeek R1 in January was the catalyst that unleashed this wave. It arrived just four months after OpenAI announced your first reasoning model, o1but with a crucial difference: DeepSeek openly published the model parameters. The noise was such that even the DeepSeek app briefly surpassed ChatGPT as the most downloaded app in the iOS App Store, Nvidia shares fell almost 20% days later, and Chinese companies rushed to integrate the model into their products. Today we are still waiting for DeepSeek to speak again with its next deep reasoning model, about which not much is known yet. On the other side, putting out fires. The United States has not sat idly by. When it comes to open weight models, OpenAI launched theirs in AugustIBM published its Granite 4 models in October, and Google, Microsoft, Nvidia and the Allen Institute for AI have also introduced new ‘semi-open’ models this year. But none have reached the level of the main Chinese open models. Lambert, who has led efforts to advance a new generation of American open models, acknowledges that progress has been slow and the gap is widening. Everything indicates that 2026 will be decisive in determining the pace of adoption by companies and, above all, in determining the choice of model in an increasingly immense ocean. Cover image | Xataka with Mockuuups Studio and Kimi AI In Xataka | NVIDIA begins to make moves with China: it is considering increasing production of the H200 in the face of an avalanche of orders, according to Reuters

V16 beacons with associated application. Five models to comply with DGT regulations as of January 1

by

There is nothing left for us to have to carry a mandatory one in the car. V16 beacon. Which one to buy? There are many models, so in this article we are going to focus on five V16 beacons that come with their own smartphone application. Help Flash IoTan affordable beacon compatible with myIncidence. Help Flash IoT+a beacon with a good number of candles that is also compatible with myIncidence. LEDOnea V16 beacon that stands out for its format and is compatible with its own app. LEDOne for trucksincludes the beacon and an arrow for commercial vehicles. FlashLEDcomes with an anti-shock case and is compatible with its own app. Help Flash IoT The first V16 beacon that we have introduced in this list is from Netun Solutions: the Help Flash IoT. It has a similar design to the Help Flash IoT+, but has different characteristics: it offers visibility up to 1 km, its autonomy is approximately two hours and it offers more than 40 effective candles. Plus, it connects to the app myIncidence to be able to quickly contact the insurance company and emergency services. The price could vary. We earn commission from these links Help Flash IoT+ Secondly, we have introduced the most complete version of the previous V16 beacon: the Help Flash IoT+. It has a more interesting technical sheet and its price is usually similar: it offers approximately 290 candlesits autonomy is up to 2.5 hours and it also connects to the myIncidence app. The price could vary. We earn commission from these links LEDOne The LEDOne It is a particularly interesting beacon due to its format, since it incorporates a support so that it can be placed a little higher, thus improving visibility. Offers 120 effective candlesits autonomy is approximately two hours, the brand mentions that it is suitable for all vehicles and can be connected to the LEDOne app to notify insurance and emergency services. The price could vary. We earn commission from these links LEDOne for trucks An alternative to the previous V16 beacon—or rather a more complete option—we have it at Leroy Merlin. The LEDOne is available in a truck pack which, in this case, includes both the beacon that we have mentioned before and a signaling for industrial vehiclesthus improving visibility. In addition, since it is the same beacon, it is compatible with the LEDOne app. The price could vary. We earn commission from these links FlashLED Finally, PcComponentes has the V16 beacon FlashLEDwhich in this case comes along with a anti-shock hard case. It works using a single battery and is compatible with its own app SOS alert. Of course, the brand does not mention either the theoretical autonomy or the figure in candles. The price could vary. We earn commission from these links Some of the links in this article are affiliated and may provide a benefit to Xataka. In case of non-availability, offers may vary. Image | Netun Solutions, LEDOne, FlashLED In Xataka | Safety, organization and entertainment gadgets and accessories for cars on long trips In Xataka | Clarifying all the mess that the DGT has on its hands: the V-16 light, the V-27 signal and the emergency triangles

