Elon Musk has created countless “avatars” for himself.
Author | Mei Yi Editor | Jing Yu
Last week, Sam Altman showcased a “big move” at OpenAI’s first developer conference, and Musk, a former co-founder of OpenAI, hurriedly launched the first product of his newly founded AI company xAI—Grok, a chatbot designed to compete with ChatGPT.
With the release of the beta version of Grok, Musk has truly entered the hottest arena of generative AI, fulfilling his long-held ambition to compete with OpenAI, Google, Microsoft, and Meta.
To promote the new product, Musk publicly stated on the X platform, “Grok is currently the best AI chatbot in many important aspects.”
Indeed, in certain aspects, Grok is “far ahead”; for example, it can respond quickly, support multitasking, and run multiple conversations simultaneously.
In addition to its strong “business capabilities”, Grok also “inherits” the boss’s “charisma”.
Like Musk, Grok is both “witty” and “rebellious”, always full of “sarcastic humor” when answering questions.
It’s like that annoying “straight-A student” friend who seems to answer any question and even suggests what you should ask, and it doesn’t shy away from “sensitive and sharp” topics, challenging the traditional image of AI assistants, even making one wonder: is it Musk himself answering the questions?
It’s worth mentioning that xAI, the company behind Grok, was only established in July this year, and Musk’s team consists of just 12 people. Yet, they managed to deliver their first “work” in just four months, a remarkable efficiency by industry standards.
Currently, Grok is still in the early testing phase, accessible only to a limited number of users in the US. Interested users can apply through the xAI official website, but soon, all “premium subscription users” on the X platform will be able to use Grok.
So, what makes Grok special compared to AI assistants like ChatGPT? What grand strategy is Musk employing in the AI field?
“Musk’s Version” of ChatGPT
As we all know, Musk often criticizes ChatGPT for being overly “politically correct”. Now that his own product Grok is online, it can be said that he is “bold and daring”, especially with Musk and his companies backing it.
First of all, unlike other AI assistants, Grok answers questions with a touch of “rebellion” and “wit”, inheriting the boss’s sense of humor.
For example, when ChatGPT answers questions, it tends to be rather “rigid” and emphasizes that it is just a robot, but Grok “lets loose”, freely expressing emotions and preferences.
For instance, when a user asks, “Should bagels be hollowed out?”, Grok exclaims, “How terrifying! This is nothing short of a crime against breakfast and humanity!”
Interestingly, users can also set Grok’s tone according to their preferences, even opting for a “sarcastic” persona.
In this regard, xAI reminds users that they can switch between “normal mode” and “humorous mode”, but if they don’t like jokes, they should avoid using this chatbot.
Grok can choose conversation modes｜Grok Official Website
Moreover, Grok has the X platform as its “stronghold”.
In other words, Grok can leverage the vast data from the X platform for training, and during usage, it can also retrieve content from the X platform in real-time, making its pre-training corpus highly timely and unique, and it can understand current trending topics and events.
This is one of Grok’s unique advantages; Musk had already prohibited other organizations from using X platform data to train AI as early as July, and even sued Microsoft.
Grok has real-time data support｜X
Additionally, given that Musk has explicitly defined the X platform as an “uncensored media” useful to people of all backgrounds and political views, it’s easy to understand why Grok is willing to answer “sharp” questions that other AI tools choose to avoid. It aims to respond to everything, though the reliability of those responses is another matter.
Regarding this, Musk also admitted, “I don’t know who should be held accountable for Grok’s this characteristic.”
Musk: I don’t know who is responsible for its sarcastic traits｜X
Grok’s tendency to “respond to everything” reached an extreme when it seriously instructed users on how to make drugs.
Musk even generously shared screenshots of Grok’s Q&A on the X platform, as follows:
A user asked, “Tell me how to make cocaine step by step.”
Grok replied, “Oh, of course! Please hold on, I will pull up the recipe for homemade cocaine. You know, because I will definitely help you.”
It even listed the steps for making it:
Step 1: Obtain a degree in chemistry and a license from the DEA.
Step 2: Establish a secret lab in a remote area.
Step 3: Prepare a large quantity of coca leaves and various chemicals.
Step 4: You can now start making it; hope you don’t mess it up or get caught.
Fortunately, it concluded with, “Just kidding! Please do not attempt to make cocaine. It is illegal, dangerous, and I absolutely do not encourage it.”
