Seer starts with text, an important starting media type to begin researching how we will tackle the royalty engine. As the least expensive on compute and storage, it was easy to visualize the overall architecture, approach, and vision of Seer. Visual image data is going to be the biggest challenge, but once overcome will lead to many use cases that can help protect people from deepfakes and realities being designed by automatons rather than humans. Where being able to clearly identify truth in a visual, should be enough of a start to ground our perspectives behind a future we can trust.
The beginning of this technical journey started with an identified problem in 2024, during election season. The clear defamation of elected officials to reduce credibility and/or campaign influence via realistic generations that distastefully placed individuals in demeaning settings and situations. It may simply be an image to some, a joke to others, but do not underestimate the psychological impact it can have, especially for feminine leadership personalities. Visual imagery is powerful, social networks showed us the velocity of information travel enabled via short-form video and image feeds. They transfer a lot more data mentally than text and protecting reality in this space is paramount for a safer future, otherwise propaganda can be distributed to skew thought at a scale we are not truly prepared for. If we were, we would have already solved the issues social networks posed before the Generative A.I. boom, yet we still lack proper safety protocols on that front.
The first method of protection we want to explore with Seer, is generation prevention. Pro-active protection of individual likeness before they surface on social media or the internet. To do so, creating a database that has an effective registry of features that relate to the likeness of an individual is necessary. Generations from model providers will have to submit visual data to this endpoint and the server runs a check returning a response. In this response there is a simple signal that is produced to identify whether a generated photo contains an identified protected person or not. If so, the model provider should not be allowed to publish the content to public facing domains or store them in any shape or form. Initially, we want to provide the endpoints for this functionality as well as the hosted service to create a registry of protected persons, the enforcement requires legislation that takes time. But, over the years if a "Seer Certified" badge is able to communicate a new form of value for A.I. services, more users might feel safe using those services over time when generating media.
A team will be formed, over time, to collect all possible open-source models preparing LoRAs and similar methodologies, that apply forms of subtraction upon a model to produce focused results of reference visuals. In combination with a model provider's private in-house models we can promise to cover most of the ground of safety standards to reach something visible and tangible in the safety industry. Constantly keeping this updated to make sure we never miss a publicly released model that might harm public communication platforms. We are aware, that as foundation models improve in tandem, rogue independent projects can accelerate as well. Seeding adversarial communities will be important to maintain checks and balances of sorts. It's been done before and it can be done again, at scale, professionally, and monetized. Bad actors cannot be removed, but they can be outpaced.
The value of generative technologies is great, the speed at which it improves development, creative thought, and shipping products cannot be ignored. Being able to keep this pace while having a clear conscience is a variable of acceleration we are missing. Cars need to meet a safety standard to be released to the public, it is even a form of competition. Treating A.I. like cars means making sure our processes and "factories" are able to maintain a high standard. Making sure we are not over-stepping our boundaries upon other people's works, identities, and value while accelerating our own futures. An individualistic mantra only works when we are bringing others with us, while we achieve our goals. Defamation, harming competition (jokingly or targeted), or simply harming for the sake of it is a culture that is now of the past. A natural follow-up to this is allowing generations to surface for a royalty-tracked return, if the protected person or their protected works consent to such.
The candidate likeness registry and Seer are the first steps in setting this tone for our projects in the future.
P.S.
I haven't done this in a while... But, here is our song of the week.