GitHub Copilot (Gen-AI) je koristan, ali ne i srebrni metak

Ovdje su moje misli nakon korištenja GitHub Copilot Pro u stvarnom svijetu 8 razvoj u posljednja tri mjeseca. ASP.NET Ovdje su moje misli nakon korištenja GitHub Copilot Pro u stvarnom svijetu 8 razvoj u posljednja tri mjeseca. ASP.NET Here are my thoughts after using GitHub Copilot Pro in real-world 8 development over the past three months. Since technology evolves rapidly, it's worth noting that these impressions are based on my experience as of March 2025.. ASP.NET ASP.NET 1. Free trial prompted me to subscribe to GitHub Copilot Pro I have read about AI code generators and watched some demo videos in the past, but I was not convinced that they are really production-ready. Bez razloga, prije 3 mjeseca, zbog GitHub Copilot Free račun je automatski aktiviran u mom u mom ASP.NET8 projektu, takozvani prijedlozi koda počeli pojavljivati. Sljedeći članak Visual Studio 2022 "duh tekst" Visual Studio 2022 „Ghost Text“ u trenucima to je bio briljantan Za one koji su nepoznati, predstavljeni korisniku u sivo-prozirnom tekstu koji se pojavljuje bez razloga kao predviđanje AI-a o tome što će korisnik učiniti sljedeće. "fantomski tekst" je GitHub Copilot (GHC) prijedlozi “ghost text” is GitHub Copilot (GHC) suggestions U nekoliko dana, odlučio sam se pretplatiti na punu pretplatu kako bih provjerio taj alat. GitHub Copilot PRO Sljedeći članak Kada sam prije mnogo godina studirao filozofiju u srednjoj školi, učili su me o konceptu "Svijeta razgovora" u svakom implicitnom razgovoru, a teme u razgovoru obično se odnose na trenutni "Svijet razgovora". GitHub Copilot (GHC) je alat Gen-AI koji je prilično koristan u zadaćama ograničenog opsega. GHC vrlo brzo uči vaš stil programiranja (kako se nosim i registrujem iznimke) bolje od bilo kojeg drugog suradnika, i slijedi vaš stil u pružanju predviđenog "spokojenog teksta" koda.GHC vrlo brzo uči vaš stil programiranja (kako se nosim i registrujem iznimke) bolje od bilo kojeg drugog suradnika, i slijedi vaš stil u pružanju predviđenog "duhovnog teksta" koda. GHC je opasno jer misli da je pametniji od vas i da će promijeniti malo stvari na svoj način, bez da to primijetite. U preuzimanju baze podataka, u slučaju iznimke, vraćao sam null. GHC me je pozvao s "duhovnim tekstom" tijekom kodiranja da završim svoju metodu, a ja sam samo imao kratak pogled i prihvatio ga. stvorio je bug; umjesto null tijekom iznimke, vraćao je prazan objekt, a moj kod drugdje je bio slomljen. Moral je: . Čak i nakon 5 pokušaja, GHC je proizvodio neki JS kod koji radi nešto drugo GitHub Copilot PRO GitHub Copilot PRO 2. Preparing for GitHub Copilot - AI usage 2.1 Obrazovanje Uvijek ozbiljno shvaćam alate, tako da unaprijed čitam priručnike, kako bih mogao koristiti alate do svog punog potencijala i biti svjestan ograničenja. Slušao sam oko 10 sati videa na GitHub Copilot, teme kao što su “prompt inženjering”, “što je kontekst”, itd. Napravio sam svoj “cheat sheet” prompt zapovijedi i kratke ključeve. nakon 10 sati obuke, bio sam spreman da ga isprobam u svom stvarnom životu profesionalnog kodiranja u 8/C#/Bootstrap/EF8/JS okruženju. ASP.NET ASP.NET 2.2 Prompt Engineering in general In my opinion, . One of the first definitions I heard of AI systems 20 years ago was that AI would be achieved when we would be able to talk to computer systems in natural language. “Prompt Engineering” is a defeat for AI “Prompt Engineering” is a defeat for AI Now they tell you that GHC is an AI system, but you can not really talk to it in “natural language”; you need to use “prompt engineering”, which is really and use symbols like /, #, and @. That looks to me like some mixture of natural language and programming language. They want to sell you their AI systems they have NOW, and 5 years from now, they will probably be telling you “now we have REAL AI, no prompt engineering needed anymore”. a sublanguage of natural language, podjezični jezik prirodnog jezika, So, the expression “prompt engineering” is coming from the period when the only way of interacting with an AI system was via a command prompt. Then, some “art” or “science” ( I would call it “pseudo-science”) in creating commands would help you make those AI systems work better. I have read several such articles, which are all “common sense”, but since the target AI system is always a “black box”, there are to show if one author's recommendations are better than another person's list of rules. Also, systems evolved and changed over time, so strictly speaking, those authors would need to test their recommendations against the new generation of systems again. Typically, they do not do that, but offer “common sense” rationale which is based on the perception of AI as another human intellect. And what is “common sense” for humans might not be the same for AI systems. So, I am a bit skeptical and do not fully believe all the recommendations regarding “prompt engineering” that are out there because there are no real metrics and tests against different generations of AI systems. They offer just “common sense” and anecdotal proof from a few command executions. no real metrics no real metrikama 2.3. Prompt Engineering u GitHub Copilot So, when talking about “prompt engineering” in the context of GitHub Copilot (GHC) system, that includes not only the command line interface, but also some interaction via the Visual Studio GUI. That is basically “the user interface of GitHub Copilot”. Ako planirate učinkovito koristiti GitHub Copilot, trebate se upoznati s GitHub Copilot UI-om. 2.4 Universe of the conversation 2.5 Što je „kontekst“ u svijetu AI Tech tvrtke koje rade na AI-u izumile su pojam "kontekst", koji ima slično značenje kao i gore spomenuti filozofski pojam.Htjela bih zadržati pojmove odvojeno, jer tehnološke tvrtke vole primorati svoje definicije onoga kako svijet treba izgledati, u nastojanju da prodaju svoje proizvode i dionice.Također, vjerojatno će biti definicija AI-Context-2025 i nova definicija AI-Context-2026, i tako dalje, kako se tehnologija razvija. So, the current context definition as of March 2025 (you can call it AI-Context-2025 ) would be: additional information the user needs to supply to the AI system for it to understand what it is required to do. 2.6 What is “context” in the GitHub Copilot In training videos for GitHub Copilot, there has been a lot of emphasis on providing the proper “context” for your requests. To me, it looks like they ask for explicitly enumerating the files that contain relevant code. I would assume that “implied context” would be your Visual Studio project/solution, but it is not, in this moment of time, at least. U stvari, postoji malo GUI kontrolno polje u GitHub Copilot VS2022 možete kliknuti kako biste potvrdili da želite da trenutni otvoreni dokument bude uključen u "kontekst" svakog vašeg zahtjeva. (uzgred, nazivaju ga "prompt engineering" i kliknite na GUI kontrolna polja... možda bi "GUI engineering" bilo bolje ime 😉 ). So, if you want to use GitHub Copilot efficiently, there is a certain procedure on how to use it and a recommended command prompt/GUI interface for it. So, be it. I have learned/read all the instructions, and wanted to see the AI thing generate nice code for my VS2022 project. The way I get it, they want you to be very specific in your request and enumerate all the relevant files with code. I see it as similar to giving instructions to another programmer, with certain specificity. It is not difficult, compared to the numerous programming languages developers learn. 3. dojmovi nakon 1. tjedna GHC je samo alat za pomoć kodovima. To nije tako "pametno" i "pametno", ali je dobro s ponavljajućim zadatcima i može uštedjeti neko vrijeme tipkanjem. It is a waste of time to “chat to it”, faster is to go to Google and read for yourself to solve some original problem. But once I create a nice pattern to solve something, it can save some typing, because it learns that pattern and reapplies it again automatically. It also tends to generate a lot of “trash code,” so a human must filter what is generated, but it is not difficult to use the “delete” button, and keep just “good snippets”. Let’s say, based on what I have seen so far, I expect it can save me 5% of the time in typing. 4. Impressions after 1.5 months It is useful sometimes, but only for local scope problems, cannot see the bigger picture. GitHub-Copilot (Gen-AI) is helpful, but not great. GitHub-Copilot (Gen-AI) je koristan, ali ne i veliki. , but sometimes it makes too many mistakes, and when asked, gives answers in several pages of text, especially because the . Sometimes it is brilliant wasting your time, verbose answers it gives are often off-topic Ponekad je sjajno wasting your time, verbose answers it gives are often off-topic U ozbiljnim problemima je beskorisno; bolje je pročitati članak StackOverflow sam i shvatiti ga. My “personal feeling” is , it is ”, and since it is a machine with huge memory of millions of lines of code memorized, . “it does not know it well” “trying to guess it out guesses are sometimes brilliant, sometimes off-topic „to ne zna dobro“ “trying to guess it out guises su ponekad briljantni, ponekad off-topic 5. dojmovi nakon 3 mjeseca So, yes, sometimes GHC is brilliant in predicting what you are about to code, and that you just need to accept into your code. GHC is sometimes brilliant. prompts you with “ghost text” GHC very fast learns your style of programming (how I handle and log exceptions) better than any coworker around, and follows your style in offering you predicted “ghost text” code. I like that so much, every person around tends to have his/her own style in making logs, etc. GHC learns fast. GHC is dangerous because it thinks it is smarter than you and will change a bit things its own way, without you noticing it. In database retrieval, in case of an exception, I was returning null. GHC prompted me with “ghost text” during coding to finish my method, and I just had a short look and accepted it. It created a bug; instead of null during an exception, it was returning an empty object, and my code elsewhere was breaking. Moral is: . GHC dangerously adds its own style. read carefully the code suggestions that it is offering to you before accepting them Hard to explain, but simply, GHC does not understand C# syntax well. It creates brilliant snippets of C#, but with small syntax errors. You need to . That is definitely not what we are accustomed to getting from machine/automated code assistants. GHC can “generate code” but does not “write code”. do a separate syntax check on the code you get from GHC First of all, it fails on nullability like string vs string? type. You need to . Not difficult, but you would expect to delegate exactly that kind of work