[m8002?]

Here is a video explaining how shareholder value became the driving force for corporates. Thanks to 'Neutron' Jack. Till then it was a combination of Customers, shareholders and employees.

https://youtu.be/j0dGNDJqUM4?si=H_2Jr0vhBRJUQZYH

The problem is also in part because customers demanding more for less/cheap conveniently forgetting that it will cost them as employees.

[Samurai]

Originally Posted by Samurai

Now let's take the next logical step with this...

Computer languages were created so that human can tell computers what to do. It would be wasteful to force AI to code in human friendly language. That means ideally AI should convert user requirement directly into machine language, that only CPUs understand. Also, AI should be able to discern the intent of the user requirements exactly.

Last year I was talking about how it would be wasteful for AI to write human readable code.

I saw this video talking about how that will happen now because of NVIDIA's Blackwell chip.

https://www.youtube.com/watch?v=f0Y0ILHwwaM

But he missed the points I addressed last year:

Originally Posted by Samurai

Now remember that AI is probabilistic. That means AI generated software would work 99% of the time. How do you determine which 1% it missed, considering you can neither review nor test the code?
.
.
Most IT work output is deterministic, and not probabilistic. No banking customer will be happy with a banking software that is accurate 99% of the time.
.
.
An IT customer would want their features work exactly as required, every time. You can't hide behind AI when customer yells at you.

Ultimately, the problem is this: Chip creators and AI champions (like Altman) don't understand IT services. They think technology can solve everything and people can be removed from the equation.

[AltoLXI]

Originally Posted by Samurai

Now remember that AI is probabilistic. That means AI generated software would work 99% of the time. How do you determine which 1% it missed, considering you can neither review nor test the code?

Just a wild guess. Why not generate the software in multiple iterations and get a diff of what it missed? The 1% it missed probably will be generated in subsequent runs and can be patched. Run the automated tests each time till the requirements are 100% met.

[Samurai]

Originally Posted by AltoLXI

Just a wild guess. Why not generate the software in multiple iterations and get a diff of what it missed? The 1% it missed probably will be generated in subsequent runs and can be patched. Run the automated tests each time till the requirements are 100% met.

Who will write the tests?

[prajwalmr62]

Originally Posted by Samurai

Who will write the tests?

What tests?

I agree to your point. It is always the last 10% that is difficult to achieve. All these news of AI achievements are from those who have vested interest in AI hype. IT has this habit of over promising and never delivering. I fear this is going to be one of those. AI has a lot of hype but we still haven't figured out where we can use it yet.

Ground reality is this is going to be like self-driving Tesla. It will end up being one more fancy tool for the driver in optimistic scenario and in worse case, just a waste of money.

[AltoLXI]

Originally Posted by Samurai

Who will write the tests?

AI will write the test cases - in fact all the possible test cases.

[PreludeSH]

Originally Posted by Samurai

Last year I was talking about how it would be wasteful for AI to write human readable code.

I saw this video talking about how that will happen now because of NVIDIA's Blackwell chip.

https://www.Youtube.com/watch?v=f0Y0ILHwwaM

But he missed the points I addressed last year:


Ultimately, the problem is this: Chip creators and AI champions (like Altman) don't understand IT services. They think technology can solve everything and people can be removed from the equation.

I think the deterministic part needs to be taken care of by test cases ? Again Humans will still be involved since they bring in requirements and have functional test cases against those.

The corner cases will have to be dealt with testing and more testing. Since with automation we can do this repeatedly, it can ultimately lead to good software.

In real time software where we care a lot about cache, prefetching, memory barriers etc under multithreading, we have to do a lot of optimizations manually. Even this could be done in AI if we find a way to analyze the performance through logs and processor counters and then program based on this analysis. Becomes a little iterative which ML can do over few iterations.

[AltoLXI]

Originally Posted by PreludeSH

I think the deterministic part needs to be taken care of by test cases ? Again Humans will still be involved since they bring in requirements and have functional test cases against those.

In fact, we are already doing this without using AI or humans in our organization. Trust me, the coverage is 100% . And only a matter of time before it gets automated.

[padmrajravi]

Originally Posted by Samurai

Last year I was talking about how it would be wasteful for AI to write human readable code.

I saw this video talking about how that will happen now because of NVIDIA's Blackwell chip.

https://www.Youtube.com/watch?v=f0Y0ILHwwaM

But he missed the points I addressed last year:


Ultimately, the problem is this: Chip creators and AI champions (like Altman) don't understand IT services. They think technology can solve everything and people can be removed from the equation.

If it is generating code correctly 99% of the time, isn't it already doing better than humans ? Human developed code also has bugs at times and it is the QA who catches these bugs and initiates the fix sequence discussing with the developers. Machine generated code does not mean a manual verification of the feature set can be avoided. We will still need that process and the second iteration of machine generation can take feedback from the feature set validation and the originally developed code, just like how it happens in case of human generated code. In the long term, this requirement for manual testing may get faded away.

