Vered Horesh

The through-line

Horesh's consistent argument is that generative AI only becomes usable by enterprises once the questions of ownership, consent and compensation are settled at the source. The barriers are not model quality but "responsibility, privacy, inclusion, biases" [1], and Bria AI is framed as a platform for developers building for enterprises so that those enterprises "can actually unlock the great premise of GenAI" [1]. The position is presented as one held from inception rather than adopted under pressure: four and a half years ago, explaining that images could be generated "out of thin air" was hard enough, and adding that it would be done responsibly, "respecting copyrights, privacy, safety", required real commitment [1]. The shift Horesh describes is external rather than internal, with the market and regulators moving toward a stance the company already occupied: "the market is really catching up, I would say, with our initial vision" [1].

On compensating every participant in the AI supply chain

The clearest statement of principle is structural. AI, in Horesh's account, "is built on three pillars, right? It's the compute, it's the AI expertise, and the content" [1], and from that follows a conclusion treated as obvious rather than novel: "You have to compensate all the participants in this joint endeavor" [1]. This is why licensing sits at the base of the product. Models are trained only on data licensed from data partners, with Getty Images, Envato, Alamy and Deposit photos among nearly 20 partners and "almost 1 billion images under contract explicitly licensed for GenAI training" [1].

The compensation claim is made operational through attribution. Horesh describes technology covered by five patents that can attribute each synthetic image generated by any of the models back to the original data owners in the training set "that impacted the most, this specific generation" [1]. The point is that tracing is not the endpoint: "we don't stop there, we also allocate the revenue that is coming toward this specific synthetic generation with the original images" [1]. Horesh treats this lineage as generative of new possibilities rather than merely defensive, a way of making the economics of training data legible.

On regulation as inevitable rather than adversarial

Regulation is described as something that "was inevitable" given how the technology is built [1], and the framing throughout is that a company already licensing its data and attributing its outputs has little to adjust. The same respect extended to partner data is extended to customers: proprietary data used by customers to inform or fine-tune models is held to the same standard of IP and privacy protection [1].

On the enterprise stack and the creative use case

The product focus is the creative space: image generation and image modification, with a first text to motion model imminent and a first video model expected within the year [1]. On infrastructure, Horesh describes Bria as "SageMaker native" on the grounds that "all our models are trained to date on SageMaker" [1], calling it a powerful environment for the company and a good one for customers to take their models into, which is the reasoning behind launching on Jumpstart [1]. The logic is continuity with where data scientists already work, so that customers can bring their own data and customise on familiar ground, with SageMaker described as "synergetic to all of what we offer" [1].

Takeaways

• The obstacles to enterprise adoption of generative AI are responsibility, privacy, inclusion and bias, not capability [1].
• AI rests on three pillars, compute, AI expertise and content, and all three sets of contributors must be compensated [1].
• Bria AI trains only on licensed data, with roughly 20 partners including Getty Images, Envato, Alamy and Deposit photos and close to 1 billion images contracted explicitly for GenAI training [1].
• Patented attribution technology, covered by five patents, traces each generated image to the training data owners that most influenced it, and revenue from that generation is allocated back to them [1].
• Regulation was always inevitable, and the market is now catching up to a responsible-by-design position held since the company's inception [1].
• The roadmap runs from image generation and modification to a first text to motion model and then a first video model [1].
• Being "SageMaker native" is both a training choice and a customer choice, extended by listing on Jumpstart so customers can work with their own data in a familiar environment [1].