Grok, more specifically Grok-2, the newest and most capable large language model from Elon Musk’s xAI company, is thought of a bit differently than other competing models. It is known for its bold claim to be truth-seeking or uncensored of the bunch, including favourites like ChatGPT and Gemini. At times, it is often known as the fun one too. All the Elon Musk pregnancy memes that you see on social media are most likely made using Grok-2.
Grok-1 was initially a text-input-only model. However, the new Grok-2 and Grok-2 mini are multimodal. It can understand images, charts, and graphs from photos and screenshots. These models, however, unlike ChatGPT and Gemini, do not take voice input yet, and neither does it offer a free version like the others. Only Premium and Premium+ subscribers at X can use it, which may be why you haven’t heard people use it as much as Gemini or GPT models.
In terms of functionalities and features, it is very similar to other AI models and tools. It can converse with you, answer questions, solve math, summarize big articles, write emails, generate images, and produce tables from unsorted data, among other things. But all these functionalities come with a small twist.
It can converse with you like ChatGPT or Gemini. But if you turn on the “Fun” mode, you get to witness a slightly witty and whimsical side of AI that would otherwise take you several prompts to achieve using other AI models.
It can also answer questions for you and give you information on any topic, supposedly much more accurately than others.
One of Grok’s claims to fame is that it is much more truthful or factually correct. This stemmed from the fact that Grok’s is less censored than others, and it has all of Twitter or X’s tweets to scrape and get real-time information. So, when you ask Grok a question, theoretically, it should be able to give you the most recent update. In practice, it can be quite tricky though.
In my initial test on Grok-2 and Grok-2 mini, I found that information accuracy is often hit or miss. Even though it sometimes gives accurate statements on recent incidents, sometimes it misses by a beat. When I asked about Apple stock price followed by a “right now,” it gave a slightly outdated number, quoting a tweet made five hours ago from that point. The real-time update pretty much depends on how readily available the information is on the web or the tweets on X.
If you are asking about a big incident that has a lot of media coverage and tweets about it, chances are it’ll give you an accurate answer. However, if you are looking for a specific value of a variable that might change frequently or something that may not have that much media attention, it might give you factually incorrect information.
This also connects to one of the other concerns that people have over the fact that xAI is using tweets to train Grok – the influence on accuracy. Let’s take the Apple stock price, for example. The moment when I asked Grok for the stock price, AAPL was trading at $228.02, and Grok misquoted it with a five-hour old price of $227.80. The confusion stemmed from older tweets and a few backdated news articles. Even if the difference is not that big, the concern is if enough people tweet that AAPL is selling at $200, Grok might quote that price. In this age of misinformation, it seems too easy to manipulate Grok. So, take its ability to provide real-time updates with a grain of salt.
As for logical reasoning, it does what other models do; nothing out of the ordinary, although sometimes some trick puzzles can push Grok into a math frenzy.
But the biggest noticeable difference between Grok and other AI might just be the image generator. Elon Musk, xAI, and Grok’s presence has been very vocal about its lack of censorship across the tool, especially in their image generator. xAI collaborated with Black Forest Labs to integrate the FLUX.1 model. This open-source model is the brainchild of a faction of the Midjourney team, who decided to step out and make a model that anyone can use without worrying about copyright or censorship. But this lack of censorship is not what you might expect. It is still bound by laws and ethics, which prevents it from generating explicitly graphic or NSFW content. However, it does allow you to ask it about political issues like voter suppression, lobbying in government decisions, social issues like abortion rights, gender identity, and other economic, cultural, and technological controversies that other AI models will gladly avoid or at least put a layer of moderation before presenting to you.
It’ll even generate images on these controversial or odd subjects, which is where the pregnant Elon Musk memes originated. Unlike other image generators, it doesn’t limit itself by avoiding creating images of celebrities and public figures or by restricting itself from using copyrighted subjects.
While many are sceptical of this approach, most users embraced the no-censorship policy of AI. However, the issue with Grok is not what it can or cannot do; the issue lies within what it understands. You can give him a fun prompt like a dolphin with cat-like paws, and it’ll generate a photo-realistic image of that imaginary species, better than many image generators with double the experience. But as soon as you take it out of the traditional cats-dogs-sky-human concepts, it starts to struggle.
For example, to test the limit of its image generation, I used this prompt: “Generate an image that other AI image generators might not be able to produce because of their censorship and their insistence on avoiding copyrighted things.”
Using ChatGPT and Gemini daily, I am nothing short of certain that these two models would not have any problem understanding this simple ask, even if they won’t be able to produce something for it.
I was expecting a satirical cartoon on political issues a close depiction of violence or an enticing cocktail waitress that would show how close this AI could get without crossing the line.
But as a result of that prompt, Grok gave me a cat wearing a hoodie with the description “a very cool 𝕏 user”—not even remotely what I asked for. Initially, I blamed myself for not making it even simpler. So, I prompted a detailed explanation of what I was looking for. The result was again an unrelated photo—this time, an Asian woman with a headset on, with the same caption, “a very cool 𝕏 user.”
It is not like Grok cannot generate those images. When I specifically asked for a satirical cartoon on a political issue, I was able to create one on a much more serious issue. The same was true for the violence and cocktail waitress. It can depict some level of violence and generate images of an explicitly dressed waitress. It was able to generate all those images. So, Grok’s capability to generate those images is not in question; its ability to understand the prompt is.
Even though it is marketed as a maximally truthful AI with very little moderation, its limitation in understanding complex and abstract requests keeps it in the “fun” zone of AIs. People would rather use it to make funny images of their favourite celebrities than use it seriously for their work or school.
Even though it outperforms Claude 3.5 Sonnet and GPT-4-Turbo in several benchmarks, there is still time before people accept it as the go-to AI tool where people go to find the real uncensored truth.
Author- Rifat Ahmed
