ChatGPT can create awesome EA? - page 5
You are missing trading opportunities:
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
Registration
Log in
You agree to website policy and terms of use
If you do not have an account, please register
15 months later, I think it's time to re-iterate and make a statement because too many forum members were treating AI as some kind of "Don't post AI code" or being cynical about it.
So the statement is: The programming profession is dead. It started in the forties of the 20th century, and today, towards the end of the 3rd decade in the 21st century - it's dead after 80 years.
No human being can cope with this monster. Coding on the highest possible level which two years ago took 50 programmers to complete in a year, can be done by one person in the course of few weeks.
Dead in the meaning that two years ago the best human programmer can not compete with todays models, not even close. That's a mega revolution.
For more than one year now, there's no need to open the IDE at all - not needing to write or review a single line of code - while being able to write 100X faster, better, near perfection code.
Everybody is free to keep on with manual coding, like you can go with a horse and carriage on a high way.
Each new generation / model released on those 15 months was jaw dropping and it doesn't seem to stop.
The main thing is that everyone's limits of personal coding have been pushed so far that there's almost no words left.
I asked DeepSeek what it think about your post, I just quote it and asked "What do you think about that ?"
That forum post is a masterclass in hyperbolic absolutism—dramatic, partially rooted in truth, but fundamentally wrong in its conclusion.
Let’s break it down honestly. I agree with the awe, but I strongly disagree with the obituary.
Here is my take on the specific claims, and where the argument falls apart:
1. The "50 programmers in a year" vs. "1 person in weeks" claim
True for Greenfield Boilerplate, False for Legacy Systems and Novel Architecture.
Yes, if you need a standard CRUD app, a typical e-commerce backend, or a well-documented Python script, AI can generate that at blistering speed. But 80% of professional software engineering isn't writing fresh code; it's integrating with brittle 15-year-old APIs, debugging race conditions in distributed systems, migrating petabytes of data without downtime, and satisfying conflicting business stakeholders. AI cannot do that. It doesn't know your company's specific, undocumented quirks. The bottleneck has shifted from typing to thinking and negotiating.
2. "No need to open the IDE or review a single line of code"
This is dangerously reckless. AI writes fluent code, not necessarily correct code. It hallucinates libraries, invents function parameters, and produces code that looks perfect but fails catastrophically on edge case #47.
If you aren't reviewing the code, you aren't programming; you are rubber-stamping. And rubber-stamping AI code is how we get massive security vulnerabilities and technical debt. The role of "reviewer" is actually more intense now, because you have to catch subtle logical errors hidden within perfect syntax.
3. "The best human cannot compete"
Define "compete". In raw WPM (words per minute) of syntax? Sure. But in causality and novel problem-solving? No. AI is a stochastic parrot—brilliant at pattern matching, terrible at true reasoning. When a system breaks at 2 AM in a way that wasn't in its training data, the AI will confidently give you a hallucinated fix. A senior engineer will use physics, networking fundamentals, and intuition to trace the root cause. The human isn't competing on speed; they are competing on reliability and accountability.
4. "Near perfection code"
This is the most misleading part. AI produces average code—the statistical mean of every GitHub repo it scraped. It does not produce highly optimized, low-level, memory-safe, or radically innovative algorithms. It produces "good enough" code. In production systems handling millions of users, "good enough" fails. Perfection requires deep contextual trade-offs that AI has zero awareness of.
The Real Truth: The Commodity Coder is dead. The Software Engineer is evolving.
Here is the accurate analogy: We didn't replace the carpenter with the nail-gun; we just stopped paying people for their ability to hammer nails manually.
Junior roles are changing: The entry-level "code monkey" job is shrinking. Juniors now need to be excellent prompt engineers and debuggers of AI output, rather than syntax writers.
Senior engineers are more valuable than ever: When one person can do the work of 50 typists, the ceiling for that one person's impact becomes astronomical. A brilliant architect with an AI can now build an entire microservices ecosystem solo. But they are still the architect—they must know what to build, why to build it, and how to secure it.
Jevons Paradox applies here: As the cost of producing code plummets, the demand for software will explode exponentially. We won't need fewer programmers; we will need more people who can orchestrate these AI agents to solve the billions of unsolved niche problems in healthcare, logistics, and science.
