The AI system developed by Google Deepmind, Google’s major AI laboratory, looks beyond the average gold medicine to solve international gealmic problems. The system is called alphageometry2, as a system version, alfageometry, who deepmind released in January Last. In the receiving research, Deepmind researchers in the flow of the flow of ai ai can complete the 84% geometry problem for the past 25 years, the IMO) contest, Math contest for high school students. Why is it too deep about the high school mathematical competition? Yes, the lab thinks the key to more find a new way to find new ways to solve geometry problems with challenging geometry. Prove the eormatics, or logical teorems explain why the Teorema (eg theorem) is true, requires all the steps and the ability to select steps to the solution. The troubleshooting skills can – if right to DeepMind – be useful components for the general-purpose model. Indeed, the summer summer, DeepMind demoed systems are combined with alfagerty2 for the formal math consideration, including the other mathematical and science approach – sample, to help counting Complex techniques. Alphageometry2 has some elements of core, including the language models of the Gemini Model AI family and “symbolic engine.” The Gemini model helps the symbolic engine, which uses the mathematical rule to reduce the solution to the problem, reaching evidence that can be done for the Geometry’sorem provided. Draw a typical geometry problem in IMO exam. Image of credits: Google (open in the new window of Olympiad geometry based on the “Gemini alphagetry2 alphability that makes construction can be useful to increase the diagram, which is the machine reference to Creating pieces. Actually, gemini alfogiotry2 alfogiotry2 alfogiotry2 steps and construction in the form of Logisensives to perform multiple searchs for similar solutions and stores useful possibilities At the General Science Base. Alphageometry2 consider the “solved” when it comes to the symbolic model of evidence. Because the data of geometry training can be used. So DeepMind makes your own synthetic data To train alphageometry2 language models, producing more than 300 million teallegys and proof of different complexity. The Deepmind team was elected from the IMO competition for the past 25 years (since 224), including a geometry equation that requires geometry objects that make geometry objects that make geometry objects. He then “Translateden” is 50 a bigger problem. (For technical reasons, some issues must be divided into two.) According to the paper, alphageometry2 resolved 42 of 50 issues, cleaning the score on average 40.9. Given, there are limitations. The technical aquirk prevents alfageometry2 from solving problems with the number of variables points, nonlinear equations, and inequality. And alphageometry2 is not technical system of the first AI AI to achieve the gold level of geometry, even when the first gets problems with this size. Alphageometry2 is also worse in other issues about the more difficult IMO issues. For the challenge added, the DeepMind team selects a problem – 29 with the total – the reminder for IMO exams with mathematicsists, but not visible in the competition. Alphageometry2 can only handle this 20. Still, the results of the study may be debate material for what AI systems should be built in symbols manipulation – which is better. Alphageometry2 with hybrid approaches: Gemini has a nerve network architecture, while symbolic machines are rules based on. The Neural Network Engineering Providence says smart behavior, from the recognition of speech to the image generation, it can be out of no variety of data and computing. About the symbolic system, the task of the job by establishing a special symbol rule of a specific project for a particular job, nerve networks trying to overcome the statistical tasks and learn from the example. Nervous networks are a powerful AI system sign like the Operanai model “O1”. However, claim with symbolic ai, he is not the end of all; Symbolic Ai may be a better position to encode efficiently with the knowledge of the world, in a complex scenario method, and “explain” how they come to the answers, the supporters denied. “This drew to see contrast between the very good benchmark, including more recent,” Carnegie Mellon Professional Company University that has specialists in AI, telling TechCrunch. “I don’t think of all the smoke and mirror, but it is important that we still don’t know what the system will be expected to know and the better and risk it may show that two Approach – symbols of manipulation and nerve tissue – combined as a paved road advanced in search for regular gender generations. Indeed, according to the Deepmind paper, O1, who also has nervous network architecture, it cannot solve imo that there is an alphageometry2 able to answer. This may not be like that. On paper, team DeepMind says you find evidence that alphageometry2 language models can produce partial solutions for problems without help from symbolic machines. “[The] Support results that large language models can be sufficient without depending on an external tool [like symbolic engines]”Team Deepmind wrote on paper,” but up to [model] The speed has been good and hallitut has been resolved, the tool will be important for the mathematical application. ”
