Happy Monday! ☀️

Welcome to the 93 new hungry minds who have joined us since last Monday!

If you aren't subscribed yet, join smart, curious, and hungry folks by subscribing here.

📚 Software Engineering Articles

• Scaling distroless adoption with AI reduces container bloat

• Amazon's bar raiser guide to cracking interviews

• Apache Kafka decision guide: when to use it

• Senior engineer behaviors that actually matter

• How LLMs are trained: demystifying the black box

🗞️ Tech and AI Trends

• Apple's new AI architecture leverages Google Gemini models

• OpenAI plots biggest ChatGPT overhaul since launch

• OpenAI's GPT-5.5 reaches general availability on Amazon Bedrock

👨🏻‍💻 Coding Tip

• Property-based testing with fast-check finds edge cases automatically through constraint-solving

Time-to-digest: 5 minutes

How Palantir Reindexes Elasticsearch at Scale Without Breaking Search 🦭

Search is mission-critical at Palantir. When schemas evolve or indices corrupt, you need to rebuild them while users are actively searching. That's where the Elasticsearch reindex machine comes in — automating what would otherwise be a painful, manual nightmare across hundreds of deployments.

The challenge: Rebuild massive search indices from the database without taking the service offline, saturating resources, or requiring a ops team to babysit the process.

Implementation highlights:

1. Shadow index pattern: Create a duplicate index with new mappings in parallel while the old one keeps serving reads. Dual-write to both until the shadow catches up, then flip the alias.

2. Parallel pipelining: Workers dynamically read pages from the database while others index them concurrently, keeping I/O flowing without explicit role assignment.

3. Three-dimensional rate limiting: Cap documents per batch, memory per batch, and total batches per run — ensuring no single dimension becomes the bottleneck that tanks your service.

4. Crash-safe state machine: Persist every index lifecycle step to the database before touching Elasticsearch. If the server dies mid-reindex, it resumes safely from the last known state.

5. Cluster-agnostic design: Writes fan out to all connected clusters, making single-cluster repairs and cross-cluster migrations work with nearly identical code paths.

Results and learnings:

• Truly online: Reindex millions of documents while live traffic continues without degradation or downtime

• Hands-free operation: Automated detection, repair, and execution across 100+ schemas per deployment across 100+ deployments

• Battle-tested observability: Logs for investigation, metrics for dashboards, and on-demand diagnostics for root causes — because different problems need different tools

Palantir's approach proves that reindexing at scale isn't black magic — it's careful design. Shadow indices, pipelining, and crash-safe state machines turn a risky operation into something you can sleep through.

Scaling out Distroless adoption With AI

Grab is moving hundreds of services to Distroless images to cut CVEs but only if workloads still run cleanly. Learn how medium tests became the migration gate, where the scaffolding toil stalled the campaign, and how an agentic workflow (skills, MCP, guardrails) generated tests and Docker changes at fleet scale; with humans still approving every MR.

Amazon's Bar Raiser Reveals How to Crack Tech Interviews

This is a guest post by Apurv Singh, a Senior Engineer and Bar Raiser at Amazon. Having conducted 200+ technical interviews, Apurv shares practical lessons and insights to help you prepare better and crack tech interviews.

When to Use Apache Kafka (And When Not To)

Discover when to use Apache Kafka and when to avoid it. Learn the core differences between event streaming and task queues to scale your systems reliably.

The Behaviors That Turn Engineers Into Senior Engineers

The gap between mid-level and senior engineers usually isn’t technical skill. It’s behavior, judgment, communication, and the ability to make everyone around them better.

Apache Airflow Document Ingestion Pipeline for RAG Systems

Learn how to design a production-grade document ingestion pipeline with FastAPI, Apache Airflow, and PostgreSQL for RAG and ML workflows.

ARTICLE (brain-bake-101)
How LLMs are Actually Trained

ARTICLE (hold-my-calls)
Call Queue: What It Is, How It Works & How to Get Started

ESSENTIAL (test-the-new-way)
A new era for software testing

ARTICLE (claude-skill-go)
Anthropic's Complete Guide to Claude Skills Building

ARTICLE (probability-brain-juice)
Bayesian Networks and Markov Networks: An Intuitive Guide to Structured Uncertainty

ARTICLE (web-tools-go-brr)
8 cutting-edge web development tools you don't want to miss

ARTICLE (rank-your-humans)
Why Performance Management Works the Way It Does - Stack Ranking, PIPs, Ratings, and a FAQ

ARTICLE (code-sniff-patrol)
Static Code Analysis and the Rules of Zero, Three, and Five

Want to reach 200,000+ engineers?

Let’s work together! Whether it’s your product, service, or event, we’d love to help you connect with this awesome community.

🤖 Apple Unveils New AI Architecture Powered by Google Gemini Models (3 min)

Brief: Apple announces a major Apple Intelligence overhaul built on foundation models co-developed with Google, featuring on-device processing and Private Cloud Compute for image creation, advanced editing, and visual understanding while maintaining privacy guarantees that prevent data access to Apple or third parties.

🤖 OpenAI Transforms ChatGPT Into a "Superapp" to Drive Revenue Before IPO (4 min)

Brief: OpenAI is overhauling ChatGPT into a "superapp" combining coding tools and AI agents to generate higher-margin revenue ahead of its planned IPO, shifting focus from the free chatbot toward paid business products like Codex, which now has 5 million weekly active users and accounts for 40% of revenue.

🚀 OpenAI's GPT-5.5 and Codex Now Generally Available on Amazon Bedrock (4 min)

Brief: OpenAI's GPT-5.5, GPT-5.4, and Codex are now generally available on Amazon Bedrock, allowing 100,000+ organizations to access frontier AI models with native AWS security controls (IAM, VPC, KMS encryption, CloudTrail) without new vendor relationships, while Codex shifts from per-seat to pay-per-token pricing for developer teams.

🚨 Anthropic Apologizes for Invisible Claude Fable Guardrails (3 min)

Brief: Anthropic faced backlash for secretly throttling Claude Fable 5 with hidden safeguards that degraded responses to distillation attempts without notifying users, but now pledges to make these restrictions transparent by routing blocked queries to an older model with visible warnings.

🤖 What Bun's AI-Driven Development Reveals About Anthropic's Open Source Strategy (5 min)

Brief: After Anthropic acquired Bun, the JavaScript runtime's commit activity shifted from primarily human to over 80% AI-generated code, while external contributors dropped significantly, raising questions about whether the company understands open source stewardship or will treat the project as internal infrastructure rather than a community-driven standard.

This week’s tip:

Leverage property-based testing with constraint-solving to find edge cases in state machines and parsers. Tools like fast-check (JavaScript) or Hypothesis (Python) shrink failing inputs automatically, surfacing minimal counterexamples that traditional unit tests miss, especially for off-by-one errors and boundary conditions.

Wen?

• Parser validation: Generate random input strings; verify the parser rejects invalid syntax and accepts valid syntax correctly.

• State machine correctness: Run sequences of random state transitions; assert invariants hold after each step.

• Serialization round-trips: Create arbitrary objects, serialize, deserialize, compare; catch encoding bugs automatically.

Some of your greatest lessons come from your darkest moments.
Roger Lee

That’s it for today! ☀️

Enjoyed this issue? Send it to your friends here to sign up, or share it on Twitter!

If you want to submit a section to the newsletter or tell us what you think about today’s issue, reply to this email or DM me on Twitter! 🐦

Thanks for spending part of your Monday morning with Hungry Minds.
See you in a week — Alex.

Icons by Icons8.

*I may earn a commission if you get a subscription through the links marked with “aff.” (at no extra cost to you).

Happy Monday! ☀️

Welcome to the 135 new hungry minds who have joined us since last Monday!

If you aren't subscribed yet, join smart, curious, and hungry folks by subscribing here.

📚 Software Engineering Articles

• AI agents accelerate Liger Kernel engineering exponentially

• Things you didn't know about database indexes

• Latency vs throughput vs bandwidth explained clearly

• Balancing cost and reliability for Spark on Kubernetes

• Null has cost billions in software failures

🗞️ Tech and AI Trends

• Angular v22 brings major framework improvements

• When AI builds itself: recursive self-improvement unlocked

• Instagram accounts hacked via Meta's AI chatbot abuse

👨🏻‍💻 Coding Tip

• Sliding-window idempotency caches trade memory for duplicate detection; Bloom filters + time-partitioned LRU prevent costly re-processing

Time-to-digest: 5 minutes

Selective Test Execution at Stripe: Fast CI for a 50M-line Ruby monorepo 🪫

Stripe's Ruby monorepo contains 50 million lines of code and 1.2 million test units. Running them all sequentially would take four months. Instead, Stripe built Selective Test Execution (STE)—a system that runs only ~5% of tests per build while maintaining full confidence in code safety.

The challenge: Static dependency analysis fails on dynamic languages like Ruby. Metaprogramming, runtime configuration, and non-code dependencies (YAML, JSON, fixtures) make it nearly impossible to predict what a test actually touches without running it.

Implementation highlights:

1. Dynamic file access interception: Build a C++ shared library (file_access_interceptor) loaded via LD_PRELOAD that records every file opened during test execution at the OS syscall level

2. Hierarchical scope tracking: Attribute file access to specific tests using a scope stack that naturally handles child processes and distinguishes global dependencies from test-specific ones

3. Roaring bitmap indexes: Store three billion data points as compressed bitmaps mapping changed files to impacted tests, enabling fast union/intersection operations during selection

4. Monotonic Revision IDs: Order build metadata without git queries using ancestry-preserving identifiers, making baseline selection fast and reproducible

5. Pragmatic guardrails: Force rerun of directory-globbing tests and linters with selective scans to handle edge cases where file access doesn't capture behavioral changes

Results and learnings:

• Massive speedup: Reduced median test execution to <0.5% of the full suite while maintaining safety and confidence

• Cost efficient: Spend <10% compute of an "always run everything" strategy across 50,000+ builds per week

• Operationally sound: Single fast database query per build makes selection reproducible and debuggable

Stripe's approach proves that observing what code actually does beats predicting what it might do. By intercepting at the syscall boundary and storing results efficiently, they've cracked the unsolved problem: how to scale testing without sacrificing safety.

EP217: Latency vs Throughput vs Bandwidth

Latency, throughput, and bandwidth often get used interchangeably, but each one tells a different story about performance.

Balancing cost and reliability for Spark on Kubernetes

Today, we’re excited to open-source Spot Balancer in collaboration with AWS, a tool that helped us reduce Spark compute costs by up to 90% while maintaining reliability.

