Musk Claims Newest AI Model, Grok 3, Can Beat Rivals

Elon Musk has announced the release of X’s newest AI model, Grok 3, which he claims far surpasses the abilities of its previous models as well as his rivals’ most recent efforts in the genAI space.

He based these claims on internal early testing the company did on its newest Grok model, which included tests on maths, science, and coding.

Musk also announced that X would be releasing a new “Deep Search” product which would act as a search engine, scanning the internet and X to deliver more detailed answers to a queries.

Grok 3 will be available to premium X subscribers first, though the model will also be available via a separate subscription.

The model was originally meant to be released in 2024, but the company missed the initial deadline.

Training on its Colossus supercomputer, which was powered by 100,000 Nvidia H100 GPUs, and 200 million GPU-hours for training, Musk claims Grok 3 was trained on ten times more computing power than the company’s previous Grok 2 model, and was training on a vast data set which included synthetic data. 

Grok 3 consists of various Grok models, it should be noted. Grok 3 mini is a small version of Grok 3, which has faster responses, despite some loss of accuracy. Some of the models are still in Beta, and will not be included in the current release.

According to xAI, not only did Grok 3 perform well on PhD level sciences and a range of math queries, but also scored well in the crowdsourced tests on Chatbot Arena compared to its main rivals.

Musk’s company claims that two of its models, Grok 3 mini Reasoning and Grok 3 Reasoning, can ‘reason’ similar to DeepSeek’s R1 model and OpenAI’s smaller o3-mini, and even surpasses o3-mini-high, across several benchmarks, including maths.

He did explain that the ‘thought’s behind these reasoning models would be obscured as to protect it from distillation, a term used to describe how AI developers take knowledge and processes from other models. DeepSeek was recently accused of doing this to OpenAI’s models.

Recommended reading

Musk says that ‘voice mode’ will be available with the model in as soon as a week, and that Grok 3 models will be made available in xAI’s enterprise API some weeks after that, as well as DeepSearch.

Following this, Musk says he will make Grok 2 open-source in the next few months, once Grok 3 is “mature and stable.”

“Grok 3 is an order of magnitude more capable than Grok 2,” Musk said in a live-streamed presentation of the model. He called the new system “maximally truth-seeking Ai, even if that truth is sometimes at odds with that is politically correct.”

This has been a longstanding aim of Musk – to create a chatbot series that is anti-woke, not politically correct and adheres to his views of what free speech should be. The Grok series, in practice, however, was found to steer away politically-charged queries and even leaned left in one study.

Musk largely blamed this on training data, and aims to make Grok more “politically neutral,” despite Musk’s own personal forays into the political landscape appear anything but.

The controversial CEO recently made a bid to buy OpenAI, which was ultimately rejected by the company’s board.

Related