Sentient Foundation hires Mozilla AI veteran Julie Belião as VP of Product to build model provenance tools
Key Takeaways
- •Sentient Foundation appointed Julie Belião, the former VP of Product at Mozilla.ai and a Google engineer with more than a decade of multilingual AI work, as its Vice President of Product.
- •Belião's mandate covers recursive self-improvement and model provenance, with the goal of building products that prove what an AI model was trained on and where it originated, since Sentient says no standard proof currently exists.
- •At Mozilla.ai, Belião ran the open-source plan for 61 repositories and co-authored a 2025 paper with Mozilla and EleutherAI arguing that AI training data should be licensed, well documented, and of clearly established origin.
- •Sentient's October Fingerprinting Leaderboard attacked ten proposed fingerprinting methods and found nearly all could be defeated without degrading the model, including its own Perinucleus method, which had placed 24,576 fingerprints into Llama-3.1-8B.
- •The appointment comes amid reported extraction attempts: OpenAI disrupted a distillation campaign linked to people tied to Moonshot AI, while Anthropic traced more than 16 million Claude requests from three Chinese labs using about 24,000 fake accounts.
Sentient Foundation has appointed Julie Belião, former VP of Product at Mozilla.ai and a veteran Google engineer, as its Vice President of Product. Belião will lead the nonprofit's effort to turn its research on proving where AI models come from into software that developers can actually use.
What Belião will oversee
Her portfolio spans two areas: recursive self-improvement, meaning AI systems that improve themselves, and model provenance. The foundation says there is not yet a standard way to prove what an AI model was trained on or where it originated, and Belião's mandate is to build working products that can do so.
A decade of open-source and multilingual AI work
During her time at Mozilla.ai, Belião ran the open-source plan for 61 repositories and helped grow the organization's developer community. In 2025, she co-authored "Towards Best Practices for Open Datasets for LLM Training" with Mozilla and EleutherAI. The Mozilla Foundation lists her as a collaborator on the paper alongside EleutherAI executive director Stella Biderman and Mozilla's Mark Surman. The paper argued that AI training data should be licensed, well documented and have a clearly established origin. That same premise — that origin should be provable — is the question her new role now poses at the model level, where Sentient says no standard proof currently exists.
Before Mozilla, Belião spent more than ten years working on multilingual AI, contributing speech and language data work for Google Assistant, Google Translate and YouTube, all products of Google (NASDAQ: GOOGL). She holds a PhD in computer science from Université Paris Nanterre.
Why model provenance matters now
The appointment lands amid a string of reported attempts to extract proprietary model capabilities. In late September, OpenAI said it stopped a coordinated effort to pull the step-by-step reasoning out of its AI models, linking the campaign to people tied to Moonshot AI, the maker of Kimi. In February, Anthropic reported that three Chinese AI labs sent more than 16 million requests to its Claude model using about 24,000 fake accounts, and said the labs sought to extract Claude's abilities. Anthropic attributed more than 3.4 million of those exchanges to Moonshot.
The problem persists more broadly: when a model is copied, fine-tuned or distilled into another system, buyers often cannot tell what trained it.
Sentient took on the challenge in October with the launch of its Fingerprinting Leaderboard, an open test of whether methods for proving a model's origin can survive copying, fine-tuning and deliberate attempts to remove them. Sentient's researchers built attacks against ten proposed fingerprinting methods, and the results showed that nearly all could be defeated without degrading the model — including Sentient's own Perinucleus method, which had placed 24,576 fingerprints into Llama-3.1-8B.
"Fingerprinting puts the proof inside the model," explained Sachi Kamiya, the foundation's director of venture and growth. She compared the technique to the hidden marks Renaissance painters used to claim their work.
The research builds on Sentient's OML work, short for Open-access, Monetizable, and Loyal AI Model Serving, conducted with Princeton University and the University of Washington. The project aims to let creators share open models while still being able to prove ownership later.
About Sentient
Sentient launched in February as a Singapore-based nonprofit that pledges to keep artificial general intelligence open-source. In June, it committed $42 million in grants and investments to open-source builders. Its projects, including ROMA, OpenDeepSearch and its OML fingerprinting work, have earned more than 12,500 GitHub stars, and researchers at Google (NASDAQ: GOOGL), Microsoft (NASDAQ: MSFT) and Meta (NASDAQ: META) have cited them.
The appointment now sets up a research-to-product test with a benchmark of Sentient's own making: its October leaderboard found nearly every proposed fingerprinting method — including its own — could be defeated, and the foundation says no standard way yet exists to prove what a model was trained on or where it originated. Turning that research into working provenance products is the gap Belião has been hired to close.
Source: Cryptopolitan
