---
title: "Latest reads"
description: "The engineering internet, summarised so you can actually read it."
---

# Latest reads
> The engineering internet, summarised so you can actually read it.

## Articles

### [The agentic marketing stack starts with the data layer](https://yomu.fyi/post/the-agentic-marketing-stack-starts-with-the-data-layer.md)
- Company: [Databricks](https://yomu.fyi/company/databricks.md)
- Author: Aly McGue
- Published: Jul 10, 2026
- Type: Explainer

The piece argues that agentic marketing depends on modernizing the data and infrastructure foundation first, rather than starting with isolated AI use cases. At Acxiom, moving core products and client solutions from on-premises Hadoop to Databricks reportedly improved workload runtimes by 80 to 90 percent, reducing jobs that took 50+ hours or as long as 90+ hours to 2–3 hours. The migration also reduced manual pipeline and infrastructure work, freeing engineering capacity for products and client outcomes. Those capabilities now support agentic audience planning, media buying, campaign activation, performance analytics, code generation, testing, and ad variation, with a goal of connecting the full marketing value chain. Because workflows handle PII, generated content passes through legal approval, defined controls, and human oversight; Acxiom is also embedding its data in client platforms and privacy-safe clean rooms to make decisions more transparent and native to customer environments.


### [Reducing High-Bandwidth Memory Bottlenecks in JAX-Based LLM Training with Host Offloading](https://yomu.fyi/post/reducing-high-bandwidth-memory-bottlenecks-in-jax-based-llm-training-w.md)
- Company: [NVIDIA](https://yomu.fyi/company/nvidia-developer-blog.md)
- Author: Tanya Lenz
- Published: Jul 10, 2026
- Type: Benchmark

Large language model training workloads frequently face GPU high-bandwidth memory (HBM) capacity constraints before compute limits are reached. To alleviate this memory pressure in JAX training workloads, host offloading transfers selected forward-pass activations to pinned host memory over high-speed NVLink-C2C interconnects and reloads them during the backward pass instead of recomputing them. Benchmarks executed on 128 GPUs within NVIDIA GB200 NVL72 systems using MaxText demonstrated significant improvements. For DeepSeek-V3 671B, host offloading combined with the Latency Hiding Scheduler and pipelined offloading reached 908.2 TFLOPs/s/device, surpassing activation rematerialization throughput by 57% and unlocking larger batch configurations. Similarly, QKV activation offloading on Llama 3.1 405B achieved a 2.9% throughput increase when transfer latency was hidden behind computation.


### [Amazon and University of Michigan give robots a sense of touch](https://yomu.fyi/post/amazon-and-university-of-michigan-give-robots-a-sense-of-touch.md)
- Company: amazon.science
- Author: Mani Nambi, Nima Fazeli
- Published: Jul 10, 2026
- Type: Problem & solution

Robotic manipulation in contact-rich settings requires accurate tactile force perception, but existing physics simulators either run too slowly for reinforcement learning or oversimplify shear forces. HydroShear addresses this tactile reality gap by incorporating path-dependent force tracking into hydroelastic contact models. By remembering object motion history across soft elastomer sensors, the GPU-parallelizable simulator computes full 3-D contact deformations, friction, and slipping dynamics. Reinforcement learning policies trained entirely in HydroShear transferred directly to a real Franka robot equipped with GelSight Mini sensors without fine-tuning. Across four challenging manipulation tasks, HydroShear achieved a 93 percent average success rate, outperforming baseline models TacSL and FOTS.


### [Introducing Feature Views](https://yomu.fyi/post/introducing-feature-views.md)
- Company: [Databricks](https://yomu.fyi/company/databricks.md)
- Author: Nick Joung, Ian Ackerman, Julia Powell
- Published: Jul 10, 2026
- Type: Announcement

Feature Views are presented as a framework for creating managed feature pipelines in Databricks, addressing duplicated logic, training/serving skew, backfills, online-store plumbing, and governance for real-time ML. A Feature View defines the source, entity, time-series column, and computation once; Databricks uses it to generate point-in-time-accurate historical data, then materializes pipelines for batch or streaming production use. Registered in Unity Catalog, materialized features receive lineage and access controls, while MLflow records model dependencies and Model Serving retrieves required features without custom lookup code. For Kafka-sourced streaming features, Spark Realtime Mode, Lakebase, and Model Serving support rolling-window updates with reported end-to-end p99 latency of 200ms from event to online availability. The announcement says Feature Views are in Public Preview and can be developed through the Feature Engineering Client SDK and Genie Code, with streaming materialization requiring an Enterprise-tier workspace in a region supporting Lakebase.


