Loading…
Browse by type
Explainer reads
Explainer posts on Yomu, summarised newest first.
- 1,318 posts
- 20 companies
The Good Stuff
Paid picks we use ourselves. Links carry a referral code — a subscription through one earns us a commission at no extra cost to you.
Explainer posts
Architectural Choices in China's Open-Source AI Ecosystem: Building Beyond DeepSeek
Chinese open-source artificial intelligence development has shifted focus toward architectural choices, ecosystem toolchains, and domestic hardware integration following the January 2025 DeepSeek release. Leading groups have adopted Mixture-of-Experts architectures like Kimi K2 and Qwen3 to balance capability, dynamic compute allocation, and operating costs. Development expanded beyond text into multimodal domains including speech, video, and 3D systems while prioritizing smaller 0.5B to 30B distilled models under permissive Apache 2.0 licensing. Furthermore, model releases increasingly provide day-zero alignment with domestic inference and training infrastructure, including Huawei Ascend, Cambricon, and Baidu Kunlun chips. Production serving engines such as Moonshot AI's Mooncake and Baidu's FastDeploy 2.0 reflect a transition from standalone models to comprehensive system-level design.
The safety of speed: How we ship code 180 times per day
Intercom describes how it averages 180 production ships per workday in January 2026 while targeting 99.8+% availability, with an average 12-minute merge-to-production cycle. The post focuses on its Ruby on Rails monolith, deployed to three data-hosting regions through independent pipelines, and frames small batches as a way to limit accumulated risk. An automated pipeline builds a deployable slug, runs parallel CI, verifies it in pre-production with boot, CI, and Datadog Synthetics gates, then performs a staggered rollout with process-level rolling restarts. Safety also depends on engineer ownership, feature flags, production testing on a single host, outcome-focused heartbeat metrics, and rapid automatic or manual rollbacks. Incident reviews then harden the system, supporting the conclusion that speed is used to keep changes small, feedback immediate, and recovery fast rather than to trade stability for velocity.
Dropbox uses Cursor to index over 550,000 files and build an AI-native SDLC
Dropbox adopted Cursor to increase software-development velocity across a demanding environment: its own data centers handle more than 300,000 requests per second, while thousands of engineers maintain a monorepo exceeding 550,000 files. Engineers first experimented organically in 2024, then AI champions, easier sign-up, and a company-wide hackathon helped broaden adoption; CTO Ali Dasdan’s two-hour “smart finder” experience reinforced executive support. Cursor indexed the monorepo by scanning non-ignored files, splitting code into structured chunks, and generating embeddings for a semantic index that supports code generation and editing. More than 90 percent of engineers now use AI tools weekly, with Cursor appearing across coding, review, testing, documentation, and migrations. Dropbox reports that PR throughput and cycle time have moved into the upper tier of industry benchmarks, while describing an ongoing redesign of its software-development lifecycle.
SkyPilot at Shopify: Multi-cloud GPUs without the pain
Shopify uses SkyPilot to run machine-learning training workloads across GPU clusters on Nebius and GCP, addressing scarce hardware, fragmented capacity, and provider-specific interfaces. Persistent Kubernetes clusters remain managed by Shopify, while SkyPilot launches jobs and a company plugin routes requests, validates labels, injects configuration, and applies organizational policies. H200 workloads go to Nebius for InfiniBand-equipped distributed training, while L4 and CPU-only jobs go to GCP; labels also connect jobs to cost ownership, Kueue quotas, priority classes, and development-environment rules. The plugin automatically configures InfiniBand access, shared caches, and temporary storage, reducing manual setup and speeding repeated workloads. The resulting YAML-based interface preserves direct resource control while hiding cloud-selection decisions, giving teams common workflows as providers, prices, and GPU availability change.
