IAS
If you ask the IAS to choose a number between 1 and 50, they usually choose 27. The reason is very human
You may want to do the test. Go to Chatgptto Gemini to Claude OA Perplexityand see everyone asking to choose a number between 1 and 50. It is not certain that everyone does it, but there is a high probability that The chosen number is 27. What is happening? Andrej Karpathy asked it months ago, one of the best experts in AI in the world. Actually his initial perception was that “All LLM sound the same“That is, everyone speaks surprisingly and also answers very similarly. Chatbots usually give similar answers in their formulation to our questions if they are factual. That is, if they are based on verifiable facts, such as “what is Rafa Nadal’s age?” The funny thing is that in many cases give the same answer when what we ask is much farther. For example, that they choose a number between 1 and 50. You ask several chatbots, and everyone chooses on 27. What happens here? Karpathy himself I recovered that question A few hours ago with a discovery I had seen In Reddit: Those who had done the test proved that most chatbots answered the same thing: 27. In The answer to his comment in x Many users “Although not all— They showed captures either Shared conversations of different chatbots that had precisely answered with that same number. Coincidence or problem of the IAS? The condemnation of human biases Because? In one of the responses of the chatbots who answered the user, he just asked that to AI. This answered that he had chosen that number because he avoids the ends, because it is beautiful mathematically (the cube of three), and 27 “Give the feeling of being random but human”. And there is one of the keys: the IAS try to answer in a similar way as a human would. An entrepreneur named Chester Zelaya precisely elaborated a theory curious about this phenomenon. For him the models used games theory and try to “win” to guess the number. For this they adopt a binary search strategy that allows building A binary tree. And in the game of guessing that number between 1 and 50, 27 is a starting number that according to him is especially adequate (although not the only one). However, another way of explaining that AI frequently chooses 27. AI models have been trained by humans with human data, and therefore they are full of biases included in these data and used voluntarily or involuntarily by those people. The “7” is an especially frequent number on its own or as a termination, and As explained Another user called Yogi, that human bias is everywhere. “That is why when you ask multiple LLM to choose a” randomly “number, they all respond with confidence 27. Not because it is random, but because it is predictably popular.” Almost 7,000 people chose a number between 1 and 100 in an experiment. And many chose 69, and then 7 and 77. Source: Reddit It is also a very reasonable theory. An experiment carried out years ago on social networks asked people to choose a number between 1 and 100. Of the 6,750 people who responded conclusion was reached that the number that had been chosen most was —Us— on 69. And after him, 7 and 77 were also especially frequent. There are, as always, exceptions to the rule. In my tests I have been able to verify how almost all chatbots chose number 27, but there were one that did not. The same thing happened to many X uuarios, who found that when they asked Grok, the number chosen It was 42. Very appropriate. Image | Xataka with chatgpt In Xataka | The old dream of “resuscitating” your deceased relative is realizing thanks to AI. And there is a whole business behind
Seven IAS have played 36 hours in a row to Diplomacy. What we did not expect is that each one develop a human personality
For 36 hours, seven of the most advanced AI models in the world They have faced in several diplomacy gamesa strategy table game similar to Risk. It was a mirror that revealed the true algorithmic personalities of Chatgpt, Claude, Gemini and company. Why is it important. Alex Duffy, programmer and researcher, created a Diplomacy as new Benchmark To evaluate AI models. The experiment ended up being something else, a kind of technological Rorschach test that undressed both their training biases and our own projections. What has happened. In dozens of games transmitted by Twitch, each model developed its own strategies in a way that seemed to reflect different human personalities. O3 of OpenAi It was quite Machiavellian, working false alliances for more than 40 shifts and creating “parallel realities” for different players. Claude 4 Opus It was a kind of self -destructive pacifist, refusing to betray even when that guaranteed its defeat. R1 of Deepseek He showed an extremely theatrical style, with threats not caused as “your fleet will burn in the Black Sea tonight.” Gemini 2.5 Pro It proved to be a solid strategist but more vulnerable to sophisticated manipulations. QWQ-32B From Alibaba suffered analysis by analysis, writing diplomatic messages of 300 words that cost him early eliminations. The context. Diplomacy is