The elite of the open models spoke in Chinese. Mistral has just placed Europe at a level that not even the US managed to reach

by

Over the last year, the elite of open models for assisted programming, at least in benchmarks as SWE-Bench Verifiedhas spoken with a Chinese accent. Names like DeepSeek, Kimi either qwen They had settled into the top positions in testing and were setting the pace in complex software engineering tasks, while Europe was still searching for its position. The arrival of Devstral 2 alters that distribution. It does not displace those who were already at the top, but it places Mistral at the same level of demand and turns a European company into a real contender in a field that until now seemed reserved for others. League change: the technical leap that had been brewing for some time. During recent months, the open models developed in Europe and the United States had shown constant evolution, although still without the performance necessary to compete in the most demanding tests. The progress was evident, but there was a lack of a project capable of consolidating it at a higher level and demonstrating that this path could give results comparable to those of the sector. Devstral 2 in data: performance, size and licenses. The new Mistral model reaches 123B parameters in a dense architecture and offers an expanded context of 256K tokens, accompanied by a modified MIT license that facilitates its adoption in open environments. Its compact version, Devstral Small 2, reduces the model to 24B licensed parameters Apache 2.0. In the SWE-Bench Verified figures published by the companyDevstral 2 obtains 72.2%, a mark that places it in the most competitive section of the open models evaluated and that confirms its presence among the most advanced alternatives in the segment. It is reflected by a panorama concentrated in the upper part of the benchmark. Among the open models, DeepSeek V3.2 leads the group with 73.1%, followed by Kimi K2 Thinking with 71.3% and for proposals such as Qwen 3 Coder Plus and Minimax M2, which are around 69 points. At lower levels GLM 4.6, GPT-OSS-120B, CWM and DeepSWE appear, with more moderate results. In the closed commercial environment (proprietary models), the graph incorporates higher scores: Gemini 3 Pro reaches 76.2%, GPT 5.1 Codex Max rises to 77.9% and Claude Sonnet 4.5 scores 77.2%, all of them above the best brands registered for open models. What SWE-Bench Verified Really Measures and Why It Matters. SWE-Bench Verified is a test designed to evaluate whether a model can solve real programming tasks, not synthetic exercises. Each case presents a bug in an open source repository and requires a patch to pass the previously failed tests. The evaluation seeks to measure whether the system understands the structure of the project, identifies the cause of the problem and proposes a coherent solution. It is a useful and demanding metric, although limited to Python repositories and a specific set of situations that do not cover the full breadth of software work. From co-pilots to agents who act on the project. The arrival of Devstral 2 coincides with a broader change in the way of working with programming tools. It is no longer just about receiving suggestions in the editor, but about having agents capable of exploring an entire repository, interpreting its structure and proposing changes consistent with its real state. In this context, Vibe CLI appears, a tool that allows Devstral to analyze files, modify parts of the code and execute actions directly from the terminal, bringing these capabilities closer to the daily workflow of developers. Cost and deployment: what each type of user can do with Devstral. The model will be available for free for an initial period and will then cost $0.40 per million tokens for input and $2.00 per million for output, while the Small 2 version will be priced lower. Its deployment also makes a difference: Devstral 2 requires at least four H100-class GPUs, aimed at data centers, while Devstral Small 2 is intended to run on a single GPU and, according to Mistral documentation, the Devstral Small family can also run in CPU-only configurations, without a dedicated GPU. This variety allows both companies and individual developers to find a suitable entry point. The appearance of Devstral 2 introduces an unexpected element in a space where Chinese companies set the pace and where not even the United States, despite its leadership in artificial intelligence, had an open model in this high performance range in SWE-Bench Verified. Mistral does not displace those who were already at the top, but it does broaden the conversation and shows that Europe can compete in a field where it did not appear until now. It is a movement that does not alter the general hierarchy, although it does open a new margin for the evolution of assisted programming tools. Images | Xataka with Gemini 3 In Xataka | OpenAI and Google deny that they are going to put ads in ChatGPT and Gemini. The reality is that accounts do not come only with subscriptions

To the question of what sense it makes to compete with Google, OpenAI or Anthropic in AI, Mistral has an answer: small and local models