to the machine/automated code assistants. I could only guess that maybe it mixes C#-.NET-Framework snippets with C#-.NET-Core code snippets. GHC does not understand C# syntax well. check it with the compiler and polish it by yourself It inserts code snippets into your VS2022 code project, but it messes with the open/closed curly brackets of code blocks. You need to count them by yourself and add/remove what is needed. Wastes your time and sometimes creates quite a mess. Such a . GHC does not understand C# syntax well. disappointment that the generated code is not immediately compilation-ready It offers me code that looks nice, and I accept it. But C# class object does not have that method at all. It was close, its parent object has that method, so I figured that by myself. OK, it was close, and put me on the right track. But that is what machines are supposed to be for. No human can remember all the methods from maybe 10.000 C# classes that are .NET8 API that I use in my coding. I would expect the machine to help me with that. But, no, . It gives you a small homework to check that manually by yourself. It looks like , but it doesn't. GHC does not understand C# methods/properties well. even GHC is not sure what method is on which class The code it gives to you will not compile, but is “close”. GHC hallucinates that the C# class has some method I am using EF8 in my VS2022 project, and have a class Customer. I started to write some access method to the Customer DB-table, and GHC offers me GetCustomer method prediction in the “ghost text”. Concept is solid, and yes, that DB-table has a Primary Key, but it is NOT called CustomerId. It looks like , but it doesn’t. Predicted , and I need to go check into the class Customer what is exactly the exact name of the Primary Key is. It is such a disappointment that a machine/automated code assistant, . It is just another class in the VS2022 project. So, the code snippet is conceptually close, but I need to . And I had such a situation 100+ times. GHC does not understand C# methods/properties well. GHC simply hallucinates that such a class property with that name should exist code will not compile GHC, can NOT check by itself proper property/method names fix manually by myself syntax errors Impression is that GHC is shallow, it does not see much more than a few pages of code, and it is trying to “guess out” what the rest of the VS2022 project looks like. GHC does not understand C# methods/properties well. It wastes your time with its guesses. I do not have time to read more than 2 proposals of generated code. I need to do some work. After 1-2 wrong guesses, I ignore GHC and write code by myself. I tried to “chat with GHC” and sometimes after 5-7 chat interactions, it will still be as stupid as at the beginning of the conversation. Feels like talking to a non-AI chatbot on some site selling tickets. It repeats its text all the time, and . I tried that approach a few times, to “refine” my request in interactive chat, but it feels like talking to an idiot. I do not do that anymore, I respect and save my time and energy. I ignore GHC and do some straight coding and get results old-fashioned way. So, the conclusion is: . GHC chats tend to waste your time. GHC has NO FOCUS on what you really need I give GHC 1-2 chances to say what it has to say, then ignore it so as not to waste my time, and do coding manually GHC is not always a good source of information. It can be brilliant and give you better info and code samples than you hoped for. But there is . It can fail. If you try to get advice on some very common topic, just a bit specific to your needs, it will . Better go to Google and find an article on Stack Overflow by yourself. GHC is a text generator, and that machine can generate a huge amount of text and code samples as a response to your question. . For a question you ask, it gives you text and code samples that require maybe 15 minutes to read and understand. It is almost as if it is assigning you a task: read this, and then we will talk again. And that answer after answer. That GHC thing can do that forever, to coach you how HTTP works, for example, and you ask about just one header. The problem is, it , . All that is . It is faster to go to Google and filter/open a couple of links to find the answer to your problem. So, the conclusion is: . GHC chats tend to waste your time. UNCERTAINITY waste your time in endless chat sessions I was overwhelmed with the quantity of code samples it gives you does not have focus that is not what I asked off-topic garbage I ask GHC and read 1-2 answers, then go to Google, not to waste my time It looks like GHC does not follow closely user's instructions closely, it thinks it is smarter than the user, and it . So, I want HTML Razor form created in a certain way/style uniform across the application, but not with GHC, when I tell that, it still gives me a HTML form like from the textbook. That happened in other situations as well. GHC does not follow user instructions closely. follows instructions loosely, and gives what it thinks is better for the user ASP.NET So, software companies have a brilliant idea: they will charge for the product, but at the same time kindly ask users to participate for free in product testing. So, like many other products, some dialogs pop up asking you for your feedback. The question is whether anyone will read what you write, even if you fill the form. So many products are popping up dialogs telling you about new features and asking for 5-star evaluations of every result they produce, or even written comments. Such dialogs just clutter my workspace, and I typically have more serious things to do than to answer several questions evaluating each and every service I got. Basically, they want users in mass to help train their AI system by evaluating each answer. GHC feedback form is a waste of time. I needed a method to Trim() all the strings in the EF class; I wanted a generic solution. I knew Reflection was the way to go, but in one line command request, GHC generated me the perfect 30-line method. Those are situations where GHC can not be beaten by humans. It would take me time to find the proper Reflection method, read some APIs, etc. This way, I got it in 20 seconds. But, , sometimes useless even in such scenarios. GHC excels in small, limited scope tasks. GHC is sometimes brilliant So, I had a simple problem, limited scope, I needed a 10-liner method to wait until jQuery has been loaded, because I was using the “defer” attribute on jQuery loading. Looked like a perfect task for GHC. But I was shocked. . It could not understand what I wanted. It was just keeping generating some snippets that are seemingly related to the problem, but in reality were useless. Like you gave a task to a Junior programmer who “claims” to know some JS, and it keeps bringing you random code samples he finds on the internet, but no one is doing the simple thing you asked for. Later, I found a snippet on StackOverflow by myself. Maybe 3 clicks on Google and I found it. . GHC fails in a bit elaborate JavaScript task. Even after 5 attempts, GHC was producing some JS code that is doing something else GHC is so inferior sometimes I was using GHC to add some comments to my methods, and the ; they do not see the big picture of what that method is trying to achieve. That is more important than line-by-line comments, now we assign this or that. It can infer from English names of methods the purposes of the method, but still not very impressive most of the time. GHC produces shallow C# comments. produced comments are a bit shallow GHC, when asked to create comments for a method, can erase some lines of real code. Be careful and read carefully what is given to you as an answer. It is a big pain, you need to manually re-read the whole method, and you can not trust GHC. It looks like , it sees just text. It happened to me, and I looked very carefully at why. I had some property set, and in the comment above, it was an old setting that I tested and commented out. . There is a probability that I have such a thing at many places in the code. So, for a simple request to just add some comments. All generated commented code needs to be reviewed manually. GHC doesn’t understand C# comments well. GHC does not understand syntax well, and what is a comment or not The GHC thing erased both of them, the code and the old code commented high risk of unwanted code manipulation It looks like there is a going on, with Tech companies pointing the finger at developers for failures of the AI-Gen products. Simply, . But there is a huge “sales effort” to sell the current generation of AI tools, however imperfect they are, and to generate some revenue NOW. So, for failures and limitations, finger-pointing is going in the direction of users: “tools are great, you just do not know how to use them”. One such blame game is the “context” setup for GHC. GHC “context” story doesn’t work as advertised. “blame” game current Gen-AI tools are not as useful as advertised in real-life situations Tech companies say: It is the user's failure when AI tools fail, not true that current AI tools have problems. I am tired of reading/listening to: “you didn’t specify CONTEXT well enough”. That is just an excuse, because I did follow all the advice, and the GHC thing is still stupid and wasting my time generating useless code snippets that do not even compile, not to mention that they are off topic. I was reading about “prompt engineering” and “context”, but after my best effort, I believe that Let’s wait for 6 months and try newer tools. It is a waste of time to try to squeeze more from GHC at this time. So, the truth is that at the present time (March 2025), GHC is quite stupid and no amount of “context info given” can help it to solve the intermediate level task. GHC “context” story doesn’t work as advertised. GHC thing is not smart enough, and can’t do the job. I had code throwing an exception and wanted to use GHC to solve it. I positioned on the Exception in Visual Studio and invoked GHC with /FIX. It was generating text, including a recommendation to “log more to better understand code defects”, but no specific answer. It was easy, I just copied the Exception text into Google, and 3 rd link was the explanation. Not to mention that there was more text on Google on why EF9 is failing, etc. . GHC can’t solve a simple Exception. GHC can be so inferior in regular tasks . I was doing some LINQ and copied some objects from classes that inherit from each other. I wanted to copy all the properties that match. There were like 25 properties. I started assigning 2-3 properties and hoped for GHC to pick the pattern and insert code for the rest. No, it wouldn’t. It added a few properties