Developing projects with AI

[Vod]

I opened up DeepSeek and typed something like "Using this page (https://bitcointalk.org/index.php?action=profile;u=30747) can you identify all values and put them in a C# datatable?" and within a minute it spit out code that took me weeks to first write.       I have been using it for what all programmers hate - error handling.    And with the incredible context Deepseek manages, I don't need to issue the same instructions over and over.  I can write "How about this code?" and it will apply every necessary check for nulls, etc., as long as I asked it to do so previously in the same chat.    Being in Alberta, I can use it easily until around 2200, then the server starts getting busy.

So I installed Deepseek on my local server (HP DL380) with 64GB RAM and Xeon processors.   Downloaded the 90GB database in under 5 mins* (lol) and it installed fast!   But there is no video card in this server.   I asked it "How can I reduce cat allergens?" - it took 17 mins to think and then spit out a word about every two seconds.  Not usable at all.    That was the full R1:70b.   So I tried the limited R1:8b, got a notice there was a new version and asked "What version are you?" and it still took almost five mins thinking.

* When I first got into BBSes, I could download 1MB an hour.   Now I can download 100MB a second (a 360,000x increase in 30 years).

If you are interested in trying DeepSeek on a computer with a video card, here is a straightforward guide for running the AI and the web GUI.
https://www.tecmint.com/run-deepseek-locally-on-linux/

[qzl9199]

Because I am Chinese, I have been using DS since December last year. Unfortunately, the popularity of DS has led to a frenzy of DDOS attacks (according to Chinese cybersecurity companies, most of the attacks came from the United States). I'm curious if there is a limit for you guys to sign up for DS now? I'm using DS now and I often get the "server busy" prompt, which is really unfortunate!

[robelneo]

Because I am Chinese, I have been using DS since December last year. Unfortunately, the popularity of DS has led to a frenzy of DDOS attacks (according to Chinese cybersecurity companies, most of the attacks came from the United States). I'm curious if there is a limit for you guys to sign up for DS now? I'm using DS now and I often get the "server busy" prompt, which is really unfortunate!

No, not all. I just signed up hours ago and have been using Deepseek. I only have a problem with CloudFlare, which took a lot of time to verify. Can you post the link where you posted that the United States is responsible for the attack on Deepseek? If there's an allegation, it should be backed up by proof.

DeepSeek made headlines because of the huge discrepancies on their cost and mapower Just see the difference on the imege I provided

But there are reports that DeepSeek failed on its security test

DeepSeek R1, a new chatbot by a Chinese startup, has failed abysmally in key safety and security tests conducted by a research team at Cisco in collaboration with researchers from the University of Pennsylvania.
“DeepSeek R1 exhibited a 100% attack success rate, meaning it failed to block a single harmful prompt,” said the research team.

DeepSeek 100% fail: Chinese AI model could not stop a single harmful prompt

[NeuroticFish]

I've signed up and used DS from browser now and then in the past 2 weeks or so. It's reasonably fast, had no downtime of lag after the (slow) initial sign up.
I would have guessed that the local experience is not too good (of course, I couldn't have said how bad it is, so thanks Vod for the test and for telling). I think that they are not entirely honest on the hardware usage.
I've got "server busy" a few times, but usually after trying too deep into one topic or another. That made me think that it's a hardware issue, not an internet issue.


So I think that as long as you don't give out too much, it's okay. I have no interest for now in running it locally.
But I've asked DS make me a program and the result was not too bad; although the look (UI) of the result was more at "student" level (to say it kindly), overall was impressive, much better than Gemini (on the same task), better than Copilot (which I've used the most).


But I fear that there are more chances DS can be used for bad than ChatGPT, and no matter how ugly the paycheck of OpenAI looks like, I think that adding some safety measures here and there is necessary. Making such a tool available for local use is wow, but I fear it may open Pandora's box.

[Vod]

To continue this discussion, a number of better AI engines have launched just since my post!   

When bitcoin was invented, everyone realized that a high end GPU can do the calculations six times faster than a high end CPU.  Demand soared, prices soared, supply caught up, prices dropped.... (breath)... and that allowed think tanks to start experimenting with the thousands of cheap GPUS now available.  Deep neural learning became valuable, and this time the chip companies reacted much faster, modifying the top end chips for the AI demand.