First Block with Gerry Giacomán Colyer, Co-founder and CEO of Clara
First Block features Gerry Giacomán Colyer, Clara’s co-founder and CEO, discussing the company’s founding, rapid growth, and expansion of AI agents for finance operations across Latin America. Clara’s bank partner shut down only months after launch, prompting the team to secure direct issuer licenses with Visa and MasterCard in two weeks and build more robust infrastructure from the outset. The episode connects that resilience to product-led growth: a self-serve onboarding flow supported more than 20,000 customers across three countries, from pre-revenue startups to the Mexican Stock Exchange. The company is now applying its transaction data across Latin American businesses to a financial analyst agent that automates policy enforcement and approval workflows. It also covers the founding lesson from a previous company’s financial-tooling difficulties, early milestones, scaling decisions, and a roadmap across markets, alongside questions about efficiency, resilient infrastructure, and distinctive data advantages.
AI in Investment Management: 2026 Outlook (Part II)
Two Sigma leadership outlines the evolving role of artificial intelligence across quantitative investing workflows heading into 2026. The firm is embedding frontier large language models into internal systems, incident management, and feature generation pipelines to accelerate research tasks that previously took months into days. Technical focus across the broader field is shifting from raw parameter scaling toward efficiency optimizations, multimodal unified representations, and mechanistic interpretability circuits. In forecasting pipelines, rapid automated hypothesis generation introduces severe risks of overfitting and compromised backtesting, particularly when pre-trained models already contain historical regime knowledge prior to their cutoff dates. Consequently, engineering success requires strong institutional research discipline, production monitoring, and skepticism alongside the adoption of automated agentic tooling.
Claude's constitution
Claude’s constitution describes how explicit principles can guide a language model’s behavior instead of relying on values implicitly determined through large-scale human feedback. The approach uses AI feedback in two training phases: the model first critiques and revises responses using the principles, then reinforcement learning selects the more harmless output using AI-generated evaluations. The post says Constitutional AI produced a Pareto improvement in tests, with the model responding more appropriately to adversarial inputs while remaining helpful and not evasive; harmlessness results came without human harmlessness data. The constitution draws on the UN Declaration of Human Rights, safety practices, other AI-lab principles, global platform guidelines, and attempts to include non-Western perspectives, while remaining an iterative, adjustable document.
One Year Since the “DeepSeek Moment”
The January 2025 release of DeepSeek's R1 model transformed China's artificial intelligence sector from a closed ecosystem into an organic open-source ecosystem. By releasing under the MIT license and sharing reasoning paths and post-training methods, DeepSeek turned complex reasoning into a reusable, downloadable engineering asset suitable for distillation in compute-constrained environments. Chinese technology organizations subsequently shifted from comparing standalone benchmark scores to developing ecosystem interfaces, application scenarios, and system-level infrastructure. Major players rapidly expanded their open footprints, driving downloads of newly created Chinese models on Hugging Face past those of any other nation. This momentum catalyzed global open-source adoption across international markets while prompting Western organizations to develop competing open-weight alternatives or build directly upon open Chinese foundations.
Differential Transformer V2
Differential Transformer V2 introduces an architectural refinement to differential attention that pairs and subtracts query heads within the same grouped-query attention group. The architecture doubles query heads while preserving key-value head counts, which avoids increasing key-value cache demands and removes the need for custom kernels. DIFF V2 replaces the per-head RMSNorm of its predecessor with a per-token, per-head projected lambda passed through a sigmoid function. This formulation expands the context root-mean-square range to between zero and root two, preventing massive gradient magnifications, mitigating attention sinks, and stabilizing large learning rate pretraining. Ablations confirm that sharing keys and values across subtracted heads and constraining lambda scaling are necessary to prevent training instability and elevated loss.
How Etsy Uses LLMs to Improve Search Relevance
Etsy’s Search Relevance team addresses a limitation of engagement-based search signals: popular listings can attract clicks even when they are not the best match for a query. It introduces semantic relevance, measured through a framework combining human-curated golden labels, an LLM annotator, cascaded models, and production integrations, with categories for relevant, partially relevant, and irrelevant query-listing pairs. The data pipeline uses sampled search-log pairs, dual human annotation, an evolving guideline, few-shot chain-of-thought prompting with o3 in LangGraph, and self-consistency sampling; a Qwen 3 VL 4B teacher then labels data at scale before distilling into a BERT-based two-tower student. The student supports real-time use with less than 10ms additional latency, while the framework evaluates and enhances search relevance alongside engagement signals.
