AI Summary
→ WHAT IT COVERS Shopify CTO Mikhail Parakhin details the company's AI adoption explosion in 2026, covering internal tooling including Tangle (ML experiment orchestration), Tangent (auto-research loops), and SimGym (customer behavior simulation), alongside infrastructure decisions around token budgets, PR review bottlenecks, and Liquid AI model deployment for sub-30ms search latency. → KEY INSIGHTS - **AI Adoption Phase Transition:** Shopify hit near 100% daily active AI tool usage across all employees, with a clear inflection point in December 2025 when model quality crossed a threshold. CLI-based tools like Claude Code and Codex are growing faster than IDE-integrated tools like Cursor and GitHub Copilot, signaling a shift toward non-visual, agent-driven workflows. Token consumption distribution is increasingly skewed toward top 10% users. - **Token Budget Strategy:** Rather than capping token usage, Shopify funds unlimited tokens per employee with a floor requirement — staff are discouraged from using anything below Opus 4.6 or GPT-5.4 Pro. The productive pattern is fewer agents in a structured critique loop (one agent generates, a different model critiques, the first revises), not many parallel agents. This increases latency but measurably improves output quality. - **PR Review as the Agentic Bottleneck:** AI-generated code volume has increased PR merge rates by 30% month-on-month, but CICD pipelines are creaking under the load. The key metric to track is the ratio of tokens spent during code generation versus expensive model tokens spent on PR review. Shopify uses pro-level models for review rather than lightweight tools, accepting longer review times to prevent downstream deployment rollbacks. - **Tangle and Auto-Research (Tangent):** Tangle is a content-hash-based ML experiment orchestration system where unchanged outputs are never recomputed, even across different teams — creating automatic cross-team efficiency gains. Tangent runs autonomous multi-experiment research loops on top of Tangle. Shopify improved search throughput from 800 QPS to 4,200 QPS on identical hardware using Tangent's automated code optimization loop, with no manual engineering intervention. - **SimGym Customer Simulation:** SimGym achieves 0.7+ correlation with real add-to-cart conversion events by training simulated buyer agents on decades of Shopify merchant behavioral data. New merchants without historical data get generic simulations; merchants with prior customer data get personalized agent distributions replicating their specific buyer mix. The system runs in headless browsers using multimodal models to capture visual layout effects that pure HTML analysis misses. - **Liquid AI for Production Inference:** Shopify runs Liquid AI models at 30ms end-to-end latency for real-time search query understanding using a 300-million-parameter model — a use case where standard transformer serving stacks and quantization tools like QServe don't perform adequately. For high-throughput offline tasks like product taxonomy classification, Shopify distills large frontier models into Liquid's architecture, where it consistently outperforms Qwen at equivalent parameter counts. → NOTABLE MOMENT Parakhin ran an auto-research experiment on a problem he considered fully optimized after years of manual tuning — expecting to prove the approach had limits. After 400 automated experiments over several weeks, one yielded a genuine improvement. He abandoned his skepticism and became an advocate, crediting Andrej Karpathy for popularizing the method. 💼 SPONSORS None detected 🏷️ AI Tooling Adoption, ML Experiment Orchestration, Agentic Code Review, Customer Behavior Simulation, Liquid Neural Networks, Auto-Research Loops