out of order (which is a problem, you cannot follow which are added and which are not), then invented and it was stacked. I was trying to issue commands from the command prompt, but it misunderstood me, inserted the whole C# class in place, etc. I was trying that several times in different places, because I have many big DTOs, so such copying is often the task. I got bored and . I was tired of giving instructions to a seemingly stupid tool. Strange that . At times, IntelliSense is much smarter and useful than GHC. GHC can’t enumerate properties from the plain C# class names of a few properties that do not exist (hallucinate?), I could not make it do it for me. copied properties manually by myself you cannot delegate such a simple task to GHC . So, I was eager to use generative AI in my VS2022/C# project and was looking for the right opportunity. I noticed one, I was creating HTML/Razor/ASP.NET8/Bootstrap views that present some HTML tables corresponding to DB tables. It was a bit elaborate; there was breadcrumb UI inside the view, some AJAX, etc. So, I manually made one MVC Action/View for the Customers table and wanted to do completely the same for Contracts. All the same, just another DB table (corresponding to EF8 class), another EF class name with new properties. The HTML table I wanted to generate was supposed to look the same, Bootstrap style. So, typical “ I have seen several videos emphasizing the importance of the “context”. So . I even spell-checked my small paragraph task for GHC. I spend some . Then I supplied it to the GHC. The first form it generated did NOT follow the template I explicitly mentioned. GHC thing gave me something like a textbook-looking HTML form, not one customized in Bootstrap for my application. I asked for a new generation of code files. The second one looked like the form I already created, but out of 15 DB table properties, it created only 4. Later, I figured . It guessed right because I was disciplinarily following the same naming conventions in different EF classes, so guesses by GHC worked. I finished the file manually, by manually adding 10+ missing properties. One would . GHC can’t do a little project of 4 C# files follow the PATTERN in 4 files problem”. in a separate text editor, I started to write precise instructions for GitHub Copilot. I explained the task, said which file I want as a template, and marked the file name with #. I explained where the MVC model will come from, which file, etc time preparing task instructions in a clear manner GHC can NOT read dependent files and is most of the time guessing property names expect that such a repetitive task could be delegated to the GHC, but no, GHC can NOT do the simple repetitive job . I tried again similar situation of generating some files in my VS2022/C# project based on a clear pattern-prototype file. Results are not very good, although that was the situation that was as advertised, arranged for GHC to succeed. Not complicated logic, clear name replacement needed, just a bit complicated HTML table corresponding to another DB table. But, it looks like GHC can not read EF8 properties from a 3 rd file and enumerate them to create a similar-looking HTML table. Also, it . Sometimes even changes the file code without reason, so a line-by-line manual check of the generated file is needed. Also, command line chatting to GHC and asking for changes consumes time, and since the , in the . GHC can’t do a little project of 4 C# files introduces a lot of errors into the generated file, by hallucinating about C# properties result is not guaranteed and contains errors end is not an efficient way of software development . Looking back, . Any person would understand the given written task and the context specified. I concluded that , even when given a very template/pattern-oriented task. I do not want to hear anymore “proper context” stories, because I put an effort to do everything right. And it is a waste of time and effort to go and ask GHC to generate files again and again and re-read every new generated code in my VS2022/C# project. I tried that, and GHC can generate another file every time you ask, and I was tired of looking into failed attempts. GHC can’t do a little project of 4 C# files I could not be more precise regarding “context” specification GHC is just not smart enough to handle a small generation of 4 files . Now, when I have a similar situation to generate a new HTML view based on the template I have, . I just use a , in my VS2022/C# project and it is the same amount of time and with . Because when I use Find/Replace, I know exactly what I am getting, and GHC thing would change strings, but on its own initiative, do some changes to code sometimes, so I was never sure if it changed or erased a line or two of original template code. It is boring work, but no unexpected surprises from “smart” GHC “improving” my code. GHC can’t do a little project of 4 C# files a pattern-based task, I do not ask GHC anymore text editor with Find/Replace more certainty . I couldn’t remember Bootstrap class names, and I wanted to enlarge the button left and right. I asked GHC and it gave me Bootstrap class names. I applied them, and it enlarged the button top and bottom. Strange, how it can make such a mistake. I needed help exactly with all the Bootstrap CSS classes, I can not memorize them all. That is such a simple question, and GHC can not do it. GHC makes silly Bootstrap mistakes So, yes, sometimes GHC is brilliant in predicting what you are about to code, and that you just need to accept into your code. GHC is sometimes brilliant. prompts you with “ghost text” So, yes, sometimes GHC is brilliant in predicting what you are about to code, and that you just need to accept into your code. GHC is sometimes brilliant. prompts you with “ghost text” GHC je ponekad briljantan. prompts you with “ghost text” GHC learns fast. GHC is dangerous because it thinks it is smarter than you and will change a bit things its own way, without you noticing it. In database retrieval, in case of an exception, I was returning null. GHC prompted me with “ghost text” during coding to finish my method, and I just had a short look and accepted it. It created a bug; instead of null during an exception, it was returning an empty object, and my code elsewhere was breaking. Moral is: . GHC dangerously adds its own style. read carefully the code suggestions that it is offering to you before accepting them čitajte pažljivo prijedloge koda koje vam nudi prije nego što ih prihvatite GHC opasno dodaje svoj vlastiti stil. read carefully the code suggestions that it is offering to you before accepting them Teško je objasniti, ali jednostavno, GHC ne razumije C# sintazu dobro. Stvara briljantne snippete C#, ali s malim sintazijskim pogreškama. Morate . To definitivno nije ono što smo navikli dobiti od strojnih/automatiziranih pomoćnika kodiranja. GHC može “generirati kod”, ali ne “pisati kod”. provjeriti zasebnu sintazu kod koji dobivate od GHC Hard to explain, but simply, GHC does not understand C# syntax well. It creates brilliant snippets of C#, but with small syntax errors. You need to . That is definitely not what we are accustomed to getting from machine/automated code assistants. GHC can “generate code” but does not “write code”. do a separate syntax check on the code you get from GHC GHC can “generate code” but does not “write code”. do a separate syntax check on the code you get from GHC First of all, it fails on nullability like string vs string? type. You need to . Not difficult, but you would expect to delegate exactly that kind of work to the machine/automated code assistants. I could only guess that maybe it mixes C#-.NET-Framework snippets with C#-.NET-Core code snippets. GHC does not understand C# syntax well. check it with the compiler and polish it by yourself First of all, it fails on nullability like string vs string? type. You need to . Not difficult, but you would expect to delegate exactly that kind of work to the machine/automated code assistants. I could only guess that maybe it mixes C#-.NET-Framework snippets with C#-.NET-Core code snippets. GHC does not understand C# syntax well. check it with the compiler and polish it by yourself GHC does not understand C# syntax well. check it with the compiler and polish it by yourself It inserts code snippets into your VS2022 code project, but it messes with the open/closed curly brackets of code blocks. You need to count them by yourself and add/remove what is needed. Wastes your time and sometimes creates quite a mess. Such a . GHC does not understand C# syntax well. disappointment that the generated code is not immediately compilation-ready It inserts code snippets into your VS2022 code project, but it messes with the open/closed curly brackets of code blocks. You need to count them by yourself and add/remove what is needed. Wastes your time and sometimes creates quite a mess. Such a . GHC does not understand C# syntax well. disappointment that the generated code is not immediately compilation-ready GHC ne razumije C# sintaksu dobro. disappointment that the generated code is not immediately compilation-ready It offers me code that looks nice, and I accept it. But C# class object does not have that method at all. It was close, its parent object has that method, so I figured that by myself. OK, it was close, and put me on the right track. But that is what machines are supposed to be for. No human can remember all the methods from maybe 10.000 C# classes that are .NET8 API that I use in my coding. I would expect the machine to help me with that. But, no, . It gives you a small homework to check that manually by yourself. It looks like , but it doesn't. GHC does not understand C# methods/properties well. even GHC is not sure what method is on which class The code it gives to you will not compile, but is “close”. GHC hallucinates that the C# class has some method It offers me code that looks nice, and I accept it. But C# class object does not have that method at all. It was close, its parent object has that method, so I figured that by myself. OK, it was close, and put me on the right track. But that is what machines are supposed to be for. No human can remember all the methods from maybe 10.000 C# classes that are .NET8 API that I use in my coding. I would expect the machine to help me with that. But, no, . It gives you a small homework to check that manually by yourself. It looks like , but it doesn't. GHC does not understand C# methods/properties well. even GHC is not sure what method is on which class The code it gives to you will not compile, but is “close”. GHC hallucinates that the C# class has some method GHC ne razumije C# metode / svojstva dobro. even GHC is not sure what method