### [Kernel Fusion in NVIDIA CUDA: Optimizing Memory Traffic and Launch Overhead](https://yomu.fyi/post/kernel-fusion-in-nvidia-cuda-optimizing-memory-traffic-and-launch-over.md)
- Company: [NVIDIA](https://yomu.fyi/company/nvidia-developer-blog.md)
- Author: Michelle Horton
- Published: Jul 10, 2026
- Type: Explainer

GPU compute often outpaces memory bandwidth, causing separate kernels to suffer from intermediate global memory round-trips and launch overhead. Kernel fusion resolves this by executing multiple operations within a single kernel to keep intermediate data inside registers. Using the sum(abs(x)) operation as a benchmark on an NVIDIA GeForce RTX 4090, a naive two-kernel setup moved 3 GB of global memory over 3.51 ms. In contrast, manual CUDA C++ fusion, PyTorch compiler fusion via torch.compile, and explicit Python fusion using cuda.compute all achieved a 3x speedup. Across all three approaches, global memory transfers dropped to 1 GB by computing absolute values inline during reduction.


### [Inspiration – Tools & craft – Notion](https://yomu.fyi/post/inspiration-tools-craft-notion.md)
- Company: [Notion](https://yomu.fyi/company/notion.md)
- Published: Jul 10, 2026
- Type: Explainer

Notion’s Inspiration page gathers perspectives on the future of work from people and teams creating companies, communities, and creative practices. Its TechFirst Block pairs Ivan Zhao’s claim that AI has arrived as “infinite minds” with stories about organizations using Notion to change how work gets done. Featured examples include Brainlabs’ company-wide AI operating layer, Kim Shui’s Notion-organized New York Fashion Week preparations, and Her Tech Circle’s reconstruction of a community around a shared workspace. Additional profiles cover Thred’s effort to make offices less lonely, Sarah Jutras’s work teaching self-employment through Freelance for Life, and how leading AI companies implement AI tools effectively; the page also identifies contributors such as founders, marketers, designers, and community leaders.


### [AI Model Co-Design: Hardware-Friendly LLM Design](https://yomu.fyi/post/ai-model-co-design-hardware-friendly-llm-design.md)
- Company: [NVIDIA](https://yomu.fyi/company/nvidia-developer-blog.md)
- Author: Elizabeth Goodman
- Published: Jul 10, 2026
- Type: Explainer

Efficient large language model serving requires co-designing architectures to balance datacenter token throughput and user interactivity within hardware constraints. Roofline analysis shows that linear layers become memory-bound when hidden or intermediate projection dimensions are small, leaving arithmetic hardware underutilized even at large token counts. For prefill workloads, Chunked Pipeline Parallelism reduces first-token latency on long contexts when models use regular, repeatable layer patterns that split into balanced stages. For latency-critical decoding, decoupling attention and feed-forward network parallelization enables independent optimization, applying expert parallelism for feed-forward networks and Helix parallelism to shard the key-value cache across sequences. Following architectural design rules—such as dimension alignment, width prioritization, and low-precision NVFP4 compatibility—allows models to achieve higher utilization on modern accelerator hardware.


### [Better tools made Copilot code review worse. Here’s how we actually improved it.](https://yomu.fyi/post/better-tools-made-copilot-code-review-worse-here-s-how-we-actually-imp.md)
- Company: [Github](https://yomu.fyi/company/github.md)
- Author: Napalys Klicius
- Published: Jul 10, 2026
- Type: Problem & solution

When migrating GitHub Copilot code review from custom navigation tools to shared Unix-style CLI utilities—grep, glob, and view—benchmarks revealed higher review costs and fewer caught issues. Rather than an issue with the underlying tools, trace analysis showed the agent used general-purpose coding assistant instructions that triggered expansive repository browsing loops. In response, the team rewrote the tool guidance to enforce a review-specific workflow anchored to pull request diffs. The updated instructions direct the agent to narrow candidate call sites and files using batched discovery before reading minimal line ranges with view. This workflow tuning reduced average review costs by approximately 20% in production while maintaining review quality.