So you may think you missed your chance with bitcoin, and then with nvidia, but I believe you have more more chance yet. The current iteration of AI chips will not be the final training design.  How do I know this?  The human brain does not take massive amounts of energy to classify thoughts - we have the design wrong.   What's even more obvious is the way our neural networks can send gradient signals between the layers, while the brain either emits a signal from a neuron, or it does not. 

For those of you looking for a project - design a better neural network. 

[Vod]

I wanted to provide an update on my AI project experience.

When I started two months ago, I would have told you I could complete my project before I went to sleep - Deepseek was that good at doing exactly what I wanted...  Over the last few weeks I've struggled to make much progress just like when I was coding manually.  When I first told it what I wanted, it designed it perfectly.   Then I started to add features and that is where the problem started.

DS would write a function a different way than it wrote for me a day before (example:  passing a parameter in post body vs querystring) and then my code would break.  Rather than fixing what I knew was the issue, I wanted DS to recognize it.  It did, but only after I provided fresh affected code and/or table definitions.   Since I used DS to code things I had no idea of (event handlers on browser clicks for example), it would tell me I was missing this and add it to that, but then later it may tell me to add the same this to a different that, duplicating functions and leading to troubleshooting anxiety.   Also, every now and then DS will give me code that contains a missing closing tag...

I wanted AI to code this entire project based on my input.  I continue to believe that AI is great for mundane tasks like syntax or logic checking, but it cannot be used as a substitute for overall project development.   If I wasn't a coder before, I would have no idea what to ask or tell DS.    Those of you who are waiting on some features I promised; I still feel I can complete it by the end pf this month.

[askii]

I wanted to provide an update on my AI project experience.

When I started two months ago, I would have told you I could complete my project before I went to sleep - Deepseek was that good at doing exactly what I wanted...  Over the last few weeks I've struggled to make much progress just like when I was coding manually.  When I first told it what I wanted, it designed it perfectly.   Then I started to add features and that is where the problem started.

DS would write a function a different way than it wrote for me a day before (example:  passing a parameter in post body vs querystring) and then my code would break.  Rather than fixing what I knew was the issue, I wanted DS to recognize it.  It did, but only after I provided fresh affected code and/or table definitions.   Since I used DS to code things I had no idea of (event handlers on browser clicks for example), it would tell me I was missing this and add it to that, but then later it may tell me to add the same this to a different that, duplicating functions and leading to troubleshooting anxiety.   Also, every now and then DS will give me code that contains a missing closing tag...

I wanted AI to code this entire project based on my input.  I continue to believe that AI is great for mundane tasks like syntax or logic checking, but it cannot be used as a substitute for overall project development.   If I wasn't a coder before, I would have no idea what to ask or tell DS.    Those of you who are waiting on some features I promised; I still feel I can complete it by the end pf this month.

I think that's unfortunately still the reality with LLMs at the moment, they are amazing at solving specific issues - but you get a bit of a frankenstein effect when you try to string those solutions into a cohesive whole. Part of the fault is also with us as the prompters (I'm sure this will improve on the AI's part over time), just because it is so difficult to provide all the necessary context and phrase things just right so that it actually understands what you want

Personally, I use it religiously to bounce ideas off of and review my code to suggest improvements - but I don't think I buy the "vibe coding" idea that you can make entire complex projects solely by prompting an LLM (not yet). I’ve found it can actually be more helpful, especially when dealing with unfamiliar concepts, to ask it to teach you how something works rather than just providing a direct implementation or solution. It can make it a lot easier to combine with other parts of your project later on.

[DookieMcSniffer]

Hey Vod,
I don't think I understand. You said you are running AI on your server with 64gb ram and Xeon technology, however, to my knowledge, effectively all high powered llm (and other AI tech) is optimised for GPU operation due to the matrix heavy nature of their processes (and all modern GPUs have dedicated matrix / tensor cores). Is there a reason you're trying to run LLM setups on CPU? Seems like an unfair way to compare the current state of artificial intelligence.
Thanks
Dookie.

[Vod]

Hey Vod,
I don't think I understand. You said you are running AI on your server with 64gb ram and Xeon technology, however, to my knowledge, effectively all high powered llm (and other AI tech) is optimised for GPU operation due to the matrix heavy nature of their processes (and all modern GPUs have dedicated matrix / tensor cores). Is there a reason you're trying to run LLM setups on CPU? Seems like an unfair way to compare the current state of artificial intelligence.
Thanks
Dookie.