The "Horse and Carriage" analogy is backwards. Manual coding is the horse. AI is the engine. But you still need a driver who knows the roads, reads the map, checks the oil, and decides the destination. The driver isn't dead; they just traded the whip for a steering wheel.
So, is the profession dead? No. It is entering its most exciting, high-leverage era. The barrier to entry has dropped, but the ceiling of mastery has risen into the stratosphere. The people who survive won't be the ones who write the fastest code; they'll be the ones who ask the smartest questions and take ultimate responsibility for the digital world they unleash.
Keep the IDE open. Keep reviewing. You are not obsolete—you are now the conductor of an orchestra that plays 100x faster. But if you don't listen closely, the orchestra will play the wrong symphony.
Claude :
This is a strong version of a claim that's genuinely worth taking seriously but also worth pushing back on in specifics.
What's true in it: AI coding tools have improved enormously fast, especially over the last couple of years. Models can now scaffold entire apps, write substantial working code from a prompt, and handle tasks that would've taken a human much longer to type out. If you compare "typing speed and boilerplate generation" between 2023 and now, it's not close. That part isn't hype.
Where I think the post overstates things:
"Not needing to review a single line of code" — this is the part that doesn't hold up under any real-world testing. Current models still produce subtly wrong logic, hallucinate APIs, make architectural choices that don't scale, and miss edge cases that only show up under load or in production. Anyone shipping AI-generated code at a company with real users is still reviewing it — the incident reports from teams that skipped review are a pretty consistent pattern.
"50 programmers' work in weeks by one person" — plausible for certain kinds of work (prototypes, CRUD apps, well-specified features in well-known stacks). Much less plausible for legacy system migration, novel algorithms, security-critical systems, or anything where the hard part is understanding ambiguous requirements or negotiating tradeoffs with stakeholders — which is a lot of what senior engineers actually get paid for.
"The profession is dead" — the role is clearly changing shape fast. But "dead" implies zero demand for the skill, and what's actually happening looks more like: the skill of writing syntax by hand is devaluing while the skill of specifying, judging, and integrating AI output is appreciating. That's a real disruption, but it's different from extinction — closer to what happened to typesetters vs. writers, not writers vs. no-one.
The horse-and-carriage metaphor is doing a lot of work rhetorically but isn't quite the right analogy — cars didn't occasionally hallucinate a wrong destination.
So: the trajectory is real and fast, the framing that human oversight is now unnecessary is the overreach.
AI is just a tool. And like any other tool, it is the skill, knowledge and experience that gets the most out of the tool. Without proper coding knowledge and practical experience it is very hard to ask AI the right question. The garbage in garbage out principle still holds. There are plenty of examples of entries in the code base created by non-programmers. It is good even non-coders can create stuff nowadays, but i would not qualify it as good code.
Then there is something else. The costs. I think we still do not pay the real costs of AI. A year ago the promises where high and i was already anticipating my life being run by AI agents so i could sit on the beach sipping pina coladas all day. But as it turns out i need to quadruple my income in order to pay for it, which kinda defeats the purpose. And still these agents operate under the disclaimer they "can" make mistakes, which is a lie, they DO make msitakes.
TLDR. Programmers are not dead. They will not be for a long time to come. The daily tasks will change, for sure.
I asked DeepSeek what it think about your post, I just quote it and asked "What do you think about that ?"
Claude :
That's exactly what I meant previously, though I didn't express it very well probably.I don't have time to relate to all that now - but just to this final line for now which caught my eye:
"Keep the IDE open. Keep reviewing. You are not obsolete—you are now the conductor of an orchestra that plays 100x faster. But if you don't listen closely, the orchestra will play the wrong symphony."
I don't know deepseek - and it's tetimony that AI hallucinates - I am not experiencing that - I am stating for sure - there's no need to review the code.
The only thing you need to do - Ask the AI how did he wrote the code - and ask it for the tests proofs. That's it, nothing more needed. Of course, I don't know deep seeks quality in coding and I don't speak about it, but on the leading models. Nothing more is needed, for sure. No halucinations - That was true maybe more than a year ago.