### [ClickHouse on Docker Hardened Images](https://yomu.fyi/post/clickhouse-on-docker-hardened-images.md)
- Company: [Clickhouse](https://yomu.fyi/company/clickhouse.md)
- Published: Jul 10, 2026
- Type: Announcement

ClickHouse is now available as Docker Hardened Images (DHI), stripped-down builds for the ClickHouse server, ClickHouse Keeper, Kubernetes Operator and Helm chart, and metrics exporter. The standard image’s Ubuntu 22.04 base includes unused packages such as Perl, wget, and apt, whose known CVEs can block enterprise deployments even though ClickHouse does not use them. DHI replaces that base with a minimal image that has no package manager or network tools, runs as non-root, provides SLSA Level 3 provenance, and is maintained for CVE patching; Docker Scout reported 0 medium and 14 low findings for DHI, versus 8 medium and 11 low across 111 packages in the standard image. Migration preserves volumes and configuration, while development tooling remains available through a dev variant or temporary docker debug attachment.


### [One Driver, One Format, Every Language: ADBC](https://yomu.fyi/post/one-driver-one-format-every-language-adbc.md)
- Company: [Clickhouse](https://yomu.fyi/company/clickhouse.md)
- Published: Jul 10, 2026
- Type: Announcement

ClickHouse announces an official ADBC driver, a modern Arrow-native alternative to ODBC and JDBC for analytics and AI applications. It provides end-to-end columnar movement: query results leave ClickHouse as Apache Arrow and arrive in applications without a row-oriented conversion or staging step. A single Rust-built implementation connects through ClickHouse's HTTP interface, supports TLS, and serves languages and tools with ADBC bindings, including Ruby, R, and C. Users can install precompiled, signed binaries through the ADBC Driver Foundry and connect with a clickhouse:// URI; dbt Fusion distributes the driver directly. The driver also supports streaming Arrow inserts, while schema discovery, improved bulk ingest, usage examples, and JSON support remain on the roadmap.


### [Accelerating End-to-End Co-Folding Performance with NVIDIA BioNeMo Agent Toolkit](https://yomu.fyi/post/accelerating-end-to-end-co-folding-performance-with-nvidia-bionemo-age.md)
- Company: [NVIDIA](https://yomu.fyi/company/nvidia-developer-blog.md)
- Author: Elizabeth Goodman
- Published: Jul 10, 2026
- Type: Problem & solution

Biomolecular structure prediction and co-folding pipelines encounter latency and memory ceilings during sequence alignment and multi-residue inference. NVIDIA addresses these constraints through the BioNeMo Agent Toolkit, which connects AI agents to accelerated biological computing components. The pipeline offloads multiple sequence alignment to MMseqs2-GPU, while cuEquivariance and the OpenFold3 NIM accelerate core geometric attention kernels to cut inference latency up to fourfold. To handle large molecular assemblies beyond single-GPU memory limits, the Fold-CP framework distributes context across multiple GPUs with O(N²/P) memory scaling. Together, these tools enable the prediction of complexes up to 32,000 tokens on 64 NVIDIA B300 GPUs, making large-scale virtual screening and ribosome-scale modeling computationally feasible.


### [The ambulatory intelligence gap](https://yomu.fyi/post/the-ambulatory-intelligence-gap.md)
- Company: [Databricks](https://yomu.fyi/company/databricks.md)
- Author: Morgan Wilkie, Bryan Smith, Holly Burke, Mary Ellen
- Published: Jul 10, 2026
- Type: Announcement

Health systems face an ambulatory intelligence gap: access, provider capacity, referral retention, panel management, and financial performance are interconnected, while information remains scattered across disconnected systems. Health Catalyst’s Ambulatory Intelligence combines AI with nearly two decades of healthcare improvement expertise, deploying directly in a customer’s Databricks workspace so sensitive data stays within the health system’s environment. It uses a medallion-based semantic layer, Unity Catalog for governance, Lakebase for low-latency serving, and Genie alongside dashboards to help leaders investigate why metrics change. The solution ships with prebuilt metrics across Access Optimization, Revenue Intelligence, Panel Management, and Referral Insights, plus cross-domain scorecards and configurable terminology and workflows. Reported outcomes from supported improvement work include increased revenue and encounters at Thibodaux Regional, 55,000 closed care gaps at INTEGRIS Health, and higher outpatient visits with fewer cancellations without reschedules at WakeMed; future plans include models based on prior outcomes and agentic capabilities.