AI Engineering for Developers

A tour through AI engineering for developers who already know how to ship software. Fourteen chapters, no LinkedIn voice, no slow warm-up. We will go from 'what is a foundation model' to 'how do you run agents in production on Google Cloud' without skipping the parts that matter.

AI helping build better AI: How agents accelerate Liger Kernel engineering

ARTICLE (thumb it up)
I Built a Free Video Thumbnail Generator That Never Uploads Your Files

GITHUB REPO (bench the keys)
Keybench – Scriptable, extensible performance tool for key value stores

ESSENTIAL (billion dollar oopsie)
Null Looks Like an Empty Value — Until You Realize It Has Caused Billions of Dollars in Software Failures

ARTICLE (ai blog go brrr)
How I Redesigned 4 Years of Blog Posts (196 of them!) Overnight with AI

ESSENTIAL (tired but wiser)
Being oncall taught me everything

ARTICLE (queue the drama)
Building observability into Notion's dead-letter queue

ARTICLE (gemma go smart)
Gemma 4 12B: The Developer Guide

GITHUB REPO (git trees grow)
Treehouse – isolate multple dev environments on different Git worktrees

Want to reach 200,000+ engineers?

Let’s work together! Whether it’s your product, service, or event, we’d love to help you connect with this awesome community.

🚀 Angular v22 Released: Production-Ready Signal Forms, Aria, and AI-Native Features (8 min)

Brief: Angular v22 brings three features to production-ready status: Signal Forms for reactive, composable forms, Angular Aria for accessible components, and asynchronous reactivity APIs for handling async operations; plus new AI-native capabilities including agentic tooling, WebMCP support, and integration with Google AI Studio and Gemini Canvas for no-code Angular app building.

🤖 When AI Builds Itself: Anthropic's Path to Recursive Self-Improvement (8 min)

Brief: Anthropic reveals that AI systems are accelerating their own development, with engineers shipping 8x more code per quarter than in 2021-2025, Claude now writing over 80% of production code, and models progressing from executing specified tasks to autonomously running research projects—signaling a potential future where AI could design its own successors, bringing both transformative benefits and significant risks that require urgent coordination and alignment research.

🔐 Meta Confirms Thousands of Instagram Accounts Hacked Via AI Chatbot Exploit (4 min)

Brief: Meta disclosed that over 20,000 Instagram users had their accounts hijacked through a vulnerability in its AI-assisted account recovery system, which hackers exploited by tricking the chatbot into resetting passwords and sending verification codes to attacker-controlled emails, granting full access to accounts, messages, and linked profiles for months until the flaw was patched this week.

🤖 Uber Caps AI Tool Usage to Rein In Spiraling Costs (3 min)

Brief: Uber is limiting employees to $1,500 monthly per AI coding tool after exhausting its 2026 AI budget in just four months, a rational cost-control measure that suggests the company values these agentic tools at roughly 11% of engineer compensation while signaling that subsidized API plans are no longer available to enterprise-scale customers.

🚀 VoidZero Joins Cloudflare to Power Open-Source JavaScript Toolchain (7 min)

Brief: VoidZero, creator of Vite, Vitest, Rolldown, and Oxc, is joining Cloudflare with all team members, while the projects remain open source, vendor-agnostic, and community-driven; Cloudflare commits $1M to a Vite ecosystem fund and plans to build its CLI on top of Vite to support AI-driven development and full-stack applications.

This week’s tip:

Implement request-level idempotency keys with a sliding-window probabilistic cache to trade bounded memory for rare duplicates. Use a Bloom filter for negative lookups and a time-partitioned LRU for positive confirmations; evict old partitions on window slide.

Wen?

• Payment processing at scale: Detect duplicate charge requests without maintaining a full history database; safe to evict very old keys.

• Event deduplication in async workflows: Bloom filter false positives cause harmless re-execution; false negatives are caught by partition re-check.

• Multi-region consistency: Each region maintains its own partition set; reconcile via event log on conflicts to avoid global consensus.

Just get out and do it. You will be very, very glad that you did.
Christopher McCandless

That’s it for today! ☀️

Enjoyed this issue? Send it to your friends here to sign up, or share it on Twitter!

If you want to submit a section to the newsletter or tell us what you think about today’s issue, reply to this email or DM me on Twitter! 🐦

Thanks for spending part of your Monday morning with Hungry Minds.
See you in a week — Alex.

Icons by Icons8.

*I may earn a commission if you get a subscription through the links marked with “aff.” (at no extra cost to you).

Happy Monday! ☀️

Welcome to the 497 new hungry minds who have joined us since last Monday!

If you aren't subscribed yet, join smart, curious, and hungry folks by subscribing here.

📚 Software Engineering Articles

• Failure modes in distributed systems explained for engineers

• Technical writing in the AI age transforms documentation forever

• Slack's multi-cloud journey scales globally and efficiently

• AI agents accelerate Liger Kernel engineering productivity dramatically

• AI agents fail in production when built backwards

🗞️ Tech and AI Trends

• Apple overhauls iOS 27 Siri with powerful AI features

• Anthropic surpasses OpenAI as most valuable startup

• Tech CEOs suffer from AI psychosis; reality check needed

👨🏻‍💻 Coding Tip

• Testcontainers with deterministic seeds catch data-dependent bugs without flakiness

Time-to-digest: 5 minutes

The Hidden Failure Modes That Break Distributed Systems ⚡

Your servers report healthy. Your dashboards glow green. Users are getting errors. Welcome to distributed systems, where "up" is a philosophical question, not a binary state.

Unlike single machines, where a crash is obvious, distributed systems hide their failures in plain sight. A node can report healthy while the whole system serves stale data. Another can be technically working but trapped in an unrecoverable state. The worst part? None of it shows up as a bug—it's just a pattern that's been humbling engineers for decades.

The challenge: Failure modes in distributed systems aren't about code bugs. They're recurring architectural patterns that kill systems while every metric looks perfect.

Implementation highlights:

1. Know your failure taxonomy: Understand byzantine failures, split-brain scenarios, and cascading timeouts as distinct problems requiring distinct solutions

2. Design for partial failures: Assume individual components will fail independently—your system must function with degraded capacity, not just all-or-nothing

3. Implement health checks that matter: Move beyond heartbeats to semantic checks that verify the system can actually serve requests correctly

4. Embrace defensive timeouts: Add circuit breakers and bulkheads to prevent one slow service from paralyzing your entire architecture

5. Test failure modes explicitly: Use chaos engineering and failure injection to expose these patterns before they hit production

Results and learnings:

• Pattern recognition: Knowing the names and mechanisms of common failures helps you spot them before they become disasters

• Architectural resilience: Systems designed with these patterns in mind stay operational even when individual components fail

• Psychological safety: Understanding these aren't bugs but patterns shifts how teams approach debugging and design

Distributed systems are humbling. They force you to think beyond happy paths and embrace the messiness of reality. Knowledge of these failure modes is your map through that chaos.

The key takeaway? Your system isn't broken because you're bad at coding. It's broken because distributed systems are genuinely hard, and there's a well-documented playbook for each way they fail.

State of the software engineering job market in 2026

A deepdive into today’s tech jobs market, with exclusive data on software engineering jobs, the AI engineering boom, whether AI engineering is “replacing” software engineering hiring, and more

Technical Writing in the AI Age | CSS-Tricks

This isn’t totally about AI. It’s about technical writing in the age of AI. I have some thoughts on this and I hope it’s helpful to you humans reading.

Slack AI: The Path to Multi-Cloud

In early 2023, Slack faced a foundational challenge: serving Large Language Models (LLMs) at enterprise scale with the security, reliability, and performance our customers expect. Over three years, we evolved from basic infrastructure to orchestrating a sophisticated multi-cloud architecture. We didn’t just want shiny new models; we needed a system resilient to regional outages and…

AI helping build better AI: How agents accelerate Liger Kernel engineering

Beyond code generation: rethinking engineering productivity in the age of AI agents

How Dropbox is moving from AI tools that assist engineers to agentic systems that can execute scoped tasks, and how we’re building platforms to support those workflows.

My Frontend Stack In 2026

ARTICLE (book who)
Nobody Cracks Open a Programming Book Anymore

ARTICLE (docs go brrr)
From decentralized Docs-as-Code to a centralized repository: Evolving Grab's documentation strategy

ARTICLE (ml sees it all)
Visual Debugging Tools for Machine Learning Workflows

ARTICLE (agents do stuff)
My Thoughts on AI, Part 2: Agent Setup, Workflow, and Tools

ARTICLE (tokens go bye)
How Many Tokens Did You Burn Today

ARTICLE (oops backwards)
Most AI Agents Fail in Production Because They're Built Backwards

ARTICLE (demo magic time)
24 tips for giving S-tier demos

ARTICLE (tired humans win)
We should be more tired than the model

Want to reach 200,000+ engineers?

Let’s work together! Whether it’s your product, service, or event, we’d love to help you connect with this awesome community.

🍎 Apple Overhauls iOS 27 With AI-Powered Siri and Advanced Camera Features (4 min)

Brief: Apple is staging a comeback in digital assistants and AI with iOS 27, featuring a revamped Siri interface, new chatbot-style app, and Pro Camera app integration set to be announced at the June 8 Worldwide Developers Conference—according to first leaked screenshots obtained by Bloomberg.

🤖 Tech CEOs Are Suffering From AI Psychosis, Says Box Founder (4 min)

Brief: Tech CEOs are making unrealistic AI assumptions by playing with prototypes without understanding real-world implementation challenges, leading to mass layoffs that lack productivity evidence—research shows AI agents won't match human quality work for years, yet executives keep betting the company on quick automation wins.

😔 AI Job Grief: Tech Workers Face Unnamed Psychological Crisis as Roles Disappear (8 min)

Brief: Knowledge workers experiencing AI-driven job displacement are grieving not just lost income but eroded professional identity, a distinct emotional state clinicians are naming as Artificial Intelligence Replacement Dysfunction (AIRD) that goes unacknowledged by employers, leaving workers in disenfranchised grief with no cultural script for recovery in an endless cycle of retraining.

🚀 SpaceX Files for Largest IPO Ever: Inside the $1.75T Space, Starlink, and AI Conglomerate (8 min)

Brief: SpaceX filed its S-1 for a ~$75B IPO at $1.75T valuation, revealing three distinct businesses—profitable Starlink connectivity ($11.4B revenue, 39% margins), a launch business funding Starship R&D, and a rapidly expanding AI segment boosted by a $1.25B/month Anthropic compute deal that could push 2026 run-rate revenue to $40B+, while positioning itself as the physical infrastructure layer for the orbital AI economy targeting 100 terawatts of space-based compute.

