# Google Takes the Gold. OpenAI under fire.

Source: https://www.youtube.com/watch?v=36HchiQGU4U
Recap page: https://rapidrecap.app/video/36HchiQGU4U
Generated: 2025-07-22T03:09:29.462+00:00

---
## Quick Overview

Google DeepMind's advanced Gemini with DeepThink model officially achieved a gold-medal standard at the International Mathematical Olympiad (IMO) 2025, solving five out of six exceptionally difficult problems with a score of 35 out of 42 points, matching OpenAI's similar achievement. This marks a significant breakthrough as both models utilized large language models (LLMs) to directly interpret and solve complex mathematical problems within the competition's time limits, unlike previous specialized AI models that required manual translation of problems.

**Key Points:**
- Google DeepMind's Gemini with DeepThink achieved a gold-medal standard at IMO 2025, scoring 35/42 points by solving 5 out of 6 problems.
- OpenAI's model also achieved a gold-medal standard at IMO 2025 with an identical score of 35/42 points.
- Both Google DeepMind and OpenAI utilized large language models (LLMs) that could directly interpret and solve problems from natural language, a significant advancement.
- Google DeepMind respected the IMO Board's request to delay result announcements until after the closing ceremony, unlike OpenAI.
- The Gemini DeepThink model incorporates advanced reasoning techniques like parallel thinking and novel reinforcement learning, trained on curated high-quality math solutions.
- AI progress in complex reasoning is accelerating faster than many experts, including Eliezer Yudkowsky, had predicted for 2025.
- The development of 'RL systems' and self-verification/self-generation of learning curricula are seen as key drivers for achieving AGI, moving beyond fixed model checkpoints.

![Screenshot at 0:00: A man with a beard and headphones looks at the camera, with the Google logo and the word "WON" next to a gold medal icon on a black background.](https://ss.rapidrecap.app/screens/36HchiQGU4U/00-00-00.png)

**Context:** The International Mathematical Olympiad (IMO) is an annual mathematics competition for pre-collegiate students. In recent years, AI companies like Google DeepMind and OpenAI have been developing large language models (LLMs) capable of solving complex mathematical problems. This video discusses the latest achievements of these AI models in the IMO 2025, highlighting their ability to achieve gold-medal standards and the implications for the future of artificial general intelligence (AGI).

## Detailed Analysis

Google DeepMind's advanced Gemini with DeepThink model has officially achieved a gold-medal standard at the International Mathematical Olympiad (IMO) 2025, scoring 35 out of a possible 42 points by perfectly solving five of the six exceptionally difficult problems. This achievement mirrors OpenAI's performance, as their model also solved five out of six problems with the same score. Both companies utilized large language models (LLMs) for this feat, a significant advancement from Google's 2024 silver medal, which was achieved using specialized AI models like AlphaProof and AlphaGeometry that required manual translation of problems into formal mathematical language. The Gemini DeepThink model incorporates advanced reasoning techniques, including parallel thinking and novel reinforcement learning, allowing it to explore and combine multiple possible solutions before arriving at a final answer. While Google DeepMind respected the IMO Board's request to announce results after the closing ceremony, OpenAI announced theirs beforehand, leading to some controversy. The video highlights that the true breakthrough lies in the LLMs' ability to understand and solve problems directly from natural language, indicating a shift towards more general-purpose AI in complex reasoning tasks. Discussions around the computational cost and efficiency of these models, as well as the concept of 'fluid intelligence' and self-verification in AI, suggest a new wave of AI progress driven by reinforcement learning and self-generated curricula, akin to DeepMind's AlphaZero.

### IMO 2025 Results

- Google DeepMind's Gemini with DeepThink achieved a gold-medal standard at the International Mathematical Olympiad (IMO) 2025, scoring 35/42 points by solving 5 out of 6 problems
- OpenAI's model also achieved the same gold-medal standard with an identical score and number of problems solved
- Both models utilized large language models (LLMs) to solve the problems, a significant leap from previous specialized AI models.

### Controversy and Ethics

- OpenAI announced their results before the IMO closing ceremony, which was deemed rude and inappropriate by some organizers, including the IMO President
- Google DeepMind respected the IMO Board's request to wait for official verification and student recognition before announcing their results
- OpenAI's model was graded by former IMO medalists, while Google's went through official IMO channels.

### Gemini DeepThink Capabilities

- The advanced version of Gemini DeepThink uses an enhanced reasoning mode incorporating latest research techniques, including parallel thinking
- This setup enables the model to simultaneously explore and combine multiple possible solutions before giving a final answer, rather than pursuing a single, linear chain of thought
- Gemini was additionally trained on novel reinforcement learning techniques that leverage multi-step reasoning, problem-solving, and theorem-proving data, and provided access to a curated corpus of high-quality solutions to mathematics problems.

