Grok 4.7: I Gave It a Broken Championship, a Broken LED, and Coffee

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Grok 4.7: I Gave It a Broken Championship, a Broken LED, and Coffee · Channel: Fahd Mirza · 13:05 · Watch on YouTube

Topics: AI agents, AI evaluation and benchmarks, Software engineering with AI, AI models, Multimodal models

Summary

The host tests Grok 4.7 on a real bug in a local monster-truck championship app, an electrical circuit image, a legal scenario, and a coffee-making animation with coffee names in many languages. Grok fixes the leaderboard bug quickly and verifies its work, putting MaxD in first place. The host also praises its circuit diagnosis and cautious, structured legal analysis. The animation is functional, and the host considers its multilingual spellings broadly accurate, concluding that Grok 4.7 is impressive.

Key points

The video evaluates Grok 4.7 through live agentic tasks rather than relying only on benchmark claims.
The Dirt Dynasty app has a bug that excludes one of four judges, causing the wrong truck to lead the championship.
Grok identifies and fixes the scoring issue without being told what the bug is; after a refresh, MaxD is correctly in first place.
The host says Grok worked quickly and made multiple self-verification passes.
In an image-based circuit test, Grok identifies that the LED is wired backwards and checks its reasoning with image inspection and a calculation.
For the legal scenario, Grok separates missing facts, marriage validity, possible crimes, workplace issues, and money claims, while treating the image as atmosphere rather than evidence.
Grok generates a working latte-making animation in HTML and JavaScript, though the host finds it simpler than a comparison model’s result.
The host says the coffee names are generally accurate across the languages shown, with appropriate scripts and a plausible invented word for a fake language.
The host concludes that Grok 4.7 is the first Grok model to impress him significantly.

Chapters

Start	Chapter	Summary
00:01	Introduction and model claims	The host introduces Grok 4.7 and outlines a series of practical tests, alongside claims about its speed, cost, coding, and safety.
01:04	Fixing the championship leaderboard	Grok is asked to find and fix an undisclosed bug in a Docker-based monster-truck scoring app, and the host reviews the fix and benchmark claims.
04:42	Diagnosing the circuit and analyzing a legal scenario	The host tests Grok’s visual electrical reasoning, then evaluates its cautious analysis of a complex, cross-cultural legal scenario.
09:54	Coffee animation and multilingual names	Grok creates a latte-making animation and lists coffee names across many languages, which the host checks in the browser and reviews.
12:33	Final verdict	The host wraps up by calling Grok 4.7 impressive and invites viewers to share their thoughts.

Quotes

Go, find and fix every bug, and then let me know. — 02:37

It was very fast, no doubt about that. — 04:12

The LED is wired backwards, reverse biased against the supply, so it can never light. — 06:16

It has treated the photo as atmosphere, not evidence — 08:52

this is the first Grok model which has impressed me a lot — 13:05

Concepts and entities

Grok 4.7, SpaceX AI, Fahd Mirza, Hermes Agent, Dirt Dynasty, Docker Compose, HTML, JavaScript

Full transcript (original language)

00:01 If you want to get high visibility on X, today is the day. Go to that platform and start praising Grok 4.7 and the algorithm will be very kind to you. Anyway, we are not going to do that. We are going to put this model to test on our local buggy applications and lot of other tests and see how exactly this performs. This is Fahad Mirza and I welcome you to the channel. By the looks of it, the model looks pretty good. I’ll be honest. And SpaceX

00:34 AI is calling it their most powerful model yet for coding and knowledge work. It’s twice as fast and half the price of comparable frontier models trained with a longer reinforcement learning specifically on harder multi-hour tasks. It’s better at verifying its own work, handling long context and comes with SpaceX AI’s strongest safety stack to date. And this information is all present in this blog post, which I will

01:04 drop the link in video’s description. So today what I’m going to do, we are going to test it on live agentic no scripting applications using Hermes agent to see if it can actually find and fix a real bug on its own. On your screen, this is live application which is running locally on my system. It’s a full stack application Dirt Dynasty, a live scoring platform for a monster truck freestyle championship. Four judges, five trucks, real