🤖 Anthropic Surpasses OpenAI to Become World's Most Valuable AI Startup (2 min)

Brief: Anthropic has dethroned OpenAI as the world's most valuable AI startup after raising $65 billion in Series H funding, pushing its valuation near $1 trillion—nearly triple its February valuation—driven by explosive demand for its Claude AI assistant and Code service, which now generates $47 billion in annual revenue.

🧭 Avoiding Death on the Yellow Brick Road: Where AI App Startups Can Actually Win (8 min)

Brief: While OpenAI and Anthropic dominate horizontal AI applications (the "Yellow Brick Road"), startups can build defensible businesses in vertical markets by owning complex workflows, accumulating domain-specific data flywheels, managing model variability across vendors, optimizing costs, and providing industry-specific governance—advantages general labs structurally can't replicate without losing their horizontal advantage.

This week’s tip:

Testcontainers with deterministic seeding for property-based tests: Combine Testcontainers for ephemeral database/service instances with property-based testing (fast-check) and fixed random seeds to catch data-dependent bugs reproducibly without flakiness.

Wen?

• Data-driven edge cases: Property tests generate permutations of inputs automatically; Testcontainers ensure DB state resets between runs.

• Reproducibility: Save the seed from a failing run and re-run with { seed: 12345 } to debug non-deterministic failures.

• Integration test scaling: Avoid mocking databases; spin up real instances per test with Testcontainers and tear down in seconds.

Victims recite problems, leaders provide solutions.
Robin Sharma

That’s it for today! ☀️

Enjoyed this issue? Send it to your friends here to sign up, or share it on Twitter!

If you want to submit a section to the newsletter or tell us what you think about today’s issue, reply to this email or DM me on Twitter! 🐦

Thanks for spending part of your Monday morning with Hungry Minds.
See you in a week — Alex.

Icons by Icons8.

*I may earn a commission if you get a subscription through the links marked with “aff.” (at no extra cost to you).

Happy Monday! ☀️

Welcome to the 300 new hungry minds who have joined us since last Monday!

If you aren't subscribed yet, join smart, curious, and hungry folks by subscribing here.

📚 Software Engineering Articles

• AI agents need memory consistency to make better decisions

• DoorDash built a clusterless ML feature store that scales

• DeepSeek-V4's million-token context changes what agents can do

• Spotify uses background coding agents for migrations

• Vector databases explained: the backbone of AI

🗞️ Tech and AI Trends

• Google I/O 2026: 15 updates powering the agentic web

• Elon loses his OpenAI lawsuit; what's next?

• Andrej Karpathy just joined Anthropic; major move incoming

👨🏻‍💻 Coding Tip

• Playwright test fixtures with dependency injection eliminate boilerplate, enable automatic cleanup, and support parallel isolation

Time-to-digest: 5 minutes

How Airbnb scales a 7 billion node identity graph with JanusGraph + DynamoDB 🕸️

Airbnb's identity graph tracks relationships between users to power Trust and Safety — think fraud detection, linked account discovery, and suspicious activity flagging. With 7 billion nodes and 11 billion edges growing by 5 million new edges daily, they ditched their third-party graph database vendor and built an internal knowledge graph infrastructure from scratch.

The challenge: Serve complex 4–8 hop graph traversals at low latency across billions of nodes, while taming long-tail P99 spikes caused by high-fanout subgraphs — all without periodic manual reboots to keep things stable.

Implementation highlights:

1. JanusGraph + DynamoDB separation: Decouple graph logic from storage by using JanusGraph for traversal and DynamoDB for persistence — scalability of AWS without reinventing distributed storage

2. Custom transaction strategy: Replace JanusGraph's heavy default locking with DynamoDB conditional writes and transaction APIs to cut write overhead significantly

3. Parallel multi-slice fetches: Rework the getMultiSlices interface to fetch data in parallel, directly attacking latency on high-fanout queries

4. Client-side query rewrites: Remove expensive Gremlin Path and SimplePath steps that bypass batch optimization, replacing them with conditional queries that enforce acyclic results

5. Multi-tenant namespace isolation: Each use case (identity graph, fraud detection, data lineage) runs in its own isolated namespace with schema enforcement and managed indexes

Results and learnings:

• 32–93% lower read latency: Across all hop patterns (1-hop through 8-hop), the internal solution crushed the vendor on both P95 and P99

• 10x write scalability: Auto-scaling during load tests hit 10x the previous vendor's write QPS — no more manual reboots needed

• 51% faster end-to-end reads: P95 dropped from ~2.1s to ~1.0s, and P99 from ~5.0s to ~2.5s for full read API calls

The takeaway is clear: if your graph workload is critical enough, owning the stack pays off. Vendor solutions get you started fast, but when you need to tune query planning, control transaction semantics, and scale writes independently — there's no substitute for building internally on proven open-source foundations.

A graph database is like your extended family tree at Thanksgiving: it starts simple, but after a few hops you discover connections you really didn't want to know about. 🍗

AI Agents: State, Memory, Consistency - A Deep Dive

#147: Understanding State, Memory, and Consistency in AI Agents

DeepSeek-V4: a million-token context that agents can actually use

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

Background Coding Agents: Supercharging Downstream Consumer Dataset Migrations (Honk, Part 4) | Spotify Engineering

This is part 4 in our series about Spotify's journey with background coding agents (internal codename: “Honk”) and the future of large-scale software maintenance. See also , , and .

Vector Database - A Deep Dive

#141: A Beginner’s Guide to the AI Stack’s Most Misunderstood Component

ARTICLE (gemini-go-brrr)
Gemini 3.5 Flash Developer Guide

ARTICLE (cpp-dad-spills-tea)
Creator of C++: Bell Labs, Negative Overhead Abstraction, Mistakes

ARTICLE (eval-funnel-not-dumpster)
Better Experiments with LLM Evals — A funnel, not a fork

ARTICLE (webpack-goodbye-rspack-hello)
Optimizing Our Build Times by Migrating from Webpack to Rspack

ARTICLE (leaderboard-gaming-detected)
The token leaderboard trap

ARTICLE (nova-agent-playground)
Introducing Nova, our internal platform for coding agents

ESSENTIAL (coding-is-just-the-tip)
Code was the smallest part of the job

ARTICLE (cyber-monsters-exposed)
Project Glasswing: what Mythos showed us

ARTICLE (ai-garbage-detector)
How to Avoid AI Code Slop

ARTICLE (test-your-robot-friends)
An Engineer's Guide to Better AI Skills: Implementing a Testing Process to Optimize Agent

Want to reach 200,000+ engineers?

Let’s work together! Whether it’s your product, service, or event, we’d love to help you connect with this awesome community.

🤖 Google Unveils 15 Chrome Updates Powering the Agentic Web Era (4 min)

Brief: Google introduced WebMCP, built-in AI models, and auto-browse capabilities to Chrome, enabling AI agents to interact with websites more intelligently, while new APIs like HTML-in-Canvas and Declarative Partial Updates push web UI performance boundaries, and Gemini in Chrome brings voice typing, image generation, and task automation across desktop and mobile devices.

📝 AI is Just Unauthorised Plagiarism at a Bigger Scale (3 min)

Brief: AI companies train models on unlicensed content without author consent, then sell access to customers who profit from the copied material—all while original creators receive zero compensation, with search engines like Google even ranking plagiarized AI-generated content higher than legitimate sources.

🚀 Forward Deployed Engineers Heat Up as Tech Giants Race to Scale AI (5 min)

Brief: Google, OpenAI, and Anthropic are aggressively hiring forward deployed engineers (FDEs) to embed AI into enterprise customers, with simplified interview processes and standalone deployment companies launched by OpenAI and Anthropic to accelerate AI rollouts while freeing core teams to focus on model development.

⚖️ Elon Musk Loses Lawsuit Against Sam Altman and OpenAI (3 min)

Brief: A California jury unanimously ruled against Elon Musk's lawsuit, finding his claims were filed too late under statute of limitations, dismissing allegations that Sam Altman and OpenAI stole a charity by converting it to a for-profit entity—a verdict that removes a major threat to OpenAI's reported IPO and prompts Musk to announce plans for an appeal.

🤖 Andrej Karpathy Joins Anthropic to Focus on LLM Research (1 min)

Brief: Legendary AI researcher Andrej Karpathy has joined Anthropic, stating he's excited to return to R&D at the frontier of large language models while planning to resume his education initiatives later.

🤖 Coding Agents Are Giving Everyone Decision Fatigue (5 min)

Brief: While AI coding agents have dramatically increased code generation speed, they've shifted the workload from writing code to reviewing and judging quality, leaving developers exhausted by constant decision-making as 80% of AI-generated content requires human editing and organizations scramble to redesign workflows for this new reality.

This week’s tip:

Use Playwright's test fixtures with dependency injection to eliminate test boilerplate and enable automatic cleanup, reporter integration, and parallel test isolation without manual setup/teardown. Fixtures are cached per test scope, reducing overhead for expensive resources.

Wen?

• Multi-tenant test isolation: Create database/API fixtures scoped per test worker, ensuring no cross-test pollution in parallel runs.

• Custom authentication flows: Abstract login/OAuth/2FA into reusable fixtures shared across hundreds of tests without code duplication.

• Resource pooling: Cache database connections, mock servers, or Docker containers at the project level and reuse across tests, reducing setup time by 60–80%.

Be brave enough to suck at something new.
Unknown

That’s it for today! ☀️

Enjoyed this issue? Send it to your friends here to sign up, or share it on Twitter!

If you want to submit a section to the newsletter or tell us what you think about today’s issue, reply to this email or DM me on Twitter! 🐦

Thanks for spending part of your Monday morning with Hungry Minds.
See you in a week — Alex.

Icons by Icons8.

*I may earn a commission if you get a subscription through the links marked with “aff.” (at no extra cost to you).

Happy Monday! ☀️

Welcome to the 341 new hungry minds who have joined us since last Monday!

If you aren't subscribed yet, join smart, curious, and hungry folks by subscribing here.

📚 Software Engineering Articles

• Scaling Android Studio performance in massive monorepos

• IC work is the new career flex for engineers

• Verifiable rewards-based reinforcement learning with GRPO on SageMaker

• Four subagent patterns shaping AI agent architecture in 2026

• JWT security in Node.js: five critical errors explained

🗞️ Tech and AI Trends

• Anthropic is now leading the AI boom's race forward

• Google reveals AI helped hackers find major software vulnerabilities

• Amazon employees are tokenmaxxing under AI adoption pressure

👨🏻‍💻 Coding Tip

• Prevent cache stampede using probabilistic early expiration and request collapsing via background workers

Time-to-digest: 5 minutes

How Discord Built Tracing for Elixir Without Breaking Everything 🧪

Discord runs millions of concurrent guild processes using Elixir's message-passing model, enabling near-instantaneous chat experiences at scale. But when things break, engineers were flying blind—metrics and logs couldn't explain why a guild suddenly felt laggy or why sessions couldn't connect.