is on which class The code it gives to you will not compile, but is “close”. GHC hallucinates that the C# class has some method I am using EF8 in my VS2022 project, and have a class Customer. I started to write some access method to the Customer DB-table, and GHC offers me GetCustomer method prediction in the “ghost text”. Concept is solid, and yes, that DB-table has a Primary Key, but it is NOT called CustomerId. It looks like , but it doesn’t. Predicted , and I need to go check into the class Customer what is exactly the exact name of the Primary Key is. It is such a disappointment that a machine/automated code assistant, . It is just another class in the VS2022 project. So, the code snippet is conceptually close, but I need to . And I had such a situation 100+ times. GHC does not understand C# methods/properties well. GHC simply hallucinates that such a class property with that name should exist code will not compile GHC, can NOT check by itself proper property/method names fix manually by myself syntax errors Ja koristim EF8 u mom VS2022 projektu, i imam klasa klijent. Počeo sam pisati neki pristup metoda na klijent DB-tabeli, a GHC mi nudi GetCustomer metoda predviđanja u "ghost tekst". Koncept je čvrst, i da, da DB-tabela ima primarni ključ, ali to se ne zove CustomerId. Izgleda kao , ali to ne. Predviđeno , i moram ići provjeriti u razred klijent što je točno točno ime primarni ključ. To je tako razočaran GHC ne razumije C# metode / svojstva dobro. GHC jednostavno halucinira da takva klasa svojstvo s tim imenom treba postojati kod neće kompilirati GHC ne razumije C# metode / svojstva dobro. GHC jednostavno halucinira da takva klasna svojstva s tim imenom trebaju postojati kod neće kompilirati GHC, ne može samostalno provjeriti pravo ime svojstva/metoda fix manually by myself syntax errors Dojam je da je GHC plitko, ne vidi mnogo više od nekoliko stranica koda, i pokušava "ugledati" kako izgleda ostatak VS2022 projekta. GHC ne razumije C# metode / svojstva dobro. Impression is that GHC is shallow, it does not see much more than a few pages of code, and it is trying to “guess out” what the rest of the VS2022 project looks like. GHC does not understand C# methods/properties well. GHC does not understand C# methods/properties well. To gubi vaše vrijeme sa svojim nagađanjima. Nemam vremena da čitam više od 2 prijedloga generiranog koda. Moram učiniti neki posao. Nakon 1-2 pogrešnih nagađanja, ignorišem GHC i napišem kod sam. Pokušao sam "čitati s GHC" i ponekad nakon 5-7 interakcija, još uvijek će biti glupo kao na početku razgovora. Osjeća se kao da razgovaram s chatbotom koji nije AI na nekim stranicama koje prodaju karte. To ponavlja svoj tekst cijelo vrijeme, a . Pokušao sam to pristupiti nekoliko puta, da "prefinim" moju molbu u interaktivnom razgovoru, ali to se GHC razgovori imaju tendenciju da troše vaše vrijeme. C nema fokus na ono što stvarno trebate Troši vaše vrijeme svojim nagađanjima. Nemam vremena pročitati više od 2 prijedloga generiranog koda. Moram učiniti neki posao. Nakon 1-2 pogrešnih nagađanja, ignorišem GHC i napišem kod sam. Pokušao sam "čitati s GHC-om" i ponekad nakon 5-7 interakcija, još uvijek će biti glupo kao na početku razgovora. Osjeća se kao da razgovaram s chatbotom bez AI-a na nekim stranicama koje prodaju ulaznice. To ponavlja svoj tekst cijelo vrijeme, a . Pokušao sam to približiti nekoliko puta, da "opravim" moju molbu u interaktivnom razgovoru, ali to GHC razgovori imaju tendenciju da troše vaše vrijeme. GHC nema fokus na ono što stvarno trebate GHC chats tend to waste your time. GHC nema fokus na ono što zaista trebate I give GHC 1-2 chances to say what it has to say, then ignore it so as not to waste my time, and do coding manually GHC is not always a good source of information. It can be brilliant and give you better info and code samples than you hoped for. But there is . It can fail. If you try to get advice on some very common topic, just a bit specific to your needs, it will . Better go to Google and find an article on Stack Overflow by yourself. GHC is a text generator, and that machine can generate a huge amount of text and code samples as a response to your question. . For a question you ask, it gives you text and code samples that require maybe 15 minutes to read and understand. It is almost as if it is assigning you a task: read this, and then we will talk again. And that answer after answer. That GHC thing can do that forever, to coach you how HTTP works, for example, and you ask about just one header. The problem is, it , . All that is . It is faster to go to Google and filter/open a couple of links to find the answer to your problem. So, the conclusion is: . GHC chats tend to waste your time. UNCERTAINITY waste your time in endless chat sessions I was overwhelmed with the quantity of code samples it gives you does not have focus that is not what I asked off-topic garbage I ask GHC and read 1-2 answers, then go to Google, not to waste my time GHC is not always a good source of information. It can be brilliant and give you better info and code samples than you hoped for. But there is . It can fail. If you try to get advice on some very common topic, just a bit specific to your needs, it will . Better go to Google and find an article on Stack Overflow by yourself. GHC is a text generator, and that machine can generate a huge amount of text and code samples as a response to your question. . For a question you ask, it gives you text and code samples that require maybe 15 minutes to read and understand. It is almost as if it is assigning you a task: read this, and then we will talk again. And that answer after answer. That GHC thing can do that forever, to coach you how HTTP works, for example, and you ask about just one header. The problem is, it , . All that is . It is faster to go to Google and filter/open a couple of links to find the answer to your problem. So, the conclusion is: . GHC chats tend to waste your time. UNCERTAINITY waste your time in endless chat sessions I was overwhelmed with the quantity of code samples it gives you does not have focus that is not what I asked off-topic garbage I ask GHC and read 1-2 answers, then go to Google, not to waste my time GHC razgovori imaju tendenciju gubiti vrijeme. UNCERTAINITY traćite svoje vrijeme na beskrajne chat sesije I was overwhelmed with the quantity of code samples it gives you nema fokus that is not what I asked off-topic garbage I ask GHC and read 1-2 answers, then go to Google, not to waste my time It looks like GHC does not follow closely user's instructions closely, it thinks it is smarter than the user, and it . So, I want HTML Razor form created in a certain way/style uniform across the application, but not with GHC, when I tell that, it still gives me a HTML form like from the textbook. That happened in other situations as well. GHC does not follow user instructions closely. follows instructions loosely, and gives what it thinks is better for the user ASP.NET Čini se da GHC ne slijedi upute korisnika, misli da je pametniji od korisnika, a . GHC ne slijedi upute korisnika. sledi upute slabo, i daje ono što misli da je bolje za korisnika GHC does not follow user instructions closely. follows instructions loosely, and gives what it thinks is better for the user ASP.NET So, software companies have a brilliant idea: they will charge for the product, but at the same time kindly ask users to participate for free in product testing. So, like many other products, some dialogs pop up asking you for your feedback. The question is whether anyone will read what you write, even if you fill the form. So many products are popping up dialogs telling you about new features and asking for 5-star evaluations of every result they produce, or even written comments. Such dialogs just clutter my workspace, and I typically have more serious things to do than to answer several questions evaluating each and every service I got. Basically, they want users in mass to help train their AI system by evaluating each answer. GHC feedback form is a waste of time. So, software companies have a brilliant idea: they will charge for the product, but at the same time kindly ask users to participate for free in product testing. So, like many other products, some dialogs pop up asking you for your feedback. The question is whether anyone will read what you write, even if you fill the form. So many products are popping up dialogs telling you about new features and asking for 5-star evaluations of every result they produce, or even written comments. Such dialogs just clutter my workspace, and I typically have more serious things to do than to answer several questions evaluating each and every service I got. Basically, they want users in mass to help train their AI system by evaluating each answer. GHC feedback form is a waste of time. GHC povratni obrazac je gubljenje vremena. I needed a method to Trim() all the strings in the EF class; I wanted a generic solution. I knew Reflection was the way to go, but in one line command request, GHC generated me the perfect 30-line method. Those are situations where GHC can not be beaten by humans. It would take me time to find the proper Reflection method, read some APIs, etc. This way, I got it in 20 seconds. But, , sometimes useless even in such scenarios. GHC excels in small, limited scope tasks. GHC is sometimes brilliant I needed a method to Trim() all the strings in the EF class; I wanted a generic solution. I knew Reflection was the way to go, but in one line command request, GHC generated me the perfect 30-line method. Those are situations where GHC can not be beaten by humans. It would take me time to find the proper Reflection method, read some APIs, etc. This way, I got it in 20 seconds. But, , sometimes useless even in such scenarios. GHC excels in small, limited scope tasks. GHC is sometimes brilliant GHC excels in small, limited scope tasks. GHC is sometimes brilliant So, I had a simple problem, limited scope, I needed a 10-liner method to wait until jQuery has been loaded, because I was using the “defer” attribute on jQuery loading. Looked like a perfect task for GHC. But I was shocked. . It could not understand what I wanted. It was just keeping generating some snippets that are seemingly related to the problem, but in reality were useless. Like you gave a task to a Junior programmer who “claims” to know some JS, and it keeps bringing you random code samples he finds on the internet, but no one is doing the simple thing you asked for. Later, I found a snippet on StackOverflow by myself. Maybe 3 clicks on Google and I found it. . GHC fails in a bit elaborate JavaScript task. Even after 5 attempts, GHC was producing some JS code that is doing something else GHC is