Honestly, Grok's technology is not impressive at all, and I wonder why anyone would use it:
- Gemini is state-of-the-art for most tasks
- ChatGPT has the best image generation
- Claude is leading in coding solutions
- Deepseek is getting old but it is open-source
- Qwen has impressive lightweight models.
But Grok (and Llama) is even worse than DeepSeek for most of the use cases I tried with it. The only thing it has going for is money behind its infamous founders. Other than that, their existence would be barely acknowledged.
I like it! For me it has replaced Sonnet (3.5 at the time, but 3.7 doesn't seem better to me, from my brief tests) for general web usage -- fast, the ability to query x nee twitter is very nice, & I find the code it produces tends to be a bit better than Sonnet. (Though perhaps that depends a lot on the domain...I'm doing mostly C# in Unity.)
For tough queries o3 is unmatched in my experience.
Llama is arguably the reason open weight LLM’s are a thing, with the leak of Llama 1 and subsequent release of Llama 2. Llama 3 was a huge push for quality, size, context length, and multi-modality. Llama 4 Maverick is clearly better than it looks if a fine tune can put it at the top of LMArena human preferences leaderboard.
Grok 3 mini is quite a decent agentic model and competitive with frontier models at a fraction of the cost; see livebench.ai.
Although Deepseek is old, I find the V3 (without reason) still to be the best non reasoning model out there.
Now, ChatGPT main advantage for me right now it's search + o4-mini. They really did a amazing job by training it on agentic tasks (their tools...) and the search with reasoning works amazing.
Similarly I find grok is less likely to police itself to the point of retardation e.g. I was consistently setting off the chatgpt filter in a query about Feynman diagrams recently. Why?
Before the release of Gemini 2.5 Grok 3 was the best coding AI IME, especially when you used reasoning. It also complained the least about things you asked it to do. Gemini for instance still won’t tell you how to use yt-dlp.
Indeed. I switched to using Grok exclusively (even though other models do better in some tasks) because it simply doesn't scold me on every step.
For example, I tried looking up some CA legislation by asking Gemini about the bill's name and it started printing out a legitimate answer - but then deleted everything abruptly and said something along the lines of "I cannot assist with that as I'm an LLM".
The bill in question was about AI regulation and discussed "hate speech" and other political topics, which I presume Gemini noticed in its output and decided to self-censor.
Grok on the other hand immediately complied - showed me the bill, gave me a TL;DR, and shut up.
Another example is: I found a bunch of old HDDs from old laptops. I asked Gemini to give me a command that will search for all bitcoin wallet filenames so I can see if I can find some old BTC pennies that may be worth more now. Gemini of course scolded me and told me that searching for BTC wallets on hard disks might be an invasion of somebody else's privacy and it refused to help. Grok on the other hand cooperated and shut up.
And yes, I might have worded my prompt carelessly (e.g. "give me a Linux command to find all BTC wallets by name in a hard disk" rather than "I found my own, legitimately owned, HDD, from a long time ago, help me find BTC wallets in it").
But I shouldn't have to walk on eggshells talking to smart sand, and I won't.
At least two times they had unauthorized changes to their prompts to inject far right content that showed up on random content. imagine you're using it for a chat bot and it starts spouting off white nationalist content like "great replacement" theory.
What was the other time? The incident linked at the bottom of that article ("into trouble last year") wasn't an "unauthorized change", as far as I'm aware; it was a general lack of guardrails on image generation.
While I'm sure the same rogue "employee" was responsible for both, they are separate incidents. Musk's AI service was pushing "white genocide" lies as answers to unrelated prompts. It was only spouting holocaust denial lies when asked directly.
I’ve found 3.7 to be garbage. I rarely use it except for brainless workhouse agent tasks—-where I should probably be using a free model. It really mangles code if you let it do anything slightly complicated.
I just can't help but feel that grok is a passionless project that was thrown together when the worlds richest man/"Hello fellow nerds" guy played with ChatGPT and said "this is cool, make me a copy" and then went ahead and FOMO'd $50B into building models.
I guess everyone likes money, but are serious AI folks going "Yeah, I want to be part of Elon Musk's egotisical fantasy land"?
- Gemini is state-of-the-art for most tasks
- ChatGPT has the best image generation
- Claude is leading in coding solutions
- Deepseek is getting old but it is open-source
- Qwen has impressive lightweight models.
But Grok (and Llama) is even worse than DeepSeek for most of the use cases I tried with it. The only thing it has going for is money behind its infamous founders. Other than that, their existence would be barely acknowledged.