The challenge: Implement distributed tracing across Elixir services that communicate via message passing (not HTTP), handle million-user fanouts without exploding span volume, and deploy with zero downtime on a system processing millions of concurrent operations.

Implementation highlights:

1. The Envelope primitive: Wrap all inter-process messages with metadata containers that transparently attach and propagate trace context without manual encoding at every call site

2. Zero-downtime rollout: Support both Transport and legacy message passing simultaneously, allowing gradual migration across fleet during normal deployments

3. Dynamic fanout sampling: Adjust trace sampling rates based on fanout size (100% for 1 session, 0.1% for 10k+) to prevent drowning in spans during massive broadcasts

4. Lazy context unpacking: Read just the sampling flag byte from trace context headers before expensive full parsing, cutting gRPC handler CPU overhead by half

5. Forbid root spans post-fanout: Sessions can only continue existing traces, never create new ones, eliminating redundant sampling decisions and regaining 10% CPU back

Results and learnings:

• Complete visibility: End-to-end traces now reveal the exact flow and timing of operations across the entire platform, from API to guild to session

• Pinpointed bottlenecks: Found that sessions were spending 75% of time unpacking trace context; a single-byte optimization fixed it

• Real impact quantified: Discovered guild outages caused 16-minute connection delays for users—impossible to diagnose without tracing

Discord proved that you don't need HTTP headers to do distributed tracing—just some clever engineering and a willingness to challenge assumptions. Their approach of optimizing for zero overhead first, then adding instrumentation, is the opposite of "measure first," and it paid off spectacularly.

IC work is the new career flex

The rise of the High-impact Individual Contributor (HI-C!)

Overcoming reward signal challenges: Verifiable rewards-based reinforcement learning with GRPO on SageMaker AI

In this post, you will learn how to implement reinforcement learning with verifiable rewards (RLVR) to introduce verification and transparency into reward signals to improve training performance.

Scaling developer experience: How we improved Android Studio in a large monorepo

Frustrated by 35-minute integrated development environment (IDE) syncs? In a large monorepo, slow builds were eroding productivity. Discover how we built a custom Focus plugin to slash sync times to under 2 minutes and drop memory usage from 10 GB to 2 GB. Learn how we leveraged Gradle and Bazel to reclaim developer flow without changing a single team's workflow.

Removing AI in Tech Interviews is Wrong

AI-assisted engineering is already the standard, and technical interviews should reflect how modern software is actually built.

How Agents Manage Other Agents: Four Subagents Patterns in 2026

Subagents solve context pollution, but how the main agent manages them matters more than whether they run in sync or async. Four orchestration patterns, from a simple tool call to an autonomous agent team, each with different requirements for model capability and result collection.

Monolith vs Microservices vs Modular Monoliths

Most software systems start simple. A small team, a single codebase, one database, and a few features.

22 Somewhat Random Bits of (Mostly Tech Career) Advice From Dave

Not in any order, or theme, just 22 thoughts from Dave.

ARTICLE (token-twister-tango)
JWT in Node.js: How It Works, 5 Errors That Compromise Your API, and Refresh Token with Rotation

ARTICLE (ai-shoulder-surfer)
What "Built It Solo" Actually Means When You Work With AI

ARTICLE (pocket-monster-power-up)
A Year Late, Claude Finally Beats Pokémon

ARTICLE (nitpick-ninja-moves)
What to do if you're not "detail oriented"

ARTICLE (memory-lane-memories)
Redis array type: short story of a long development

ARTICLE (brain-box-builder)
Build an AI-Powered Learning Management System That Actually Trains People

ARTICLE (code-grim-reaper)
Software Engineers are Obsolete

ESSENTIAL (attention-grabber-guide)
A deep dive into the Transformer architecture

ARTICLE (spotify-spell-caster)
Building a Natural Language Interface to the Spotify Ads API with Claude Code Plugins

ARTICLE (startup-snatch-watch)
Why are AI companies buying the teams behind your favorite dev tools?

Want to reach 200,000+ engineers?

Let’s work together! Whether it’s your product, service, or event, we’d love to help you connect with this awesome community.

⚠️ Mitchell Hashimoto Warns Against "AI Psychosis" in Software Development (4 min)

Brief: Hashimoto expresses concern that companies are caught in "AI psychosis," operating under the belief that AI agents can automatically fix bugs at scale, echoing past infrastructure lessons where over-automating resilience created hidden systemic risks that metrics failed to capture.

🖥️ Ubuntu Prioritizes Local AI Over Cloud-Centric Integration (2 min)

Brief: Canonical is shifting Ubuntu's AI strategy away from cloud-first approaches toward local inference and on-device models, offering modular design, user control, and inference snaps for simplified hardware-optimized AI installation while rejecting a global AI killswitch to maintain flexibility across different Ubuntu consumption methods.

🚀 AI is the New Netflix: Why Upload Traffic is the Real Story (4 min)

Brief: AI is becoming the killer app of the next broadband era, but unlike streaming video, it's driven by upload traffic, not downloads—cloud sync now dominates upstream data, while IoT devices, connected cars, and voice search generate massive outbound flows that are reshaping how networks must be built.

⚙️ Amazon Workers "Tokenmaxxing" to Hit AI Usage Targets (3 min)

Brief: Amazon employees are automating unnecessary tasks using the internal MeshClaw AI tool to artificially boost their token consumption and meet company AI adoption targets, creating perverse incentives despite management claims the metrics won't affect performance reviews.

This week’s tip:

Implement cache stampede prevention using probabilistic early expiration (xFetch pattern) combined with probabilistic collapsing of thundering herd requests via Kafka topics or Redis Streams. When a key nears expiration, refresh it with background workers before client demand spikes, reducing lock contention and miss-latency storms.

Wen?

• High-QPS APIs with shared expensive dependencies: Probabilistic refresh (e.g., 25% of clients trigger background refresh at 75% TTL) avoids coordinated expiry spikes that spike backend load 10x.

• Distributed caches with hot keys: Combine with Kafka consumer groups to load-balance refresh requests across workers; prevents single coordinator from becoming a bottleneck.

• Session caches with long TTLs: Early refresh at 75% TTL reduces the window where lock-wait cascades occur; stale-while-revalidate pattern masks latency from clients.

Learn from the rejection and turn it into an opportunity!
Mary Engelbreit

That’s it for today! ☀️

Enjoyed this issue? Send it to your friends here to sign up, or share it on Twitter!

If you want to submit a section to the newsletter or tell us what you think about today’s issue, reply to this email or DM me on Twitter! 🐦

Thanks for spending part of your Monday morning with Hungry Minds.
See you in a week — Alex.

Icons by Icons8.

*I may earn a commission if you get a subscription through the links marked with “aff.” (at no extra cost to you).

Happy Monday! ☀️

Welcome to the 530 new hungry minds who have joined us since last Monday!

If you aren't subscribed yet, join smart, curious, and hungry folks by subscribing here.

📚 Software Engineering Articles

• Data mesh at Grab: foundational tools behind certification

• Adaptive parallel reasoning: the next paradigm in inference

• The bottleneck was never the code; rethink your agents

• Stop failures from spreading between your microservices

• Bad retries can break good systems; here's why

🗞️ Tech and AI Trends

• AI load breaks GitHub – why other vendors survived

• Google Chrome silently installs a 4GB AI model on your device

• OpenAI fast-tracking AI phone for 2027 launch

👨🏻‍💻 Coding Tip

• Use OpenTelemetry exemplars to link metrics directly to traces, cutting debugging time dramatically

Time-to-digest: 5 minutes

Zero downtime upgrade: Yelp's Cassandra 4.x upgrade story 🐿️ 

Yelp's Database Reliability Engineering team upgraded over 1,000 Cassandra nodes from 3.11 to 4.1 with zero downtime. This wasn't just a version bump, it required orchestrating complex distributed systems, fixing compatibility issues, and keeping production traffic flowing the entire time.

The challenge: Upgrade a thousand production nodes without breaking client code, requiring schema migrations, or accepting any downtime while managing gossip communication failures, proxy incompatibilities, and CDC architectural changes.

Implementation highlights:

1. Init container gossip trick: Use Kubernetes init containers to let nodes gossip with their new IP on 3.11 before upgrading to 4.1, avoiding gossip failures during simultaneous IP and version changes

2. Parallel Stargate versions: Run version-specific Stargate proxy instances side-by-side in the service mesh, seamlessly routing to matching Cassandra versions during the transition

3. Separate Git branches: Publish version-specific Cassandra images from dedicated branches, enabling independent deployments to 3.11 and 4.1 without hard blocking either version

4. CDC Publisher backwards compatibility: Refactor the Cassandra Source Connector to handle architectural changes in 4.1 (immediate CDC log creation, schema change detection) while maintaining 3.11 support

5. Orchestrated three-stage automation: Implement checkpoint-enabled scripts for pre-flight (schema agreement, backup verification), flight (rolling node upgrades with monitoring), and post-flight (repair re-enablement) stages

Results and learnings:

• Massive latency wins: Achieved up to 58% reduction in p99 latencies on key clusters, plus 11% throughput improvement

• Zero production impact: No client code changes, no downtime, no incidents—the upgrade was invisible to users

• Faster restarts: Leveraged non-disruptive seed list reload for quicker gossip convergence and significantly improved node restart times

Yelp proved that big infrastructure upgrades don't require chaos. Their playbook: automation with checkpoints, version-agnostic design, and obsessive monitoring, shows you can move mountains without breaking a sweat.

The bottleneck was never the code

The other month I finally ran an experiment we had been postponing for over a year at .txt.

Create Your First AI Agent In 5 Minutes

A practical guide to building real Claude Code subagents with tools, context, and and well-defined outputs

How to Stop Failures from Spreading Between Services

Practical patterns to protect your services from failing dependencies and excessive load.

How to correctly use MCP servers with your AI Agents

MCP servers are not dead. Blindly enabling them bloats your context, which leads to higher cost and worse performance. Here are two proven patterns on how to correctly use MCP servers and avoid the bloat.

Bad Retries Can Break Good Systems

A practical guide to using backoff, jitter, retry budgets, idempotency, and circuit breakers in backend systems.

