​Let’s Talk About the AI Engineer – Project Prometheus

How might an AI engineer approach a problem differently from how a human engineer would?

If you have been paying attention, then you have heard about the USD 12 Billion that Jeff Bezos and Vik Bajaj raised to develop an AI general engineer. One that can do what any other engineer can do – so it should be much like a global engineer. Read more here if you are not aware.
If possible, then what would that mean for professional engineers like us?
Bezos thinks this will help augment engineers and give them more time. Others think that all engineers will eventually be replaced. And what has been achieved thus far is not yet in the public domain.
In this article, I am not going to talk specifically about what Prometheus will or will not be able to do. Instead, I am going to use this as an opportunity to talk about the following:

1. What would be needed to train such an AI system?
2. What that means for the nature of such an AI system?
3. What would that mean for engineers working with such an AI system?
4. Should we even expect such an AI system to think the same way as human engineers?

I am not going to assume that I will know exactly what is being done on Project Prometheus. Nor that I will have the exact answers, but this conversation will help you and me think more about what’s needed to make an AI engineer and what it is to be an engineer in general.
What would be needed to train an AI engineer?
In my book on being a global engineer I noted what has been found about the way the best engineers think. In summary, they do 3 things really well:

1. They frame problems in a way that makes them easier to solve. They do not accept the problem as is. They look at it from different angles to find another way to define it – and then find an angle that makes the solution obvious and easier to implement.
2. They think systemically. They do not focus on just the challenge in front of them. They look at all peripheral elements for both challenges in accomplishing their goal and opportunities to help achieve those goals.
3. They use first principles. They do not make arbitrary decisions about the values of key parameters. They calculate the optimum value based on theory. They also use theory to understand the nature of any challenge they face.

Would an AI engineer explicitly be trained on these three attributes?
If so, then would we explicitly state these as needs or would we train the system to engage in these actions by default?
Not only that, but would we consider other, very human aspect of engineers?
One example is fixation. Where someone, in this case an engineer, stays focused on an idea that has come to them. It might be because it is the first one that showed promise, something they are excited about, or something that seems obvious because they have been working in a certain field for so long.
Fixation can, at first, seem like a bad thing. However, it has also been found to provide drive – and engineers have developed remarkable innovations by overcoming the challenges caused by this fixation. Sometimes better than what would be expected by someone objective who could see that fixation taking effect.
Would we want the AI engineer to show this tendency toward fixation to explore ideas fully?
Which brings us to another attribute: co-evolution
Many engineers talk about the need to iterate a design. I agree with the need, but I do not think the word “iterate” explains the depth of what is going on. The word “co-evolution”, I think, elicits the true deeper meaning of what is happening. As a solution is implemented in some sort of trial, new information is generated about the problem. So as we generate the solution, we better understand the problem. The solution and the problem evolve together.
As we co-evolve both the solution and the problem, we expand our understanding of the solution and problem space. It is like exploring a new land – but one where the terrain is the challenge we are solving. Some solutions find a path through rocky terrain that opens new areas to explore that we never knew about.
Do we need to find a way to train an AI engineer to explore like this?
If so, then is it all through simulation or do we need to run the physical tests under its instructions and then report back (expecting questions about how well we ran the tests)?
What does this mean for the nature of such an AI system?
Considering the above, do you feel that there is data enough to train an AI system to be a general engineer?
Personally, while I do think AI engineers will eventually be a thing – and eventually be better than us, I am not sure it will happen as a result of ingesting large amounts of data.
I think there will at least need to be some kind of reinforcement learning. Where the AI will adjust through trial and error to provide better outcomes.
That means the system will need some kind of objective measure of success. It will need to be able to assess how well an idea has performed. Maybe based on a numerical goal. Maybe on some other system that is more “reflective” so it can ponder how it could have done better – comparing what it actually did with other things it could have done.
This is something that many engineers do. You have possibly thought back to things you did years ago and thought “now why didn’t I do it like this?” I know I have done that.
So the AI engineer would need to have some kind of inner monologue. Is that possible?
Maybe it will actually need to be a team of AI engineers – each with different parameters from their own training – interacting with each other. Comparing the performance of each other’s ideas and then adjusting their own parameters to think in a better way.
So it might be that we never have a single AI engineer in the true sense, but a collection of AI engineers working together and then presenting output as if it is from a single AI engineer.
This considers the “cognitive” aspects. But what about the physical or real-world execution?