AI Trends to Watch In 2026
eBay’s leaders identify five AI-related trends they expect to shape commerce and engineering in 2026, building on the company’s 2025 work in search, recommendations, seller tools, and eBay Live. They predict AI will absorb complexity across translation, compliance, payments, logistics, and risk while enabling continuous verification and agentic commerce that can negotiate, verify provenance, and execute cross-border transactions. Other forecasts include using AI to surface pre-loved goods, identify recyclable materials, deliver company-specific context to models for developer productivity, and make shopping more emotionally engaging. The final theme is uncertainty: because generative models introduce unpredictable behavior, product teams will need guardrails and faster learning from observable failures rather than relying on fully predictable flows.
Building a better Bugbot
Bugbot is a code review agent built to identify logic bugs, performance issues, and security vulnerabilities in pull requests before production, as increasingly capable coding agents increased review workload. Its development moved from qualitative assessments to systematic experimentation, using an AI-driven resolution-rate metric, the BugBench benchmark, and online evaluations of real pull requests. An early production flow ran eight parallel passes with randomized diff ordering, majority voting, category filters, validation, and deduplication, while later versions adopted a fully agentic design with tools and dynamic context. Across 40 major experiments, resolution rate rose from 52% to over 70%, bugs flagged per run from 0.4 to 0.7, and resolved bugs per pull request from roughly 0.2 to about 0.5. Newer versions caught more bugs without a comparable rise in false positives, and planned capabilities include Autofix, self-verification, deep research, and continuous scanning.
Scaling long-running autonomous coding
The post reports experiments running coding agents autonomously for weeks, with hundreds concurrently contributing to projects that typically take human teams months. It describes initial dynamic coordination through shared files, locks, and optimistic concurrency control, then explains why flat peer structures caused bottlenecks, fragility, risk aversion, and stalled progress. The successful design separates recursive planners from task-focused workers, uses a judge between cycles, and lets workers push changes to a shared branch with minimal conflicts. In tests, agents built a browser with more than 1 million lines across 1,000 files, migrated Solid to React over three weeks with +266K/-193K edits, and delivered a 25x faster Rust video renderer; other projects remain underway. The authors say GPT-5.2 models perform better on extended autonomous work, role-specific models and carefully tuned prompts matter, and moderate structure outperforms both flat peer coordination and elaborate controls, though drift and scheduling remain unsolved.
The Design & Implementation of Sprites
Sprites are Linux virtual machines designed as fast, disposable computers with root access, persistent 100GB durable filesystems, automatic sleeping, and low inactive cost. To make creation feel instantaneous, the platform removes user-facing OCI container images and keeps pools of preloaded standard Fly Machines, so a new Sprite mainly performs startup work rather than pulling and unpacking an image. Its storage stack follows a hacked-up JuiceFS model, splitting immutable data chunks from metadata, keeping metadata durable with Litestream, and using a sparse NVMe volume as a dm-cache-like read-through cache for object storage; checkpoints and restores primarily shuffle metadata. This design supports migration and recovery without depending on local storage, while low-cost sleeping and object-backed disks suit interactive work more than warm, containerized production workloads, and billing applies only to storage blocks actually written.
AI in Investment Management: 2026 Outlook (Part I)
Senior leaders and researchers at Two Sigma analyze the expanding role of artificial intelligence across quantitative investment management for 2026. Rapid model improvements are inverting traditional quantitative research workflows by vastly expanding hypothesis generation and shifting operational bottlenecks toward rapid evaluation. Rather than relying on large language models to execute trades independently, firms are integrating agentic workflows as an underlying operating system across data pipelines and portfolio rooms. However, autonomous agents present alignment challenges because they optimize proxy objective functions relentlessly without intrinsic contextual awareness. Sustained success relies heavily on human supervision, rigorous safety monitoring, and disciplined research execution rather than mere compute scale or model complexity.