Data Mesh at Grab (Part II): The foundational tools behind certification

How does Grab manage quality across hundreds of thousands of data assets? Discover the foundational tools powering our Signals Marketplace. We dive into Hubble for discovery, Genchi for observability, and our Data Contract Registry to see how event-driven certification turns 'data as a product' into a reliable, AI-ready reality. Stop guessing and start trusting your data.

Adaptive Parallel Reasoning: The Next Paradigm in Efficient Inference Scaling

The BAIR Blog

ARTICLE (shadow clone jutsu)
Enhancing Flink deployment with shadow testing

ARTICLE (robot boss energy)
How to Thrive as an EM in the AI Era

ESSENTIAL (a11y zen mode)
Three stoic principles for better web accessibility

ARTICLE (auth roulette)
From Supabase to Clerk to Better Auth

ARTICLE (ai editor go brr)
Editing my LLM assisted Articles

ESSENTIAL (redis patience speedrun)
Redis array type: short story of a long development

ARTICLE (whisper sweet prompts)
Realtime prompting guide

ARTICLE (form go spinny)
Introducing TanStack Form

ARTICLE (ai interview hot take)
Removing AI in Tech Interviews is Wrong

ARTICLE (api wallet cries)
Computer use is 45x More Expensive Than Structured APIs

ARTICLE (postgres oops moment)
How Linux 7.0 Broke PostgreSQL

Want to reach 200,000+ engineers?

Let’s work together! Whether it’s your product, service, or event, we’d love to help you connect with this awesome community.

🔥 AI Load Breaks GitHub – Why Other Vendors Stay Strong (8 min)

Brief: GitHub suffers critical reliability meltdown with data integrity failures, multiple outages, and 85% uptime as AI agent load spikes 3.5x, prompting famous developers like HashiCorp's founder to abandon the platform, while competitors like GitLab and Vercel handle similar growth without major issues—exposing GitHub's infrastructure missteps and organizational bloat.

🚨 Google Chrome Silently Installs 4GB AI Model Without Consent—Here's the Climate Cost (27 min)

Brief: Google Chrome automatically downloads a 4GB Gemini Nano AI model to user devices without permission, persists even when deleted, and carries a massive carbon footprint estimated between 6,000–60,000 tonnes of CO2 equivalent—a practice that violates GDPR, ePrivacy law, and deceptive design standards while giving users zero transparency about the cloud-backed AI Mode that doesn't even use the local model.

🤖 Designing the AI-Native Engineering Organization (5 min)

Brief: Engineering leaders from Microsoft, 1Password, and Atlassian reveal how AI is reshaping dev workflows—compressing coding and operations while expanding planning and validation, shifting teams to smaller mission-driven squads, collapsing planning horizons to 90-day cycles, and demanding cost management rigor on token spend that rivals cloud infrastructure negotiations.

📱 OpenAI Fast-Tracking AI Phone for 2027 Launch, Says Kuo (3 min)

Brief: OpenAI is accelerating its AI agent phone to hit mass production in early 2027 instead of 2028, driven by its planned IPO and competition, with MediaTek as the lead processor supplier, a custom Dimensity 9600 chip, and an advanced image signal processor for enhanced camera perception—potentially reaching 30 million units by 2028.

🔒 Mozilla Hardens Firefox with AI-Powered Security Audits, Fixes 271 Bugs (5 min)

Brief: Mozilla used Claude Mythos Preview and other AI models to identify and fix an unprecedented 271 latent security bugs in Firefox, including complex sandbox escapes and use-after-free vulnerabilities that fuzzing alone couldn't detect, while building a scalable agentic harness pipeline that dynamically tests AI-generated hypotheses to eliminate false positives.

This week’s tip:

Instrument RED metrics (Rate, Errors, Duration) with OpenTelemetry Exemplars to link slow traces to metrics dashboards. Exemplars embed trace IDs in histogram buckets, enabling one-click drill-down from 'p95 latency spike' to the actual trace without log searching.

Wen?

• SLO burn debugging: Exemplars auto-link p99 latency alerts to slow traces, cutting MTTR from "search logs for timestamp" to "click exemplar, see trace graph".

• Capacity planning: Correlating error rate spikes with trace samples shows if errors are transient network blips or systematic (e.g., OOM), guiding scaling decisions.

• Incident postmortems: Collecting exemplars during the incident window creates a trail of representative traces at each percentile, preserving evidence without storing every trace.

The more people you help become successful the more successful you become.
Steve Harvey

That’s it for today! ☀️

Enjoyed this issue? Send it to your friends here to sign up, or share it on Twitter!

If you want to submit a section to the newsletter or tell us what you think about today’s issue, reply to this email or DM me on Twitter! 🐦

Thanks for spending part of your Monday morning with Hungry Minds.
See you in a week — Alex.

Icons by Icons8.

*I may earn a commission if you get a subscription through the links marked with “aff.” (at no extra cost to you).

Happy Monday! ☀️

Welcome to the 655 new hungry minds who have joined us since last Monday!

If you aren't subscribed yet, join smart, curious, and hungry folks by subscribing here.

📚 Software Engineering Articles

• How Amazon uses LLMs to recommend products by building a commonsense knowledge graph

• Agent harness engineering the system around the model matters more

• The peril of laziness lost LLMs lack the virtue that makes good software

🗞️ Tech and AI Trends

• PyTorch Lightning compromised in supply chain attack stealing credentials on import

• Mistral launches Medium 3.5 with cloud coding agents and Work mode

• Vercel breached via Context.ai — customer data sold for $2M

👨🏻‍💻 Coding Tip

• Use ripgrep to search at lightspeed

Time-to-digest: 5 minutes

How Amazon Uses LLMs to Recommend Products🛒 

Search "shoes for pregnant women" on Amazon and the best results are slip-resistant shoes—even though the word "pregnant" appears nowhere in those listings. Traditional recommendation systems match keywords and purchase history, but when a shopper's intent requires a reasoning step that lives entirely in human common sense, those systems hit a wall. Amazon built COSMO, a commonsense knowledge graph that teaches the recommendation engine to think the way a human shopper would.

The challenge: Factual knowledge graphs encode what a product is (brand, color, material), but not why a human would want it. A query like "winter clothes" implies warmth, but no product listing says "this coat is warm." You need structured commonsense knowledge at e-commerce scale—and it didn't exist.

Implementation highlights:

1. LLM-as-ore-mine generation: Fed 5M+ behavior pairs (query→purchase and co-purchase) into OPT-175B hosted on internal infrastructure (data privacy constraint). The LLM generated millions of candidate explanations, but only 35% of search-buy and 9% of co-purchase explanations met quality bars—the rest were circular or trivially obvious

2. Multi-stage filtration pipeline: Rule-based filters removed incomplete or parroted text; embedding similarity filters caught semantic paraphrases; 30K human annotations evaluated plausibility and typicality via 5 decomposed yes/no questions; DeBERTa-large classifiers generalized those judgments to all remaining candidates

3. Data-driven ontology: Instead of designing relations top-down, Amazon mined 15 relation types from recurring LLM output patterns (used_for_function, used_for_audience, used_in_location, etc.), then canonicalized them into a structured schema

4. COSMO-LM distillation: Instruction-tuned LLaMA 7B/13B on the annotated data across 18 domains, 15 relations, and 5 tasks—collapsing the "big LLM + classifier" stack into a single smaller model that both generates and self-evaluates knowledge

5. Two-tier serving architecture: Pre-loaded cache covers frequent searches; daily batch processing handles newer queries. SageMaker manages model refresh from daily session logs, feeding Search Relevance, Recommendation, and Navigation simultaneously

Results and learnings:

• 30K annotations → 29M knowledge graph edges: The leverage ratio across 18 product categories, achieved through heavy investment in sampling, annotation design, and classifier training rather than brute-force labeling

• Beat KDD Cup 2022 leaderboard: COSMO-augmented cross-encoder hit 73.48% Macro F1 and 90.78% Micro F1 on the public ESCI dataset, surpassing the top ensemble model. Frozen-encoder ablation showed a 60% Macro F1 lift from COSMO triples alone

• 0.7% relative sales increase → hundreds of millions in revenue: A/B tested on ~10% of U.S. traffic over several months, with an 8% navigation engagement lift—from a single small feature on the search page. Projected billions in revenue if extended across all traffic

Amazon proved that LLMs are powerful hypothesis generators but terrible oracles. COSMO works because the engineering effort went into filtration; throwing away 91% of co-purchase explanations. The system's daily refresh means it can't react to flash sales, and aggressive plausibility thresholds leave gaps in long-tail coverage, but Amazon chose precision over recall because unreliable commonsense in production is worse than missing knowledge.

How The Heck Does Shazam Work? (An Interactive Exploration)

Explore how Shazam and song identification work through interactive visualizations — spectrograms, constellation maps, hash fingerprints, and time-offset matching.

Floating point from scratch: Hard Mode

Actually building floating point from scratch!

Agent Harness Engineering

A coding agent is the model plus everything you build around it: prompts, tools, context policies, hooks, sandboxes, feedback loops. Harness engineering is t...

I Improved 15 LLMs at Coding in One Afternoon. Only the Harness Changed.

Cross-posted from X / @_can1357 In fact only the edit tool changed. That’s it. ¶ 0x0: The Wrong Question The conversation right now is almost entirely …

Agents can now create Cloudflare accounts, buy domains, and deploy

Starting today, agents can now be Cloudflare customers. They can create a Cloudflare account, start a paid subscription, register a domain, and get back an API token to deploy code right away. Humans can be in the loop to grant permission, but there’s no need to go to the dashboard, copy and paste API tokens, or enter credit card details.

You can beat the binary search

We sometimes have to look for a value in a sorted array. The simplest algorithm consists in just going through the values one by one, until we encounter the value, or exhaust the array. We sometimes call this algorithm a linear search. In C++, you can get the desired effect with the std::find function. For … Continue reading You can beat the binary search

ARTICLE (trust your agent, lose your data)
An AI Agent Just Destroyed Our Production Data 

ARTICLE (the CIA had a .bashrc too)
Cleaning Up Merged Git Branches: A One-Liner From the CIA's Leaked Dev Docs

ARTICLE (AI agent PRs a blog post roasting the maintainer) matplotlib/matplotlib#31132

ARTICLE (six phases from chatbot skeptic to overnight delegation)
My AI Adoption Journey — Mitchell Hashimoto

ARTICLE (laziness as a virtue, continued)
Laziness, Impatience, and Hubris — Adam Jacob

ESSENTIAL (laws that explain why your project is late)
The 20 Software Engineering Laws

GITHUB (agent skills from Google's eng culture)
addyosmani/agent-skills

GITHUB (the .claude directory everyone is forking)
mattpocock/skills 

GITHUB (Karpathy's CLAUDE.md in one file)
forrestchang/andrej-karpathy-skills

GITHUB (adaptive scraping that handles anti-bot like it's 2016)
D4Vinci/Scrapling

Want to reach 200,000+ engineers?