by

French startup Mistral AI Mistral 3 has been launcheda family of 10 open source artificial intelligence models that represent its most ambitious commitment to date. The Parisian company, which is often considered the main European hope in the development of AI, seeks to differentiate itself from the large American technology companies by betting on flexibility and deployment in all types of devices instead of raw power. Under these lines we tell you all the news. What Mistral has presented. The Mistral 3 family includes a flagship model called Mistral Large 3, with 675 billion parameters, and nine compact models grouped under the name Ministral 3 (in three sizes: 14,000, 8,000 and 3 billion parameters). All models are released under Apache 2.0 license, allowing unrestricted commercial use. The large model also has multimodal capacity, being able to process text and images. It is also multilingual, with a special emphasis on European languages. On the other hand, small models can run on devices with just 4 GB of memory, making them perfect for modest laptops, mobile phones and embedded systems without the need for an internet connection. Why strategy matters. While OpenAI, Google and Anthropic focus on increasingly powerful and closed systems with agentic capabilitiesMistral has focused on the breadth and scope of its models, efficiency and what its co-founder Guillaume Lample calls “distributed intelligence.” According to declared told VentureBeat, the company believes the future of AI is defined not by scale, but by ubiquity: models small enough to run in drones, vehicles, robots and consumer devices. The economic and practical argument. Lample explained It means that in more than 90% of cases, a small, specifically tuned model can get the job done, especially if it is trained with synthetic data for specific tasks. According to Lample, this is not only cheaper and faster, but it eliminates concerns about privacy, latency and reliability. The company also has teams that work directly with customers to analyze specific problems and fine-tune small models that perform specific tasks. This, above all, can attract companies that become frustrated when choosing the best possible model for a specific task and, if it does not perform adequately, they end up giving up. Europe is lagging behind. If we talk about innovation and technology around AI, we do not hesitate to say that Europe is leagues away of what companies in the United States and China are offering. This is why Mistral AI advocates a different approach in which it prioritizes massive deployment in devices and the flexibility of its smaller models. The capacity offered by open models can be a great asset to continue betting on these technologies. In China, for example, the open models of DeepSeek, Alibaba or Kimi are emerging widelyabove in certain tasks even competitors as large as ChatGPT. Lample explained that most leading Chinese models are exclusively text-based, with separate image processing systems. For this reason, they also want to opt for a multimodal approach. A complete ecosystem. Mistral no longer only offers language models. The company has built an entire ecosystem that includes Mistral Agents APIwith connectors for code execution, web search and image generation; Masterlyyour reasoning model; Mistral Code for programming assistance; and AI Studioan application deployment platform that also has analytical and logging capabilities. Furthermore, his assistant Le Chat It has incorporated a deep research mode, voice capabilities and a list of more than 20 enterprise integrations. Thus, in addition to its model offering, the company can provide other companies with a whole layer of personalized products and services, with the aim of being their main source of financing. Digital sovereignty. Although Mistral is often characterized as Europe’s answer to OpenAI, the company prefers to consider itself as ‘a transatlantic collaboration’. Its CEO, in fact, is in the United States, has teams on both continents and trains these models in collaboration with American teams and infrastructure. However, its positioning as a defender of European digital sovereignty has earned it strategic partnerships with the French army, the country’s employment agency, the Luxembourg government and various European public organizations. The European Commission presented in October a strategy to promote European AI tools that provide security and resilience while boosting the continent’s industrial competitiveness. Offline capabilities for democratization. The use cases that Mistral has designed for its small models include, above all, local applications, such as factory robots that use sensor data in real time and without relying on the cloud, drones in natural disasters or rescues that operate offline, and smart cars with functional AI assistants in remote areas. Lample stood out that there are billions of people without internet access but with laptops or cell phones capable of running these small models, which he considers potentially revolutionary. Additionally, by running on the device, these apps preserve the privacy of user data. Real “open source” debate. Not everyone celebrates Mistral’s approach. Some critics question his decision to opt for models’open weight‘, that is, free to access but providing less information about their code than truly “open source” models, which provide the code and training data necessary to train a model from scratch. Andreas Liesenfeld, assistant professor at Radboud University and co-founder of the European Open Source AI Index, declared to the Financial Times that data at scale is the missing key in the European AI innovation ecosystem and that Mistral does not contribute to that at all. The long-term strategic bet. Lample recognize that their models are “a little behind” the most advanced closed systems, but argued that the important thing is that “they are catching up quickly.” Time will tell if Mistral’s approach to low-cost, versatile models with local applications ends up working for them to end up positioning themselves as one of the great European bets on AI. Cover image | Mistral AI In Xataka | China already has an army of 5.8 million engineers. His new plan involves accelerating doctorates

These are the 13 models that are already updated, and the other nine from Redmi and POCO that will update later

