Cameras, clout and creepiness—photography in the era of gen AI

Greetings from the United Club near gate F9 at ORD. After 16 hours of travel I'm almost home from the Qualcomm Snapdragon Summit. I've got so much coverage to get through and so little brain capacity left to get me through that coverage, but you've got to keep pushing. This is maybe (definitely) a weird piece, but it's definitely (maybe) a weird subject so here we go...

Get ready for some big, big numbers. According to research published in June by Bloomberg Intelligence, the gen AI market “is poised to explode,” growing from a market size of around $40 billion in 2022 to $1.3 trillion in the next decade, a CAGR of 42%. The reckoning from Bloomberg Intelligence researchers is that the first growth wave will come from training infrastructure “gradually shifting to inference devices for [LLMs], digital ads, specialized software and services.” They also call out additional growth in the $280 billion range around “specialized assistants” supported by the public cloud.

Qualcomm’s focus certainly includes hybrid cloud but is more focused around on device gen AI tools, including various types of specialized assistants to borrow the phrasing. What that might look like is a gen AI agent that’s got domain specificity in travel planning, or an agent optimized to help you pick out a new car, or an agent that’s tailored to come up with an educational curriculum to help you add some formalized structure to your otherwise amateur study of economics. But enough about me…

The thing about all of that is the requirement for highly personal data to make it effective and to make the assistants get more effective over time. “We’re at an interesting point here…visioning for this future,” Qualcomm SVP Chris Patrick said during a group Q&A session at the event in Maui. “We definitely expect this intelligent AI agent, this AI assistant that works for you, to be a big part of your life moving forward…but there’s a lot of work yet to do.” Talking broadly about the concerns around managing permissions between data streams that feed into various AI agents, he said it all “needs to be carefully managed. The user always has to opt into these capabilities. Those are problems yet to be solved.” And Qualcomm will work internally and with partners to solve those problems and not rush things into market.

?But there is something of an apparent market tension there. OEMs want the silicon and software to put into their hardware to deliver these ideally valuable, differentiated applications to their customers. “We actually have to hold them back kind of because they want to launch these things extremely quickly,” Ziad Asghar, a Qualcomm senior vice president across Snapdragon technologies and roadmap, said in the same session. Device OEMs “want these capabilities today. We don’t need to incentivize them. They’re coming to us with asks, with requests, for more capabilities…The value of our solutions actually goes up because we’re able to offer experiences that others perhaps cannot.”

?Back to this idea of responsibly rolling out gen AI tools and broadly around incentivization, I’m reminded of the 2014 book Who Owns the Future by Jaron Lanier, a cerebral Silicon Valley guy who has had his hands in a lot of things over a lot of years. It just so happens I tucked that book into my suitcase when I was packing for the trip to Maui because I thought it might come in handy.

From a chapter called “Creepy” that’s not about gen AI, rather about pervasive machine vision systems continuously tracking peoples’ movements but still feels highly relevant to this discussion, Lanier writes: “Despite their real potential for harm, I remain of the opinion that these tools are just tools. There is nothing inherently evil in a machine vision algorithm. However, it is also inadequate to say that the only level on which to address the ethical use of tools is personal responsibility. No, what we have to look at is economic incentives…The only effective point to intervene, to fight creepiness, is in the fundamental economic model. If the economic model tends to bring out non-creepy developments, then only true creeps will want to be creepy…Society has always had to deal with those challenges. Legit companies and professionals should not be motivated to go creepy.”

Beyond assistants, another focus of what Qualcomm is doing with gen AI as it touches its mobile platforms is more feature-rich camera technologies, things like Video Object Eraser by Arcsoft, for instance. “There is a lot more to do with camera,” Asghar said. “It’s really a very broad space and we are not getting anywhere close to the full capabilities.” This is a tricky one for me because the pictures you’re taking with your smartphone aren’t really pictures in that they’re not one-for-one encapsulations of what the world looked like at a point in time. And we’ll continue to go down that route much, much faster with the ongoing development and commercialization of gen AI.

But Qualcomm is working on that. Baked into the Snapdragon 8 Gen 3 Mobile Platform is technology from a company called Truepic. “Truepic’s mission is to restore authenticity and transparency to digital content on the internet,” according to its website. The idea here is that Truepic photo capture aligns with C2PA, an open technical standard that essentially records layers of metadata to trace the origin and modification of media, to let someone know whether a photo is original or if it has been modified with gen AI. The characterization that came up in the Q&A session was of a sword and shield. “We want people to be able to fully use generative AI capabilities,” Asghar said, while enabling them to decipher the provenance of media. The goal is “to make sure that becomes the established norm.”

Maybe I’m overly cynical but to use that tech or something similar requires a degree of technical know-how that many of the people using gen AI to modify images perhaps don’t have, and perhaps don’t care to have. Again, maybe overly cynical, but I think there’s an argument that the collective mainstream aesthetic has changed. The goal isn’t to capture and share media that accurately portrays the world as it is; in fact, I see a preference in many individuals to capture media, modify it, and share something that portrays the world as they want other people to think it is for them. Gen AI is after all maybe the ultimate use case for data, and data is a proxy for the people that create it. And those people may prefer to see that data actualized through the filter of gen AI instead of through the often harsh lens of reality.