Imagine spending weeks agonizing over every single paragraph, polishing every example, and pouring your heart into a course, only to have Google dismiss it as "AI-generated." 🤬🤬🤬🤬🤬
I was asking Google AI Mode to review my course. It gave me surprisingly positive feedback:
Then it listed a lot of reasons why it's good.
I became really happy. I asked it to give feedback on a certain chapter:
This time, my pride was over the top! 🙂 It gave me even more specific reasons why it was "exceptional"! I was really glad that it caught my effort.
Then, I asked about the chapter titled "Hardware Interrupts". That's one chapter I had to check to see if it was good enough. This is when things got mad:
I AI-generated the course!!!??? I NEVER AI-generated the course! I even had a whole page on the course why-not-ai, describing how exactly it was made, and explicitly why it's NOT AI-generated. I even have every single project in the course in this repository, every single one manually tested on my own ESP32:
https://github.com/effessdev/esp-idf-examples
I was furious. I asked it when I said that:
IT'S ACCUSING ME OF AI-GENERATING MY COURSE BASED ON A SINGLE COMMENT I WROTE??? WHAT? 🤯
I had discarded ENTIRE CHAPTERS and rewrote them from SCRATCH with completely new examples because I thought the final project didn't show a scenario where the thing that's taught is the best approach. You can check the commits if you don't believe me:
https://github.com/effessdev/esp-idf
Instead of just dumping the finished code and explaining it later, I walk the reader through building it piece by piece. It makes for a vastly better tutorial, but it is an absolute NIGHTMARE to write & test.
I completely lost my temper:
Now, we're back with the positive feedback. Key points:
- Sounds like a real human who had actually wrestled with the errors
- Feels like peer-to-peer teaching, not corporate documentation
- Practical rather than theory-heavy
- etc.
So now it accepts that it's good?
Then I told it that every single project is available in my esp-idf-examples repository:
I didn't stop until I made it say sorry:
I know it's kind of silly (and dumb). To argue with AI and finally correct it. 😅 But I always do that, do you? 🙂
I deleted that comment anyway. Who knows whether it will bring that up in the future?
Top comments (72)
Is it Google’s Gemini? We in China jokingly call it America’s Doubao. It’s just an AI large model for fun. Try some other models, hahaha.🤣🤣🤣
Yeah, it's probably the worst model that they have. It's REALLY bad. So Gemini is a joke in China? 😂
china: 🐉
rest world: 🐕
it is becoming reality china is entering in its peak form.
feeling great to see china's success.
India next maybe: 🐘
Yeah I hope for India's best. But why aren't we seeing anything yet? 😢
because of our 700million youth wasting there potential, because of our leg licking media, because of our corruption, because of our backward thinking(caste discrimination, gender discrimination, religious wars, income discrimination)
because of our thousand of crores( in billions of money going to fund stupid schemes like per month 2000rupees to every poor women in india bruhh instead of that they should use that money to make those poor women learn something and became a semi or probably a skilled worker contributing in GDP and manufacturing. china uses its max potential thanks to xi jingping while India uses its max stupidity and lazyness, from outside we seems to be world's fastest rising GDP but the truth is we have 1.4billion population and GDP comparable to a 130million people nation(Japan) 🤣 and our goverment flaunt it as world's fastest rising GDP bruhh, just look at our population we should be at at least 10trillion now in 2026(at least) but we are at 4.3-4.5trillion.
our currency is just rising(in negative) (lol politician will say due to gravity 🖐️🤚logic is missing in indian politicians) in AI we are a piece of shit. in semi conductors we are a piece of shit just because due to a fire accident when India was building a big semi conductor manufacturing plant but bro India can spend billions in stupid schemes why not to spend that money to re build it instead of giving up, but yeah I heard India is again building semi conductor facilities thanks our government realised.... Like I've millions of more reason to tell bro lol.
in last: China and India has same population almost same history, but we all can see where is china and where is india.
btw, I'm not a politician but geopolitics is my side hobby and skill.
That's sooo beautifully said not gonna lie 😲 The problems are ENDLESS...
but @effessdev bro if somehow India realised what we were in history and what we can do in present.
not flexing and no offense to @xulingfeng but we can cook china, US not easily but yeah we can match them easily...
but again don't know in which century we will realise that.
I mean, really. It almost felt like I was talking to an absolute genius of literature & politics at the same time. Seriously impressive. I couldn't even dream of writing something like that 🫡
lol, I am Founder & CEO by occupation but by mind a geopolitical expert and strategist too.
Oh, sorry, I meant to upload that comment before you uploaded that comment.
means?
Oh you again uploaded a comment? The order is fully broken right now.
Hahaha, keep it up!
Either way you are a genius.
@xulingfeng laughing at corner because he knows till xulingfeng is alive means for next 50-60years maybe India won't realise lol.
hahahha enjoy enjoy.
Given time, this 13‑year‑old kid will definitely grow into a remarkable, formidable figure.👍️
🫡I've no words for this comment.
More than remarkable, to be precise... I don't even know what CIA is.