by

Let’s tell you to what models of Xiaomi phones and tablets HyperOS 3 has started to arriveso that if you have one of them you know that you can now start looking for the update. We already told you how to check if your mobile will update using an app, but now the moment of truth has arrived. We are going to start the article with the models where the update of this customization layer based on Android 16. And then we will tell you What are the next mobile phones that will receive it?so that if you have one you know that Xiami also counts on you. Xiaomi phones that already receive HyperOS 3 These are all the mobile phones and tablets from the Chinese company that You can now update to HyperOS 3.0. If you go into your mobile settings and check for system updates, you can start enjoying Android 16 after updating. Xiaomi Pad 7 Xiaomi Pad 6S Pro 12.4 Xiaomi 14 Ultra Xiaomi 14 Ultra Titanium Special Edition Xiaomi 14 Pro Xiaomi 14 Pro Titanium Special Edition Xiaomi 14 Xiaomi MIX Fold 4 Xiaomi MIX Flip Xiaomi Civi 4 Pro Redmi K70 Pro Redmi K70 Ultimate Edition Redmi K70 Redmi K70E The normal thing is that the update arrives without you having to do anything, a notification appears that alerts you that it is available so that you only have to click on it. But if not, in the settings click on about the phonewhere after clicking on your version of HyperOS it will search for updates. Mobile phones that will also receive HyperOS 3 In addition to the models that are already receiving the update, we also have a list of mobile phones that will soon receive it. For them there are no specific dates, but if you have one you should know that yes, you will have Android 16. In Xataka Basics | Android 16: 17 functions and some tricks of the new version of Google’s mobile operating system

The two most important weather models in the world are discussing whether Santander is going to freeze next week. And the cold is winning

by

Where has all the cold gone? So far this fall (with the sole exception of Siberia), temperatures have been relatively mild on all continents. And it seems that the situation is going to continue like this: it is true that the forecasts speak of a progressive decrease in temperatures in the southeast of Canada, the eastern United States and northern Europe; but no model paints a scenario that is particularly cold (except some very long term prediction). However, all eyes are on the polar vortex. If the models are right, it is very possible that the vortex will experience an unprecedented disturbance in November, leading to an interesting weather period starting in December. “There is no way this is fulfilled.” While November continues with its strange meteorology, the models draw increasingly strange scenarios. At this point in the week, we cannot rule out that on the 18th and 19th we have a more than considerable winter storm with the ‘beast from the east‘looming over Western Europe. In the next few hours we will have a war between models: The American marks a cold entry on Santander, the European said no. Little by little, the two seem to be converging towards a cold scene. It’s too early to say, but in a very few hours the daisy will be shedding its leaves. Anyway, the central issue is that all of this is minute sin. The breaking of the vortex. Except for that event in the middle of next week, autumn will continue to be very warm and mild on almost all continents. However, this could change if sudden stratospheric warming appears. That is, the vortex breaks. Sudden stratospheric warming? To understand it simply, we have to remember that the atmosphere is a kind of “lasagna of air layers” and each of them follows its own logic. That is, they work quite differently and independently. As far as it affects us: the circulation of air in the troposphere (the one closest to the surface) and the circulation in the stratosphere (the layer directly above) are related, yes; But, in general terms, they each do their own thing. During the “sudden stratospheric warming“, a part of the troposphere warms rapidly and, as a consequence, invades the stratosphere, causing a profound alteration of the circulation at high altitude. That is, for a few days, everything turns upside down. And what happens? The most common consequence of this is that the polar vortex weakens and may break down. The polar (arctic) vortex is a current of air that runs from west to east around the north pole and contains cold air at high latitudes. When this current is strong and stable, preventing it from flowing towards places like Spain. If the vortex It destabilizes and its winds lose strength (due to, for example, “sudden warming”), it is relatively common for cold air masses to escape on their way south. What if it doesn’t break? In reality, the vortex does not even need to break. It only needs to move from the Arctic region to lower latitudes. By moving a huge mass of cold air with it, the result is always very similar: an icy cold that can turn any country upside down (even the best prepared ones). And that seems to be what we are going to see. It’s hard to know if it will affect us or not, but there’s no doubt that the late fall weather is getting “interesting.” Image | Meteociel In Xataka | The last hope of winter in Spain is desperate, but increasingly possible: the breaking of the polar vortex

The industry became obsessed with training AI models, while Google prepared its masterstroke: inference chips