Building the Universal Commerce Protocol
Universal Commerce Protocol (UCP) is an open standard co-developed by Shopify and Google to let AI agents discover merchant capabilities, negotiate what both sides support, and transact with merchants. It addresses commerce complexity arising from cart-, buyer-, and market-dependent payment rules, discount combinations, and fulfillment variations. UCP uses layered extensibility: a shopping service supplies transaction primitives, capabilities such as Checkout and Orders are independently versioned, and extensions add domain-specific schemas without changing the core. Profiles and reverse-domain namespaces support decentralized capability discovery, while checkout states and continue_url enable structured handoff when human input or unsupported capabilities are required. Payment handlers are negotiated dynamically per transaction, and the public protocol is presented as ready for merchants, agents, and providers to build against.
Why team size no longer matters
Lean teams competing with larger companies cannot match their rivals’ hiring capacity or budgets, prompting a reassessment of success based on individual impact rather than headcount. Adding tools, integrations, and AI can create a “multi-tool trap,” scattering information and accelerating mistakes when infrastructure and direction are weak. The recommended response is to consolidate work and knowledge into connected systems, invest in documentation such as roadmaps, specifications, requirements, and strategic memos, and choose infrastructure improvements over reflexive hiring. These foundations give AI usable context, preserve decisions through organizational change, and help individuals operate as effective generalists; the conclusion is that scalable infrastructure can let one person do work that once required three, while making each future hire more impactful.
Lyft’s Feature Store: Architecture, Optimization, and Evolution
Lyft built a centralized Feature Store to optimize feature management, deployment, and data consistency for machine learning workflows across its rideshare platform. The system ingests batch features defined by Spark SQL and JSON metadata into Airflow DAGs, which simultaneously write data to Hive for model training and to an online serving layer for low-latency inference. Real-time streaming features are processed with Apache Flink and Apache Kafka before being ingested into the serving infrastructure. The online layer, dsfeatures, couples DynamoDB with a ValKey write-through LRU cache and OpenSearch to serve embeddings. Infrastructure optimizations reduced standard P95 read latency by a third while supporting over a trillion additional read and write operations.
Dynamic context discovery
Cursor describes dynamic context discovery, a coding-agent pattern that provides fewer details initially and lets the agent retrieve relevant information instead of injecting everything as static context. The approach uses files as a simple interface: long tool responses, chat history, Agent Skills, MCP tool descriptions, and integrated terminal output are stored or synchronized so the agent can inspect only what it needs with tools such as tail, grep, rg, and jq. For MCP, only tool names are included initially, while descriptions are synced into one folder per server; an A/B test found a 46.9% reduction in agent tokens for runs that called an MCP tool, with high variance based on the number of MCPs. Cursor says the design reduces context bloat, preserves information truncation could lose, supports recovery after summarization, and can expose MCP re-authentication status. The changes are planned to become available to all users in the coming weeks.
New research: Customer service team evolution
An analysis of 166 interviews examines how customer service roles, workflows, and team structures changed after adopting Fin or similar AI agents. Using interviews with Intercom customers or prospects, the researchers cleaned conversation transcripts with Python and coded them through the OpenAI API against a closed codebook. Workflow disruption was reported by 94.58% of participants, while 82.53% reported changed responsibilities; 27.71% reported headcount or hiring effects, and 6.02% reported changed team structures or reporting lines. The findings describe hybrid human–AI operations in which AI handles triage, routing, translation, repetitive responses, and some refunds, while people increasingly oversee quality, testing, metrics, and complex or sensitive interactions, although organizational strategies and required skills remain uneven.
Top topics
- Architecture830
- Performance812
- LLMs648
- Machine Learning617
- Open Source561
- AI Agents542
- AI454
- Developer Experience446
Weekly digest
The best engineering reads, once a week.