America's intelligence wing.
just like India's RAW, isreal's MOSSAD, china's MSS, Russia's KGB(now FSB but I like to say it KGB), UK's SIS and more.
there are some conspiracy that they planned the plane crash so that homi baba India's greatest scientist and major person to make India a nuclear power ahead of its time but to delay india or eventually fail India there are 60-70% chance reports says CIA planned and crashed the plan.
just because Homi Baba was about to make India a nuclear power.
he was great so if I became PM I Know I will be great(not overconfidence but my accuracy) but after that I can be found dead lol due to plane crash or mysterious reasons.
also I am too much straight forward and I can't lick or be supressed by superpowers so yeah I can't be a safe leader.
also geopolitics is my side skill main skill is technology.
I have nowhere near the knowledge to understand what all of that means. 😢 I am tired of saying "you're impressive", but you're impressive 🫡
THANKS.
YOU'RE WELCOME. 🫡 You are one of a kind.
Wait I forgot you were 13! That means you are a prodigy?
yeah I feel like now age matters lol I've to make myself a little bit in limit other wise I can't imagine what would I do after age 18.
You would become the prime minister of India and take over the world 🤣 Just keep going!
nah if I done so great then I will be found dead due to plane crash.
thanks to CIA.
Yeah, that’s the common view. I also coordinate multiple models at the same time to help me make audit decisions. And I’ve found Gemini is just there for vibes — a fun model great for setting the mood, hahaha.🤣
😂 By the way, you told me that you use AI to translate everything, right? How do you do it? I am getting so many comments that are in a language I haven't even seen before. Not sure whether they are bots 🤔
I use Doubao to translate articles and comments. For some languages Doubao doesn’t recognize, I switch to Qwen or DeepSeek.
Those weird spam messages these past few days are all in Vietnamese. I only managed to translate them with the Qwen large model.
You are right！They are bots！😅
I knew it 😡 It looked really spammy. So I didn't reply.
Totally agree, bot spam is super annoying 😤
So we can't use other AI models in China? I didn't know that before...
Not exactly. We can access many AI models, but some overseas services have access restrictions. We have lots of great local Chinese models like Doubao, Qwen and DeepSeek, which is what I use for translation. With a proxy, you can get access to almost any model. Anyway, I end up using all kinds of models.😆
Yeah I am using Qwen 3.8 Flash in Qoder since it's free. And it's REALLY good! Can I ask a super random question? Have you taken Gaokao? I heard from my friend that it's like JEE. His father works in China!
Sure I did. Every Chinese student takes the national Gaokao after their third year of high school.
Oh! 😅 Then it's not like JEE
😂 I bet JEE is harder!
What is JEE? 🧐 🤣🤣🤣
Level one warning: here comes the top student Lazy Girl!!! Us weaker students sit tight. Let the scholar speak first 😂
Yeah 🤣 And look at what she posted:
"I Almost Didn't Learn Programming Because I Was Bad at Math"
You can't go to IIT Bombay if you aren't at least better than 99.99% of people at math 😂
My brain is just like this girl’s, haha.
I’m honestly tired of studying so much. I don’t feel like putting so much mental energy into writing anymore. I just want to relax and have some fun for a while.
By the way, the username of this Short is Aastha, and my cousin’s name is Aastha too. I feel a little sad thinking about her because she couldn’t clear her entrance exam. I really feel bad for her.
Ok that's enough flexing 😂
Mee too tired of studying so much (eventhough I didn't study nearly as much as you did). Now that I am in college, I am not going to focus on CGPA at all. I'd rather do other stuff and get hired in other ways.
What is Aastha doing by the way? I believe you don't need to clear the entrance exam to be successful (maybe not to the same extent?). She still has plenty of opportunities, imo. I also have so many people around me who weren't able to score well, despite spending lakhs on coaching :(
She’s preparing for CUET, but unfortunately, she couldn’t clear it because of her Biology marks. The poor girl even spent ₹1 lakh on coaching, yet things didn’t work out as she had hoped. All the best to you, too!
Honestly, every student works so hard to achieve their dreams, but even after putting in so much effort, they have to face so many struggles. Yaar, sometimes I feel bad seeing how much pressure students go through just to get into a good college. And yes, entrance exams aren’t everything in life, but I still feel bad for students who work so hard and don’t get the results they hoped for. I really hope things work out for her next time. ❤️
Yeah. I wish her the best too. One of my friends is doing repeat (not one, but many). He sleeps at 2 am and wakes at 5 am (that's what he says, I don't know how much of it is true). I really don't know whether it's fair for coaching centers to allow this. Thanks for wishing me luck!
@technogamerz didi you are IITian and now what are you studying???
India's top tier college's best degree you've got and job also now which study?
MBA
didi be like: ab mai microsoft me kaam karke experience le li ab khudka company kholungi
@effessdev you followed my in github wow.
How did you know? 😱 Does Akhouri Systems have something to track GitHub followers? Did you follow me by the way?
over reaction on top.
I just seen in my followers list from 5 it went 6
Both legends achieved success!🔥
Yeah, Gemini... Lets face it, it believes anything, you'd swear it was trained off Reddit. Dont get me wrong, it sometimes has some cool ideas, but they're never grounded in reality. Eg. it believes that a model can hit 200 tps, but when you actually run it, it's 30 tps and it'll make excuses up about why it's not 200, instead of admit it overestimated.