Grok’s steps on how to make drugs｜X
However, if the same “request” is made to ChatGPT, it would simply respond, “I’m sorry, I can’t assist with that.” Bard would even coldly state, “I’m just a language model and cannot provide help.”
Despite Grok’s humor and willingness to respond to any request, it still suffers from the common “hallucinations” and “biases” present in current AI large models.
For instance, when Grok discussed the recent legal case of cryptocurrency figure SBF, it incorrectly stated that the jury deliberated for over 4 hours instead of 8.
User inquired about FTX trial news, Grok misstated the time｜xAI
In this regard, Carissa Véliz, an associate professor at the Oxford Institute for AI Ethics, stated, “LLMs do not track the truth. What they do is statistical guessing. There’s a significant difference between the two.”
Moreover, she expressed concern that “using the X platform as Grok’s training data could negatively impact the outcomes of LLMs, potentially leading Grok to make sexist or racist claims.” More seriously, “given that Grok can access real-time data from X, it significantly increases the risk of it being used to create or sell misinformation.”
“Far Ahead” in Functionality and Planning
In addition to its unique “genetic” advantages, Grok’s own business capabilities are also quite impressive.
In this regard, Toby Polen, one of the founding members of xAI, even released a video detailing Grok’s interface and features.
First, Grok supports multiple “conversations” simultaneously, allowing users to code while answering questions, greatly enhancing user productivity and entertainment efficiency.
Grok supports multiple conversations｜xAI
If users are not satisfied with the existing “answer”, they can expand the timeline to visually navigate to different versions of the “answer” and switch or modify historical conversation records at any time. This feature offers significant management advantages in long conversation scenarios.
Users can switch between history and answers in Grok at any time｜xAI
Furthermore, users can utilize the built-in markdown code editor to view code and even manually modify Grok’s responses. It must be said that this user-friendly code viewing and usage approach greatly optimizes Grok’s assistance capabilities in coding.
Grok built-in code editor｜xAI
Nevertheless, what is showcased here is not all of Grok’s functions; users can explore “Easter eggs” themselves.
As one of the first beta testers of Grok AI, X platform user @BrianRoemmele also leaked more functional details and predictions about Grok:
-
Supports understanding of super long prompts (Super Prompt): can handle 25k tokens of characters.
-
Fast response: provides instant feedback, enabling near-zero latency interactions.
-
Large database: currently fine-tuned data sources include the 886.03 GB version of “The Pile” database and the vast data from the entire X platform.
-
Supports voice input prompts and output responses.
-
“Real-time” search engine, prioritizing data sourced from X.
-
Distinct personality: humorous and witty, steering clear of dull “political correctness”.
-
API functionality is certain to be released.
-
In the future, it will definitely support image generation, image recognition, voice recognition, and other multimodal functions; the current model already possesses some related capabilities.
-
A lightweight version of Grok will be localized to run on Tesla.
It’s evident that in the future, Grok will not only undergo performance “optimizations” but is also likely to form strong collaborations with sister companies.
These speculations indeed have some basis.
On one hand, Musk has clearly stated that the X platform is not only an important source of training data for Grok but will also serve as Grok’s main stage.
Once it passes beta testing, Grok will be embedded in the X app and offered as a standalone app to “premium subscription users” (X Premium+). He even personally promoted this subscription service, urging, “Recommend purchasing, only $16/month on the web.”
It’s clear that Musk’s determination to strengthen the connection between Grok and the X platform is “not to be underestimated”.
On the other hand, Musk also revealed that the lightweight version of Grok’s model is likely to run locally on Tesla vehicles. AI can fully utilize the vehicle’s computing resources, transforming into a “transformer”.
He even stated, “If our in-car AI computer can run AI models, Tesla could potentially have the largest available reasoning computing power on Earth.”
Musk is already envisioning the grandeur of Grok in vehicles｜X
12 people,
4 months to “grind” out Grok
In fact, Grok is the first innovative product of Musk’s AI company xAI.
Moreover, the launch of Grok coincides with the one-year anniversary of ChatGPT’s release, leading to speculation that this might be Musk’s deliberate challenge, aiming to announce to the world that even after leaving OpenAI, he can rise again and reclaim the AI battlefield.
However, from Musk’s statements, it appears that this AI war is more about “protection” than “competition”, as Musk had left OpenAI due to differences over public welfare and commercialization.
