Discussing AI Frontiers: Combat Harm at Scale on Tinder

As a CI senior AI engineer, I'll be discussing AI Frontiers and trust & safety, specifically combating multifaceted harm on Tinder at scale.

1. Speaker is VB, a senior AI engineer at Tinder, working on trust and safety for the past five years.
2. Trust and Safety (TNS) is about preventing risk, reducing risk, detecting harm, and mitigating harm to protect users and companies.
3. Tinder, as the largest dating app, encounters various types of violative behavior, such as social media in profile, hate speech, harassment, and scams.
4. Generative AI (gen) can cause problems for the trust and safety industry:
a. Enables rapid generation of content, spreading misinformation and low-quality spam
b. Lowers the bar for impersonation and catfishing, leading to malicious interpersonal harm
c. Scales up organized spam and scam operations
5. Existing signals used for detecting fraudulent activities will be less effective due to gen's ability to create text and images rapidly.
6. Opportunities with AI labs pre-training and open-sourcing large language models (LLMs):
a. Powerful latent semantic capability and global language coverage
b. Fine-tuning can achieve state-of-the-art performance in downstream textual detection tasks
c. Easier fine-tuning process with mature open-source libraries and tools
7. Generalization performance of LLMs slows the model degradation curve significantly, making it more difficult for bad actors to get around spam rules.
8. To use LLMs for TNS violation detection:
a. Create a high-quality training data set with hundreds to thousands of examples
b. Use GPT-type LLMs as text in, text out models with classification labels or extracted characters representing the violation
c. Assemble a data set manually, using large LLMs for data generation, or using hybrid processes for better alignment
9. Instead of directly using API LLMs like gbd4, fine-tuning your own models allows full control over model weights and refining when production performance degrades.
10. Hugging Face libraries make it easy to fine-tune models with just a few hundred lines of code without deep learning expertise.
11. Parameter-efficient fine-tuning methods, like low rank adaptation (Laura), are essential for rapid experimentation and inference optimizations.
12. Lorax, an open-source framework, efficiently serves thousands of fine-tuned models on a single GPU with marginal cost virtually zero for adding new adapters.
13. Using Laura allows Tinder to take advantage of massive inference optimizations and serve many different types of trust and safety violations on one or a few GPUs.
14. Lorax enables real-time inference at Tinder, supporting 7 billion predictions per second, tens of queries per second (QPS), and 100ish milliseconds of latency on A10 GPUs.
15. For high-frequency domains, gating requests with hero sticks can further reduce throughput.
16. LLM outputs are computationally expensive for generation but classification or extraction tasks require only a few tokens, leading to faster time-to-prediction.
17. Today's large language models offer massive improvements in precision and recall due to their higher latent semantic capability.
18. Adapter-based models get stale less quickly than traditional machine learning models, providing better defense against harm in the long run.
19. Future work includes exploring non-textual modalities, such as explicit image detection using pre-trained visual language models like lava.
20. Tinder aims to rapidly train adapters for detecting harm along the long tail of TNS violations and build a next-generation defensive mode against harm.
21. Using AI in the loop can automate training and retraining pipelines for fine-tuning adapters, taking advantage of Lorax's low marginal cost for inference.
22. The goal is to create a safer, healthier platform by leveraging gen's landscape today.

Source: AI Engineer via YouTube

❓ What do you think? What are your thoughts on the ideas shared in this video? Feel free to share your thoughts in the comments!