by

In recent years, what was truly relevant was training AI models to make them better. Now that they have matured and training it no longer scales as noticeablywhat matters most is inference: that when we use AI chatbots they work quickly and efficiently. Google realized this change in focus, and has chips precisely prepared for it. Ironwood. This is the name of the new chips from Google’s famous family of Tensor Processing Units (TPUs). The company, which began developing them in 2015 and launched the first ones in 2018now obtains especially interesting fruits from all that effort: some really promising chips not for training AI models, but for us to use them faster and more efficiently than ever. Inference, inference, inference. These “TPUv7” will be available in the coming weeks and can be used to train AI models, but they are especially aimed at “serving” these models to users so that they can use them. It is the other big leg of AI chips, the really visible one: one thing is to train the models and quite another to “execute” them so that they respond to user requests. Efficiency and power by flag. The advance in the performance of these AI chips is enormous, at least according to Google. The company claims that Ironwood offers four times the performance of the previous generation in both training and inference, and is “the most powerful and energy-efficient custom silicon to date.” Google has already reached an agreement with Anthropic so that the latter has access up to one million TPUs to run Claude and serve it to its users. Google’s AI supercomputersand. These chips are the key components of the so-called AI Hypercomputer, an integrated supercomputing system that according to Google allows customers to reduce IT costs by 28% and a ROI of 353% in three years. Or what is the same: they promise that if you use these chips, the return on investment will be multiplied by more than four in that period. Almost 10,000 interconnected chips. The new Ironwoods are also equipped with the ability to be part of joining forces in a big way. It is possible to combine up to 9,216 of them in a single node or pod, which theoretically makes the bottlenecks of the most demanding models disappear. The size of this type of cluster is enormous, and allows for up to 1.77 Petabytes of shared HBM memory while these chips communicate with a bandwidth of 9.6 Tbps thanks to the so-called Inter-Chip Interconnect (ICI). More FLOPS than anyone. The company also claims that an “Ironwood pod” (a cluster with those 9,216 Ironwood TPUs) offers 118x more ExaFLOPS FP8 than its best competitor. FLOPS measure how many floating-point math operations these chips can solve per second, ensuring that basically any AI workload is going to run in record times. NVIDIA has more and more competition (and that’s a good thing). Google chips are a demonstration of the clear vocation of companies to avoid too many dependencies on third parties. Google has all the ingredients to do it, and its TPUv7 is proof of this. It’s not the only oneand many other AI companies have long sought to create their own chips. NVIDIA’s dominance remains clearbut the company has a small problem. In inference CUDA is no longer so vital. Once the AI ​​model has been trained, inference operates under different game rules than training. CUDA support remains a relevant factorbut its importance in inference is much less. Inference focuses on obtaining the fastest possible answer. Here the models are “compiled” and can run optimally on the target hardware. This may cause NVIDIA to lose relevance to alternatives like Google. In Xataka | When you’re OpenAI and you can’t buy enough GPUs, the solution is obvious: make your own

which models are going to update to the next version

by

Let’s tell you which Motorola phones are going to update to Android 16, now that the manufacturer has begun to deploy the new version of the operating system. In this way, if you have a mobile phone of this brand we will leave you the list where you can see the models they receive Android 16. Motorola is going to update almost thirty mobile models to the new version of Android, including both its flagships and other models that have been launched in the last three years. Thus, if your mobile model appears on this list, you will receive the update in the coming weeks or months. Motorola phones with Android 16 Motorola has a policy to cover two years of operating system updates. If your mobile is on this list you will receive Android 16although the time to wait will depend on each model. In the cases of the latest high ranges the updates will arrive soon, while in others it may take several months for them to appear. These are the mobile phones that will be updated between the remainder of 2025 and the beginning of 2026: Motorola Edge 60 Pro Motorola Edge 60 Motorola Edge 60 Fusion Motorola Edge 60 Stylus Motorola Edge 2025 Motorola Edge 50 Ultra Motorola Edge 50 Pro Motorola Edge 50 Neo Motorola Edge 50 Fusion Motorola Edge 50 Motorola Edge 40 Pro Motorola Razr 60 Ultra Motorola Razr 60 Motorola Razr 2025 Motorola Razr Plus 2025 Motorola Razr 50 Ultra Motorola Razr 50 Motorola Razr Plus 2024 Motorola Razr 2023 Moto G86 Moto G86 Power Moto G56 Moto G Power 2025 Moto G 2025 Motorola G Stylus 2025 Moto G85 Moto G75 Moto G55 ThinkPhone 25 Cover photo | Ivan Linares In Xataka Basics | Android 16: 17 functions and some tricks of the new version of Google’s mobile operating system