so inferior sometimes So, I had a simple problem, limited scope, I needed a 10-liner method to wait until jQuery has been loaded, because I was using the “defer” attribute on jQuery loading. Looked like a perfect task for GHC. But I was shocked. . It could not understand what I wanted. It was just keeping generating some snippets that are seemingly related to the problem, but in reality were useless. Like you gave a task to a Junior programmer who “claims” to know some JS, and it keeps bringing you random code samples he finds on the internet, but no one is doing the simple thing you asked for. Later, I found a snippet on StackOverflow by myself. Maybe 3 clicks on Google and I found it. . GHC fails in a bit elaborate JavaScript task. Even after 5 attempts, GHC was producing some JS code that is doing something else GHC is so inferior sometimes GHC ne uspijeva u malo složenom zadatku JavaScript-a. GHC je tako niži ponekad Koristio sam GHC za dodavanje nekih komentara mojim metodama, a ; oni ne vide veliku sliku onoga što ta metoda pokušava postići. GHC proizvodi plitke C# komentare. proizvedeni komentari su malo plitki Koristio sam GHC da dodam neke komentare mojim metodama, a ; oni ne vide veliku sliku onoga što ta metoda pokušava postići. GHC proizvodi plitke C# komentare. proizvedeni komentari su malo plitki GHC produces shallow C# comments. produced comments are a bit shallow GHC, when asked to create comments for a method, can erase some lines of real code. Be careful and read carefully what is given to you as an answer. It is a big pain, you need to manually re-read the whole method, and you can not trust GHC. It looks like , it sees just text. It happened to me, and I looked very carefully at why. I had some property set, and in the comment above, it was an old setting that I tested and commented out. . There is a probability that I have such a thing at many places in the code. So, for a simple request to just add some comments. All generated commented code needs to be reviewed manually. GHC doesn’t understand C# comments well. GHC does not understand syntax well, and what is a comment or not The GHC thing erased both of them, the code and the old code commented high risk of unwanted code manipulation GHC, kada je zatraženo da stvori komentare za metodu, može izbrisati neke linije pravog koda. Budite oprezni i pažljivo pročitajte što vam je dano kao odgovor. To je velika bol, morate ručno ponovno pročitati cijelu metodu, a ne možete vjerovati GHC. Čini se da , vidi samo tekst. To se dogodilo meni, i pogledao sam vrlo pažljivo na zašto. Imao sam neke svojstva set, a u komentaru gore, to je stara postavka koju sam testirao i komentirao. . Posto GHC ne razumije C# komentare dobro. GHC ne razumije sintaksu dobro, a što je komentar ili ne GHC stvar izbrisana i od njih, kod i stari kod komentiran GHC doesn’t understand C# comments well. GHC does not understand syntax well, and what is a comment or not The GHC thing erased both of them, the code and the old code commented high risk of unwanted code manipulation It looks like there is a going on, with Tech companies pointing the finger at developers for failures of the AI-Gen products. Simply, . But there is a huge “sales effort” to sell the current generation of AI tools, however imperfect they are, and to generate some revenue NOW. So, for failures and limitations, finger-pointing is going in the direction of users: “tools are great, you just do not know how to use them”. One such blame game is the “context” setup for GHC. GHC “context” story doesn’t work as advertised. “blame” game current Gen-AI tools are not as useful as advertised in real-life situations Tech companies say: It is the user's failure when AI tools fail, not true that current AI tools have problems. Čini se da postoji igra GHC “kontekst” priča ne radi kao što je oglašeno. “krivnja” koja se odvija, s tehnološkim tvrtkama koje ukazuju na programere zbog neuspjeha AI-Gen proizvoda. Jednostavno, . Ali postoji ogroman “prodajni napor” za prodaju trenutne generacije AI alata, bez obzira na to koliko su nesavršeni, i za stvaranje nekih prihoda NOW. Dakle, za neuspjehe i ograničenja, ukazivanje prstiju ide u smjeru korisnika: “alati su veliki, jednostavno ne znate kako ih koristiti”. trenutni Gen-AI alati nisu toliko korisni kao što se oglašavaju u stvarnim situacijama Tehnološke tvrtke kažu GHC “kontekst” priča ne radi kao što je oglašeno. “blame” game current Gen-AI tools are not as useful as advertised in real-life situations Tech companies say: It is the user's failure when AI tools fail, not true that current AI tools have problems. I am tired of reading/listening to: “you didn’t specify CONTEXT well enough”. That is just an excuse, because I did follow all the advice, and the GHC thing is still stupid and wasting my time generating useless code snippets that do not even compile, not to mention that they are off topic. I was reading about “prompt engineering” and “context”, but after my best effort, I believe that Let’s wait for 6 months and try newer tools. It is a waste of time to try to squeeze more from GHC at this time. So, the truth is that at the present time (March 2025), GHC is quite stupid and no amount of “context info given” can help it to solve the intermediate level task. GHC “context” story doesn’t work as advertised. GHC thing is not smart enough, and can’t do the job. I am tired of reading/listening to: “you didn’t specify CONTEXT well enough”. That is just an excuse, because I did follow all the advice, and the GHC thing is still stupid and wasting my time generating useless code snippets that do not even compile, not to mention that they are off topic. I was reading about “prompt engineering” and “context”, but after my best effort, I believe that Let’s wait for 6 months and try newer tools. It is a waste of time to try to squeeze more from GHC at this time. So, the truth is that at the present time (March 2025), GHC is quite stupid and no amount of “context info given” can help it to solve the intermediate level task. GHC “context” story doesn’t work as advertised. GHC thing is not smart enough, and can’t do the job. Kako opisati AI sustave poput GitHub CopilotTipična dobra definicija nečega novog sastoji se od dva dijela: 1) objekta/koncepta s kojim je sličan i 2) načina na koji se razlikuje od sličnog objekta/koncepta. Može biti briljantan i brzo riješiti složene zagonetke, ali može ne uspjeti na vrlo jednostavnom zadatku. Nakon iskustva, kako sad koristim GitHub Copilot Koristi se kodom i ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno ispravno GHC “context” story doesn’t work as advertised. GHC thing is not smart enough, and can’t do the job. I had code throwing an exception and wanted to use GHC to solve it. I positioned on the Exception in Visual Studio and invoked GHC with /FIX. It was generating text, including a recommendation to “log more to better understand code defects”, but no specific answer. It was easy, I just copied the Exception text into Google, and 3 rd link was the explanation. Not to mention that there was more text on Google on why EF9 is failing, etc. . GHC can’t solve a simple Exception. GHC can be so inferior in regular tasks I had code throwing an exception and wanted to use GHC to solve it. I positioned on the Exception in Visual Studio and invoked GHC with /FIX. It was generating text, including a recommendation to “log more to better understand code defects”, but no specific answer. It was easy, I just copied the Exception text into Google, and 3 rd link was the explanation. Not to mention that there was more text on Google on why EF9 is failing, etc. . GHC can’t solve a simple Exception. GHC can be so inferior in regular tasks GHC can’t solve a simple Exception. GHC može biti toliko niži u redovitim zadatcima . I was doing some LINQ and copied some objects from classes that inherit from each other. I wanted to copy all the properties that match. There were like 25 properties. I started assigning 2-3 properties and hoped for GHC to pick the pattern and insert code for the rest. No, it wouldn’t. It added a few properties out of order (which is a problem, you cannot follow which are added and which are not), then invented and it was stacked. I was trying to issue commands from the command prompt, but it misunderstood me, inserted the whole C# class in place, etc. I was trying that several times in different places, because I have many big DTOs, so such copying is often the task. I got bored and . I was tired of giving instructions to a seemingly stupid tool. Strange that . At times, IntelliSense is much smarter and useful than GHC. GHC can’t enumerate properties from the plain C# class names of a few properties that do not exist (hallucinate?), I could not make it do it for me. copied properties manually by myself you cannot delegate such a simple task to GHC . I was doing some LINQ and copied some objects from classes that inherit from each other. I wanted to copy all the properties that match. There were like 25 properties. I started assigning 2-3 properties and hoped for GHC to pick the pattern and insert code for the rest. No, it wouldn’t. It added a few properties out of order (which is a problem, you cannot follow which are added and which are not), then invented and it was stacked. I was trying to issue commands from the command prompt, but it misunderstood me, inserted the whole C# class in place, etc. I was trying that several times in different places, because I have many big DTOs, so such copying is often the task. I got bored and . I was tired of giving instructions to a seemingly stupid tool. Strange that . At times, IntelliSense is much smarter and useful than GHC. GHC can’t enumerate properties from the plain C# class names of a few properties that do not exist (hallucinate?), I could not make it do it for me. copied properties manually by myself you cannot delegate such a simple task to GHC GHC can’t enumerate properties from the plain C# class names of a few properties that do not exist (hallucinate?), I could not make it do it for me. copied properties manually by myself you cannot delegate such a simple task to GHC . So, I was eager to use generative AI in my VS2022/C# project and was looking for the right opportunity. I noticed one, I was creating HTML/Razor/ASP.NET8/Bootstrap views that present some HTML tables corresponding to DB tables. It was a bit elaborate; there was breadcrumb UI inside the view, some AJAX, etc. So, I manually made one MVC Action/View for the Customers table and wanted to do completely the same for Contracts. All the same, just another DB table (corresponding to EF8 class), another EF class name with new properties. The HTML table I wanted to generate was supposed to look the