S10E220: Is Project Controls Behind on AI? | Yash Desai, Movar Reply

And in order to do that, that costs money.

Now we work on multi-year, multi-million, multi-billion pound projects, right? And oftentimes it's on the main contractor or subcontractors to adopt the technology because they're on site, you know.

and they want to protect their margins and so are they really going to invest in it um or is it going to have to be something that starts now in contracts where it's built in so you know the project starting now which will only start potentially being um developed and and being on site in a year two three years down the line and then who knows what technology is available then so it's a tough one to to call but i don't think the dial would shift that much in a year yes what do you reckon

I think the pace at which we've seen specifically the technology change in the last three years, I think it's really crazy, I would say.

Because if you look at specifically if you look at language models, There were separate models previously for text classification and email classification and these sorts of things which don't really exist now because a basic small language model can actually do that for you.

There were models specifically trained for object detection.

There were models, let's take an example of like on a construction site if you want to see if a specific person is wearing a PPE kit or a safety hat, you would need Again, like five years ago, you would actually have to train a computer vision, let's say a deep learning model, with a lot of images where there's people wearing hats, there's people wearing no hats, there's people wearing PPE kits, there's people wearing no PPE kits, with labels, with all of this training data.

You would have to train this model and then again identify if a person is wearing a PPE kit or not.

now like with language models just take a very small nano model which is multi model it can do this job for you quite easily.

So, again like with language the way language models have adapted and the way they have like eaten up lots of small use cases around micro decisions that people make.

I think similarly within robotics as well, again we start with micro decisions at the moment where there's small things like if you look at spills, I'm sure there's lots of things going on around spills at the moment.

There's EDMs, there's even duration monitors where they try and capture the duration of spills within CSOs.

And there's lots of different ways where you can automatically turn on and turn off these, you know, speed meters where based on the weather, based on the flow and these sorts of things.

And again, this basically is a micro decision and this decision can be taken off of data.

So if you have enough data, if you are able to train a model that is accurate enough, precise enough to take that micro decision for you, then you can you have a chance of improving the outcomes of this of the system.

Link to what you both said there.

Do you think there's a role for governments in robotics? Do you think their role should be very much around governance or is there potential room to invest in there? Open question, or is it very much a private sector thing that the potential returns on all of this stuff must be so great that they want to keep it there.

So where do you think governments can or can't play a role in this space? open question to you both.

Goyesh, keen to hear your thoughts.

I think it's again an interesting one.

Like definitely governance is an important sort of factor when it comes to robotics and agents both.

But I think how you kind of have to split this based on the types of jobs the robot is doing in the first place.

If the data it spills out is inaccurate, then you have to hold the person who's programmed that from a data perspective in the first place.

If it's doing something inaccurate or if it's doing something wrong physically and trying to like harm or break something then again the person who's programmed the hardware can be held responsible similarly to agents as well there's there's an aspect where there's a software in a specific agent but again there's yeah there's an aspect of how well we've provided the agent with information and does was it provided with enough information in the first place where you hold probably uh you know the person from that specific area or let's say if it's a planning agent you hold the planner to account you hold an ai engineer to account when a planning agent goes rogue.

So again, it's again like an open ended thing where I think there will be more frameworks.