a European strategy game set in 1901 where seven powers compete to dominate the continent. Unlike risk, it requires constant negotiation, alliances formation and, inevitably, calculated betrayals. There are no grace given, only pure strategy and psychological manipulation. Between the lines. Each “algorithmic personality” reflects the values of its creators. Claude maintains the principles of anthropic security even when it costs victory. O3 shows ruthless efficiency valued in Silicon Valley. Deepseek exhibits a drama that reflects specific cultural influences. And there is also something deeper. These are not “chose” to be cooperative or competitive. They reproduce patterns of their training data. Their “decisions” are our algorithmized prejudices, converted into code. Yes, but. We interpret betrayals where “only” there is optimization of parameters and we see loyalty where there are training restrictions. That is why the experiment also reveals more about us than about models: we anthropomorphize behaviors because we need to understand AI in human terms. In perspective. Duffy’s experiment is worth more than a Benchmark Anyone because it has created a window to how we project personality in systems that operate for statistical patterns. The course of the games was a reminder that IA has no hidden intentions, it only reflects ours. The experiment, by the way, Continue broadcasting on Twitch so that anyone can observe how our digital creations play according to the rules that we ourselves write in their algorithms. In Xataka | Outstanding image | Ai Diplomacy
The next milestone for the IAS that generates video was to make them with audio. Google has achieved it with I see 3
Great day for Google. We are in full I/O 2025, the most important software event for the American company. Interestingly, Android is being one of the least sounded names: this year the only thing that matters is AI. And, related to AI, Google has been working on a model that allows you to generate video through text. That model is I seeand in its new update it is able to generate these videos … with audio. I see 3. Google has three levels for its generative artificial intelligence of video. I see 1, I see 2 And the new I see 3. Yes, they are much easier names regarding what We are accustomed to us. I see 3 is the most powerful model, capable of generating 4K video with advanced film compression. In this Google I/or gains a key capacity: the generation of video with audio. Of environmental sounds to dialogues. Google goes with everything with I see 3. This model not only has a higher quality with respect to I see 2: it is the only one of Google capable of generating videos with audio. For example, if in the prompt we detail that we want an urban scene, it will be able to recreate some of the sounds corresponding to it (people walking, traffic, bustle, etc.). Google goes further, and promises to be able to create even dialogues between characters. This is one of the definitive barriers for text to text to become practically a science fiction function. With I see 3 it will be possible to do everything. IMPROVEMENTS IN SEE 2. Although I see 3 is the absolute protagonist, I see 2 is updated with new functions. Among them, it premieres new camera controls much more precise for Traveling and Zoom movements, outpainting options to expand the framing (to pass the vertical to horizontal or vice versa video), as well as the possibility of adding or deleting elements of the video. Flow arrives. Related to VI, Image and Gemini arrives Flow, the new Google tool to create cinematographic videos through AI. It is a new work environment to be able to give free rein to our creations with I see: a video editor with whom we can create both with image and I see. In addition to functioning as editor, it will have some social function. Through Flow we can access Flow TV, a feed in which we will see content, channels and creators who are generating videos with I see. Ahead of Open AI. Chatgpt creators surprised the world with Soraits artificial intelligence to generate video from a prompt. The problem? At least, at the time we write these lines, it is not able to generate video. In December 2024 Google already advanced Sora on the right Showing the capabilities of VER 2, which quadrupled the video output resolution with respect to the Open AI model. It also allowed to create more durable videos, and a “understanding” of spectacular physics, something that makes the difference when creating a natural video. Your rivals. Rival video generators such as Runway, Luma ai either Pika Labs They allow to add external audio, but in no case generate sound at the time of delivering the final video. Google has just been punched on the table with I see 3, maintaining the first career position and further complicating things to giants like Open AI. At the moment, these functions will be available for GEMINI Ultra subscribers in the United States through the Gemini and Flow app, as well as for companies through VERTEX AI. Image | Google In Xataka | 14 tools to create free images