Craig received a rumor that Canon is considering adding an AI subscription to our future Canon cameras. The thought of this stirs mental images of people heading to Canon HQ with pitchforks.
Maybe not so fast with the fire though. But do keep it handy if we disagree.
Our cameras in the past even in the DSLR age used machine learning for object recognition and focus. So some form of AI has been with us for some time, but the larger models in use today, are far more complex than that, and this is what we will be discussing in this article. I assume that the smaller machine learning models used for object recognition, focus tracking, etc will remain free to all users – unless it's a special use case.
No one wants their camera locked up and unable to be used unless they have an active subscription, this is something we can all agree on and something we should demand. With pitchforks and torches if necessary. So I will assume that our cameras will work fine without any generative AI processing subscription, but some features will be disabled. Samsung is already going down this route with AI added to their smartphones, stating that the AI features will have a subscription price shortly.
There are several types of AI. Generative AI loosely covers a bunch of models but interprets our input prompts and generates images, video, or text. Usually, these are large language models that interpret our text commands and process them accordingly. But there are also large vision models, that can process and interpret visual data, including live video, or images.
In this article, we are discussing large vision models.
An adage is that there is no such thing as a free lunch, and I think stepping back we have to realize some of the costs associated with AI and these large vision models.
The Upfront Costs
Before generative AI begins the model has to be created, and the large vision model (LVM) has to be prepared for training. This requires substantial upfront computing power by AI processors (GPU processors these days). The name of that game is Nvidia, who are selling their units for whatever they wish to because people are desperate to buy them at any price.
After the large vision model is developed, there needs to be extensive data available to the model for training. In the case of what we care about, that’s images and many of them. Most large image stores are now charging to use their images for training, and there’s a lot of legality surrounding using images for training without direct permission from the photographers.
But let’s assume we get through all that, then after all this is done, we have an LVM AI that we can use as consumers.
The Operational Costs
Then we get into the costs of operation, which no matter what, need to be paid for.
Usually in a data center, “data” or hard disk / solid state drives consume a vast amount of rack space, and those components are atypically low on the power usage scale. But this isn’t true when you have blade after blade of nothing but AI processors consuming far greater power – all the time. On top of the power, is the fact that the heat generated by these AI processors has to go somewhere. Fans and overall data center cooling can be quite expensive in terms of power usage.
As an example, a modern rack nowadays will usually consume around 10-30kw of power, whereas a high-capacity AI rack can consume well above that and is expected to be around 50-100kw of power.
Just to put this into perspective the average American home uses 11kw, so it’s entirely possible that in the near future a single AI rack in a data center will reach the equivalent of 10 homes worth of power. A typical data center can have hundreds of racks, reaching essentially a small suburb’s worth of power. The cost of powering such a data center could easily reach into the 10’s of millions of dollars per year in electrical and maintenance costs.
On top of the power consumption, other resources such as water may also be required if the servers or GPUs are water-cooled.
So, when we take this into account, and even ignoring the upfront hardware costs vendors have to put out to start up AI, it’s no wonder that companies are looking at this and going, yes, someone has to pay for it over time.
In terms of selling hardware that uses AI, if there is no subscription you have to judge how long the person could use that hardware, and incorporate that into the actual costs of the hardware – dramatically increasing the upfront costs to use AI.
So no matter how distasteful I find the thought of yet another subscription, depending on what the AI does this could be a worthwhile purchase on a monthly or annual basis.
AI for our mirrorless cameras
First, it’s going to come. It’s simply a matter of time. We have already quietly started down that path by using machine learning for our auto-focus systems.
I imagine that when this gets implemented, it will most definitely be in the cloud and not actually in the cameras themselves. There seems to be a large amount of resistance to putting the latest generation processors and hardware into cameras, not to mention memory, etc. Without that, AI simply won’t work well inside our cameras.