### [How Swiggy Saved 3.7 MB by Solving a Vector Drawable Mystery](https://yomu.fyi/post/how-swiggy-saved-3-7-mb-by-solving-a-vector-drawable-mystery.md)
- Company: [Swiggy](https://yomu.fyi/company/swiggy.md)
- Author: Shivam Gupta
- Published: Jul 10, 2026
- Type: Explainer

An Android APK investigation found high-resolution PNG variants beside vector drawable XML files, although those raster assets had not been added directly to the codebase. The cause was AAPT2 rasterizing vectors that used features unavailable below API 24, including gradients in android:fillColor, despite the app’s minSdkVersion being 23; the generated files could be much larger than their XML sources, and were produced even for unused resources. The described remedy was to remove unused raster-prone assets, use VectorDrawableCompat with app:srcCompat for affected vectors, and tell Gradle and AAPT to use support-library rendering rather than density-specific PNGs. Removing unused assets cut download size by more than 1 MB, while raising minSdkVersion to 24 produced a further 2.7 MB reduction, for 3.7 MB total; similar gains were observed in Toing and Instamart.


### [How Swiggy Migrated from protobuf-java to Square Wire](https://yomu.fyi/post/how-swiggy-migrated-from-protobuf-java-to-square-wire.md)
- Company: [Swiggy](https://yomu.fyi/company/swiggy.md)
- Author: Balvinder Gambhir
- Published: Jul 10, 2026
- Type: Problem & solution

Swiggy migrated from protobuf-java to Square Wire after profiling showed that the runtime and protoc-generated code consumed 6.735 MB of Android app size. protobuf-javalite was rejected because Swiggy relies extensively on google.protobuf.Any, while Wire retained support for Any and other well-known types with a smaller runtime and immutable Kotlin models. The migration required updating snake\_case property access, message construction, serialization calls, ProGuard rules, and a custom pipeline that compiled all schemas together before publishing package-specific AARs to AWS CodeArtifact. The team also fixed repeated enum fields with \[packed = true\] after discovering a silent wire-format mismatch. It reduced Swiggy’s app size by 3.52 MB and produced measurable reductions for Instamart and Toing.


### [Scaling Android CI: From 44 → 10 minutes — A Deep Dive into Build Time Optimization and Best…](https://yomu.fyi/post/scaling-android-ci-from-44-10-minutes-a-deep-dive-into-build-time-opti.md)
- Company: [Swiggy](https://yomu.fyi/company/swiggy.md)
- Author: Pratik Sahu
- Published: Jul 10, 2026
- Type: Problem & solution

An Android monorepo’s CI pipeline averaged about 35 minutes in its build phase, with a P90 of roughly 44 minutes, as repeated dependency downloads, environment-driven invalidations, variant bloat, and machine limits increased work. The team reworked caching with branch- and architecture-aware keys, release/develop fallbacks, cached SDK and dependency data, and a Gradle Remote Cache, while filtering repositories and restricting CI to required variants. It also tuned 32GB runners to a 16GB heap and 12GB Metaspace, disabled configuration-cache saves in ephemeral CI, upgraded Kotlin, AGP 8.10.1, and Gradle 8.11.1, and moved critical workflows to Apple Silicon. Gradle tasks fell from 3399 to 2745, while remote-cache hit rates reached P50 44% and P90 56%. After the Apple Silicon refresh, total CI time reached P50 10 and P90 11 minutes, average credits per build dropped from 55 to 43, and a comparable rollout window saved about 5800 credits.


### [UST is bringing Claude to physical AI](https://yomu.fyi/post/ust-is-bringing-claude-to-physical-ai.md)
- Company: [Anthropic](https://yomu.fyi/company/anthropic.md)
- Published: Jul 10, 2026
- Type: Announcement

UST is partnering with Anthropic to bring Claude into engineering and production systems used by semiconductor, automotive, manufacturing, telecom, embedded, and IoT companies. Claude Code reads chip pinouts and hardware schematics, then writes and runs regression tests while carrying design context across hours-long tasks. UST is integrating Claude as the reasoning layer in its iDEC validation pipeline, which generates tests, compares live equipment data with a digital twin, and already cuts validation cycles from four days to 48 hours, 50 to 70 percent. The partnership also applies Claude to healthcare, telecom, and banking platforms, with human approval retained for recommended actions and response workflows. UST will train 20,000 associates worldwide and build specialized deployment teams, with governance and audit controls supporting production use in high-stakes industries.