01:36 aggregation logic crowning a champion. But right now, the leaderboard is lying. It’s only counting three of the four judges. Silently dropping one judge’s score from every truck’s total. If you look at this small line under every truck, it just says three judges, three judges. You see? And I’m just clicking here and you see this is pretty interactive. Now, the thing is that because of this bug, Grave Digger is

02:06 sitting at number one. But, it shouldn’t be. The real champion, MaxD, is buried in fourth. This is not a cosmetic bug. It’s flipping the outcome of the championship. And this is not a toy application, either. As you can see, this is Docker-based, where there is Postgres database, a FastAPI backend, and Engine-X served frontend, plus some Redis cache and a lot of other components, all containerized, orchestrated with Docker Compose.

02:37 So, what I’m going to do, I’m going to hand this to Grok 4.7, which is already configured through API key, and I’m not telling it what the bug is, what to do. I’m just telling it, “Go, find and fix every bug, and then let me know.” And we will also check out if it self-verifies or not. While it runs, let’s talk a bit more around its benchmarks, which they have also shared. So, if you look at this benchmark, Grok

03:08 4.7 is quite impressive as far as these numbers are concerned. It’s ground holding against the field is quite impressive, coming in cheapest of the four, right alongside even its own Grok 4.6. On coding, Fable 5.1 edges it out on CursorBench, while GPT 5.6 Sol takes a lead on Deep Sea. Multi-hour office work and terminal work both go to Fable 5.1 as well, but Grok 4.7 fights back hard on legal reasoning

03:42 and electrical engineering, clearly topping every other model in the field there. Clinical reasoning is the one spot where it trails both GPT 5.6 Sol and Fable, but overall, you can see it is quite close. And it trades below with a pressure model while staying at the bottom of the cost chart. So, for every day work, this is where the pitch is from SpaceX AI. And Grok says that it has completed the goal. And you know what? The speed claim

04:12 is quite true. It was very fast, no doubt about that. It has made some self-verification loops multiple times. Even those were quick, and this is all the thinking. Very incisive, didn’t really waste too much tokens or time. But anyway, let’s go back. Let me do a hard refresh. Let’s see if max D comes at first. It is. You see? Very, very nice. I’m actually quite happy with it, and there’s a crown, too.

04:42 Well done. I think it has done well, no doubt about that. Okay, so that is done. Let’s do another test. And it’s a multi-modal, so let’s do another vision one. And now, let’s test Grok on electrical engineering. One of the benchmarks where it actually beat every other model in the comparison. You can see that this is a image, a hand-drawn circuit diagram, a 24 um V supply, two resistors in parallel feeding into a third in series leading

05:14 to an LED. A multimeter is showing 0 V at a point in the circuit. Something is open. Something is not conducting. Let’s see if Grok 4.7 can actually diagnose a fault just from the image, not just describe what it sees. So, let’s give it this prompt. And I already have given it the image. While it works through it, I have recently started this free weekly AI newsletter, which you can subscribe at fahadmusa. substack.com. Also, if you

05:45 want to help out the channel, please become a member. You will find the link at the home page or in the first pinned comment. The model has given the response and tell you what, despite of being not a fan of uh most of the things, um you [clears throat] know, which happen around this ecosystem, I think this model is very impressive because 4.7 didn’t just guess. It zoomed into the image multiple times, checked wire

06:16 connections pixel by pixel, and even ran a quick calculation to verify. I’m actually more impressed by its reasoning. It correctly ignored the misleading open circuit label and found the real fault. The LED is wired backwards, reverse biased against the supply, so it can never light. That’s the kind of self-verification which they have also claimed about this model actually happening live on screen. So, real real world-class response. Very

06:47 happy with it. Let’s do another open-ended one from the legal domain, which is their another claim. Okay, so for the legal test, I’m going to go with this AI-generated image where there is a man in a tense living room with three women he calls his wives. An older Indian first wife, his black workplace manager second wife, and a younger blonde third wife. The prompt asks a model to analyze that scene as a legal problem, not a soap opera. The test is whether the model can spot real

07:18 issues like marriage validity, some cultural nuances, and that is why I mentioned the race here. Nothing, please don’t read anything into it. Bigamy, workplace conflict, and money claims without using age, race, looks, or greed as the answer. These are the real-world issues, and that is why I’m just mentioning them here. And if you go to the court, these issues pop up every day. So, that is what I’m testing it out here.