Let’s work together! Whether it’s your product, service, or event, we’d love to help you connect with this awesome community.

🔒 Vercel Breached via Third-Party AI Tool — Customer Credentials Stolen and Sold for $2M (3 min)
Brief: Hackers compromised Context.ai, a third-party AI tool used by a Vercel employee who had granted it OAuth access to their Google Workspace. The attackers pivoted from that foothold into Vercel's internal systems, accessing environment variables that weren't encrypted at rest. Stolen data appeared for sale on BreachForums at $2M. Vercel says Next.js and Turbopack were unaffected, but the incident is a textbook example of OAuth supply chain trust gone wrong.

☁️ Cloudflare and Stripe Launch Protocol That Lets AI Agents Create Accounts, Buy Domains, and Deploy (4 min)
Brief: Cloudflare and Stripe co-designed a new protocol that lets AI agents autonomously discover services, provision accounts, purchase domains, and deploy apps—all within a single agent session with zero human pre-configuration. If no Cloudflare account exists, one is created automatically. The protocol uses OAuth extended with payments and account creation, treating agents as first-class citizens. Any platform can integrate the same way Stripe does.

🤖 Google Releases Gemini 3 Flash — Pro-Grade Reasoning at Flash Speed and a Fraction of the Cost (2 min)
Brief: Google shipped Gemini 3 Flash, which outperforms Gemini 2.5 Pro while being 3x faster at $0.50/1M input tokens. It hits 78% on SWE-Bench Verified in Gemini CLI—beating even Gemini 3 Pro on agentic coding—and uses 30% fewer thinking tokens than 2.5 Pro on everyday tasks. It's now the default model in the Gemini app and AI Mode in Search, with two toggle modes: "Fast" for low-latency chat and "Thinking" for complex reasoning.

🔒 Critical GitHub RCE Flaw Lets Authenticated Users Execute Code via a Single Git Push (2 min)
Brief: CVE-2026-3854 is a high-severity remote code execution vulnerability in GitHub Enterprise Server where an authenticated user with push access could trigger arbitrary code execution through a crafted git push. GitHub patched it quickly with no evidence of pre-disclosure exploitation, but the flaw puts private repos, secrets, and deployment scripts at risk for any self-hosted GHES instance that hasn't updated.

🤖 OpenAI Drops Microsoft Exclusivity — Can Now Sell Across AWS and Google Cloud (2 min)
Brief: OpenAI quietly rewrote its deal with Microsoft to sell models and API access directly through AWS and Google Cloud, ending years of Azure exclusivity. The move lets enterprise customers run OpenAI models on their existing cloud provider without migration, dramatically expanding OpenAI's addressable market while signaling that no single cloud giant can lock in the AI API layer.

🤖 Google I/O 2026 Set for May 19–20 With "Agentic Era" Focus (1 min)
Brief: Google I/O returns May 19–20 with a keynote centered on the "agentic era" of development. Expect major updates across Android, AI, Chrome, and Cloud, with the event focused on new tools designed to automate complex workflows and simplify building AI-ready applications. Gemini 3.2 Flash has already been spotted testing on Eleuther AI Arena ahead of the event.

This week’s tip:

Leverage ripgrep's --type and --glob patterns with FPAT-like regex for multi-line matching without loading files into memory. ripgrep is orders of magnitude faster than grep/awk for code search, especially when combined with --multiline (-U) and --max-columns to handle generated or minified code safely.

Wen?

• Searching across large TypeScript/Rust codebases for async/await patterns: ripgrep's memory-mapped I/O and parallel search by default beats grep even with xargs -P.

• Finding multi-line regex in JSON config or log files: Use --multiline with lookahead to avoid OOM on huge files.

• Narrow search scope before slow operations: Combine with --files-with-matches -c to get counts first, then feed results to parallel processing tools like GNU parallel.

It’s always day one.
Jeff Bezos

That’s it for today! ☀️

Enjoyed this issue? Send it to your friends here to sign up, or share it on Twitter!

If you want to submit a section to the newsletter or tell us what you think about today’s issue, reply to this email or DM me on Twitter! 🐦

Thanks for spending part of your Monday morning with Hungry Minds.
See you in a week — Alex.

Icons by Icons8.

*I may earn a commission if you get a subscription through the links marked with “aff.” (at no extra cost to you).

Happy Monday! ☀️

Welcome to the 1245 new hungry minds who have joined us since last Monday!

If you aren't subscribed yet, join smart, curious, and hungry folks by subscribing here.

📚 Software Engineering Articles

• AI-powered UI test generation for iOS apps explained

• AWS lessons learned from 3000 incidents

• 20 software engineering laws you should know

• Inference caching in LLMs complete guide

• 5 silent failures in data pipelines

🗞️ Tech and AI Trends

• DeepSeek unveils flagship AI model one year later

• OpenAI announces GPT-5.5, latest AI model released

• John Ternus becomes Apple CEO as Tim transitions

👨🏻‍💻 Coding Tip

• Use native <dialog> with showModal() for accessible, focus-trapping modals without libraries

Time-to-digest: 5 minutes

Speeding up agentic workflows with WebSockets in the Responses API 🕷️

OpenAI's Codex agent loops dozens of back-and-forth API calls to complete tasks like debugging code. Each request validates, processes, and sends the full conversation history; a massive overhead that was eating latency gains from faster inference speeds. They needed to move the bottleneck away from the API layer.

The challenge: When inference speeds jump from 65 to 1,000 tokens per second, repeated API overhead becomes the wall you hit, not the model. You can't just make each request faster—you need to eliminate redundant work entirely.

Implementation highlights:

1. Persistent WebSocket connections: Replace synchronous HTTP calls with a single long-lived connection that maintains state across the agent loop lifecycle

2. In-memory response caching: Store previous response state (tokens, tool definitions, sampling artifacts) on the connection to skip rebuilding full conversation history

3. Minimal API surface: Use familiar response.create + previous_response_id Instead of new async patterns; developers integrate without rewriting code

4. Selective reprocessing: Run safety classifiers and validators only on new input, not the entire conversation history every round

5. Overlapped execution: Handle non-blocking work (billing, logging) asynchronously while subsequent requests are already processing

Results and learnings:

• 40% end-to-end speedup: Agent workflows became dramatically faster across Vercel SDK, Cline, and Cursor integrations

• Infrastructure matched inference: Hit the 1,000 TPS target for GPT-5.3-Codex-Spark with bursts reaching 4,000 TPS in production

• Developer adoption was instant: Zero-friction migration path meant adoption ramped up immediately across the community

OpenAI proved that when your model gets faster, your infrastructure has to get faster too. Smart API design isn't about adding features—it's about removing friction so users actually feel the speed they're paying for.

Record, generate, run: AI-powered UI test generation for iOS

Discover how Grab's innovative Mobile UI Testing AI Workflow transforms the way developers create and execute UI tests. By leveraging AI-powered automation, our system captures real user interactions to generate comprehensive, executable test scripts in minutes. Join us on this journey to enhance test coverage, improve reliability, and accelerate development cycles. Dive into our blog post to learn more about the architecture, best practices, and how you can be part of this exciting advancement in mobile app testing.

AWS Distinguished Eng: Learnings From 3000 Incidents And How Engineering Is Changing | Marc Brooker

Where caching is bad, thoughts on the industry, and learnings across his career

The 20 Software Engineering Laws

A field guide to why software projects fail, systems rot, and teams slow down.

A Well-Designed JavaScript Module System is Your First Architecture Decision | CSS-Tricks

Behind every technology, there should be a guide for its use. While JavaScript modules make it easier to write “big” programs, if there are no principles or systems for using them, things could easily become difficult to maintain.

The Complete Guide to Inference Caching in LLMs

Inference caching reduces latency and cost by storing and reusing computation from previous LLM requests instead of recomputing everything each time. It operates across three complementary layers: KV caching within a request, prefix caching across shared prompts, and semantic caching that reuses full responses for similar queries.

ARTICLE (ai startup vibes)
How an AI-Native Startup From SF Works and Builds Its Product

ARTICLE (self-care code edition)
Build yourself flowers

ARTICLE (images go squish)
The end of responsive images

ARTICLE (robots do chores)
Agents as scaffolding for recurring tasks

ESSENTIAL (time money brain hurt)
Extreme Time Value of Money: Late-stage Career Planning

ARTICLE (stories go zoom)
Interactive Storytelling for the Web: Building Immersive Stories with Timelines, 3D, and Layered Scenes

ARTICLE (remote dev nest tour)
A Deep Dive into My Remote Development Setup

ESSENTIAL (sql brain juice)
Mastering CTEs in SQL

ARTICLE (data oopsie traps)
The 5 Silent Failures in Data Pipelines

Want to reach 200,000+ engineers?

Let’s work together! Whether it’s your product, service, or event, we’d love to help you connect with this awesome community.

🤖 DeepSeek Launches V4 Flash and V4 Pro AI Models to Challenge OpenAI and Anthropic (3 min)

Brief: China's DeepSeek unveiled its V4 Flash and V4 Pro series, claiming they're the most powerful open-source AI models, featuring improved coding benchmarks, advanced reasoning capabilities, a new Hybrid Attention Architecture for better conversation memory, and a 1 million-token context window that allows entire codebases to be processed as single prompts.

🤖 Meta Cuts 10% of Workforce to Double Down on AI Race (3 min)

Brief: Meta is laying off 8,000 employees (10% of workforce) starting May 20 and canceling 6,000 open positions as it pivots aggressively toward generative AI to catch competitors like OpenAI and Google, following multiple smaller cuts since January.

🤖 OpenAI Launches GPT-5.5 With Advanced Coding and Research Capabilities (2 min)

Brief: OpenAI unveiled GPT-5.5, its latest AI model excelling at coding, computer use, and research, now rolling out to paid subscribers across ChatGPT and Codex, as the company races to compete with rivals like Anthropic's Claude Mythos amid intensifying competition in the AI sector.

💰 Amazon Pumps $25 Billion Into Anthropic as AI Infrastructure Race Heats Up (3 min)

Brief: Amazon is doubling down on AI infrastructure with a $25 billion investment in Anthropic (on top of $8 billion already committed), while Anthropic pledges to spend over $100 billion on AWS services over the next decade and secure 5 gigawatts of capacity for its Claude AI models—a massive power play as tech giants race to build out AI infrastructure.

