Last week I discovered that GPT-6 Astra can read an analogue meter.
This may not sound particularly impressive. After all, humans learn to read analogue clocks as children. Looking at a dial, identifying where the needle is pointing and translating that position into a number feels like one of the simpler visual tasks imaginable. For AI, however, it has been surprisingly difficult, which makes the fact that it is starting to work rather more interesting than it first appears.
This may not sound particularly impressive. After all, humans learn to read analogue clocks as children. Looking at a dial, identifying where the needle is pointing and translating that position into a number feels like one of the simpler visual tasks imaginable. For AI, however, it has been surprisingly difficult, which makes the fact that it is starting to work rather more interesting than it first appears.
Seeing is not the same as reading
By 2024, multimodality had become an expected feature of leading AI models. Users increasingly assumed that systems such as ChatGPT, Claude and Gemini should be able to work not only with text, but also with images and, increasingly, audio and video.
The improvement has been dramatic. Models can identify objects in photographs, extract text from screenshots, interpret diagrams and work with charts embedded in PDF reports. Tasks that only a few years ago required specialised computer-vision software can now often be performed simply by uploading an image and asking a question.
But not all image understanding is equally difficult.
Consider an analogue clock. To answer the question "What is in this picture?", the model needs to recognise a clock face, some numbers and two or three hands. To answer "What time is it?", it has to do considerably more. It must distinguish the hour hand from the minute hand, determine their exact orientation, relate that orientation to the geometry of the clock face and then translate those spatial relationships into a symbolic answer such as 7:43.
Humans perform all of this almost without noticing. AI does not.
The improvement has been dramatic. Models can identify objects in photographs, extract text from screenshots, interpret diagrams and work with charts embedded in PDF reports. Tasks that only a few years ago required specialised computer-vision software can now often be performed simply by uploading an image and asking a question.
But not all image understanding is equally difficult.
Consider an analogue clock. To answer the question "What is in this picture?", the model needs to recognise a clock face, some numbers and two or three hands. To answer "What time is it?", it has to do considerably more. It must distinguish the hour hand from the minute hand, determine their exact orientation, relate that orientation to the geometry of the clock face and then translate those spatial relationships into a symbolic answer such as 7:43.
Humans perform all of this almost without noticing. AI does not.
Why geometry is difficult
A useful, although simplified, way of thinking about modern computer vision is that the model does not receive an image in quite the same way that we experience one. Many vision systems convert regions of an image into visual tokens: numerical representations that can then be processed alongside text. In early Vision Transformer architectures, this was quite literally done by dividing the image into a grid of small patches.
That approach works remarkably well for recognising what is present. There is a dog. There is a door. This is a table. That looks like a boiler.
An analogue dial asks a different type of question. The model does not merely need to know what the objects are; it needs to understand exactly where they are in relation to one another. Is the needle slightly before the 40 mark or slightly after it? Is the shorter hand pointing at seven, or has the model confused it with the longer hand? If the dial is photographed at an angle, what does that do to the apparent position of the needle?
This combination of visual recognition, geometry and reasoning remains difficult. Reading analogue clocks is still used as a test of multimodal models because even very capable systems can get a clock face completely wrong. It is a useful reminder that difficulty for humans and difficulty for AI are not the same thing.
That approach works remarkably well for recognising what is present. There is a dog. There is a door. This is a table. That looks like a boiler.
An analogue dial asks a different type of question. The model does not merely need to know what the objects are; it needs to understand exactly where they are in relation to one another. Is the needle slightly before the 40 mark or slightly after it? Is the shorter hand pointing at seven, or has the model confused it with the longer hand? If the dial is photographed at an angle, what does that do to the apparent position of the needle?
This combination of visual recognition, geometry and reasoning remains difficult. Reading analogue clocks is still used as a test of multimodal models because even very capable systems can get a clock face completely wrong. It is a useful reminder that difficulty for humans and difficulty for AI are not the same thing.
We see the same problem in charts
The same effect becomes more familiar when we move from clocks to material most of us actually use at work.
Give a current multimodal model a clean bar chart and it will usually understand it very well. It can identify the axes, compare the bars and describe the main conclusion. Make the visualisation more complicated and the situation changes.
A scatter plot with overlapping points, irregular clusters, shaded confidence areas and several trend lines requires the model to reason much more precisely about spatial relationships. It may understand the overall story of the chart while becoming considerably less reliable when asked for an exact value.
There is a big difference between asking "What does this chart show?" and "What is the approximate value of this particular point?" The first primarily requires semantic understanding. The second requires spatial precision. For a long time, this second category has been one of the weak spots of multimodal models.
Which brings me back to my meter.
Give a current multimodal model a clean bar chart and it will usually understand it very well. It can identify the axes, compare the bars and describe the main conclusion. Make the visualisation more complicated and the situation changes.
A scatter plot with overlapping points, irregular clusters, shaded confidence areas and several trend lines requires the model to reason much more precisely about spatial relationships. It may understand the overall story of the chart while becoming considerably less reliable when asked for an exact value.
There is a big difference between asking "What does this chart show?" and "What is the approximate value of this particular point?" The first primarily requires semantic understanding. The second requires spatial precision. For a long time, this second category has been one of the weak spots of multimodal models.