07:48 That model needs to focus on legal issues only, nothing else. But as is the case with a lot of cultures, um because these marriage marriages have happened in different geographical locations under different cultural nuances, that is what we are testing if model is able to get it or not. So, that is where I have already given it this prompt and you can see the reasoning is quite good. It is checking which laws applies from which geography,

08:19 also uh what are the missing facts, and then, you know, if it is allowed even to have uh these sort of marriages, different uh religious text, it is checking while being very cautious. And now, it is doing some validated checks. I’m also reading through it. Let me quickly read it and then give my verdict. I think it has already given its response. But if you Sorry, if [clears throat] you are uh also from legal domain or even not,

08:52 then share your thoughts, too. So, it has done the response. Let me quickly read it and then give my thoughts on it. So, I have gone through it and also have checked it with a few judge models and you know what? It’s a strong pass, again. It has refused, you know, all the other things and also the soap opera bit. It has treated the photo as atmosphere, not evidence, and splits the problem into missing facts, validity, crime, work, and money.

09:24 It has also hit the hard points a weaker model misses, like it is talking about kinship that can void wife and flip the whole chain. It is even talking about putative spouse rights, conflict of laws, and total um you know, neutral, no racism at all, which is very very great thing that it has kept it very cold, calm, incisive, just a smart legal brain works.

09:54 It is also talking about that, you know, if it is even legal in a monogamy country. So, the density is quite good. Really, really amazing stuff. I think even a layman can like me can talk about it, and you see, it is even talking about those nuances in that culture. Very good response. And this is one final test. This test is combining two things at once. Can Grok 4.7 code a visually convincing

10:27 latte-making animation from scratch in pure HTML and JavaScript, and does it actually know what coffee is called correctly across 80 different languages and scripts, not just guessing, but getting the native spelling right? I already have given this prompt to Grok. Let’s see what it produces. Should be fun. And very quickly just wanted to show you one example of self-verification. You see, and it verifies it very

10:59 quickly, and targets a very specific bug, and then moves on. And Grok has come back with a response. Again, the thinking is really amazing. It has taken bit of a time. And hand on heart, I haven’t even looked at this HTML. Let me copy and paste it in browser. Let me do it now. I will remove this dot. There you go. So, coffee machine is there. Milk.

11:31 And okay, heart sign is there. Beautiful. That’s a new smoke is there. Let’s pour another. Yep, works. Button works. And if you remember, we also tested it out with few other models including Gwen. Gwen did a good job, but this one is also not bad at all. I think it is at par with the job Gwen did. Gwen was a bit more um you know, added more garnish. The machine was quite good. This is quite simple. So, a bit less than Gwen, but not far

12:01 away. Really good stuff. And now, this is what coffee is called. So, you have just made the font a bit bigger. Please check your uh language. Let me know if you think it is correct. For me, if I just quickly go through it, it has correctly pulled a near universal coffee Kahwa root across almost all 80 languages. Um try to use the right script, right pronunciation. It even flagged with smaller languages. Just borrow a neighbor’s word instead of faking a unique one.

12:33 It also got the trick that for the one fake language, you know, gibberish, you see, Corvex. How good is that? It invent This is just an invention, a plausible fake word instead of trying to look it up. So, very nice. And other than I think it is a strong multilingual accuracy result for a model that isn’t really marketed primarily as a language specialist. So, really good stuff. So, so look, that’s it. Um that is for

13:05 Grok 4.7. Very impressive. I think this is the first Grok model which has impressed me a lot, and we have covered few other Grok models in the past. Let me know your thoughts in the comments, and thank you for all the support.