same, Bootstrap style. So, typical “ I have seen several videos emphasizing the importance of the “context”. So . I even spell-checked my small paragraph task for GHC. I spend some . Then I supplied it to the GHC. The first form it generated did NOT follow the template I explicitly mentioned. GHC thing gave me something like a textbook-looking HTML form, not one customized in Bootstrap for my application. I asked for a new generation of code files. The second one looked like the form I already created, but out of 15 DB table properties, it created only 4. Later, I figured . It guessed right because I was disciplinarily following the same naming conventions in different EF classes, so guesses by GHC worked. I finished the file manually, by manually adding 10+ missing properties. One would . GHC can’t do a little project of 4 C# files follow the PATTERN in 4 files problem”. in a separate text editor, I started to write precise instructions for GitHub Copilot. I explained the task, said which file I want as a template, and marked the file name with #. I explained where the MVC model will come from, which file, etc time preparing task instructions in a clear manner GHC can NOT read dependent files and is most of the time guessing property names expect that such a repetitive task could be delegated to the GHC, but no, GHC can NOT do the simple repetitive job . So, I was eager to use generative AI in my VS2022/C# project and was looking for the right opportunity. I noticed one, I was creating HTML/Razor/ASP.NET8/Bootstrap views that present some HTML tables corresponding to DB tables. It was a bit elaborate; there was breadcrumb UI inside the view, some AJAX, etc. So, I manually made one MVC Action/View for the Customers table and wanted to do completely the same for Contracts. All the same, just another DB table (corresponding to EF8 class), another EF class name with new properties. The HTML table I wanted to generate was supposed to look the same, Bootstrap style. So, typical “ I have seen several videos emphasizing the importance of the “context”. So . I even spell-checked my small paragraph task for GHC. I spend some . Then I supplied it to the GHC. The first form it generated did NOT follow the template I explicitly mentioned. GHC thing gave me something like a textbook-looking HTML form, not one customized in Bootstrap for my application. I asked for a new generation of code files. The second one looked like the form I already created, but out of 15 DB table properties, it created only 4. Later, I figured . It guessed right because I was disciplinarily following the same naming conventions in different EF classes, so guesses by GHC worked. I finished the file manually, by manually adding 10+ missing properties. One would . GHC can’t do a little project of 4 C# files follow the PATTERN in 4 files problem”. in a separate text editor, I started to write precise instructions for GitHub Copilot. I explained the task, said which file I want as a template, and marked the file name with #. I explained where the MVC model will come from, which file, etc time preparing task instructions in a clear manner GHC can NOT read dependent files and is most of the time guessing property names expect that such a repetitive task could be delegated to the GHC, but no, GHC can NOT do the simple repetitive job GHC can’t do a little project of 4 C# files follow the PATTERN in 4 files problem”. in a separate text editor, I started to write precise instructions for GitHub Copilot. I explained the task, said which file I want as a template, and marked the file name with #. I explained where the MVC model will come from, which file, etc time preparing task instructions in a clear manner GHC can NOT read dependent files and is most of the time guessing property names expect that such a repetitive task could be delegated to the GHC, but no, GHC can NOT do the simple repetitive job . I tried again similar situation of generating some files in my VS2022/C# project based on a clear pattern-prototype file. Results are not very good, although that was the situation that was as advertised, arranged for GHC to succeed. Not complicated logic, clear name replacement needed, just a bit complicated HTML table corresponding to another DB table. But, it looks like GHC can not read EF8 properties from a 3 rd file and enumerate them to create a similar-looking HTML table. Also, it . Sometimes even changes the file code without reason, so a line-by-line manual check of the generated file is needed. Also, command line chatting to GHC and asking for changes consumes time, and since the , in the . GHC can’t do a little project of 4 C# files introduces a lot of errors into the generated file, by hallucinating about C# properties result is not guaranteed and contains errors end is not an efficient way of software development . I tried again similar situation of generating some files in my VS2022/C# project based on a clear pattern-prototype file. Results are not very good, although that was the situation that was as advertised, arranged for GHC to succeed. Not complicated logic, clear name replacement needed, just a bit complicated HTML table corresponding to another DB table. But, it looks like GHC can not read EF8 properties from a 3 rd file and enumerate them to create a similar-looking HTML table. Also, it . Sometimes even changes the file code without reason, so a line-by-line manual check of the generated file is needed. Also, command line chatting to GHC and asking for changes consumes time, and since the , in the . GHC can’t do a little project of 4 C# files introduces a lot of errors into the generated file, by hallucinating about C# properties result is not guaranteed and contains errors end is not an efficient way of software development GHC can’t do a little project of 4 C# files introduces a lot of errors into the generated file, by hallucinating about C# properties resultat nije zajamčen i sadrži pogreške end nije učinkovit način razvoja softvera . Looking back, . Any person would understand the given written task and the context specified. I concluded that , even when given a very template/pattern-oriented task. I do not want to hear anymore “proper context” stories, because I put an effort to do everything right. And it is a waste of time and effort to go and ask GHC to generate files again and again and re-read every new generated code in my VS2022/C# project. I tried that, and GHC can generate another file every time you ask, and I was tired of looking into failed attempts. GHC can’t do a little project of 4 C# files I could not be more precise regarding “context” specification GHC is just not smart enough to handle a small generation of 4 files . Looking back, . Any person would understand the given written task and the context specified. I concluded that , even when given a very template/pattern-oriented task. I do not want to hear anymore “proper context” stories, because I put an effort to do everything right. And it is a waste of time and effort to go and ask GHC to generate files again and again and re-read every new generated code in my VS2022/C# project. I tried that, and GHC can generate another file every time you ask, and I was tired of looking into failed attempts. GHC can’t do a little project of 4 C# files I could not be more precise regarding “context” specification GHC is just not smart enough to handle a small generation of 4 files GHC can’t do a little project of 4 C# files I could not be more precise regarding “context” specification GHC is just not smart enough to handle a small generation of 4 files . Now, when I have a similar situation to generate a new HTML view based on the template I have, . I just use a , in my VS2022/C# project and it is the same amount of time and with . Because when I use Find/Replace, I know exactly what I am getting, and GHC thing would change strings, but on its own initiative, do some changes to code sometimes, so I was never sure if it changed or erased a line or two of original template code. It is boring work, but no unexpected surprises from “smart” GHC “improving” my code. GHC can’t do a little project of 4 C# files a pattern-based task, I do not ask GHC anymore text editor with Find/Replace more certainty . Now, when I have a similar situation to generate a new HTML view based on the template I have, . I just use a , in my VS2022/C# project and it is the same amount of time and with . Because when I use Find/Replace, I know exactly what I am getting, and GHC thing would change strings, but on its own initiative, do some changes to code sometimes, so I was never sure if it changed or erased a line or two of original template code. It is boring work, but no unexpected surprises from “smart” GHC “improving” my code. GHC can’t do a little project of 4 C# files a pattern-based task, I do not ask GHC anymore text editor with Find/Replace more certainty GHC can’t do a little project of 4 C# files a pattern-based task, I do not ask GHC anymore text editor with Find/Replace više sigurnosti . I couldn’t remember Bootstrap class names, and I wanted to enlarge the button left and right. I asked GHC and it gave me Bootstrap class names. I applied them, and it enlarged the button top and bottom. Strange, how it can make such a mistake. I needed help exactly with all the Bootstrap CSS classes, I can not memorize them all. That is such a simple question, and GHC can not do it. GHC makes silly Bootstrap mistakes . I couldn’t remember Bootstrap class names, and I wanted to enlarge the button left and right. I asked GHC and it gave me Bootstrap class names. I applied them, and it enlarged the button top and bottom. Strange, how it can make such a mistake. I needed help exactly with all the Bootstrap CSS classes, I can not memorize them all. That is such a simple question, and GHC can not do it. GHC makes silly Bootstrap mistakes GHC makes silly Bootstrap mistakes good definition 1) the object/concept it is similar to, and 2) how it is different from a similar object/concept So, when talking about Intelligent systems, people usually take humans as a reference value. They tend to say: AI-Gen system is on the level of a Junior programmer, but is better/worse in this or that. But I feel that . Humans gradually progress in their intellectual ability, they have the ability first to solve simple tasks, then more complicated and so on. for AI systems like GitHub Copilot (GHC), humans are not a good reference for AI systems like GitHub Copilot (GHC), humans are not a good reference I do not know much about autism, except for Hollywood movies like “Rain Man” (1988) with Tom Cruise. But if we are to compare GHC to humans, GHC looks like that autistic character from the movie. . It can be brilliant and