But the camera companies could get creative here. Large model AI may force camera companies to join the 2020s and put 5G into the cameras directly or via an accessory such as a grip. Maybe a grip is a good idea, as to have a generative image and auto-focus AI compute engine contained in an addon grip with a fixed price and subscription model. The processing would happen both online for smaller tasks and in the cloud for larger processing tasks.
I suspect when we start to see this, it will be something similar to what other AI imaging processing ventures have been (thinking Samsung here) – which includes post-processing, etc.
I’m still hopeful that whatever Canon and other camera companies come up with for generative AI in our cameras will be an optional service and not required. And that it will augment our cameras and imagine processing pipelines. I don’t think they have much choice, what we are seeing come out as first-generation image-processing AI on smartphones dramatically changes what cameras can do, if the interchangeable lens camera systems don’t adapt, they will certainly be left behind.
The elephant in the room, and one I think deserves its article is that at the same time, there has to be image and video authentication done at an industry level as well to ensure that images and video that we see are real images and video. We also need image and video auditing so that we know what has been modified from the sources. We need a method of determining what is real, what has been changed, and even more scary, what is entirely generated via AI.
To see what AI thinks about this I asked. Of course, the first response from Copilot was;
Am I not valuable hooman?
And suddenly Edge was acting a little odd.
Asking Chat-GPT this time and got a bit better result
Generative AI costs are influenced by model complexity, data, infrastructure, talent, and environmental factors. A subscription model helps manage these costs effectively while providing continuous access to powerful AI capabilities.
I have to fix my Edge browser now, so this seems like a good summary and an excellent time to end this article.
Richard's Note: I do want to make sure that people understand that AI is just a catchall buzzword for everything that involves machine learning. We are not talking about discrete machine learning such as what we have seen in AF systems for the last decade, or even improvements on those AF systems for skeletal anaylsis, or animal recognition. Generative AI and large vision models are far more complex and can't be placed easily inside a camera. These usually require large processing and data centers, but may move into our devices 5, or 10 years down the road with advances in AI processing chips, etc.
Header Image Source: AI (what else?)
Image credit to the Simpsons who in 20 years have predicted too much
I read yesterday that there are rumors that Microsoft is planning to build a data center that need power from no less than 3 nuclear power plants :oops:
Yes, for A.I.
it's a lot for sure. a company that i work for has patents on tackling the issue. However, it's one thing to invent it.. it's another to convince Microsoft and others that this is a viable technology.
the waste heat from AI is going to be absolutely bonkers.
that's always my hope on opinion pieces ;) it makes my cold dead heart warm again.
Yes, it takes a lot of money to build and train models, but it doesn’t necessarily take as much power as folks would think to RUN models. Various image recognition models run quite well on economy hardware like Raspberry Pis. I’ve also got an LLM running locally on my MacBook Pro, and it performs quite well without consuming too much battery power.
If Canon turns “AI” autofocus into a subscription, it will be out of greed and not necessity. Unfortunately, it seems that major manufacturers are hellbent on eroding ownership of products you “buy.”
Edited: For grammar, because I can't proofread.
My hot take....
I look at Atomos, and when I bought my Ninja V a few years ago. I was quite ignorant about recorders. You hit the Atomos Dashboard, check for the firmware updates and whatever.. and then I discovered there were a bunch of pay to play feature options. Initial reaction? You can imagine :cool:
However, a few of them I didn't need, want or honestly, even know what the hell they were.. Then the H.265 came and that was something I did want. I was quite happy to pay for it. They had spent time and resources adding a feature that probably came from customer feedback. I don't believe it's a super popular codec, but I actually I like it.
I checked for firmware updates a couple of weeks ago and the OS version 11 had hit, it's a paid upgrade. I looked at the features added and I didn't need any of them. They'll continue with any bug fixes on OS version 10. I could go buy the latest Ninja for OS11, There's just nothing at this time I need/want from the latest line. I think a Shogun is for me in the future anyway.... aging eyes!