[Samurai]

Originally Posted by AltoLXI

In fact, we are already doing this without using AI or humans in our organization. Trust me, the coverage is 100% . And only a matter of time before it gets automated.

You are talking about code coverage. With AI written code, there is no code to cover.

I am talking about feature coverage from customer's point. Why will AI test features it didn't generate in the first place?

I covered this last year:

Originally Posted by Samurai

Who is going to write these test plans? Same AI that generated the source code, after reading the user requirements? If the source code and test plans were created by the same AI entity, why bother running the tests? AI obviously would have generated code that will pass the test plans generated by it. So, there is no need to test?

The gap is between the user requirement and the software. AI doesn't know this gap, hence it can't test it. Who is going to explain this gap to the AI engine?

[whitewing]

Originally Posted by Samurai

The gap is between the user requirement and the software. AI doesn't know this gap, hence it can't test it. Who is going to explain this gap to the AI engine?

I would assume that the "prompt engineer" who nudges AI to author the code, would we "nudging" based on a certain requirement specification?
Think of BDD where the tests are authored based on the functionality (not necessarily by technical folks, but always with folks with exposure to the domain).

As an aside, we tried BDD (not using AI, just normal humans) for some web devpt parts and the experiment was not very successful.

[torque789]

Originally Posted by charanreddy

To be fair to the recruiters, an employee seeking a lower paying job is either keeping this job as a back up or using this as a transitional job until he finds something that pays more. I was genuinely looking for a work experience which added skills to my resume and was willing to take a pay cut, but HR politely said it’s a bad idea and did not consider my resume.

Precisely put. Yes recruiters are not wrong in this perspective. But it puts even genuine job seeker in suspicion and that's the bad part of it.

Personally I have taken a hit due to this. 3 years ago the major reason for looking for switch was my mental health. I could I was going downhill both mentally and physically health-wise. I got an opportunity through one of my close buddy and directly interacted with VP of a large Automotive SW services org. And the role I was referred was for Associate Director.

The VP who was referring me internally, checked with his higher ups and came back to me with the max CTC that can offer. He wanted my confirmation if I am ok with that and only then formal interview shall be scheduled. It was ~10% less than my current CTC but the role was way bigger than my current role. (Yes bigger role, I am not referring to the title). I thought for couple of days and told him I am fine with it. He went back to his higher ups and informed and asked me to wait for interview in a week. After a week he called me saying they prefer not to go ahead with my profile. At a personal level he informed me that the reasons were : I may not join or I may leave early even if I join. As a VP with such large org he could not convince the process, such is the quantum of the "I can take a pay cut" impact in Indian setup.

He was sorry for not being able to help me, however I was thankful to him for making an effort over and above than what he was needed to as he was not my direct friend.

[DigitalOne]

Originally Posted by Samurai

Why will AI test features it didn't generate in the first place?
.....
The gap is between the user requirement and the software.

You are looking at AI as a single entity. If you see an AI tool as an equivalent to a median employee, the answer becomes clear. One AI system generates the code, while other generates the test cases and tests the code generated by the first.

This would be a setup which isn't too different from existing software development processes, with the only critical difference being the AIs are working 24/7.

[Turbohead]

Originally Posted by DigitalOne

You are looking at AI as a single entity. If you see an AI tool as an equivalent to a median employee, the answer becomes clear. One AI system generates the code, while other generates the test cases and tests the code generated by the first.

This would be a setup which isn't too different from existing software development processes, with the only critical difference being the AIs are working 24/7.

So are you suggesting that a company has to subscribe to AI engines from different companies to do this successfully? I don't think that'll work for long - from my limited understanding of how Gen AI works, wouldn't the two or three systems learn from each other thereby bringing us to the orginal problem?

I mean, why would I trust my money with banks which develops and maintains the code exclusively by AI? What happens when something goes wrong? Blame the AI engine? I don't think you can remove humans who know their stuff. The whole AI system, if I'm not wrong, works on existing data and producing data from data. I can't imagine how it can be creative and bring in a brand new initiative. It can definitely enable humans to finish tasks, that are already known, very quickly though.

If I'm wrong, would like to see the reasons/articles about the same.

[Samurai]

Originally Posted by DigitalOne

You are looking at AI as a single entity. If you see an AI tool as an equivalent to a median employee, the answer becomes clear. One AI system generates the code, while other generates the test cases and tests the code generated by the first.

How does that solve the problem?

First, let's not say AI will code, at least not human readable code. It directly creates the software in opcode, by skipping the coding, compiling & linking cycle.

If the software creating AI is different than testing AI, they both must still take their input from the same user requirement, right?

If both of them interpret exactly, all tests will pass, irrespective of whether the user requirement was really met.

If they interpret differently, many tests will fail. How does the iteration work here? Who decides which AI got it wrong?

I simply don't see how the human presence can be removed here.