### Evolution of AI in Mathematics

- In 2024, Google's AlphaProof and AlphaGeometry achieved a silver medal at IMO, but problems had to be manually translated into formal mathematical language
- For IMO 2025, both Google and OpenAI's LLMs processed problems directly from natural language, producing rigorous mathematical proofs within the 4.5-hour competition time limit
- This signifies a move towards more general-purpose AI systems capable of understanding and solving complex problems without specialized manual translation.

### The Future of AI and AGI

- The video discusses the concept of AGI being more about the underlying 'RL system' or 'LLM factory' within companies rather than fixed model checkpoints
- Key advancements include self-verification and self-generation of learning curricula, allowing AI to teach itself and improve at scale
- The 'AlphaZero lesson' emphasizes that AI can achieve superhuman performance by learning through self-play and generating synthetic data, rather than relying solely on human-curated data or hand-crafted rules.

### Market Predictions and Progress

- Betting markets had predicted a 10-16% chance of an AI model winning gold at IMO 2025, indicating that the current progress is happening faster than many experts anticipated
- Elon Musk's Grok 4 model also performed well on the 'SimpleBench' benchmark, scoring 60.5% (second only to Gemini 2.5 Pro's 62.4%), demonstrating 'fluid intelligence' through massive amounts of RL compute.

### Availability of Gemini DeepThink

- Google's Gemini DeepThink model is currently being tested by trusted testers, including mathematicians
- It will eventually be rolled out to Google AI Ultra subscribers, indicating its commercialization and broader accessibility.

![Screenshot at 0:04: A research paper title reads "Advanced version of Gemini with Deep Think officially achieves gold-medal standard at the International Mathematical Olympiad".](https://ss.rapidrecap.app/screens/36HchiQGU4U/00-00-04.png)
![Screenshot at 0:25: A quote from IMO President Prof. Dr. Gregor Dolinar confirms Google DeepMind's achievement of 35/42 points, a gold medal score, praising their astonishing, clear, precise, and easy-to-follow solutions.](https://ss.rapidrecap.app/screens/36HchiQGU4U/00-00-25.png)
![Screenshot at 0:32: A tweet from Alexander Wei states that their model \(OpenAI's\) solved 5 of 6 problems on the 2025 IMO, earning 35/42 points, enough for gold, with scores finalized after unanimous consensus by former IMO medalists.](https://ss.rapidrecap.app/screens/36HchiQGU4U/00-00-32.png)
![Screenshot at 0:59: The International Mathematical Olympiad 2025 awards section shows that a score of \>= 35 points is required for a gold medal, with a maximum possible score of 42 points \(7 points per problem for 6 problems\).](https://ss.rapidrecap.app/screens/36HchiQGU4U/00-00-59.png)
![Screenshot at 1:55: A tweet from Demis Hassabis \(CEO of Google DeepMind\) explains that Google DeepMind did not announce results earlier because they respected the IMO Board's request for AI labs to share results only after official verification and student recognition.](https://ss.rapidrecap.app/screens/36HchiQGU4U/00-01-55.png)
![Screenshot at 3:20: A Google DeepMind science page title reads "AI achieves silver-medal standard solving International Mathematical Olympiad problems" from July 25, 2024, indicating their previous year's performance.](https://ss.rapidrecap.app/screens/36HchiQGU4U/00-03-20.png)
![Screenshot at 4:42: A comparison chart of Google AI Pro and Google AI Ultra plans, showing that the advanced reasoning model DeepThink will be available to Google AI Ultra subscribers.](https://ss.rapidrecap.app/screens/36HchiQGU4U/00-04-42.png)
![Screenshot at 6:04: A section of the Google DeepMind blog post highlights that their advanced Gemini model produced rigorous mathematical proofs directly from official problem descriptions within the 4.5-hour competition time limit, unlike the 2024 models that required manual translation.](https://ss.rapidrecap.app/screens/36HchiQGU4U/00-06-04.png)
![Screenshot at 9:41: A bar chart titled "Ludicrous rate of progress" shows the increasing compute dedicated to reasoning \(RL compute\) for Grok models, with Grok 4 reasoning showing a 10x increase over Grok 3.](https://ss.rapidrecap.app/screens/36HchiQGU4U/00-09-41.png)
![Screenshot at 10:25: A leaderboard from SimpleBench shows Gemini 2.5 Pro \(Google\) in 1st place with 62.4% and Grok 4 \(xAI\) in 2nd place with 60.5% in a multiple-choice text benchmark for LLMs.](https://ss.rapidrecap.app/screens/36HchiQGU4U/00-10-25.png)