# Grok 4 Real Experience ## 0 Introduction Elon Musk released Grok 4 at Beijing time noon on July 10. After my nap, right when the launch ended, I immediately used Grok 4 through an AI pooling platform. Combining these days' real experience and some valuable opinions online, sharing my real usage feelings. ## 1 Pros ### 1.1 Data source includes X platform For Gemini, native Google Search enhancement elevates its capability. For Grok, X platform's high-quality data makes it stand out among AI products. ![google search is the largest ai product in the world](https://cdn.sa.net/2025/07/12/Vh6Sjq4lKRibagm.webp) X posts are unique to Grok. Other AIs connecting to internet mostly = connecting to web pages. ![X posts are unique to Grok. Other AIs connecting to internet = connecting to web pages.](https://cdn.sa.net/2025/07/12/YaW3pjiCOxn4SF2.webp) Overall, X platform has far more English high-quality data than Chinese high-quality data. X platform attracts many Chinese users mainly for NSFW content like daily competitions, plus lots of ads and random junk. ### 1.2 Agent LLM Second LLM vendor natively providing Agent LLM [not counting Minimax, Tiangong and other agent products]. The Agent LLM I'm referring to is Chat and Agent combined, not separate. First was ChatGPT o3 model, third will be Gemini's Agent Mode [reportedly coming in a few weeks, Ultra-exclusive]. Grok 4 is second. ![Grok4 is also an Agent LLM](https://cdn.sa.net/2025/07/12/YDbW2x4IJCdFZiT.webp) This Agent LLM's biggest use for me is fact-checking. Like when I collaborate with Claude on paper writing, after first draft of a paragraph and manual verification, I generally have ChatGPT o3 model verify the expression again. ### 1.3 Low censorship Low censorship was already Grok 3's biggest selling point. Recently Grok got caught up in anti-Semitic, atomic bomb fireworks and other PR storms. Meanwhile, Western bloggers found Grok tends to reference Elon Musk's views [Musk's toy]. Hope Grok continues maintaining low censorship—sometimes politically incorrect statements are quite interesting. ![Grok4 tends to reference Elon Musk's views](https://cdn.sa.net/2025/07/12/a9E3CIRrbF6qODJ.webp) Human world's opinion texts are used to train Grok, meanwhile Grok's reference materials also come from humans. Adding guardrails to LLMs is meaningless—censorship allowing only one voice actually breeds bias. ### 1.4 Grok Task I was already looking forward to Grok 4 when Musk first announced July 4th release, but unfortunately it didn't release as scheduled. Then I set a scheduled task in Grok to check daily if Grok 4 was released. The email notifications' beautiful layout deeply attracted me—this experience is far superior to ChatGPT and Gemini. ![Grok Task email layout is beautiful](https://cdn.sa.net/2025/07/12/7XB3JLOYTstqPlG.webp) ## 2 Cons ### 2.1 Image recognition capability Chinese image recognition is poor. English uncertain, but expected image recognition can't match Gemini and other models. ### 2.2 Deep Research base ![Deep Research base model hasn't switched to Grok4 yet](https://cdn.sa.net/2025/07/12/cCQIfgkDZPuq7vY.webp) Initially thought Grok pooling site doesn't offer Grok 4 Deep Research. After seeing Super Grok review content, realized Deep Research base model doesn't support switching to Grok 4 yet. ### 2.3 Chinese experience lower than English experience Yesterday saw an L Site post—[Grok 4 isn't bad actually, but I won't use it](https://linux.do/t/topic/780642). This comment caught my attention. Perhaps one reason for the Chinese vs English community reaction gap to Grok 4 is language. Though Grok 4 sometimes converts user Chinese questions to English before proceeding with tasks. ![L Site post comment](https://cdn.sa.net/2025/07/12/HujyhcRZm1G74ao.webp) Whether to use Chinese or English for Grok 4 questions—depends entirely on context. If Chinese context works better, ask in Chinese. If English context works better, ask in English. ![Whether to use Chinese or English for Grok 4 questions depends entirely on context.](https://cdn.sa.net/2025/07/12/Y9LOVxlMdXeoEak.webp) ## 3 Summary Grok 4 currently serves as a ChatGPT o3 model competitor on my local computer. Technical questions I generally ask Claude and Gemini. Fact-checking, news, etc. can consider Grok 4. Currently ChatGPT mainly uses o3 model for fact-checking. But these are just generalities—comparing various top models is the norm. Like last night I asked Grok 4 a technical question—a US New York VPS I bought in April couldn't connect to npm service. Grok 4 suggested changing VPS DNS settings. After following its advice, VPS could connect to npm normally. Currently, no way I'll pay full price for SuperGrok. $30/month not worth it. Pooling or freeloading SuperGrok around June 20's last batch of edu discount accounts is optimal. The pooling site I use—Grok pools are usually idle, while Claude, ChatGPT, etc. get busy. Seems everyone's not interested in Grok 4. Before Grok 4, only one roommate commonly used Grok 3 model. This roommate had previously subscribed to ChatGPT and Claude memberships but has now switched entirely to freeloading. When chatting, one of his statements left a deep impression: "AI services are all pretty much the same." This is indeed true—giving weak context to top models vs giving rich context to second-tier models—maybe second-tier solves your need while top-tier can't. Besides, Grok 3 was also among top model ranks at that time. ## Supplement Updates to this Grok 4 article—see Flarum post—[Grok 4 Experience Notes](https://bbs.gujiakai.top/d/15-grok4). ## References - [XAI Grok 4 Blog](https://x.ai/news/grok-4) - [Grok: searching X for "from:elonmusk (Israel OR Palestine OR Hamas OR Gaza)"](https://simonwillison.net/2025/Jul/11/grok-musk/) - [Simon Willison Grok 4 Blog](https://simonwillison.net/2025/Jul/10/grok-4/) - [Tibor Blaho Grok 4 Summary Tweet](https://x.com/btibor91/status/1943249002109243611) - [Grok Heavy Test Tweet](https://x.com/IterIntellectus/status/1943230837685928086)