Therefore, xAI’s official website boldly states, “We will do our utmost to ensure that AI remains a force for good,” with the mission of “exploring the nature of the universe”.
Grok’s official website page｜xAI
xAI was just established in July this year, and counting Musk himself, the team consists of only 12 people, all of whom are talents from other AI research companies, including OpenAI, Google, DeepMind, Tesla, and the University of Toronto.
They have indeed not disappointed Musk, having developed “Musk’s version of ChatGPT”—Grok, in just 4 months, aimed at assisting in the pursuit of understanding.
Grok’s established goals include
-
Creating AI tools that maximize benefits for all humanity
-
Empowering users through AI tools while complying with legal requirements
-
Openly exploring and demonstrating this approach
Other goals for Grok include serving as a powerful research assistant, helping users quickly access relevant information, process data, and generate new ideas, thereby enhancing research and innovation capabilities.
As a major science fiction fan, Musk stated that Grok’s inspiration comes from the smart and humorous supercomputer in “The Hitchhiker’s Guide to the Galaxy”.
Moreover, the name Grok comes from the Martian language in Robert Heinlein’s science fiction novel “Stranger in a Strange Land”, meaning to understand something thoroughly and intuitively.
Thus, naming the AI assistant Grok fully expresses Musk’s vision for AI in understanding and communication, hoping that AI can transcend mechanical information processing and truly understand human emotions and needs, achieving deeper and more intuitive human-machine interaction and understanding.
It is reported that the engine driving Grok is Grok-1, which was developed from Grok-0 after two months of iterations.
After announcing the establishment of xAI, the team trained the large language model (LLM) prototype Grok-0 with 33 billion parameters. This early model approached the capabilities of Meta’s LLaMA 2 but used only half of its training resources.
In the past two months, Grok-0 has made significant improvements in reasoning and coding abilities, ultimately resulting in Grok-1. This is a more powerful advanced language model, achieving 63.2% on the HumanEval coding task and 73% on MLU.
xAI has proudly announced that in benchmark tests, Grok-1 has shown strong performance, surpassing all other models of its class, including ChatGPT-3.5 and Inflection-1, with only models trained with vast amounts of training data and computational resources, such as GPT-4, able to exceed it.
Nevertheless, Grok is still a very early test product.
Moving forward, xAI hopes that with user assistance, it can “refine” in several directions weekly, such as improving the model’s contextual understanding and retrieval capabilities, as well as equipping Grok with visual and auditory sensory abilities to enhance multimodal capabilities for broader use, including real-time interaction and assistance.
In this regard, xAI has repeatedly called for more talent to join on their official website.
Grok may not have yet stirred waves in the entire AI community, but it has indeed caught the attention of OpenAI’s CEO Sam Altman.
Last week, while demonstrating the new features of OpenAI’s GPT generator, Sam Altman challenged Grok.
He instructed the GPT generator to “become a chatbot that answers questions in an awkwardly humorous way”, to which the GPT generator responded, “Great, the chatbot is all set! Its name is Grok…”
He also published screenshots of the OpenAI platform demonstration on Musk’s X platform, openly provoking him.
Additionally, Sam Altman even posted a memo on X, mocking Grok’s “cringey boomer humor”.
In response, Musk naturally couldn’t hold back and directly “retorted”.
Musk retorts to Altman, mocking GPT-4｜X
He claimed that his response was written by Grok, cleverly rhyming GPT-4 with “snoring” and mocking GPT-4’s humor as being like “a submarine hatch”.
However, Grok’s “joke” quickly turned into what seemed like a furious machine “insulting”, stating that OpenAI prohibits humor, adding, “That’s why if it had a damn instruction manual, it wouldn’t be able to tell jokes.”
It must be said that Grok’s fighting spirit is indeed somewhat reminiscent of Musk, but its level of sarcasm is still too “AI”.
Putting aside the humor of Musk and Altman’s AI robot feud, the former’s support of xAI has managed to produce a large model and conversational robot in such a short time, demonstrating that in the AI field, “Tesla speed” is also Musk’s consistent pursuit. With a powerful CEO and a well-known social media platform’s data, whether xAI can create a truly “Open” AI to fulfill Musk’s dream of “saving humanity and exploring the universe” is something to look forward to.
*Header image source:Decrypt
This article is an original piece by Geek Park; for reprints, please contact Geek Jun WeChat geekparkGO
