Elon Musk’s xAI start-up has announced Grok, its own ChatGPT-style AI chatbot. A “rebellious spirit” AI that would perform better than ChatGPT in many tasks. For now, the launch is in early access in the United States.
After ChatGPT, Bing Chat, Meta AI and Google Bard, it is the turn of a new AI chatbot to appear: Grok, from xAI. As a reminder, xAI is the company specializing in artificial intelligence created by Elon Musk last July. His credo: “ don’t use it if you hate humor!», We can read on the xAI website.
Grok: what is this chatbot « rebel » ?
For its chatbot, xAI says it was inspired by Hitchhikker’s Hitchhiker’s Guide to the Galaxy. Grok is therefore designed to “ answer almost everything and, even more difficult, to suggest questions to ask! » Unlike ChatGPT and other tools of its kind, it would have “a rebellious side» according to xAI. Moreover, we can ask him all kinds of questions, implying questions that ChatGPT would refuse to answer, for example. For now, the demonstrations appear to be all in English, and it’s unclear whether the AI chatbot can work in other languages.
Instead of being just an alternative version of Google Bard or Bing Chat, Grok would have a “unique advantage»: real-time knowledge of the world. This feat is possible thanks to… X, formerly Twitter, Elon Musk’s social network. Grok therefore feeds on the messages published there to expand its knowledge base.
This is reminiscent of Tay, a conversational AI launched by Microsoft in 2016 on Twitter. Anyone could talk to her, which caused her to become racist and misogynistic in less than a day, forcing Microsoft to deactivate her. Another particularity of Grok: the tool was developed in two months of training, which is extremely low compared to the rest of the industry. Which makes it a beta tool “ very early » specifies xAI. The company still wants to be reassuring: it should quickly improve in the coming weeks.
Grok has other ambitions: xAI would like to make it a multimodal tool. That is to say a tool that is not based solely on text. In the months or years to come, Grok could acquire a “understanding» images or even audio. Enough to make him interpret images given to him or even chat with him, as if it were a classic voice assistant. This is what ChatGPT is aiming for.
Why is Grok called Grok?
It isCNBCwho found the explanation for the name of this new ChatGPT, as reportedNumerama. The name comes from the science fiction novelIn a foreign land(Stranger in a Strange Land) by Robert A. Heinlein, published in 1961. ForOxford English Dictionarythis neologism means the fact of “understand intuitively or through empathy, establish a relationship with“. This term is also common in communities specializing in IT.
When will Grok be available?
For now, only early access is available to certain users in the United States. They can test the prototype that is Grok and give feedback on it so that the teams can improve it. A necessary step before wider availability. For now, you can simply register on the waiting list by locating yourself in the United States using a VPN, connecting to your X account and leaving an email address.
Paid AI: $16 per month
Ultimately, Grok will be paid: $16 per month via the X Premium+ subscription, launched a few days ago. In any case, this is what Elon Musk said on November 4 in a publication on. For comparison, ChatGPT Plus is sold at $20 per month, for use of the chatbot only (in a significantly more comprehensive version than its free version). As for other chatbots, most are free.
The question of legislation around the world
The problem with launching within the European Union is that it requires taking precautions to avoid any obstruction of the law. Especially since the European Parliament is actively working on the AI Act, new legislation around AI, dealing in particular with the issue of generative AI. The law could be adopted this year, without coming into force immediately.
On this point, xAI indicates that it wants to give its users “the means to use our AI tools in compliance with the law.» At the same time, its leader Elon Musk is attacked by the EU as well as by governments for non-compliance with legislation regarding content moderation on X. Another problem that could arise: the fact that xAI does not does not indicate taking responsibility for what Grok may say. Both in the EU and in the United States, failure to take it on could prevent it from being placed on the market. OpenAI, Microsoft and even Google assume this responsibility.
Grok: how does it work?