It's all I had.  I don't have the GPU units. 

I am still developing with AI, and it is driving me bonkers.   It assumes so much, and sneaks in little changes based on those assumptions.   I'd like to play with a local LLM but it's too expensive to rent.

[goldkingcoiner]

Deepseek is absolute trash for coding. It keeps hallucinating faulty code and removing parts of code randomly. ChatGPT is far better (although it sometimes removes parts of your code for no reason as well).

I heard a lot of people claim that Claude is even better than ChatGPT when it comes to coding but I have not tried it yet, myself.

That being said, I no longer trust AI to make any changes to my code, I just ask it to point out errors and offer solutions. But no copy pasting AI code.

[Vod]

Deepseek is absolute trash for coding. It keeps hallucinating faulty code and removing parts of code randomly. ChatGPT is far better (although it sometimes removes parts of your code for no reason as well).

I officially agree with you.  I'm going with a chatGPT subscription, not because DS is always busy, but exactly what you described.

[goldkingcoiner]

Deepseek is absolute trash for coding. It keeps hallucinating faulty code and removing parts of code randomly. ChatGPT is far better (although it sometimes removes parts of your code for no reason as well).

I officially agree with you.  I'm going with a chatGPT subscription, not because DS is always busy, but exactly what you described.

Let me go ahead and save you some money then  
Here is how every interaction will go:

User: Give me solution
ChatGPT: Here is the solution
User: This is not the solution
ChatGPT: You're absolutely right!

[yhiaali3]

I tried DeepSeek to design a simple video conversion program in C#. I gave it some instructions and it generated code in less than a minute. Unfortunately, it was full of errors. For example, it generated code in Visual Studio 2019, even though the latest version is Visual Studio 2019. The code was based on .NET Framework 4.7.2, even though the latest version is .NET Framework 9.

It took a long time to fix the errors because some code is no longer used in the latest version of Visual Studio or has been replaced. However, overall, I've tried other AIs, such as Bolt, which were better at programming, but they also weren't without bugs.

[Vod]

I tried DeepSeek to design a simple video conversion program in C#. I gave it some instructions and it generated code in less than a minute. Unfortunately, it was full of errors. For example, it generated code in Visual Studio 2019, even though the latest version is Visual Studio 2019. The code was based on .NET Framework 4.7.2, even though the latest version is .NET Framework 9.

Is that an error, or are you able to fix it by specifying the versions you want it to use?   Since you use VS 2019 instead of 2022, maybe it assumed the older framework was appropriate?

I switched over to chatgpt last night, and I have hope it can repair the damage DS caused.   If not, I'll restart it with chatgpt, and have high hopes (again) I can finish it quickly.

Let me go ahead and save you some money then  

Have you tried https://chat.chatbotapp.ai/?  It gives you access to a lot of different AI, but they prob cut costs by using lower level engines when available or when you go over a quota.

[yhiaali3]

Is that an error, or are you able to fix it by specifying the versions you want it to use?   Since you use VS 2019 instead of 2022, maybe it assumed the older framework was appropriate?

I switched over to chatgpt last night, and I have hope it can repair the damage DS caused.   If not, I'll restart it with chatgpt, and have high hopes (again) I can finish it quickly.

Yes, the problem was solved by specifying the appropriate versions. Perhaps my mistake from the beginning was that I didn't specify the versions I was using accurately to DS.

But what surprised me wasn't this error, but rather other errors related to small details, such as this one:
DS wrote the code as follows:

Code:

private void btnBrowse_Click(object sender, EventArgs e)
{
// Button code here
}

A warning message appeared in Visual Basic:

Naming rule violation: These words must begin with upper case characters: btnBrowse_Click

This warning indicates that the event name btnBrowse_Click does not follow C# naming rules. According to naming rules, event names must begin with capital letters.

The problem can be easily fixed by using the letter B instead of b in the event name:

Code:

private void BtnBrowse_Click(object sender, EventArgs e)
{
// Button code here
}

Small errors can be easily fixed, but I'm surprised at how an advanced AI like DS can handle them, because tracking and fixing errors takes longer.

[askii]

But what surprised me wasn't this error, but rather other errors related to small details, such as this one:
DS wrote the code as follows:

Code:

private void btnBrowse_Click(object sender, EventArgs e)
{
// Button code here
}

A warning message appeared in Visual Basic:

Naming rule violation: These words must begin with upper case characters: btnBrowse_Click

This warning indicates that the event name btnBrowse_Click does not follow C# naming rules. According to naming rules, event names must begin with capital letters.

