AICTO Pulse

Building a 30% Better AI: The Taste Graph Moat

VentureBeat (GamesBeat)
VentureBeat (GamesBeat)•May 27, 2026

Why It Matters

Pinterest shows that open‑source, fine‑tuned models can deliver superior accuracy at dramatically lower cost, offering a scalable blueprint for enterprises seeking AI advantage without vendor lock‑in.

Key Takeaways

  • •Pinterest fine‑tuned open‑source models to cut costs 90% and boost accuracy 30%
  • •Custom “Pin‑CLIP” embeddings power visual‑first search and lateral discovery
  • •Navigator 1, built on Quen 3VL, delivers a voice‑first shopping assistant
  • •Open‑source stack for core experiences; closed models for internal tools
  • •Continuous gold‑set evaluation and rapid A/B testing drive model improvements

Summary

Pinterest’s CTO Matt Madreal explains how the visual discovery platform has built a 30% more accurate, 90% cheaper AI stack by fine‑tuning open‑source foundations. The company leverages a custom "Pin‑CLIP" embedding layer that unifies image and text metadata, enabling semantic search, lateral exploration, and real‑time personalization for its 620 million monthly users.

Key to the strategy is replacing generic foundation models with tailored versions. By stripping the vision encoder from Quen 3VL and injecting Pinterest‑specific multimodal embeddings, the team reduced inference latency by a factor of twenty and achieved higher relevance scores. Navigator 1, the voice‑first shopping assistant, showcases this approach, combining open‑source LLMs with proprietary embeddings to deliver a visual‑first conversational experience.

Madreal highlights concrete results: the recommender model outperforms off‑the‑shelf alternatives by 30% in accuracy, while operating at a fraction of the cost. The team runs a gold‑set of precision‑recall evaluations and pushes rapid A/B experiments, measuring engagement metrics such as pin saves, board creations, and merchant clicks. Open‑source components power core user‑facing features, whereas closed‑source models remain limited to internal productivity tools.

The broader implication is a proof point that large enterprises can achieve frontier‑level AI performance without relying on costly proprietary APIs. Pinterest’s model demonstrates that domain‑specific data, when combined with open‑source fine‑tuning, can create a sustainable competitive moat and accelerate time‑to‑market for AI‑driven products.

Original Description

Pinterest's open-source AI stack costs 90% less than frontier models — and their custom-trained recommender outperforms off-the-shelf alternatives by 30% in accuracy. Pinterest CTO Matt Madrigal breaks down exactly how they did it, and what enterprise AI teams can actually replicate.
Madrigal walks through the full architecture behind Navigator 1, Pinterest's conversational shopping assistant built on Qwen 3 VL — and the specific decision to rip out its native vision encoder and replace it with PinCLIP, Pinterest's proprietary multimodal embedding layer. That swap alone closes a 20x inference latency gap and makes the economics work at 620 million monthly active users. This is the clearest public explanation yet of how a scaled platform operationalizes the "core vs. context" principle for model selection: open-source and custom-built where it touches the user, frontier models where speed-to-prototype matters more than cost.
The conversation also covers the Taste Graph — Pinterest's knowledge graph across hundreds of billions of pins and 15 billion boards — and how post-training on that proprietary data lets a smaller, fit-for-purpose model beat a larger frontier model on production metrics. Madrigal details their eval framework: gold set benchmarks, product-level evals tied to engagement and merchant click outcomes, and a structured A/B test pipeline that runs from engineer PRs through to live user signal.
On the organizational side: how Pinterest manages a "default yes" multi-IDE policy (Cursor, Windsurf, Claude Code, Codex) without collapsing security posture, how they segment sandbox environments between ML engineers with Taste Graph access and general application developers, and why Madrigal measures AI coding ROI in token usage and experimentation velocity — not lines of code.
🎙️ GUEST: Matt Madrigal | CTO, Pinterest
🎙️ HOSTS: Matt Marshall | VentureBeat, Sam Witteveen | VentureBeat
00:00 Show Intro and Guest
01:17 Open Source Cost Breakdown
02:20 Pinterest Multimodal Roots
02:37 PinClip and Embeddings
05:46 Core vs Context Models
07:43 Navigator 1 Assistant Stack
11:52 Benchmarking and Evals
13:29 Accuracy from Proprietary Data
17:16 Taste Graph Explained
18:29 Taste Graph in Training
22:22 Fighting AI Slop
25:16 Developer Tools and Velocity
27:57 Tool Choice and Governance
28:56 Security Sandboxes and CICD
30:57 Wrap Up
If you enjoy these conversations, you need to be in Menlo Park this July.
VB Transform 2026 is VentureBeat's flagship enterprise AI event, built entirely around one question: How do you orchestrate AI autonomy at scale? July 14–15, Hotel Nia. Real projects, proprietary research, no fluff.
50% off for listeners with code BEYONDTHEPILOT: https://bit.ly/4fK4F6z
Subscribe to VentureBeat: https://www.youtube.com/@VentureBeat
#EnterpriseAI #OpenSourceAI #AIInfrastructure #LLM #MachineLearning

Comments

Want to join the conversation?

Loading comments...