LLMs ARE JUST A FUCKING TOOL
so stop using it as a lobotomy
I am fucking sick of the LLM hype. I’m not denying LLM can be useful, they really are, but the way people use them is horrible…
You open Twitter, and all you hear is how AI is going to replace us, how it writes code for you, how it’s the future of software engineering.
And there is Dario (Anthropic CEO) who has been saying for years now that AI will take our jobs in less than 6 months...
You are regressing
Let’s get something straight, right now. An LLM is a tool.
Just like a compiler. Just like a debugger. Just like a fucking drill. Now, here is the problem with tools: A tool should never, ever reduce your ability to work without it.
If I use a drill to drive a screw, I did not forget how to use a screwdriver. My hands have not atrophied.
But what I’m seeing now with developers... is people willingly outsourcing their thinking. You’re delegating the one thing that actually makes you an engineer, your brain, to a statistical parrot. I’ve watched people’s coding skills degrade in real-time.
They get complacent. They trust the machine more and more with every token generated and every PR merged. They see a blob of text that looks like code, they nod their head like a dumbass, and they paste it into the repo.
Or they ask a second LLM to perform the review, and I’m not saying LLM code review is bad, CodeRabbit is a great tool, but trusting it implicitely, removing human thinking from the loop is where issue start. You are losing your critical thinking skills because you’re letting the machine do it for you. And the moment you delegate thinking, that’s it. Game over.
You have been downgraded from engineer to a tool for a fucking chatbot, you are now the HID (Human Interface Device) for the LLM.
We have “senior” developers like Uncle Bob going on Twitter and admitting he can’t code anything anymore if Claude is down.
What the fuck?
If the cloud service goes down, you just stare at the screen? You’re the guy who wrote “clean code”! (I don’t agree with a lot of Uncle Bob takes, but nobody can deny he is (was?) a senior developer with a lot of experience)
They Know About It, And Bet On It
The worse in that labs are completely aware of the effect their product has on people, Anthropic, the lab making Claude Haiku, Claude Sonnet, and Claude Opus.
https://www.anthropic.com/research/AI-assistance-coding-skills
The sale pitch is the same as a fucking drug dealer, you will feel great right now, the high make you feel powerfull, and will become dependant of their LLM.
https://www.anthropic.com/research/AI-assistance-coding-skills
At this point it’s best to not talk about the Junior Dev, and Students that will never develop the knowledge, the critical thinking, and search skills… They are already hooked on the LLM, TikTok killed their attention span, LLM killed their ability to think…
And let’s talk about this “intelligence” that tech twitter worship. It hallucinates. It makes shit up. It got better at hiding it, “for shure” (on), but it’s still lying to you.
But even worse for coding, IT’S STUCK IN THE PAST
Those model are trained on data scraped off the web a few years ago. This resulted in funny issue, like GPT-5 requiring to have the name of the president of the USA in it’s system prompt to avoid naming an ex-president...
The world moves fast, but the training data do not.
Try to get GPT-5 to use Tailwind v4 instead of v3. It’s a pain in the ass because it just defaults to what it knows. Ask Opus 4.5 to write code that use Glibc, and it insists on using function deprecated in version 2.35. We are at 2.44…
They don’t know or care! It’s just statistical guessing based on old data patterns. And this is a big issue, historical data will always be more present than recent data in LLM training.
So what do we do?
Stagnate? MCP? AGENT.md? SKILL.md?
Should we stop innovating, from now on tech is stuck 3 years in the past and nothing new is developed. Do we bloat the context with a ton of MCP and AGENT.md just so the model has the new library in its context?
But MCP and AGENT.md context stuffing degrade the reply of an LLM, the bigger the context, the harder it is for them to retrive the neddle in the haystack.
So a Ralph loop then? Yeah... it just make the model retarded like Ralph Wiggum... (And apparently turn you into a crypto scammer if we look at the creator of the Ralph loops... but that’s unrelated...)
Oh, let’s use SKILL.md, then, the model can just call it when needed... No useless context bload, and the model can chose when to load it... Oh wait... fuck...
https://vercel.com/blog/agents-md-outperforms-skills-in-our-agent-evals
So now to fix SKILL.md you also need to stuff index reference to docs in AGENT.md. And add multiple instruction to AGENT.md to increase the chance the model chose to look at the docs...
And this can works “fine” (and I use that term loosely) if you’re doing the same generic web slop that 90% of the internet is made of. It will write TypeScript better than the average web dev (sure it’s a low bar), it’s seen a trillions of lines of React. But the moment you step out of that tiny little bubble? It falls apart.
Outside the Web Bubble
Try to write some embedded C mixed with Assembly. The thing has a seizure.