The biggest challenge for companies will probably be releasing a fully featured product and no perceived "cripple hammer", people won't respond well to removing things and putting them behind a paywall.. New products still need to be innovative and push the envelope on features and capabilities out of the box.
Take the R3, fully featured AF system. If they came up with a way to say target sports photographer with "AI" updates, based on feedback and provide a feature set that is fully developed, I think that would be worth the price of admission if it's something you want. The wedding photographer won't care and won't buy it.
There can also be a lot of external influences that change the viability of a product.. trends, other new tech.. etc. Being able to adapt and provide a way for an existing product to meet these news demands.. that's great, but it will have to come at a cost, but that cost will be significantly less than buying a new camera.
However, if they tried to charge for say that pixel shift 400mp thing, that would not be cool. I don't think it was very well done, but there could have obviously been technical limitations with the hardware that prevented it from being better. If they go this route, they will have to develop hardware with the future in mind.
I also see a potential for it to help brand retention and to grab some revenue from people that use their cameras until they die. Product cycles are a lot longer now. Sure we're all camera dorks here, but I don't believe we're the majority.
This sort of thing could also help when a competitor launches something cool while your competing product is in the middle of its life cycle. If it's a software trick or whatever... well... here you go.. $50 and now your camera has it if you want it. I do think hardware evolution is going to move slower than software over the next 5 or whatever years.
Keep firmware updates going for minor features, bug fixes, compatibility. Put more resources into adding cool features. That comes at a cost for the company, and you should pay for it if you want it, but it has to be done well. How many times have you looked at a firmware update and installed a bunch of features you don't want, need, or care about? You install it for the bug fixes and those sorts of things. I don't think it's unreasonable to separate those things. It could also make those camera breaking firmware updates less of a thing?
It could also help Canon become better at software, there have been *cough* a few misses in the past.
Being verbose aside, if it adds value to a specific demographic of shooter and doesn't hinder the other demographics feature wise or cost wise. I think that could be a good thing. There is nothing wrong with finding new streams of revenue in a shrunken market if it's done right and makes your product better and your customers happier.
I don't believe this is coming any time soon, but I wouldn't be surprised if the new product lines are going to be capable of these sorts of future developments.
I will add.. no monthly BS... I think a lot of us are tired of that model.
Isn't that something they do in Denmark or somewhere in that part of the world?
First crypto/nft demands.. I hadn't even thought about AI data centres.
Down the rabbit hole I go.
A need to rely on an Internet connection in order to make a shot is, of course, a no-go.
"Generative AI" inside the camera? A bigger screen would be better to watch out for those six-finger hands.
As someone who was hit by the ‘lightroom needs to refresh its token every 2 weeks or it stops working’ problem on vacation, I’m very sceptic about subscriptions for my camera.
I guess we’ll finally get internet time sync as a side effect
The workflows are also clunky with modern cameras, e.g. for HDR, and very computing intensive. I can't see how cameras can get out of the slump they are in right now. Getting a photo out of my cameras take so much more time and effort compared to just shooting with my phone, and with modern social media, any extra details the camera can produce is simply destroyed by low resolution uploads/compression. Sure - you can always have your own portfolio website, but the number of people you can reach will be greatly smaller just by default.
Even a basic connection to the camera from the phone app sometimes doesn't work for no reason, if they can't get that right, how can they get the other things which would naturally rely on these steps? In many ways the camera industry is way behind and the gap is only widening as far as I can see.
However, subscriptions for features that require remote processing in some far-away datacenter seem all but inevitable. DCs cost money. Electricity costs money. These are ongoing costs that require ongoing payments. I'm not sure how that would end up being implemented as in many cases (and many places) there won't be the connectivity available that such features would require.
My ultimate hope is that AI quickly evolves and we all end up in a post-scarcity world. If done properly, would be amazing.
:mad::mad::mad: