Artificial Intelligence thread

siegecrossbow

General
Staff member
Super Moderator
Give it to more years....

Most of the code out there is either mediocre or crappy. The longer and more data you use to train your model… But I guess if companies are okay with people taking shortcuts they should be with AI taking short cuts or worse as well. At the end of the day, all they care about is shareholder values.
 

tphuang

Lieutenant General
Staff member
Super Moderator
VIP Professional
Registered Member
Most of the code out there is either mediocre or crappy. The longer and more data you use to train your model… But I guess if companies are okay with people taking shortcuts they should be with AI taking short cuts or worse as well. At the end of the day, all they care about is shareholder values.

there are stuff that you can probably automate away with enough training. If you can only do javascript or python, you are probably in trouble. But if you are working in really complicated systems using C++ and have to worry about memory leaks, write long unit tests and evaluate performances, this is something AI cannot do. At least not the current model of training.

That's because there is no data to train on. How do you train on high end C++ programming if no firm in there right mind would leak their code?

A major part in data training is sorting through the data that you feed into the model or RAG or whatever. I'm dealing with this stuff daily.

Even if you have the code, but the people sorting through the code to feed into the model don't understand what the code does, how is that helpful?

People completely underestimate just how much work and intelligence is involved in sorting out good data from bad.

Like the AI work currently being done right now can probably replace Indians & Filipino labor. But that's because the people working on this stuff is far more intelligent than the people that work on the jobs to be automated away.
 

Engineer

Major
Not even AI can handle undetermined and vague business stake holder.
They already have issues just talking to a person let alone them knowing what to prompt and how to validate.
As you said, this problem has to do with the stakeholders, nothing to do with AI's ability.

My prediction will be from the early 2030s there will be an huge shortage of senior software engineer that can fix up the pile of shit called AI generated code. The industry will once again stand still for 2~5 years like what happened when they outsourced code to code sweatshop countries.
No need to fix code. By then, it would just be cheaper and easier to ask AI to generate new code every time.

This. People keep conflating software engineers with software programmers/coders, but there's more to software development than just coding the functions and there are more responsibilities software engineers have compared to the latter. When making a new software project, there are many factors to consider than just the functional aspect of your software, like non-functional aspects such as maintainability, security, compatibility, performance, etc., and building test cases to ensure the stability of the project as you add or modify it.

During the initial phase of the development process you also need to carry out requirement gathering and analysis, feasibility studies and how to keep your client's expectation/demand of the project from a simple product landing page to the next Amazon + Google + Netflix website. How is the AI going to keep the project within reasonable and realistic scope and communicate with the client, because from my experience with using ChatGPT it have a tendency to agree to everything and hallucinate answers that is not correct.

The two above is just some of the few complexities when it comes to develop a proper software product, I just don't see how you can put 100% trust in the AI to full fill all the responsibilities of software development. Which is why I don't see software engineers (which I am a part of) losing all of their jobs to AI in the future, but I do think it would significantly improve productivity and speed of projects as it gets developed in the future, and will impact the job market for programmers considerably.
These two issues are easier to solve than you might think. I saw a demo that does exactly these — project management. This is not the ChatGPT that you would use via a browser, but a dedicated software interfaces with ChatGPT via API. It also uses up to a million tokens in an hour. The problem lies in translating those requirements into codes. I would actually go as far as saying that project management would get automated first before coding would, because code must follow a very strict formatting standard spanning across multiple files.
 
Last edited:

9dashline

Captain
Registered Member
As you said, this problem has to do with the stakeholders, nothing to do with AI's ability.


No need to fix code. By then, it would just be cheaper and easier to ask AI to generate new code every time.


These two issues are easier to solve than you might think. I saw a demo that does exactly these — project management. This is not the ChatGPT that you would use via a browser, but a dedicated software interfaces with ChatGPT via API. It also uses up to a million tokens in an hour. The problem lies in translating those requirements into codes. I would actually go as far as saying that project management would get automated first before coding would, because codes must follow a very strict formatting standard spanning across multiple files.
Yup

Please, Log in or Register to view URLs content!

End to End fully automated AI Software Engineering will be a thing soon...

1million tokens a hour is pretty cheap, ranging from 3 bucks to 15 cents depending on the API inference model

Hell, starting pay at Chickfila is now $16/hr

Most SWE are about to get cashiered
 
Last edited:

Engineer

Major
Yup

Please, Log in or Register to view URLs content!

End to End fully automated AI Software Engineering will be a thing soon...

1million tokens a hour is pretty cheap, ranging from 3 bucks to 15 cents depending on the API inference model

Hell, starting pay at Chickfila is now $16/hr

Most SWE are about to get cashiered
Well, $3 is a bit low, $15 would sound more realistic; but yeah, $15 in only one hour is peanut compares to paying a senior's salary over multiple weeks.
 

tphuang

Lieutenant General
Staff member
Super Moderator
VIP Professional
Registered Member
Yup

Please, Log in or Register to view URLs content!

End to End fully automated AI Software Engineering will be a thing soon...

1million tokens a hour is pretty cheap, ranging from 3 bucks to 15 cents depending on the API inference model

Hell, starting pay at Chickfila is now $16/hr

Most SWE are about to get cashiered
again, this is just nonsense. It's very low level programming. I can see how certain IT workers in India will get automated away, but all this is capable of is simple stuff.
 

tphuang

Lieutenant General
Staff member
Super Moderator
VIP Professional
Registered Member
Please, Log in or Register to view URLs content!

Please, Log in or Register to view URLs content!

Yuanxiang Releases MoE open source large model with 255 b parameter and 36B active parameters. The model looks to have reasonable performance compared to other open source LLMs.

It is now available to try on huggingface and modelscope

Please, Log in or Register to view URLs content!

Please, Log in or Register to view URLs content!

you can also see it on GitHub
 
Top