As someone who used to work in a County Council Highways and Planning department, I found this very interesting.
A few comments - if this had been done 10 years ago, I suspect AI wouldn't have been mentioned - it would have been called "automated image classification" or something similar.
On preventative maintenance, there is a lot of stuff in the operational research literature about optimum maintenance strategies, and I doubt if repairing every small defect wherever and whenever they appear is optimal.
Perhaps "AI" or whatever you want to call it can help in developing the maintenance and inspection programmes and schedules
Maybe the exciting thing is the collection of up-to-date street-level images, rather than the AI. There's lots of council use cases even beyond the road maintenance department. Maybe those CCTV mini cars which go looking for parking violations could be replaced by generic "photograph everything" vehicles. Maybe the planning department would like proof of exactly when an unauthorised building was put up. And beyond local councils, I'm imagining what would happen if they loaded all these images into a publicly published site like google streetview (or not like google streetview, because their data is not open licensed ...like Mapillary or others: https://wiki.openstreetmap.org/wiki/Street-level_imagery_services ). Personally I would be love to be able to use very up-to-date images of streets to make updates to OpenStreetMap data for example.
Is this an example of how the umbrella term 'AI' has been toxified by ChatGPT-style bullshit? I might be misunderstanding here, but is the pothole stuff with onboard chips of a different character entirely to LLMs? Seems to be more 'sensible use of technology' rather than 'planet burning IP theft machine'? Genuine question really.
Yeah, exactly this - pothole detection apps have been around for several years now, based on tried-and-tested machine learning. It's always been a problem with AI that it's a very vague term when you boil it down. At one end it's generative AI, at the other it's just basic off-the-shelf technology, so when companies or nations talk about an AI strategy it's not always clear which parts they mean. (Personally I'd rather see a compute strategy, but that's a whole other post.)
I assume they're both built around deep neural networks, and it wouldn't surprise me if some parts of the pothole-detectors use the same Transformer architecture (the T in "GPT") as LLMs - it was used as part of the pipeline that decoded the Herculaneum scrolls recently. You can't draw a hard-and-fast line between LLMs and other ML techniques.
The difference is more around the size and complexity of model required, it's the first L in LLM that tends to make current AI problematic.
Image classification models have been around for well over a decade at this point (Google famously got into hot water over their model back int 2015).
It appears the model is simple enough to be run on-device. A small fleet of (hopefully electric) vehicles with locally processed "AI" images probably isn't going to be classed as planet burning
I'm in two minds on this, because you're absolutely right that improved technology is moving things forward on potholes, but then the big problem is that councils aren't using the technology that already exists. For example, we have machines from companies JCB that can fix a pothole cheaply in a few minutes, where the councils that use them have seen insane leaps in the amount they can clear in a season, and other councils are just... still not using them, for boring reasons of culture and procurement. So from a government policy perspective the issue isn't 'moar AI' - the private sector is doing a great job on that already as you've pointed out. The issue is - fix council procurement to actually use this stuff. Which is the sort of boring thing that would probably 3x our economy but doesn't generate the same headlines as AI.
What about using data to get ahead of the problem, so rather than fixing potholes, using data on traffic patterns, groundwater, vehicle use & type and no doubt a bunch of other stuff I’m not thinking of to know when a patch of road is likely to start getting potholes and then re-lay the bitumen of that section before the holes start to form but no so early you’re wasting resources and you get actual smooth roads not the rutted road feel you get when a road has heaps of small pothole repairs
Maybe this is an issue that AI would be able to deal with, but I can imagine that the scale of the road network and the amount of data that would need to be collected for each section of the road (this second bit is more significant I can imagine) that it would not be feasible to scale it like that
I think James is gesturing towards this idea - the data collected from the automatic pothole cameras could be used to train predictive models of road decay.
You'd probably need those automated traffic counters they have already. Remember, the Fourth Power Rule applies - the (axle) weight of traffic is the crucial factor. Taxing vehicles on that could fund the entire system (using the 4th power rule).
I can't see a good reason to oppose AI being used for transport infrastructure - but I do think residents reporting potholes as well as local councillors, desperate for a picture for their newsletter next to fixed pothole to show how great they are, is a significant source of information - and potentially more useful than inspectors because they highlight the potholes that are actually causing local people distress.
The experience of this process is that local councils are becoming increasingly slow to fix these potholes that they definitely know about and that they know are causing real people real problems. I'm assuming part of this is cut backs and part of it is that the growth of car usage (and the increasing size of cars) has meant there are more pot holes than there were back in the day.
What AI will not do is increase the capacity to fix nor solve the problem that our roads are more prone to pot holes.
Personally I was more concerned about the talk of AI deregulation and opening up of NHS data to the private sector - but I think it's fair enough to be sceptical that AI will "solve" potholes, but then again I'm not sure Keir Starmer claimed it would.
coincidentally you've published this on the same day that a report has come out saying, among other things, that councils are doing the the minimum legally required in road repairs in order to save costs.
As someone who used to work in a County Council Highways and Planning department, I found this very interesting.