• Will the AI engineer call up suppliers to confirm current capabilities and use that as an input?
• Will it have hyper-generalised transducer (maybe a humanoid robot) that allows it to build and run prototype tests?
• It will likely be able to connect to rapid prototype machines, but will humans (with their nimble hands) still be the ones putting the parts together and finding (with their comprehensive senses) the ideal place to run the assembled product?
• What is the most economic execution of the above actions at this time, and what does that mean the AI engineer should be like?

The above does seem to point to the notion put forward by Bezos that the generalised AI engineer will be working with other engineers to increase productivity and free up the time for humans. For now, at least.
What does this mean for engineers?
If you do ever find yourself working with a general AI engineer, then it would seem likely that you will be using it to do the things that would normally be time consuming for you. Explore numerous ideas, source the appropriate first principles, conduct simulations and calculations, generate test plans, come up with questions to better refine the problem statement.
You could then share your own ideas to help move things along.
At first that seems like you get to enjoy the fun parts more. And have more free time.
 
However, in the continued pursuit of profits, I am sure there would eventually be fewer engineers working for the same amount of time. History just shows that’s how things go.
But how will you interact with the AI engineer?
Again, in my book on being a global engineer I covered how words are not enough to explain engineering concepts. How engineering is very much a visual thing – even though other senses can help better understand any challenge you are facing. How do you convey what’s in your mind’s eye to the AI engineer?
Engineers are either going to have to improve their CAD skills to quickly convey their ideas or work on their ability to sketch. You might be able to use words to ask for an initial concept, but sooner or later you will need to edit some sort of visual to convey what’s in your mind.
But will the AI engineer visualise the way we do?
Should we even expect such an AI system to think the same way as human engineers?
We have developed our engineering skills by working with the brain that biological evolution gave us. This does not mean it is the only way invention can come about.
There are numerous examples in nature of animals showing the action of invention. Sometimes it is thought to be a result of instinct from evolution and sometimes from actual intent. But, as we learn more, the definition of intelligence and creativity seems to broaden.
Such ideas have also been explored in literature.
In the Children of Time series, Adrian Tchaicovsky explores how creatures with multiple brains or limited memory capacity could evolve inventiveness. In some, there is a separate brain that takes the problem as given by the main brain and then returns a solution some time after processing. Others have an ability to install and uninstall knowledge so they can use what is needed for the respective task.
In the West of Eden series, Harry Harrisson explores the nature of intelligence and inventiveness of an evolved reptilian brain. Where all technological advancement is an iteration on previous efforts with no major leaps – keeping them away from metal working and machinery, but phenomenal selective breeding.
Given the diversity of thought that we have found in nature and what we can conceive when we put the effort in, we can expect the possibility of a general AI engineer that thinks very differently from how we do while still generating excellent solutions.
Will we be able to work with such an engineer? Will we automatically think it inferior? Will we try emulating it ourselves after it reveals new ways of thinking to solve engineering problems?
In this article, I have only covered the major and some select aspects of engineering cognition. I have also only covered a select number of engineering activities associated with the full implementation of any engineering solution. I have also only alluded to the importance of engineering teams. And the diverse nature of the potential forms of intelligence was given only a cursory coverage. To deal with them properly would require an entire book.
That means there is much more to explore and consider. I currently can’t make any solid predictions about how this will all end up.
However, I do now feel convinced that the effort to make this general AI engineer will reveal much more about what it is to be an engineer. Even if it fails. So I plan on following it in detail, and I think you should too.
As you think about it now, what challenges and opportunities do you see? What would you want a general AI engineer to be like?
Share your thoughts – this is a conversation I would like to extend so we can all learn more about what it is to be an engineer.