solve complex puzzles fast, but it can fail on a very simple task I would put AI systems like GHC in their own category. Their speed and huge memory, and ability to generate a huge amount of text/code fast make them incomparable to humans. That is like an idiot that has memory and mathematical ability a million times better than any human, but is still an idiot in front of the simple problem. Can you call it stupid because it does not “write code logically” but instead probably searches millions of lines of code in its memory and finds the solution to the problem faster than you? 7.1 GitHub Copilot makes many C# mistakes Regarding code-assisted generation, GHC is such a huge and . That is definitely not what one expects to see from a machine. My feeling is that it can not reason logically at all, otherwise, it would be able to follow simple syntax rules all the time, and not create a mess with extra brackets or hallucinate about non-existing C# class methods or properties. disappointment that it can not get the C# syntax right all the time check the existence of C# properties/methods by itself disappointment that it can not get the C# syntax right all the time check the existence of C# properties/methods by itself A big shock was when GHC was asked to add comments, it deleted the active line of code because a similar line of code was commented out. That otherwise it would not delete it. It seems it just sees text of some kind, and generates “look-alike” text. More like a child with huge memory and speed toying with code, than a “pair programmer” or a “peer programmer” as it is advertised. GHC thing does not fully understand what is “active line of code,” GHC thing does not fully understand what is “active line of code,” 7.2 When to use GitHub Copilot So, I have coding tasks to do, and , but now it is time to be serious. My time is limited, and my energy needs to be focused productively. toying with GHC was fun toying with GHC was fun , review it, and accept suggestions when I like them. It does not waste my energy, text just pops up in my VS2022/C# project, sometimes it is useful, other times it can be interesting to see what AI thinks I should do now, and it's . I use “ghost test” from GHC a lot a bit of fun to read AI suggestions , review it, and accept suggestions when I like them. It does not waste my energy, text just pops up in my VS2022/C# project, sometimes it is useful, other times it can be interesting to see what AI thinks I should do now, and it's . I use “ghost test” from GHC a lot a bit of fun to read AI suggestions I use “ghost test” from GHC a lot a bit of fun to read AI suggestions I select a part of code in one file in my VS2022/C# project, and ask for some changes or comments. GHC can understand such tasks well. I . Sometimes GHC suggestions are useful, but many times it can not understand what I want or produces incorrect answers. It can be lost even after 3-4 iterations of chat prompts about a simple question like which specific Bootstrap CSS class I need. But, I have it all under control, and it does not waste my time much. focus on a limited number of lines, maybe 50, and ask for suggestions If it doesn’t know the answer well, I go to Google for the resolution. I select a part of code in one file in my VS2022/C# project, and ask for some changes or comments. GHC can understand such tasks well. I . Sometimes GHC suggestions are useful, but many times it can not understand what I want or produces incorrect answers. It can be lost even after 3-4 iterations of chat prompts about a simple question like which specific Bootstrap CSS class I need. But, I have it all under control, and it does not waste my time much. focus on a limited number of lines, maybe 50, and ask for suggestions If it doesn’t know the answer well, I go to Google for the resolution. focus on a limited number of lines, maybe 50, and ask for suggestions If it doesn’t know the answer well, I go to Google for the resolution. zatražiti . GHC je odličan za takve zadatke. Ovdje, to može biti stvarno , ali ponekad daje snippete kod koji su pogrešan odgovor. . Ja koristim GHC tekst-prompt stranicu generaciju snippeta ili male funkcije s jasnom funkcionalnošću briljantan ponekad Nema jamstva GHC će uspjeti ovdje, ali ako to učini, to može biti briljantno to ask for the . GHC is great for such tasks. Here, it can be really , but sometimes it gives code snippets that are the wrong answer. . I use the GHC text-prompt page generation of snippets or small functions with clear functionality brilliant at times No guarantees GHC will succeed here, but if it does, it can be brilliant I use the GHC text-prompt page generation of snippets or small functions with clear functionality brilliant at times No guarantees GHC will succeed here, but if it does, it can be brilliant , since it is a pain and effort to chat with GHC to do that. Even worse, the code generated after EACH GHC intervention needs to be re-read since one is never sure what else GHC will change. Answers are at best incomplete, with many errors, like C# properties and methods that do not exist (it hallucinates?). It does not follow the given pattern closely, so I need to carefully read the generated code in my VS2022/C# project, since it can delete a line of code or add some code of its own, which is wrong in this particular case. Also, , and that is a machine that can generate new iterations of sample code forever. And I am just . And if I ask it to regenerate an answer, that is to give me a new proposed solution, I need to REVIEW it all again from the beginning. I no longer try bigger changes involving 3-4 files at the same time Using GHC for such tasks is a waste of my time. it is time and energy-consuming to REVIEW each answer GHC provides a human, I might need maybe 10 minutes of full focus to review 300 lines of combined C#/Razor/CSS/JS, which GHC generates in 2 seconds , since it is a pain and effort to chat with GHC to do that. Even worse, the code generated after EACH GHC intervention needs to be re-read since one is never sure what else GHC will change. Answers are at best incomplete, with many errors, like C# properties and methods that do not exist (it hallucinates?). It does not follow the given pattern closely, so I need to carefully read the generated code in my VS2022/C# project, since it can delete a line of code or add some code of its own, which is wrong in this particular case. Also, , and that is a machine that can generate new iterations of sample code forever. And I am just . And if I ask it to regenerate an answer, that is to give me a new proposed solution, I need to REVIEW it all again from the beginning. I no longer try bigger changes involving 3-4 files at the same time Using GHC for such tasks is a waste of my time. it is time and energy-consuming to REVIEW each answer GHC provides a human, I might need maybe 10 minutes of full focus to review 300 lines of combined C#/Razor/CSS/JS, which GHC generates in 2 seconds I no longer try bigger changes involving 3-4 files at the same time Using GHC for such tasks is a waste of my time. it is time and energy-consuming to REVIEW each answer GHC provides a human, I might need maybe 10 minutes of full focus to review 300 lines of combined C#/Razor/CSS/JS, which GHC generates in 2 seconds I was thinking, why does it hallucinate about C# properties and methods that do not exist for well-known and frequently used APIs, like related to Http context, etc.? Maybe it copied a snippet from.NET4.8 Framework, and that method is no longer there for .NET Core… But, for C# classes that are part of my VS2022 project, like EF8 classes, it looks it does not read them, it just tries to guess, there must be a Primary Key with the name CustomerId, but it is actually called differently, so I need to manually open EF8 class and copy-paste proper names of the properties. I was thinking, why does it hallucinate about C# properties and methods that do not exist for well-known and frequently used APIs, like related to Http context, etc.? Maybe it copied a snippet from.NET4.8 Framework, and that method is no longer there for .NET Core… But, for C# classes that are part of my VS2022 project, like EF8 classes, it looks it does not read them, it just tries to guess, there must be a Primary Key with the name CustomerId, but it is actually called differently, so I need to manually open EF8 class and copy-paste proper names of the properties. , but believe simply that to do serious work, and it is faster to code things straight than to endlessly chat with GHC. I do not believe in “proper context” story the tool GHC is not smart enough , but believe simply that to do serious work, and it is faster to code things straight than to endlessly chat with GHC. I do not believe in “proper context” story the tool GHC is not smart enough I do not believe in “proper context” story alat GHC nije dovoljno pametan That “ by companies producing AI tools, trying to blame developers for not being able to use AI products properly. So, I have learned enough about “prompt engineering” and was making a serious effort to exploit GHC to generate code for me in my VS2022/C# project, but failed. I even planned to use GHC for tasks that suit it, to follow patterns I made (like HTML tables) and generate analogous code, but no, GHC was not good enough. Altogether, . proper context” story is a blame game it generates more mess than useful code, for assigned pattern-based tasks involving 4 files That “ by companies producing AI tools, trying to blame developers for not being able to use AI products properly. So, I have learned enough about “prompt engineering” and was making a serious effort to exploit GHC to generate code for me in my VS2022/C# project, but failed. I even planned to use GHC for tasks that suit it, to follow patterns I made (like HTML tables) and generate analogous code, but no, GHC was not good enough. Altogether, . proper context” story is a blame game it generates more mess than useful code, for assigned pattern-based tasks involving 4 files proper context” story is a blame game it generates more mess than useful code, for assigned pattern-based tasks involving 4 files Vratio sam se na stari način rukovanja datotekama UI uzoraka. Na primjer, za stvaranje "ListOfContracts" HTML (prilično napredni AJAX obrazac) na temelju vrlo sličnog "ListOfCustomers" UI obrasca. Kopiram i preimenujem datoteku. Zatim koristim urednik teksta Find-Replace kako bih izmijenio obrazac u svom VS2022/C# projektu. To je sporo, ali postoji sigurnost u tijeku, i nema neugodnih iznenađenja s manjkajućim linijama koda ili dodavanjem neočekivanih linija koda. Mislio sam da bih koristio GHC za takve zadaće koje izgledaju kao uzorci, ali to je vrijeme-efikasnije i energetski