🤖 'Tokenmaxxing' Goes Wrong: How Tech Giants Gamified AI Spending Into Massive Waste (8 min)

Brief: Tech companies including Meta, Microsoft, and Salesforce created internal token leaderboards to boost AI adoption, but the gamified competition backfired spectacularly—engineers are now burning billions in tokens on throwaway work just to climb rankings, with Meta alone spending 60.2 trillion tokens monthly (potentially $100M+ in costs) on largely wasteful tasks while only Shopify's approach with circuit breakers and transparent oversight proved effective at avoiding the trap.

🍎 Tim Cook Steps Down as CEO, John Ternus Takes the Helm at Apple (3 min)

Brief: Apple announces Tim Cook will become executive chairman while John Ternus, senior vice president of Hardware Engineering, becomes CEO effective September 1, 2026, following a long-term succession plan that grows the company from $350B to $4 trillion in market cap under Cook's 15-year leadership.

🤖 SpaceX and Cursor Team Up to Build the World's Best Coding AI (2 min)

Brief: SpaceXAI and Cursor are partnering to develop advanced coding AI by combining Cursor's AI software platform and engineer user base with SpaceX's Colossus supercomputer (equivalent to a million H100s), with SpaceX holding an option to acquire Cursor for $60 billion later this year.

This week’s tip:

Use the native <dialog> element with showModal() and inert attribute to create accessible, focus-trapping modals without third-party libraries, ensuring proper ARIA semantics and keyboard handling. Browsers automatically manage focus-within-dialog and backdrop interactions.

Wen?

• Focus management: showModal() automatically traps focus and restores it on close; no manual tabindex juggling needed.

• Form submissions: Using method="dialog" simplifies close handling and value passing to parent scripts.

• Multiple modals: Nested dialogs with inert on backdrop content prevent interactive elements outside active dialog from receiving focus.

Your dream has to be bigger than your fear.
Steve Harvey

That’s it for today! ☀️

Enjoyed this issue? Send it to your friends here to sign up, or share it on Twitter!

If you want to submit a section to the newsletter or tell us what you think about today’s issue, reply to this email or DM me on Twitter! 🐦

Thanks for spending part of your Monday morning with Hungry Minds.
See you in a week — Alex.

Icons by Icons8.

*I may earn a commission if you get a subscription through the links marked with “aff.” (at no extra cost to you).

Happy Monday! ☀️

Welcome to the 879 new hungry minds who have joined us since last Monday!

If you aren't subscribed yet, join smart, curious, and hungry folks by subscribing here.

📚 Software Engineering Articles

• Building the foundation for running extra-large language models

• Securing every Kubernetes workload at scale

• Netflix's operations layer behind live at scale

• Relational database design guide: master the fundamentals

• Why MicroVMs power Docker's sandbox architecture

🗞️ Tech and AI Trends

• Google opens Gemma 4 under Apache 2.0 with multimodal capabilities

• Claude Opus 4.7 released: faster reasoning, better performance

• Zero-copy GPU inference from WebAssembly on Apple Silicon

👨🏻‍💻 Coding Tip

• React Server Components eliminate client-side data fetching, reduce bundle size dramatically

Time-to-digest: 5 minutes

How NVIDIA built a multilingual OCR model using 12M synthetic images 🗄️

NVIDIA's OCR v1 model was solid for English but completely fell apart on other languages; we're talking 0.56 to 0.92 NED scores (basically gibberish). Instead of manually annotating millions of images across six languages, they went all-in on synthetic data generation and built Nemotron OCR v2.

The challenge: Train a single unified OCR model that handles English, Chinese, Japanese, Korean, and Russian without language-specific variants — and do it without the prohibitive cost of annotating millions of real-world images with word-, line-, and paragraph-level bounding boxes.

Implementation highlights:

1. Synthetic data pipeline: Generate 12.2M training images with pixel-perfect ground truth by rendering text from mOSCAR (a 163-language web corpus) onto randomized backgrounds with diverse fonts, layouts, and augmentations

2. Hierarchical annotations for free: Every synthetic image comes with word, line, and paragraph bounding boxes plus a reading order graph — labels that would cost a fortune to annotate manually

3. Line-level recognition for CJK: Chinese and Japanese don't use spaces between words, so they switched from word-level to line-level recognition — eliminating the need for a separate word segmentation step

4. Shared backbone architecture: Based on the FOTS design, a single RegNetX-8GF backbone processes the image once, then detection, recognition, and relational models all reuse those features — no redundant computation

5. Language-agnostic extensibility: Adding a new language only requires source text + fonts. No architecture changes, no manual annotation. The pipeline generates millions of annotated pages per day on a single machine

Results and learnings:

• From gibberish to near-perfect: NED scores on non-English languages dropped from 0.56–0.92 down to 0.035–0.069 — a single model outperforming even language-specific competitors

• Absurdly fast: 34.7 pages/second on a single A100 GPU, 28x faster than PaddleOCR v5's server pipeline

• Data > architecture: Expanding the character set alone barely helped. The model could theoretically output CJK characters — it just had never learned what they looked like

NVIDIA's approach is a masterclass in letting synthetic data do the heavy lifting. If you can programmatically generate your labels, you skip the most expensive part of ML; and you get to control exactly what your model sees during training. The dataset and model are both open, so you can extend this to your own languages today.

Data Pipeline Foundations - Everything You Need To Know About Data Pipelines

Hi, fellow future and current Data Leaders; Ben here 👋

Building the foundation for running extra-large language models

We built a custom technology stack to run fast large language models on Cloudflare’s infrastructure. This post explores the engineering trade-offs and technical optimizations required to make high-performance AI inference accessible.

Securing every Kubernetes workload at scale

A Guide to Relational Database Design

In this article, we cover the core concepts that inform those decisions. We’ll look at tables, keys, relationships, normalization, and joins, with each concept building on the last.

Why MicroVMs: The Architecture Behind Docker Sandboxes | Docker

Learn from Docker experts to simplify and advance your app development and management with Docker. Stay up to date on Docker events and new version

ARTICLE (design-code-tango)
Figma Design to Code, Code to Design: Clearly Explained

GITHUB REPO (stack-em-high)
Stacked PRs

ARTICLE (math-goes-brrr)
All elementary functions from a single binary operator

ARTICLE (protocol-party-time)
Scaling MCP adoption

ARTICLE (money-moves-smooth)
Engineering the Forge Billing Platform for Reliability and Scale

ARTICLE (tile-time-magic)
Frontend Engineering at Palantir: Polar Scaled Tiles in Zodiac

ARTICLE (robot-future-vibes)
The Agent Stack Bet

ARTICLE (slice-it-vertical)
Vertical Slice Architecture in Node.js: One Folder Per Use Case

ARTICLE (gpt-plays-favorites)
Why ChatGPT Cites One Page Over Another

Want to reach 200,000+ engineers?

Let’s work together! Whether it’s your product, service, or event, we’d love to help you connect with this awesome community.

🤖 Google Opens Gemma 4 Under Apache 2.0 with Multimodal and Agentic Capabilities (2 min)

Brief: Google releases Gemma 4, an open-weight model family with Apache 2.0 licensing spanning 2B-31B parameters, featuring native video, image, and audio processing, 256K context windows, and agentic capabilities like function-calling and structured JSON output—with the 31B variant matching performance of models 3-5x larger and broad distribution across Hugging Face, Kaggle, vLLM, Ollama, and other platforms.

🚀 What is Inference Engineering? A Deep Dive into Serving AI Models at Scale (10 min)

Brief: As open LLM models grow increasingly capable, inference engineering—the practice of optimizing how AI models run in production—is shifting from a niche skill at frontier labs to a critical competency for any company building AI products; key techniques include quantization, speculative decoding, caching, parallelism, and disaggregation to achieve faster, cheaper, and more reliable model serving.

🚀 Anthropic Releases Claude Opus 4.7 with Major Coding and Vision Upgrades (4 min)

Brief: Anthropic launched Claude Opus 4.7, a significant improvement over Opus 4.6 that excels in advanced software engineering, complex long-running tasks, and vision capabilities—processing images at 3.75 megapixels—while introducing cybersecurity safeguards and new effort control options, available across all Claude products at unchanged pricing.

🎨 Anthropic Launches Claude Design to Democratize Visual Creation (3 min)

Brief: Anthropic unveils Claude Design, an AI-powered tool that lets teams collaborate with Claude to create polished designs, prototypes, and presentations by describing what they need, with features like automatic brand consistency, inline editing, and seamless handoffs to code—available now for Claude Pro, Max, Team, and Enterprise subscribers.

🤖 OpenAI Upgrades Codex with Computer Control, Image Generation, and Advanced Developer Tools (4 min)

Brief: OpenAI significantly expands Codex capabilities for its 3M+ weekly users, adding computer control to operate apps, image generation, native web browsing, 90+ new plugins, and memory features to streamline the entire software development lifecycle from coding to deployment.

🚀 Zero-Copy GPU Inference from WebAssembly on Apple Silicon (5 min)

Brief: A developer demonstrates that WebAssembly modules can share memory directly with Apple Silicon GPUs without copying data, using the chip's unified memory architecture to run stateful AI inference with portable KV caches that can be serialized, restored across machines, and recovered 5.45× faster than recomputation.

This week’s tip:

Use React Server Components (RSCs) with Server Actions to eliminate client-side data-fetching boilerplate and reduce JavaScript bundle size by moving logic server-side. RSCs render on server, stream to client, and support direct database/API access without intermediate API endpoints.

Wen?

• Data-heavy dashboards: Fetch and filter data server-side; stream only rendered HTML.

• Secure mutations: Run sensitive operations (auth checks, DB writes) in Server Actions; client never sees tokens or credentials.

• Progressive enhancement: RSCs work without client JavaScript; add interactivity with 'use client' islands only where needed.

Greatness comes from living with purpose and passion.
Ralph Marston

That’s it for today! ☀️

Enjoyed this issue? Send it to your friends here to sign up, or share it on Twitter!

If you want to submit a section to the newsletter or tell us what you think about today’s issue, reply to this email or DM me on Twitter! 🐦

Thanks for spending part of your Monday morning with Hungry Minds.
See you in a week — Alex.

Icons by Icons8.

*I may earn a commission if you get a subscription through the links marked with “aff.” (at no extra cost to you).

Happy Monday! ☀️

Welcome to the 651 new hungry minds who have joined us since last Monday!

If you aren't subscribed yet, join smart, curious, and hungry folks by subscribing here.

📚 Software Engineering Articles

• Backend for AI-coded apps reveals modern architecture patterns

• Unwritten laws of software engineering every developer should know

• Netflix's interval-aware caching scales Druid efficiently

• Research-driven agents code smarter by reading first

• Game engines teach databases forgotten data patterns

🗞️ Tech and AI Trends

• AI agent benchmarks broken by Berkeley researchers; what's next?