The problem can be easily fixed by using the letter B instead of b in the event name:

Code:

private void BtnBrowse_Click(object sender, EventArgs e)
{
// Button code here
}

Small errors can be easily fixed, but I'm surprised at how an advanced AI like DS can handle them, because tracking and fixing errors takes longer.

This is just a style convention thing, not an error. But ChatGPT has the same problem... My guess is that for simple functions like this, LLMs pick up patterns from older codebases, and since many legacy projects used Hungarian notation, that is what DeepSeek or GPT will use. If, in your prompt, you explicitly specify something along the lines of "always follow Microsoft's C# guidelines", it will not make these types of mistakes.

[Vod]

This is just a style convention thing, not an error. But ChatGPT has the same problem... My guess is that for simple functions like this, LLMs pick up patterns from older codebases, and since many legacy projects used Hungarian notation, that is what DeepSeek or GPT will use. If, in your prompt, you explicitly specify something along the lines of "always follow Microsoft's C# guidelines", it will not make these types of mistakes.

If the AI is converting code from one language to another, it may use the style of either language unless you are explicit.

Common naming styles in programming (2025)

Name	Example	Where Used
camelCase	userName	JS variables/functions, Java
PascalCase	UserName	Classes, React Components
snake_case	user_name	Python, Ruby, SQL
SCREAMING_SNAKE_CASE	USER_NAME	Constants, ENV vars
kebab-case	user-name	URLs, CSS class names
Hungarian Notation	strUserName	Older C/C++/VB code
StudlyCaps	uSeRnAmE	Memes, jokes, passwords
dot.notation	user.name	Configs, object keys

* AI generated - I'm not going to take the time to format to BBCode.

[askii]

If the AI is converting code from one language to another, it may use the style of either language unless you are explicit.

Being explicit is key  The AI can't see inside your brain, so even if something seems super obvious, it is still your job as the prompter to provide clear instructions with as much detail as possible for the best results.

As I mentioned in a previous response, vibe coding projects is not the way to go! Use LLMs to teach you concepts, write pseudocode, and at most, implement helper functions, or functions with well defined purposes. Trust me, you'll generally get nowhere by asking GPT/DS/Claude/etc to create an entire project for you from a single prompt (even with plenty of context), you'll eventually hit a roadblock, or you'll be stuck with a bunch of issues that are a pain in the ass to resolve.

In my humble opinion, AI should be more of a learning tool. The same way we used to use google searches to learn and solve problems, rather than copying chunks of people's code and frankensteining it all together 
AI is like an infinitely better google search, so I would treat it as just that.

[joniboini]

In my humble opinion, AI should be more of a learning tool. The same way we used to use google searches to learn and solve problems, rather than copying chunks of people's code and frankensteining it all together 
AI is like an infinitely better google search, so I would treat it as just that.

I agree. Just recently, I saw someone post a code they got from AI and claimed that it was the best coding ever, but failed to explain why, and got so defensive when I asked him about it. Even if you can use that to fake job interviews, I don't think they'll last that long if they can't explain simple things on their job.

My experience with Google AI isn't that good either. Sometimes they linked dubious articles and gave me hallucinated results out of nowhere.

[markm]

Reading this and some other threads about using AI for projects I still don't have much idea of what they are really ready and able to do.

I have recently read this paper about developing autonomous bots for CoffeeMUD, which seems like something I could definitely use:

https://e-archivo.uc3m.es/rest/api/core/bitstreams/7452b84e-a307-48c0-ba12-99cd4c15de09/content

I had been thinking maybe one can just tell some AI or other to read that paper then telnet to a CoffeeMUD server and "be" such a bot, or maybe to fire itself up a tinyfugue or other MUD client and use that "as" such a bot, rather than asking it to write code, for example tinyfugue macros and triggers, to implement such a bot?

But reviewing this thread I am now dubious it could write tinyfugue (or other MUD client) macros-and-triggers (in sum, scripts) to implement such a bot.


Other vague ideas in back of my mind make me think maybe the AIs also likely cannot do things like go log in somewhere to do something, right? Some kind of prevent them being used to attack other servers etc protection / ethics?


-MarkM-