It starts injecting Arduino function call into a Motorola Coldfire codebase running a custom RTOS. It tries to call POSIX functions on bare metal... It has absolutely zero context for reality because it wasn’t trained for custom hardware, or real enterprise embedded system constraints.
It just sees “C code” and starts vomiting unrelated fragments into your editor.
So again, we start adding a AGENT.md, and tons of comments and custom prompt to force the model into barely following the project structure and respecting the limitation of the system.
At the end of the day, I could have written the code in half the time it took to fight the LLM, and the LLM code is not even optimal for embedded systems.
Now it’s time to review the code... It looks like C, a good start...
Oh... it won’t build?
Why is there an overflow in the ram segment ?!? And in half the flash pages !?!
Yeah... LLM don’t care about writing optimized code that will fit on the device...
I’m not writing a web wrapper running on a PC... I can’t be like discord and restart the app when it takes more than 3GB of ram… (They really do that...) I don’t even have a GB... let alone 3GB...
What I have is 32K of RAM and 256KB of flash, and I also need to fit a Bootloader and some logging area in that 256KB… I don’t have 6 or more cores running at 4GHz, I have a single core at 60MHz… It’s as far from web slop as you can be.
(To be fair to the LLM, if you use hobby device like Arduino, Raspberry Pi, ESP32, there is enough Open Source code online that it can do basic to medium stuff, but nothing optimised or reliable, still good enough for hobby projects)
The Fucking Cost Just Moved
LLM make code cheap, that is a fact, they can generate more LOC for the price than any engineer.
But, LOC was never a valid measurement of productivity. Rewriting the same feature so it has less LOC and is faster is productivity, but pumping out 200~300 lines for something that needed only 30~40 LOC is not...
But for LLM it’s all people look at, “look it wrote 10K lines of code today”, they don’t care that it can make big way more onerous to fix, and has huge security implications.
With the type of code I write, a small mistake could easily result in bricked devices and thousands to millions in cost, it’s not a simple, let’s roll back the Vercel build to the previous one, it’s sending service technicians in planes to customers location with JTAG programmers.
But the worst part is not the technology. The technology is just a stupid tool. The worst part is how people use it.
Look... I don’t fucking care if you use AI for your personal weekend projects.
Do it, Have fun. Make a little website, make a led blink with Arduino.
Post about it on twitter if you want, it’s cute.
But the moment that code touches production? The moment it’s connected to something a customer is going to use? The moments it handle data that should not be leaked.
Please! It’s not 2024-2025 anymore... so for 2026 stop leaking customer name, address, and credit card in your vibe coded SAAS. it makes all the industry look bad.
You need to fucking review the code and understand it
I really mean it. You review it line by line as if it was written by a retarded schizophrenic intern who hates you. Because that’s what the AI is. It’s a fast intern with zero common sense and a tendency to lie and hallucinate. If you don’t manually review the code, then you don’t own it, and “it’s the fault of the AI” is not a defence for not doing your job.
Real engineers need to understand this. When you live in the world of firmware and OS development like me, you know tools aren’t always available.
We’re used to ditching the fancy debugger and falling back to a serial port. Sometimes, when things get bad, or if it’s Wednesday, then you only have a single LED blinking at you. I once had to debug a system in the field using only a cheap portable oscilloscope and a single GPIO pin that I could pulse to follow the state of the program. If I didn’t have an oscilloscope, then the multimeter was the next in line, but with multi seconds pulse...
That’s problem solving... you know the thing you now have the AI do for you...
The ability to design, code, and debug without your usual (or favorite) tools is what separates a senior engineer from a code monkey.
The ability to work when your tool is down or unavailable is a SKILL. It is a REQUIRED skill. At this point, if you can’t function without your AI autocomplete, or write code without your IDE, we don’t want to hire you. You lack adjustability. You lack problem-solving skills. You are a liability.
Use the LLM, I sure do. Use it to speed up the boring stuff. Use it for stuff that don’t matter.
But for the love of god, don’t be dependent on it. Stop acting like it’s a replacement for a brain or the second coming of Chuck Norris, the LLM can’t count to infinity... twice.
It’s just a tool. Treat it like one, or get the hell out of my industry and my twitter feed Or mostly my twitter feed…
Source:
https://www.anthropic.com/research/AI-assistance-coding-skills
https://vercel.com/blog/agents-md-outperforms-skills-in-our-agent-evals
https://www.cnbc.com/2025/04/29/satya-nadella-says-as-much-as-30percent-of-microsoft-code-is-written-by-ai.html
https://fortune.com/2026/01/29/100-percent-of-code-at-anthropic-and-openai-is-now-ai-written-boris-cherny-roon/