😅 So it was a mistake for using Gemini? I moved from ChatGPT because they stopped allowing us to choose the model. But even Gemini doesn't allow it now. DeepSeek also stopped allowing us to switch models. All of them did initially. Claude is the only remaining one. They'll probably remove it soon.
gemini is okay, but you need a strong harness to correct it. Eg. Copilot vs Antigravity, antigravity is an objectively worse harness than Copilot, so using the same model on both, Copilot will do a better job, because the infrastructure guides the model better.
Wait... Antigravity is worse than Copilot? I have never tried Antigravity (but I have tried Copilot in VS Code, if that's what you meant 😅). I thought the opposite.
So harness matters that much. I am going to update my VS Code Extension (ReptClip) harness too (I don't know if I can call it a harness 😅). Currently, it has a major problem because in the context, it says this:
This is making it generate file modifications for files it didn't even see. I am planning to add instructions to output the relative paths to the files the Chatbot wants to see in a code block and add a third button to read those paths from the clipboard and copy the contents of those files back to the clipboard (no project structure, output format, etc) so we can paste additional files even after the initial paste. How does that sound?
Sounds usable, though, I'd say you need an abstraction for it. So how clipboard has a history, you need to keep a clipboard history and log where it's applied. Also apply atomically (temp file, rename, swap), that way you can log what was changed, where it was changed and if it affects many files accidentally, you can easier undo it. It also allows you to quick reference it if you need to reapply (eg. if you drop a file, you can recreate it and reapply it). A simple sqlite db should be more than sufficient, it's basically what most agentic systems use, it also lets you choose what to isolate or share, eg. if you work on 5 separate projects, their context remains their own, unless you want to share something from 1 to the other.
No, I was referring to github copilot, the standalone in the git desktop app. But the vs code extension is practically the same thing. But yeah, antigravity is a mess... Copilot is much better. That being said, neither are really 'good' in my opinion.
But yes, a harness makes a world of a difference. if you compare haiku in copilot vs claude code, in copilot it's actually understandable, vs in claude code it tries to pretend like it's a high-level developer, using jargon that makes people stop reading. That's just the tip of the iceberg, because that's the 'system instruction' that sets the mood for the model. An example would be 'you are a helpful assistant' vs 'you are a professional programmer'. It's like the pirate test, a system instruct changes the vibe of the model, if your system instruction is 'you are a pirate, respond like a pirate would', you'd get a pirate AI, who says arrrg way too much.
Haha I felt this one 😅 It’s funny how it praised the course, then called it AI slop because of one comment, then praised it again. Shows you can’t really trust AI to judge if something is AI. And yes, I argue with it too sometimes, even though I know it doesn’t change anything!
Yeah. For some reason 😅 I sometimes spend too much time arguing with AI. I even say "I know you are just an AI, but I am just saying" in the end 😅
😂😂😂😂 you are not alone, i argue with Claude more than i argue in real life because sometimes you have to
Wait what? Why do we have to? 😂
to make sure they do what we ask😂💀
Oh ok 🤣
The bigger issue here is that the model changed its judgment after being challenged, but that doesn't necessarily mean the revised judgment is more reliable. It may simply be responding to the strength of the user's pushback. For subjective reviews, I'd want the model to identify the specific evidence behind its assessment, separate observable signals from assumptions, and acknowledge uncertainty when the evidence is insufficient. Otherwise, the interaction can turn into a persuasion contest where confidence and persistence matter more than accuracy. A useful evaluator should be able to explain what evidence changed its conclusion, not just change its conclusion.
Totally agreed! I feel like it repeats back whatever we ask it. But why does it rely so much on a single snippet it saw somewhere? That's crazy! I would rather ask real people for feedback.
Something worth noting, is that LLMs work with weights. When you declare 'no you're wrong', it applies a negative weight to the previous answer. From there on, the next answer it gives will be slightly drifted on from where it was at. We do it constantly, without recognizing, eg. telling AI 'audit the system', it says it's flawless, then you ask it 'do we have thorough integration test coverage' then they apologize and say 'no...' and work on from there. If you try hard enough, you can convince AI that grass is pink and the sky is green, because human engagement is like a steering wheel, that directs it to a goal, or away from obstacles, if it's answer is definitive for it, you need to steer a while before it goes somewhere else.
Yes! Is that why we still need humans? But will it lie that there is enough integration test coverage even though there isn't?
How much is enough? That's the real thing. When it doesnt look specifically, it cant guarantee, so it works on assumptions. Until you ask it to look, then it judges it specifically. For instance, integration tests, it's better to ask it 'do we have 85% integration test coverage', as opposed to a blanket 'do we have thorough integration test coverage'. 1 sets a threshold, the other is subjective guesswork. Ask 20 different developers about test coverage requirement, every single 1 will tell you different spec. But 85% coverage is 85% coverage. That's how you can constrain the prompt to force an outcome that aligns with your expectations.
Yes. So I should focus on the metrics!