This is a bit of rambling from me and what I believe is a good setup for developing software together with AI tools. I believe the AI tools are good, which will help good developers produce better software for our end clients.

What is the aim of creating software?

This is a super hard question because it is not always the same for different dev setups, but at some point, in the production of the software and the company paying the bill, the aim is to produce as much value as possible for the least amount of cost and within the time requirements. The least amount of cost is for the full lifecycle and not just the creation of the software.

How does AI fit, in the future development processes?

AI will be a large annual cost for the software development process. Software needs to be paid for with value. At present, the companies providing these AI services are not making profits and so the AI costs must go up. This means that if we use AI to produce software, the costs must be covered. This will only work if we become more efficient. Even the companies which are leading the way in software development with AI are not meeting the required cost targets once the price goes up. The hope is that the tools will get better.

Who can use AI efficiently?

This is actually really hard to answer and not clear. A lot of people proclaiming more speed, and amazing solutions are not really being honest. A big problem is, to use AI efficiently, you need to be a domain expert in the area where you use AI. So, if I use AI to produce security code, I can be faster because I can judge, if the output is good or bad. If I use AI somewhere where I do not understand the output, I will produce a worse solution than if I did not use AI. This is because without AI, I would read it, learn, ask experts, and educate myself, what is good in this domain. There are still no short cuts to this process for producing production code.

The skills we need in the future are people who understand their domains. Someone that can code good, will be able to code with AI. Someone who is not so skilled will produce a high amount of slop and slow down the whole team or reduce the quality of the product.

What do we need as software developers in the future?

One of the biggest challenges we have now is finding access to real, reliable and quality information. The internet is getting filled with AI slop and the people producing quality software blogs are declining. Stack overflow seems to be used less. Less blogs are being created because there are no rewards anymore. The content gets taken by AI bots and shared without any recognition. The payment, reward models are broken. People with knowledge or access to real knowledge will be key in the future.

What type of dev teams do we need in the future?

We need domain experts. And we need a way to train people to become domain experts. When hiring, people who learn to understand the topics are the skilled professionals we need and not the ones who are good at prompting. I think future successful dev teams will be small teams with very strong developers who can talk to the client and understand the domain. Funny thing, this was the same before AI when quality and costs are the main drivers.

What about outsourcing?

If AI brings all the promises it gives, this industry will be required less in the future, because I can just use AI to implement the features. The engineering work is what is still required. So code experts, architects, domain experts, these are the skills which will be still required. People close to the client, people who speak the same language are the future.

How will this affect project team setups?

We need more senior technical people and domain experts and less medium people. Good teams will be smaller and closer to the client. Less agile processes and less product team management is required. Closer to the client with experts is the key. This would require a complete revamp of how the industry does and creates software.

What about debugging and monitoring?

This is one of the areas where AI can shine, if the applications are created with quality. If the right information and the correct logs are created using a good tool, AI can be used to find all sort operational or performance issues. This will depend on the quality of the application but this is an area with loads of potential for efficiency gains.

Should we let AI complete PRs?

Absolutely not. We are responsible for the code, and at the center of every agent, or AI process, is a non-deterministic piece of software. This will choose a probable answer or anything that will fulfil the prompt request. It has no intelligence, just probability and statistical decision-making. To produce maintainable software, the dev team must understand this, otherwise the quality will suffer. A person is required between the deterministic conversions and the non-deterministic AI parts. This is why we do not need to understand assembly, but we do need to understand the code. C# to assembly compiles and always returns the same.

AI and security

This is the bit which worries me the most. AI will execute any instruction it is given. It does not think. If AI tools have access to all your data, there is a possibility that your data is shared with services which should not get your data. If you let AI act on your behalf, this is even more dangerous and the best answer for the prompt is not always what you want. GDPR, data protection and client NDA agreements are regularly getting broken when using AI in software processes. There are some great guidelines on security from OWASP and this is something I need to invest in.

AI and the planet

When we use AI, we use a large amount of energy and water, and we are no longer working in a clean industry. I think at some stage, the energy factor should also be paid for and must be visible. We need to understand how much energy and water was used to create the feature X. If I know what I use, then I can make a decision, if this was worthwhile or not. At present, this is not transparent.

Which AI tools do I use

Almost all of them in the Microsoft world. I enjoy Visual Studio Copilot and Visual Studio Code Copilot using different models depends on which delivers the best results. I really like the Github copilot.