Join our daily and weekly newsletters for the latest updates and the exclusive content on AI coverage. Learn more
Less than two years since its launch, Xai has shipped which could undoubtedly be the Most advanced AI model to date. Grok 3 corresponds or beats the most advanced models on all key benchmarks as well as key references as well as editors Chatbot arenaAnd his training is not even over yet.
We still don’t have much details on Grok 3, because the team has not yet published a newspaper or technical report. But from what XAI shared in a presentation and on the basis of different experiences, IA experts worked on the model, we can guess how Grok 3 could affect the AI industry in the next month.
Faster
With the competition increasing between the AI laboratories (just look at the exit from Deepseek-R1), we can expect the model release cycles to become shorter. In the presentation of Grok 3, Elon Musk, the founder of XAI, said that users can “notice improvements almost every day because we continually improve the model”.
“The competitive pressure of Deepseek and Grok integrated into a political environment changing for AI – both national and international – will be earlier than the main established laboratories”, writes Nathan LambertMachine learning scientist at Allen Institute for AI. “The increase in competition and the decrease in regulations probably make us, users, will receive a much more powerful AI on much faster deadlines.”
On the one hand, this can be a good thing for users because they have constant access to the most recent and larger models rather than waiting for a month’s deployments. On the other, it can have a destabilizing effect for developers who expect a coherent behavior of the model. Previous research and empirical evidence of users has shown that various models of models can react differently at the same invitation.
Companies must develop personalized evaluations and run them regularly to ensure that new updates do not break their applications.
Scale laws
The recent outing of Deepseek-R1 has undermined the massive spending that large companies make to create large clusters of calculations. But the sudden rise of XAI is a justification for the massive technological companies of investments in AC Accelerators. Grok 3 was formed in record time thanks to XAI Collosus SuperClusiveter in Memphis.
“We have no details, but it is reasonably sure to take a point of data for scaling still helps for performance (but perhaps not on costs),” writes Lambert. “XAI’s approach and messaging were to get the largest online cluster as soon as possible. The explanation of the Occam razor until we have more details is that the scaling has helped, but it is possible that most Grok’s performance comes from techniques other than setting up naive scale. »»
Other analysts have stressed that Xai’s ability to evolve its computer cluster was the key to the success of Grok 3. However, Musk alluded That there is more than just work evolution here. We will have to wait until the paper gets all the details.
Open Source Culture
There is an increasing change to models of large open language (LLMS). XAI has already opened Grok 1. According to Musk, the general company policy is to open each model with the exception of the latest version. Thus, when Grok 3 is completely released, Grok 2 is open-source. (Sam Altman was also amusing The idea of opening certain models of Openai.)
XAI will also refrain from showing the tokens in full chain of thoughts (COT) of the Grok 3 reasoning to prevent competitors from copying it. Rather, it will show a detailed overview of the model’s reasoning trace (like Openai did with o3-mini). The complete COT will only be available once XAI Open Sources Grok 3, which will probably come after the release of Grok 4.
Make your own room check
Despite the impressive reference results, the reactions to Grok 3 have been mixed. Former scientist Openai and Tesla ai Andrej Karpathy Placed his reasoning capacities to “around the state of the art”, with O1-Pro, but also stressed that it is late on other peak models on certain tasks such as the creation of graphics Evolutionary vectorlands Compositions or navigation of ethical problems.
Other users have underlined Defects in the coding capacities of Grok 3 Compared to other models, although there are also many cases of Grok 3 impressive coding exploits.
Based on my own experience with the leading models, I advise you to make your own verification and atmosphere search. I never judge a model based on an invitation to a blow. Have a set of tests that reflect the type of tasks you do in your organization (see a Some examples here). There is a good chance that, with the right approach, that you can make the most of these advanced models.