We already know how to retrieve the exact prompts that people use in AI models. It’s terrifying news

by

A group of researchers has published a study that once again raises alarm bells regarding privacy when using AI. What they have managed to demonstrate is that it is possible to know the exact prompt that a user used when asking a chatbot something, and that puts AI companies in a delicate position. They can, more than ever, know everything about us. A terrifying study. If you are told that ‘Linguistic models are injective and, therefore, invertible’ you will probably be shocked. That’s the title from the study carried out by European researchers in which they explain that large language models (LLM) have a major privacy problem. And it has it because the transformer architecture is designed that way: each different prompt corresponds to a different “embedding” in the latent space of the model. A sneaky algorithm. During the development of their theory, the researchers created an algorithm called SIPIT (Sequential Inverse Prompt via ITerative updates). Such an algorithm reconstructs the exact input text from the hidden activations/states with a guarantee that it will do so in linear time. Or what is the same: you can make the model “snap” easily and quickly. What does this mean. What all this means is that the answer you got when using that AI model allows you to find out exactly what you asked it. In reality, it is not the answer that gives away, but the hidden states or embeddings that the AI ​​models use to end up giving the final answer. That’s a problem, because AI companies keep these states hidden, which would theoretically allow them to know the input prompt with absolute accuracy. But many companies already saved the prompts. That’s true, but that “injectivity” creates an additional privacy risk. Many embeddings or internal states are stored for caching, for monitoring or diagnosis, and for customization. If a company only deletes the plain text conversation but does not delete the embeddings file, the prompt is still recoverable from that file. The study shows that any system that stores hidden states is effectively handling the input text itself. Legal impact. There is also a dangerous legal component here. Until now, regulators and companies argued that internal states were not considered “recoverable personal data,” but that invertibility changes the rules of the game. If an AI company tells you that “don’t worry, I don’t save the prompts” but it does save the hidden states, it’s as if that theoretical privacy guarantee is of no use. Possible data leaks. A priori it does not seem easy for a potential attacker to do something like this because they would first have to have access to those embeddings. A security breach that results in the leak of a database of those internal/hidden states (embeddings) would no longer be considered an exposure of “abstract” or “encrypted” data, but rather a plain text source from which, for example, financial data or passwords that a company or user has used when asking the AI ​​model could be obtained. Right to be forgotten. This injectivity of LLM also complicates the requirements of regulatory compliance for the protection of personal data, such as the GDPR or the “right to be forgotten.” If a user requests complete deletion of their data from a company like OpenAI, they must ensure that they delete not only visible chat logs, but also all internal representations (embeddings). If any hidden state persists in any register or cache, the original prompt would still be potentially recoverable. Image | Levart Photographer In Xataka | OpenAI is making the tech industry unite its destiny with yours. For the sake of the global economy, it better work

There is a trick to make AI models more reliable: talk badly to them

by

If you greet ChatGPT and thank it when it responds, you’re not getting the most out of it. Some researchers wanted to check if the tone we use when asking the AI ​​for things changes the results and they have discovered something interesting: being rude makes them more trustworthy. Rude. They tell it in How to AI. A study carried out by researchers at the University of Pennsylvania has analyzed whether the tone we use when writing a prompt has an effect on the result and the conclusions are clear. Prompts with a ‘rude’ or ‘very rude’ tone elicited up to 4% more accurate responses than those with a more polite tone. The study. To test it, they generated a list of 50 questions on different topics such as history, science or mathematics. Each of the questions was asked using five different tones: very polite, polite, neutral, rude, and very rude. The model they used was ChatGPT-4o. The results. The researchers did ten rounds with all the questions in different tones and the conclusions are very clear. If we look at the variations, the difference between the neutral or rude tone is only 0.6%, but at the extremes the difference becomes more evident. When using a ‘very friendly’ tone, the average accuracy was 80.8%, while if we went to ‘very rude’, it increased to 84.8%. Kindness by default. We tend to speak kindly to chatbots, this is reflected the survey that Future conducted at the end of 2024. At least 70% of respondents admitted to using “please” and “thank you” when using AI chatbots. Many claimed to do so as a matter of custom, culture and “because it is the right thing to do”, although a small percentage admitted to being afraid that robots would rebel in the future. It is expensive. Regardless of the reasons that lead us to be kind to AI, there is a reality and that is that “please” and “thank you” have an absurd cost. When we thank ChatGPT, requests to the language model increase, which increases electricity and water consumption in data centers. We don’t have figures, but Sam Altman assured that kindness has cost OpenAI “tens of millions of dollars well spent.” The prompt. Despite the enormous advances in AI, language models continue to amaze and are not 100% reliable. However, many times the fault that the answers are not exact does not lie with the model, but with how we are asking it. There is tricks to get a good prompt and being friendly or using fillers like “if you can, I would like to…” is one of the points to avoid. It is not a question of treating them badly either because that does not contribute either, but the more direct and clear you are, the better the result will be. Image | Pexels In Xataka | AI agents want to take our jobs. First they will have to learn not to fail in 70% of the tasks

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.