Google AI won the “only gold medal” at the IMO, and Silicon Valley lined up to offer congratulations. Altman is just a total embarrassment.

Google's Gemini won a gold medal at the IMO, and it's an officially recognized one.

After being scored by the official IMO judges, the new Gemini model answered 5 out of 6 questions correctly and won a gold medal with a score of 35 points.

The version of Gemini that won the gold medal is an advanced one equipped with a new thinking mode. It will be available to Google AI Ultra subscribers later - those who pay $1400 per month.

Last year, it took three days to win a silver medal. This year, it took only 4.5 hours to win a gold medal. DeepMind's math performance has made rapid progress.

In addition to the congratulatory messages sent by Demis Hassabis, the CEO of DeepMind, and Sundar Pichai, the CEO of Google, Elon Musk also sent a tweet to express his congratulations.

DeepMind has received congratulations from all sectors, which is both decent and considerate.

However, the more praise DeepMind gets, the more outshined OpenAI appears. Similarly, when AI participates in the IMO, it's okay to work secretly, but it also steals the limelight from human teenagers for marketing purposes.

Under the leadership of Sam Altman, OpenAI has only been embarrassing itself recently.

DeepMind officially announces that its AI won a gold medal at the IMO

According to DeepMind's announcement, the new Gemini model answered 5 out of the 6 questions in this year's IMO correctly and scored 35 points.

Moreover, this result has been personally certified by Gregor Dolinar, the president of the IMO:

We can confirm that Google DeepMind has reached the long - awaited milestone, scoring 35 points (out of a possible 42) - a gold medal. Their solutions are amazing in many ways. The IMO graders found these solutions clear, precise, and mostly easy to understand.

Not only were the scoring rules in line with the IMO's undisclosed regulations, but the time standard used was also exactly the same as that for human contestants - the answers had to be submitted within 4.5 hours.

Last year, DeepMind's silver medal was achieved by using AlphaProof and AlphaGeometry together over three days.

This time, not only did it take less time, but the answering process was also completed end - to - end in natural language throughout, without the need to modify the question format specifically as in last year.

Thang Luong, a senior scientist at DeepMind and the leader of the IMO team, said that this is a paradigm shift compared to last year.

In addition, Junehyuk Jung, a researcher at DeepMind and a professor at Brown University, introduced that in this year's third question, many human contestants used solutions at the postgraduate level (the IMO is a high - school competition), but Gemini gave a self - consistent proof using only elementary number theory knowledge.

Jung also said that for the last question that Gemini didn't answer correctly, it started in the wrong direction from the beginning. However, only five human contestants answered this question correctly.

By the way, Google has fully published the answers to the 5 questions that Gemini answered correctly (the link is at the end of the article). You can verify them if you're interested~

This time, DeepMind used a brand - new, as - yet - unpublished model equipped with the Deep Think reasoning mode for the competition.

DeepMind introduced that Gemini Deep Think is an enhanced reasoning mode for complex problems, integrating some of its latest research technologies, including parallel thinking.

This mode enables the model to explore and combine multiple possible solutions simultaneously and finally arrive at the final answer, rather than pursuing a single linear thinking chain.

To fully utilize the reasoning ability of Deep Think, DeepMind additionally trained this version of Gemini to make it capable of using more multi - step reasoning, problem - solving, and theorem - proving data.

In addition, DeepMind also provided Gemini with a carefully selected high - quality math question bank and added some general tips and techniques on how to answer IMO questions to its instructions.

As Demis Hassabis said, the model has been provided to mathematicians and other groups for small - scale testing, and it will be available to Google AI Ultra subscribers later.

Of course, DeepMind's official announcement today has also received various praises.

DeepMind announces its results and is well - received in Silicon Valley

Including praises from its competitors, DeepMind has been well - received throughout Silicon Valley.

In addition to Elon Musk's simple "Congrats" at the beginning, a Meta AI scientist praised Gemini for providing clear, precise, and self - study - friendly solutions.

The reason for its popularity is not only its excellent performance but also Google's respect for the IMO rules.

Gemini participated in the challenge at the official invitation of the IMO, and all the rules and standards were the same as those for humans.

Moreover, DeepMind also complied with the official requirements of the IMO and didn't rush to announce the results.

Carefully studying Demis Hassabis's tweet, he didn't mention OpenAI at all, but it was clearly targeted at OpenAI.

Hassabis's three replies mainly emphasized the following points: The official announcement was not premature, the result was certified by the IMO, and the model will also be available in the future.

These three points are in sharp contrast to OpenAI, which was criticized for prematurely announcing its gold - medal result.

The IMO jury and coordinators generally believe that it is "rude and inappropriate" for AI developers to announce their results during the IMO (especially before the closing ceremony).

In addition to the issue of the announcement time, there is also a controversy over whether OpenAI actually won a gold medal.

Thang Luong, who leads DeepMind's super - reasoning team, added that there is actually an official scoring guide within the IMO that is not available to the public. Without scoring based on this guide, there is no qualification to receive a medal.

There were 6 questions in this year's IMO, with 7 points for each question. The gold - medal line is 35 points. OpenAI's self - reported score just passed the line. Even a small deduction in the answering process could drop OpenAI from a gold medal to a silver medal.

Some netizens commented: OpenAI will do anything for hype as always. It has no official score, no patience, and no sense of shame.

In addition, the model OpenAI used for the competition is an unpublished version and may not be made public in the future.

In short, it is in sharp contrast to Google in every aspect.

An OpenAI researcher responds to the doubts

In response to these doubts, Noam, an OpenAI researcher, sent his blessings to Google and also responded:

First, Noam explained that OpenAI didn't cooperate with the IMO official not because it didn't receive an invitation, but because OpenAI itself declined.

Second, regarding the answering process and the authority of the scoring, Noam said that no RAG or any tools were used, and the scoring was done by 3 IMO winners.

At the same time, Noam also made OpenAI's model answers public, saying that anyone can check them.

Finally, regarding the publicity time, Noam said that before announcing the results, he had communicated with the members of the IMO board. The latter asked to wait until after the award ceremony to release the news, and OpenAI agreed.

OpenAI officially announced the results at 1 a.m. Pacific Time, when the award ceremony had already ended. OpenAI was not required to wait until a later time to release the news.