učinkoviti I returned to the old method of handling UI patterns files. For example, to create “ListOfContracts” HTML (quite advanced AJAX form) based on a very similar “ListOfCustomers” UI form. I copy and rename the file. Then I use a text editor Find-Replace to modify the form in my VS2022/C# project. It is slow, but there is a certainty in progress, and no nasty surprises with missing lines of code or added unexpected lines of code. I was thinking I would use GHC for such pattern-looking tasks, but it is more time-efficient and energy-efficient to code it directly manually. I was thinking I would use GHC for such pattern-looking tasks, but it is more time-efficient and energy-efficient to code it directly manually. Its answers tend to be verbose, and if it knows it, I will get an answer in 2 attempts, which can be brilliant at times. If it does not give me a good answer in 2 attempts, I will go read Google for the same problem. That GHC thing can generate a huge amount of text and code samples and coach you on nice things, but I have a specific problem to solve, not time to chat endlessly. I ask GHC for help with problems, but read max 2 chat replies. GHC has a serious focus problem, answers are frequently off-topic. Its answers tend to be verbose, and if it knows it, I will get an answer in 2 attempts, which can be brilliant at times. If it does not give me a good answer in 2 attempts, I will go read Google for the same problem. That GHC thing can generate a huge amount of text and code samples and coach you on nice things, but I have a specific problem to solve, not time to chat endlessly. I ask GHC for help with problems, but read max 2 chat replies. GHC has a serious focus problem, answers are frequently off-topic. I ask GHC for help with problems, but read max 2 chat replies. GHC has a serious focus problem, answers are frequently off-topic. 8. Marketing for AI products is very strong Marketing za proizvode AI od strane tehnoloških tvrtki je vrlo jak, tako da se mora uložiti u napor da se u sadašnjem trenutku ostane usmjeren na sposobnosti proizvoda AI. stay grounded u pogledu mogućnosti AI proizvoda Marketing is strong, because Tech companies have put money into it and want some sales and revenue NOW. Even if actual AI products are not “production-ready”, they want to Just to collect he money, marketing will create any story it needs. A good story that sells does not need to accurately reflect technical reality. sell them on a “vision story”. They are talking about working, you and an AI-guy like GHC. Maybe in the future, but as of March 2025, I see GHC as just another code-assistance tool. “pair programming”, meaning you will not work alone but in “pair”, there will be 2 of you They are talking about . As of March 2025, GHC is still having big problems with C# syntax and hallucinates a bit of what some .NET Core API C# class precisely looks like. It can’t read the file in the VS2022 project that is there, but instead, , while it should be the opposite; humans should delegate such boring tasks to AI tools. “peer programming”, meaning AI-guy will be equal to you in your task GHC gives a human task to check syntax and finish its work To make developers feel good about themselves, there are imaginative created about . At the end of the day, you are just a guy in front of a keyboard and monitor, not in a fancy airplane traveling the world. And that GHC “AI buddy” you have (as of March 2025) looks a bit autistic to me, talks much, sometimes brilliant, sometimes stupid, and better check everything he does/says because it may sound like he knows it, but can be wrong. marketing stories “piloting and aircraft” and “having a copilot” Many are being demoed publicly in a bit of arranged, non-realistic scenarios. , like “factorials” or “Fibonacci numbers”. These are well-known programming textbook tasks. AI systems excel in math and might even just pull a code sample from their huge memory of code snippets. Such demos do not show how such systems would hold up when faced with real-life 300+ files of code software development projects. AI-code generators They limit the demo to a generation of a small 50-line function Samo za prikupljanje novca, marketing će stvoriti bilo koju priču koja mu je potrebna. Marketing je jak, jer su tehnološke tvrtke uložile novac u njega i žele neke prodaje i prihode SADA. Čak i ako stvarni proizvodi AI nisu "spremni za proizvodnju", žele ih prodati na "vizija priču". Samo da bi prikupili novac, marketing će stvoriti bilo koju priču koja mu je potrebna. Marketing je jak, jer su tehnološke tvrtke uložile novac u njega i žele neke prodaje i prihode SADA.Čak i ako stvarni proizvodi AI nisu "spremni za proizvodnju", žele ih prodati na "viziji". prodajte ih na „priča o viziji“. They are talking about working, you and an AI-guy like GHC. Maybe in the future, but as of March 2025, I see GHC as just another code-assistance tool. “pair programming”, meaning you will not work alone but in “pair”, there will be 2 of you They are talking about working, you and an AI-guy like GHC. Maybe in the future, but as of March 2025, I see GHC as just another code-assistance tool. “pair programming”, meaning you will not work alone but in “pair”, there will be 2 of you "parno programiranje", što znači da nećete raditi sami, već u "par", bit će 2 od vas They are talking about . As of March 2025, GHC is still having big problems with C# syntax and hallucinates a bit of what some .NET Core API C# class precisely looks like. It can’t read the file in the VS2022 project that is there, but instead, , while it should be the opposite; humans should delegate such boring tasks to AI tools. “peer programming”, meaning AI-guy will be equal to you in your task GHC gives a human task to check syntax and finish its work They are talking about . As of March 2025, GHC is still having big problems with C# syntax and hallucinates a bit of what some .NET Core API C# class precisely looks like. It can’t read the file in the VS2022 project that is there, but instead, , while it should be the opposite; humans should delegate such boring tasks to AI tools. “peer programming”, meaning AI-guy will be equal to you in your task GHC gives a human task to check syntax and finish its work “peer programming”, meaning AI-guy will be equal to you in your task GHC gives a human task to check syntax and finish its work Da bi se programeri osjećali dobro o sebi, postoje maštovite stvorene o . Na kraju dana, vi ste samo tip ispred tipkovnice i monitora, a ne u fantastičnom zrakoplovu koji putuje svijetom. I taj GHC "AI prijatelj" koji imate (od ožujka 2025.) izgleda malo autički prema meni, govori puno, ponekad briljantno, ponekad glupo, i bolje provjerite sve što radi / kaže jer može zvučati kao da to zna, ali može biti pogrešno. marketinške priče "pilotiranju i zrakoplovu" i "imanju kopilota" Da bi se programeri osjećali dobro o sebi, postoje maštovite stvorene o . Na kraju dana, vi ste samo tip ispred tipkovnice i monitora, a ne u fantastičnom zrakoplovu koji putuje svijetom. i taj GHC "AI prijatelj" koji imate (od ožujka 2025.) izgleda malo autički prema meni, govori puno, ponekad briljantno, ponekad glupo, i bolje provjerava sve što radi / kaže jer može zvučati kao da to zna, ali može biti pogrešno. marketinške priče "pilotiranju i zrakoplovu" i "imanju kopilota" marketing stories “piloting and aircraft” and “having a copilot” Many are being demoed publicly in a bit of arranged, non-realistic scenarios. , like “factorials” or “Fibonacci numbers”. These are well-known programming textbook tasks. AI systems excel in math and might even just pull a code sample from their huge memory of code snippets. Such demos do not show how such systems would hold up when faced with real-life 300+ files of code software development projects. AI-code generators They limit the demo to a generation of a small 50-line function Many are being demoed publicly in a bit of arranged, non-realistic scenarios. , like “factorials” or “Fibonacci numbers”. These are well-known programming textbook tasks. AI systems excel in math and might even just pull a code sample from their huge memory of code snippets. Such demos do not show how such systems would hold up when faced with real-life 300+ files of code software development projects. AI-code generators They limit the demo to a generation of a small 50-line function AI-code generators They limit the demo to a generation of a small 50-line function 9. Conclusion It does save me a measurable amount of time by providing code snippets and suggestions sometimes. GitHub Copilot (GHC) (as of March 2025) is a useful tool, and I will continue to use it in my programming. GitHub Copilot (GHC) (as of March 2025) is a useful tool, and I will continue to use it in my programming. The technology level currently is that . In such scenarios, results are incomplete . GitHub Copilot (GHC) can not be trusted with a bit complicated task involving several files at the same time and not time-efficient compared to direct manual programming GitHub Copilot (GHC) can not be trusted with a bit complicated task involving several files at the same time and not time-efficient compared to direct manual programming A serious problem is that . The , requiring a lot of . GitHub Copilot (GHC) tends to hallucinate about C# methods and properties that do not exist GHC-generated code does not compile right away manual work to finish it GitHub Copilot (GHC) tends to hallucinate about C# methods and properties that do not exist GHC-generated code does not compile right away manual work to finish it Things are evolving quickly, so the notes above are already almost outdated—apparently, a new tool called GPT-4o was released just a few days ago. It seems that AI tech companies are quietly upgrading these tools in the background, making them better and better without users receiving any explicit notice. Some reports in the media suggest that large language models (LLMs) have reached their maximum and that by pure scaling, they are not becoming better. Maybe we will need to wait for some theoretical progress in AI science before we see real progress with AI tools. Version of . We can expect that over time it will become better. But, still needs to be seen/verified in reality. GitHub Copilot from March 2025 is useful, though not as great as advertised GitHub Copilot od ožujka 2025. je koristan, iako ne tako dobar kao što je oglašeno At the end, of course, ask for a second opinion on everything that was said in the article above.

GitHub Copilot (Gen-AI) je koristan, ali ne i srebrni metak

About Author

KOMENTARI

VIJESI OZNAKE

OVAJ ČLANAK JE PREDSTAVLJEN U

Related Stories

Light-Mode

Classic

Newspaper

Dark-Mode

Neon Noir

Minty

HN StartUps