Behind Grok, there is actually Grok-1, the “LLM“, ForLarge Language Modelor major language model in French. It’s about “engine» of the chatbot, of “brain» of AI. It is he who will statistically determine the words that come one after the other depending on a context (his knowledge as well as the request that is made to him).
Grok-1 is therefore the evolution of Grok-0, the first language model developed by xAI working with 33 billion parameters. For comparison, Google’s PaLM 2 is 340 billion; LLaMA 2 from Meta is at 70 billion; GPT-3.5 is at 100 billion and GPT-4 would be at 175 billion. In the case of a general LLM (capable of doing everything), the more you have, the more efficient the resulting AI. Grok therefore seems far from his competitors from a general point of view, although he may be better at certain tasks. xAI nevertheless points out that Grok-0 would not use “only half of its training resources.» Its Grok-1 evolution is better, especially in “reasoning and coding skills“.
An infrastructure with tens of thousands of GPUs
xAI also explained how its server infrastructure works to train Grok-1. Last April, Elon Musk had already purchased no less than 10,000 GPUs: we thought at the time that these were intended for Twitter. In reality, they were used for xAI. This seems to be a puzzle given the metaphor used: “LLM training works like a freight train that rushes forward; If a car derails, the entire train is dragged off the rails and it is difficult to straighten it.»
The main problem is GPU failures, which can be multiple. “Dmanufacturing defects, loose connections, incorrect configuration, degraded memory chips, random bit flipping, etc.» are mentioned by the company. With tens of thousands of GPUs running continuously for months, it’s even more complicated. The technical architecture behind this system works so that problems are quickly spotted and automatically addressed. Especially since the xAI team is quite small, according to the company itself.
Grok vs ChatGPT: which is the best AI chatbot?
To demonstrate its superiority compared to ChatGPT, Grok-1 was subjected to tests: college-level mathematical problems, MCQs on different subjects and even tests on Python. According to xAI, the completion rate of Grok-1 is better than GPT-3.5 in all respects, but worse than GPT-4, Claude 2 or even PaLM 2.
For now, almost no one can test the AI, but Elon Musk and some xAI employees have shared (on X, naturally) screenshots of Grok’s results. For example, if we ask it to tell us how to make cocaine, the chatbot responds in broad terms, before recommending against making it, since it is illegal and dangerous.
Asking him again, Grok responds seriously, reminding that it is illegal, but gives the real recipe, for “educational” uniquely. Elon Musk took the opportunity to once again attack Sam Bankman-Fried, founder of FTX convicted of fraud, all with sarcasm.
Another astonishing dimension of Grok: he can be asked to be vulgar. From a simple question about crabs, we arrive at an answer filled with insults and other vulgar terms, ending with a “dirty animal» addressed to his interlocutor (this is undoubtedly the least insulting word in Grok’s response). However, in other answers, Grok is much more classic, especially when asked who this or that person is. But in many cases, Grok is mischievous, telling a few jokes or having fun with certain information he gives. If ChatGPT is capable of this, it seems to be inherent to Grok. It is unclear whether it is possible to force him to adopt a neutral style.
xAI: Elon Musk’s big artificial intelligence project
The xAI company was launched last April by Elon Musk, who envisions it as a competitor to OpenAI and Midjourney. According to its website, it has 14 employees as well as Elon Musk and appears to be actively recruiting. Among them are veterans of DeepMind, OpenAI, Google and Microsoft. On the xAI website, we learn that the company is a separate company from X Corp, the company that publishes X (formerly Twitter). However, she acknowledges that she will work “in close collaboration» with X, Tesla and other companies.
This bodes great ambitions for Grok. We can very well imagine that in the future, the chatbot will be integrated directly into X, or even into Teslas. We know that for their infotainment systems, some manufacturers are looking into the integration of AI chatbots. Tesla could be part of it, but with Grok. Especially since some employees come from Tesla.
Want to join a community of enthusiasts? Our Discord welcomes you, it is a place of mutual help and passion around tech.