### [Side Chats and Conversation Search](https://yomu.fyi/post/side-chats-and-conversation-search.md)
- Company: [Cursor](https://yomu.fyi/company/cursor.md)
- Published: Jul 10, 2026
- Type: Announcement

An update adds side chats for exploring tangents alongside a main agent conversation without interrupting it. Each side chat is a durable, full agent conversation with main-chat context, can be revisited, followed up on, and at-mentioned back into the main thread. In the Agents Window, users can search agent transcripts through Cmd+K using a local index designed to scale to thousands of conversations, or search within a conversation with Cmd+F. Redesigned project and repo pickers support creating projects and connecting GitHub, GitLab, or Azure DevOps without leaving the picker, with search scoped to the current location. New cloud agent hooks expose prompts, responses, thinking, subagents, compaction, and turn completion for observation and control.


### [Scaling Grab's Data Lake: Our journey to Apache Iceberg adoption](https://yomu.fyi/post/scaling-grab-s-data-lake-our-journey-to-apache-iceberg-adoption.md)
- Company: [Grab](https://yomu.fyi/company/grab.md)
- Author: Rahul Penti
- Published: Jul 10, 2026
- Type: Problem & solution

Grab is transitioning its petabyte-scale data lake from Hive Parquet tables managed through the Hive Metastore to a table-centric architecture centered on Apache Iceberg, while retaining storage-compute separation on S3. The move addresses catalog listing latency, severe small-file fragmentation, manual partition registration, absent native ACID operations, and catalog state becoming stale after direct S3 changes. Rather than converting the entire lake at once, Grab prioritizes high-value tables and uses features such as Z-ordering; representative workloads cut one query from 70 seconds to 6, reduced daily S3 API costs by up to 95%, and halved cluster resource usage. Its open-sourced UnifiedSparkCatalog detects table formats through metastore metadata or path inference and routes operations to the appropriate Spark catalog without format-specific prefixes. Deployment also required resolving Hive lock contention, timestamp incompatibilities, and storage-tier cost spikes.


### [Ask, build, compose: What our 5th Genie Hackathon taught us about Databricks Genie](https://yomu.fyi/post/ask-build-compose-what-our-5th-genie-hackathon-taught-us-about-databri.md)
- Company: [Databricks](https://yomu.fyi/company/databricks.md)
- Author: Shruti Prasanna, Rob Bajra
- Published: Jul 9, 2026
- Type: Explainer

The fifth Databricks hackathon used Databricks Genie to test three ways of working with governed data: asking, building, and composing. Genie Agents give business users domain-specific natural-language access to curated data, while Genie Code helps analysts create metric views, functions, pipelines, and dashboards inside Databricks. The composition track uses Conversation APIs and a managed MCP server to make Genie an agent tool alongside custom logic, Model Serving, and other MCP servers, with Unity Catalog governing access. Projects included supervisory routing across 190 tables, governance and analytics products built in days, and multi-agent systems such as ShipBob’s overnight operations brief. Across the tracks, the stated lesson is that shared governance and semantic context let business users, builders, and engineers use Genie at different levels without abandoning grounded, permissioned data access.


### [From Day 1 to Production: Building Lyft’s Analytics & Rides Intelligence Assistant as Onboarding…](https://yomu.fyi/post/from-day-1-to-production-building-lyft-s-analytics-rides-intelligence.md)
- Company: [Lyft](https://yomu.fyi/company/lyft.md)
- Author: Sagarbaronia
- Published: Jul 9, 2026
- Type: Problem & solution

Lyft developed Aria, an AI-powered assistant allowing approved employees to query ride analytics via natural language through a backend with LangGraph orchestration and SQL generation. However, Aria's user interface was limited to a Streamlit prototype that lacked standard service framework integration, Lyft authentication, and multi-user scalability. To resolve these limitations, a new engineer built a production-grade web client from scratch using Lyft's internal Node.js framework and Next.js as a three-week onboarding assignment. The implementation involved configuring Envoy routing, resolving authentication plugin incompatibilities, debugging service connectivity with Grafana, and managing interface state with an XState state machine. Aria was successfully hardened and launched into production, expanding natural language analytics access across the company.


[Newer posts](https://yomu.fyi/posts/page/96.md) · [Older posts](https://yomu.fyi/posts/page/98.md)