Four bets made from the start.
Yobi was founded on four hypotheses about what it takes to build the best models:
The best models require differentiated data
Public text has been consumed. Yobi trains on proprietary behavioral data secured through exclusive, multi-year partnerships: signal the open internet doesn’t hold.
The best models require real R&D
Not a fine-tune of someone else’s work. A homegrown research culture, building and training original architectures end to end.
The best models require serious infrastructure
Custom-built training and serving systems that make population-scale prediction economically viable, at millions of decisions per second.
The best models require no privacy compromise
Consumer consent and privacy are constraints the architecture was designed around, not policies added after.
The next action is not the next word.
Next-token prediction is a remarkable objective for language. It’s the wrong inductive bias for forecasting behavior. Predicting what a person will do is decision-making under uncertainty, and expected value isn’t encoded in the probability of the next word. So Yobi trained for a different objective: given a real sequence of behavior, predict the action that follows.
A vocabulary a thousand times larger
Language models work from a few hundred thousand base tokens. The vocabulary of human behavior, every product, place, publisher, and action, is orders of magnitude larger. A different problem needs a different model.
Trained on outcomes, not text
The model learns from discrete behavioral outcomes across modalities: ecommerce transactions, web and publisher browsing, physical card swipes. Language models play a supporting role in the stack; the model itself takes behavior in and puts prediction out.
Built for people who change
Behavior shifts daily, so the architecture is built to represent new people and new patterns as they appear, not just the world as it looked at training time.
One embedding space. Any behavioral question.
Yobi’s foundation model combines graph neural network and transformer architectures, trained on behavioral sequences at population scale, with hundreds of billions of parameters. What it produces is the core of everything Yobi offers: a single shared space where people, places, products, publishers, and concepts all live together.
Broadly predictive by design
The test of a foundation model is transfer: a campaign for a product nowhere near the training data still performs. Representations learned once carry to new tasks, new verticals, and new outcomes.
Graphs where language models have none
Behavior raises identity questions text never does: connecting anonymous signals in ways that respect privacy. Graph architectures handle what attention alone can’t.
Fine-tuning as the last mile
Pre-training is amortized across the full dataset; adapting to a new outcome takes a light, fast fine-tune. That’s what makes a custom model per campaign economically viable, and what makes the model a foundation rather than a one-off.
Millions of decisions per second.
The model doesn’t live in a lab. It runs live, evaluating context and predicting outcomes in the milliseconds an ad decision allows, at millions of queries per second.
Real-time by architecture
Representations are pre-computed and served from optimized lookup systems, so prediction fits inside real-time latency budgets without sacrificing model depth.
Learning in flight
Impression and conversion data flow back into the fine-tuned models throughout a campaign, so predictions sharpen while the media is still running.
No legacy heuristics
Yobi builds without decades of accumulated rules and workarounds. Where the industry maintains walls of heuristics, Yobi trains models against the actual objective.
Private by design.
The model is not anchored to personal identity. It’s anchored to behavioral sequences, and an anonymous session is as valid a training input as any other. What comes out is signal without the record: behaviors in, predictions out, no interpretable profile of any individual.
Consent, end to end
Data enters through consented, governed partnerships, and consent governs every attribute downstream.
An active research agenda
Differential privacy and privacy-preserving machine learning aren’t compliance checkboxes at Yobi; they’re a research area the team works in and intends to contribute to, because earning trust with sensitive data is the whole bet.
Minimum interpretable surface
The mapping into the embedding space is one-way by design. Maximum predictive signal, minimum exposure.
Built by people who publish in AI and Recommender Systems.
Selected works from the Yobi team.
Centaur — a language model tuned on 10M+ human choices across 160 experiments — predicts and simulates behavior in entirely new settings.
Lets models explore reasoning as a search tree — lifting GPT-4 on Game of 24 from 4% to 74%.
An open tournament on human economic choices — folding behavioral decision theories into ML features delivered the best predictions.
Embeds a network of users, tweets, and entities into representations that power personalized recommendation and ranking at scale.
Frames engagement prediction on a billion-row Twitter dataset as a fairness-constrained problem — the basis of the RecSys 2021 Challenge.
A node-classification method that holds up on both graph types — designed for billion-node real-world networks.
Defines a family of set functions uniting submodularity with deep architectures, and shows how to learn them from data.
A consented, self-labeled video dataset of 3,011 people for measuring AI fairness across age, gender, skin tone, and lighting.