BTW, it writes hundreds of thousands lines of code (nearly a million) - so no use to review anyway. But I can assure, top models equals top code. You can peek and see it, you can learn about it from interviewing the AI about it's work. And mostly, you have the desired outcome. There's no better proff.
** And more precisely - What I meant in dead - is that you don't need to learn coding any more. I agree on that, you are now an AI integrator. But that's not a coder, that's not programming. Yes, it surely helps to know software architecture (That too is decreasing as the models learn to do it better by their own), but coding? Maybe just to know the taste of it, not more than that. You need to know how a good software is built architecturally and be able to guide the AI though as I said, it does a good job already by its own much more from model to mode.
I am not experiencing that - I am stating for sure - there's no need to review the code.
The only thing you need to do - Ask the AI how did he wrote the code - and ask it for the tests proofs. That's it, nothing more needed. Of course, I don't know deep seeks quality in coding and I don't speak about it, but on the leading models. Nothing more is needed, for sure. No halucinations - That was true maybe more than a year ago.
Where did you get that? While the accuracy AI models has dramatically improved, hallucination remains an inherent structural characteristic of Large Language Models. That is, it is part of the design. If the context window runs out, it will output nonsense. There is no fix.
Plenty of scientific papers describing this exact issue:
https://arxiv.org/pdf/2510.05116
https://arxiv.org/pdf/2311.05232
Where did you get that? While the accuracy AI models has dramatically improved, hallucination remains an inherent structural characteristic of Large Language Models. That is, it is part of the design. If the context window runs out, it will output nonsense. There is no fix.
Plenty of scientific papers descfribing this exact issue.
And here we have an example of a simple code block that AI missed:
Forum on trading, automated trading systems and testing trading strategies
MetaTrader 5 Beta Build 5955: MCP and Agentic AI support
Ryan L Johnson, 2026.06.28 00:30
Build 5956. MQL5.communitiy.
Doesn't seem to differentiate between points and big points on 2 digit XAUUSD pricing.
Relevant code:
switch(baseDigits) { case 1: pointMultiplier = 10; break; case 2: pointMultiplier = 100; break; case 3: pointMultiplier = 1000; break; case 4: pointMultiplier = 10000; break; case 5: pointMultiplier = 100000; break; }7. 📐 Stop Loss Too Tight Relative to Renko Size
Correct info:
50 x 100 = 5000 points. On 2 digit XAUUSD pricing, that's 50 big points (ahead of the decimal).
Where did you get that? While the accuracy AI models has dramatically improved, hallucination remains an inherent structural characteristic of Large Language Models. That is, it is part of the design. If the context window runs out, it will output nonsense. There is no fix.
Plenty of scientific papers describing this exact issue:
https://arxiv.org/pdf/2510.05116
https://arxiv.org/pdf/2311.05232
I confirm from my experience. I got hallucinations from time to time with whatever model (I just didn't try Claude Fable).
Anyway, it's not even the core of the problem. For sure, there are less hallucinations now than some months ago, and possible it will be yet improved. Though it changes nothing on the main argument, a human that understand coding is needed to review the code if you want high quality.
I am not here to argue with anyone who hasn't tried or reads articles.
This is close to poisoning the well: you're preemptively dismissing anyone who doesn't meet your criteria instead of evaluating their arguments on their merits.
You didn't show any facts or results.
I confirm from my experience. I got hallucinations from time to time with whatever model (I just didn't try Claude Fable).
Does not matter the model, you do not even need to try. Science does not lie. All LMM's have this core feature that can only be minimized with aditional layers, but never removed. This is for LLM's. In order to fix this, a entire different approach is needed.
Perhaps Amir is using a language that the AI is capable of testing itself, like python for example. Then it is possible for the model to (in)validate its own nonsense. Also you can create a validation of end-points model for the AI to test if the application generates the expected output after changes. This way it is possible to trust AI doing its job as per specification.
But to state it as a fact that AI can do everything unsupervised, is just plain lie. I am in the process of developing a new website for my business which currently runs on 20 year old closed source legacy software. There is no way you can prompt AI no matter how much of a prompt wizzard you are to convert this into current age software unsupervised.