A few comments - if this had been done 10 years ago, I suspect AI wouldn't have been mentioned - it would have been called "automated image classification" or something similar.
On preventative maintenance, there is a lot of stuff in the operational research literature about optimum maintenance strategies, and I doubt if repairing every small defect wherever and whenever they appear is optimal.
Perhaps "AI" or whatever you want to call it can help in developing the maintenance and inspection programmes and schedules
Maybe the exciting thing is the collection of up-to-date street-level images, rather than the AI. There's lots of council use cases even beyond the road maintenance department. Maybe those CCTV mini cars which go looking for parking violations could be replaced by generic "photograph everything" vehicles. Maybe the planning department would like proof of exactly when an unauthorised building was put up. And beyond local councils, I'm imagining what would happen if they loaded all these images into a publicly published site like google streetview (or not like google streetview, because their data is not open licensed ...like Mapillary or others: https://wiki.openstreetmap.org/wiki/Street-level_imagery_services ). Personally I would be love to be able to use very up-to-date images of streets to make updates to OpenStreetMap data for example.
Is this an example of how the umbrella term 'AI' has been toxified by ChatGPT-style bullshit? I might be misunderstanding here, but is the pothole stuff with onboard chips of a different character entirely to LLMs? Seems to be more 'sensible use of technology' rather than 'planet burning IP theft machine'? Genuine question really.
Yeah, exactly this - pothole detection apps have been around for several years now, based on tried-and-tested machine learning. It's always been a problem with AI that it's a very vague term when you boil it down. At one end it's generative AI, at the other it's just basic off-the-shelf technology, so when companies or nations talk about an AI strategy it's not always clear which parts they mean. (Personally I'd rather see a compute strategy, but that's a whole other post.)
I assume they're both built around deep neural networks, and it wouldn't surprise me if some parts of the pothole-detectors use the same Transformer architecture (the T in "GPT") as LLMs - it was used as part of the pipeline that decoded the Herculaneum scrolls recently. You can't draw a hard-and-fast line between LLMs and other ML techniques.
The difference is more around the size and complexity of model required, it's the first L in LLM that tends to make current AI problematic.
Image classification models have been around for well over a decade at this point (Google famously got into hot water over their model back int 2015).
It appears the model is simple enough to be run on-device. A small fleet of (hopefully electric) vehicles with locally processed "AI" images probably isn't going to be classed as planet burning
I'm in two minds on this, because you're absolutely right that improved technology is moving things forward on potholes, but then the big problem is that councils aren't using the technology that already exists. For example, we have machines from companies JCB that can fix a pothole cheaply in a few minutes, where the councils that use them have seen insane leaps in the amount they can clear in a season, and other councils are just... still not using them, for boring reasons of culture and procurement. So from a government policy perspective the issue isn't 'moar AI' - the private sector is doing a great job on that already as you've pointed out. The issue is - fix council procurement to actually use this stuff. Which is the sort of boring thing that would probably 3x our economy but doesn't generate the same headlines as AI.
What about using data to get ahead of the problem, so rather than fixing potholes, using data on traffic patterns, groundwater, vehicle use & type and no doubt a bunch of other stuff I’m not thinking of to know when a patch of road is likely to start getting potholes and then re-lay the bitumen of that section before the holes start to form but no so early you’re wasting resources and you get actual smooth roads not the rutted road feel you get when a road has heaps of small pothole repairs
Maybe this is an issue that AI would be able to deal with, but I can imagine that the scale of the road network and the amount of data that would need to be collected for each section of the road (this second bit is more significant I can imagine) that it would not be feasible to scale it like that
I think James is gesturing towards this idea - the data collected from the automatic pothole cameras could be used to train predictive models of road decay.
You'd probably need those automated traffic counters they have already. Remember, the Fourth Power Rule applies - the (axle) weight of traffic is the crucial factor. Taxing vehicles on that could fund the entire system (using the 4th power rule).
I can't see a good reason to oppose AI being used for transport infrastructure - but I do think residents reporting potholes as well as local councillors, desperate for a picture for their newsletter next to fixed pothole to show how great they are, is a significant source of information - and potentially more useful than inspectors because they highlight the potholes that are actually causing local people distress.
The experience of this process is that local councils are becoming increasingly slow to fix these potholes that they definitely know about and that they know are causing real people real problems. I'm assuming part of this is cut backs and part of it is that the growth of car usage (and the increasing size of cars) has meant there are more pot holes than there were back in the day.
What AI will not do is increase the capacity to fix nor solve the problem that our roads are more prone to pot holes.
Personally I was more concerned about the talk of AI deregulation and opening up of NHS data to the private sector - but I think it's fair enough to be sceptical that AI will "solve" potholes, but then again I'm not sure Keir Starmer claimed it would.
coincidentally you've published this on the same day that a report has come out saying, among other things, that councils are doing the the minimum legally required in road repairs in order to save costs.
https://www.constructionenquirer.com/2025/01/17/state-of-englands-local-roads-a-national-embarrassment/