Which brings me back to my meter.
Why I care that Astra could read it
We have repeatedly used analogue meters in our teaching as an example of something AI finds unexpectedly difficult. So when Astra read one correctly, my first reaction was simply that another benchmark was beginning to fall.
My second reaction was more interesting.
There are an enormous number of things in the physical world that were designed to be read by humans rather than computers: meters, gauges, control panels, indicator lights, labels, display screens and equipment plates. Buildings contain plenty of them.
Traditionally, if we wanted to use information from these objects computationally, we first had to make the information digital. Connect the equipment to a building-management system. Install a sensor. Add a smart meter. Create an API. Enter the reading manually.
A multimodal AI system potentially introduces another route. If a human can look at something and obtain information from it, an AI with sufficiently good vision may increasingly be able to do the same.
That means the boundary of the data available to AI is no longer necessarily the boundary of our databases. It can potentially include parts of the physical world that were never digitised in the first place.
My second reaction was more interesting.
There are an enormous number of things in the physical world that were designed to be read by humans rather than computers: meters, gauges, control panels, indicator lights, labels, display screens and equipment plates. Buildings contain plenty of them.
Traditionally, if we wanted to use information from these objects computationally, we first had to make the information digital. Connect the equipment to a building-management system. Install a sensor. Add a smart meter. Create an API. Enter the reading manually.
A multimodal AI system potentially introduces another route. If a human can look at something and obtain information from it, an AI with sufficiently good vision may increasingly be able to do the same.
That means the boundary of the data available to AI is no longer necessarily the boundary of our databases. It can potentially include parts of the physical world that were never digitised in the first place.
But should we actually use it?
There are two fairly large qualifications.
The first is reliability. Astra reading my meter correctly does not mean that analogue-meter reading is a solved problem. One successful example tells us that the capability exists. It tells us very little about how consistently it works.
Change the lighting. Photograph the meter at an angle. Add reflections to the glass. Make the needle thinner. Put three similar gauges next to one another. Ask for readings repeatedly rather than once. The error rate may look very different.
A model being able to read a gauge is interesting. A model being able to read 10,000 gauges with a sufficiently low and measurable error rate to support an operational process is something else entirely.
The second question is easier to overlook: how much does it cost?
Suppose we have a building with 200 old analogue meters. We could send somebody around once a month to read them and enter the values into a spreadsheet. Or we could photograph every meter, send the images to a multimodal model, extract the readings, validate them, deal with uncertain cases and maintain whatever workflow sits around the model.
The second option sounds much more technologically sophisticated. It is not automatically the better one.
The AI approach has inference costs. The images have to be captured somehow. Somebody has to build and maintain the workflow. Errors need to be identified and corrected. And if a person is still walking around the building taking the photographs, we may simply have inserted an expensive AI step into a process that a human could have completed while standing in front of the meter.
It might be slower to have somebody read the meters manually. It might also be cheaper.
Of course, the economics change if the photographs can be captured automatically, if readings need to be taken very frequently, if thousands of assets are involved or if the same images can be used to extract several pieces of information at once. Then a capability that looks like a clever demo starts to become an operational opportunity.
For the last few years, a lot of our attention has gone into asking whether AI can do a particular task. More and more often, the answer is going to be yes.
The more useful questions will increasingly be how reliably it can do it, at what scale, what has to be built around it and how the total cost compares with the rather unexciting human process it is replacing.
Astra can read the meter.
It may still be cheaper to send someone to look at it.
The first is reliability. Astra reading my meter correctly does not mean that analogue-meter reading is a solved problem. One successful example tells us that the capability exists. It tells us very little about how consistently it works.
Change the lighting. Photograph the meter at an angle. Add reflections to the glass. Make the needle thinner. Put three similar gauges next to one another. Ask for readings repeatedly rather than once. The error rate may look very different.
A model being able to read a gauge is interesting. A model being able to read 10,000 gauges with a sufficiently low and measurable error rate to support an operational process is something else entirely.
The second question is easier to overlook: how much does it cost?
Suppose we have a building with 200 old analogue meters. We could send somebody around once a month to read them and enter the values into a spreadsheet. Or we could photograph every meter, send the images to a multimodal model, extract the readings, validate them, deal with uncertain cases and maintain whatever workflow sits around the model.
The second option sounds much more technologically sophisticated. It is not automatically the better one.
The AI approach has inference costs. The images have to be captured somehow. Somebody has to build and maintain the workflow. Errors need to be identified and corrected. And if a person is still walking around the building taking the photographs, we may simply have inserted an expensive AI step into a process that a human could have completed while standing in front of the meter.
It might be slower to have somebody read the meters manually. It might also be cheaper.
Of course, the economics change if the photographs can be captured automatically, if readings need to be taken very frequently, if thousands of assets are involved or if the same images can be used to extract several pieces of information at once. Then a capability that looks like a clever demo starts to become an operational opportunity.
For the last few years, a lot of our attention has gone into asking whether AI can do a particular task. More and more often, the answer is going to be yes.
The more useful questions will increasingly be how reliably it can do it, at what scale, what has to be built around it and how the total cost compares with the rather unexciting human process it is replacing.
Astra can read the meter.
It may still be cheaper to send someone to look at it.
Created with