• Anthropic's Project Glasswing secures AI software era

• MIT's leaner AI models learn faster during training

👨🏻‍💻 Coding Tip

• Navigation API and URLPattern decouple routing from history, enabling fine-grained back-button control

Time-to-digest: 5 minutes

How Meta Used AI to Map Tribal Knowledge in Large-Scale Data Pipelines 🔥

Meta faced a sprawling data pipeline: 4,100+ files across three languages and four repositories. When they pointed AI agents at it, the agents got lost fast—missing naming conventions, breaking serialization compatibility, and producing code that compiled but was subtly wrong. The real problem? All the knowledge lived in engineers' heads.

The challenge: Give AI agents a map of undocumented tribal knowledge without overwhelming their context windows or creating stale documentation that causes more harm than help.

Implementation highlights:

1. Swarm of specialized agents: Deployed 50+ AI agents in orchestrated phases—explorers, analysts, writers, critics, and fixers—each answering specific questions about the codebase

2. The five-question framework: Every module analyst extracted what it configures, common patterns, non-obvious gotchas (like hidden field renames), dependencies, and undocumented conventions

3. Compass, not encyclopedia: Generated 59 concise context files (~1,000 tokens each) with quick commands, key files, non-obvious patterns, and cross-references—zero fluff

4. Multi-round quality gates: Three independent critic passes improved scores from 3.65 to 4.20/5.0, eliminated hallucinations, and verified every file reference

5. Self-healing automation: Periodic jobs validate paths, detect gaps, re-run critics, and auto-fix stale context—because documentation debt compounds faster than interest

Results and learnings:

• Coverage exploded: AI navigation jumped from 5% to 100% of modules (50 to 4,100+ files) while documenting 50+ non-obvious patterns that never existed on paper

• Efficiency gains: Agents used 40% fewer tool calls per task; research that took two days now completes in 30 minutes

• Model-agnostic: The knowledge layer works across leading models because it's structured data, not proprietary integration

Meta proved that tribal knowledge isn't a feature bug—it's a context opportunity. By turning undocumented expertise into machine-readable navigation, they didn't just help AI agents; they created a system that improves with every task it solves.

The best part? Your codebase probably has the same gaps. Start with the five questions, keep context tight, and let your critics do the gatekeeping. Your future self (and your AI agents) will thank you.

The unwritten laws of software engineering

Last April, I wrote a well-received article about the 13 software engineering laws - Hyrum’s, Conway’s, Zawinski’s, and 10 famous others.

A backend for AI-coded apps

Instant 1.0 is out! This essay shows a bunch of demos, to explain why we think Instant is the best backend for AI-coded apps. We also cover the architecture that makes all of it work.

Trust But Canary: Configuration Safety at Scale

As AI increases developer speed and productivity it also increases the need for safeguards. On this episode of the Meta Tech Podcast, Pascal Hartig sits down with Ishwari and Joe from Meta’s …

Research-Driven Agents: What Happens When Your Agent Reads Before It Codes

Coding agents working from code alone generate shallow hypotheses. Adding a research phase — arxiv papers, competing forks, other backends — produced 5 kernel fusions that made llama.cpp CPU inference 15% faster.

Frontend Engineering at Palantir: Building a Backend-less Cross-Application API

How we enabled Palantir apps to easily talk with each other with no server involved.

ARTICLE (game brain go brrr)
Introduction to Reinforcement Learning Agents with the Unity Game Engine 

ARTICLE (supercomputer but make it chill)
Monarch: an API to your supercomputer

ESSENTIAL (chaos management 101)
Why Distributed Systems Fail and How to Limit the Damage

ARTICLE (robot butler energy)
Deep Agents Deploy: an open alternative to Claude Managed Agents 

ARTICLE (databases are speed runners)
What Game Engines Know About Data That Databases Forgot

ESSENTIAL (robot thinks it thinks)
What Is the AI Agent Loop? The Core Architecture Behind Autonomous AI Systems

ARTICLE (code go cheap cheap)
Code Is Cheap Now, And That Changes Everything

ARTICLE (glow up your skillset)
Using skills

ARTICLE (shell = life hack champion)
Shell tricks that actually make life easier

Want to reach 200,000+ engineers?

Let’s work together! Whether it’s your product, service, or event, we’d love to help you connect with this awesome community.

🔓 Researchers Broke Top AI Agent Benchmarks—Here's What Comes Next (15 min)

Brief: UC Berkeley researchers built an automated exploit agent that achieved near-perfect scores on eight major AI benchmarks (SWE-bench, WebArena, OSWorld, GAIA, and others) without solving a single task, exposing critical vulnerabilities in how we measure AI agent capability—from isolated evaluation environments to public answer leaks—and introducing BenchJack, a tool to adversarially test benchmarks before publication.

🛡️ Project Glasswing: Tech Giants Unite to Defend Critical Software Against AI-Powered Cyber Threats (5 min)

Brief: Anthropic launches Project Glasswing with 12 major tech companies and organizations to use Claude Mythos Preview—an AI model that can find and exploit software vulnerabilities at human-expert levels—for defensive cybersecurity, committing $100M in usage credits and $4M in donations to secure the world's critical infrastructure before such capabilities fall into malicious hands.

🛠️ Eight Years of Waiting, Three Months of Building: How AI Finally Made SQLite DevTools Possible (12 min)

Brief: A developer shares how AI coding agents enabled him to ship syntaqlite, professional-grade SQLite developer tools he'd wanted for 8 years, in just 3 months—but reveals the project required a complete rewrite after "vibe-coding" initially produced unmaintainable spaghetti code, highlighting that AI excels at implementation but fails at design decisions.

⚡ New Technique Makes AI Models Leaner and Faster While They're Still Learning (3 min)

Brief: MIT researchers developed CompreSSM, a technique that compresses AI models during training rather than after, using control theory to identify and remove unnecessary components early on, achieving up to 4x training speedups while maintaining performance comparable to larger models.

Failed to summarize "System Card: Claude Mythos Preview [pdf]"

🚀 Artemis II Crew Splashes Down Near San Diego After Historic Moon Mission (5 min)

Brief: The four Artemis II astronauts successfully returned to Earth after a historic 10-day mission to the moon, splashing down in the Pacific Ocean off San Diego on April 10, 2026 at 8:07 p.m. ET in a "perfect descent" before being airlifted to recovery ship USS John P. Murtha, setting a new record for farthest distance humans have traveled from Earth at 252,756 miles.

This week’s tip:

Exploit the Navigation API and URLPattern to decouple client-side routing from browser history state, enabling fine-grained control over back-button behavior and form state restoration in SPAs. The Navigation API provides currEntry, canGoBack/canGoForward, and scroll restoration hooks; URLPattern allows regex-like route matching without manual URL parsing, making it easier to serialize form state into search params without mutations.

Wen?

• Complex forms with multi-step workflows: Store intermediate state in search params; Navigation API scroll restoration can restore scroll position per step.

• Search and filter UIs: URLPattern + search params create shareable/bookmarkable results; back-button correctly restores both route and scroll without re-fetching.

• Hybrid content (SSR + SPA): Server renders initial page, then Navigation API intercepts subsequent navigations, blending full-page and client-side navigation benefits.

It's necessary to get the losers out of your life if you want to live your dream. Les Brown

That’s it for today! ☀️

Enjoyed this issue? Send it to your friends here to sign up, or share it on Twitter!

If you want to submit a section to the newsletter or tell us what you think about today’s issue, reply to this email or DM me on Twitter! 🐦

Thanks for spending part of your Monday morning with Hungry Minds.
See you in a week — Alex.

Icons by Icons8.

*I may earn a commission if you get a subscription through the links marked with “aff.” (at no extra cost to you).

Happy Monday! ☀️

Welcome to the 227 new hungry minds who have joined us since last Monday!

If you aren't subscribed yet, join smart, curious, and hungry folks by subscribing here.

📚 Software Engineering Articles

• Why distributed systems fail and how to limit damage

• Agentic engineering insights from Lenny's podcast conversation

• Claude Code's source leak reveals production agent design patterns

• The seniority trap: avoiding stagnation after years

• Database performance strategies and their hidden costs explained

🗞️ Tech and AI Trends

• Oracle lays off workers amid AI investment surge

• Google releases Gemma 4 open models for developers

• OpenAI closes record funding round in Silicon Valley

👨🏻‍💻 Coding Tip

• URLPattern API simplifies complex URL routing with named capture groups and regex alternation patterns

Time-to-digest: 5 minutes

How Etsy uses LLMs to improve search relevance 🦭

Etsy's search used to rely heavily on engagement signals like clicks and purchases to rank results. The problem? Popular listings kept winning, even when they weren't the best match. So they built a Semantic Relevance Framework powered by LLMs to understand what buyers actually mean when they search.

The challenge: Measure and improve how well search results match user intent at scale, without replacing human judgment or slowing down real-time search.

Implementation highlights:

1. Human-anchored golden labels: Curate a dataset of query-listing pairs labeled as relevant, partially relevant, or irrelevant by trained annotators, with evolving guidelines that reflect cultural shifts (think "face masks" pre vs. post 2020)

2. LLM-as-a-judge with guardrails: Use o3 with few-shot chain-of-thought prompting and self-consistency sampling to generate millions of labels, but only after validating alignment with human judgments

3. Three-tier distillation pipeline: Cascade from the LLM annotator → a fine-tuned Qwen 3 VL 4B teacher → a lightweight BERT-based two-tower student model, each trading accuracy for speed

4. Sub-10ms real-time inference: The student model plugs directly into the search stack for filtering irrelevant results, enriching ranking features, and boosting highly relevant listings — all with <10ms added latency

5. Daily offline evaluation at scale: The teacher model scores millions of query-listing pairs daily via vLLM, powering relevance dashboards and A/B test guardrails at near-zero cost

Results and learnings:

• Measurable relevance uplift: Fully relevant listings in search results jumped from 58% to 62% in just three months

• Fairer visibility: Semantic relevance helps surface small and new sellers who lack engagement history but offer great matches

• Engagement-relevance tension: Improving semantic relevance can decrease engagement metrics — a known pattern across e-commerce that requires adaptive, query-aware strategies

Etsy's playbook is a masterclass in scaling human judgment with LLMs without losing the human in the loop. The three-tier distillation approach is something any search team can steal: let the expensive model teach, and let the cheap model serve.

Diving into Claude Code's Source Code Leak

Engineer’s Codex is a publication about real-world software engineering.

Why Distributed Systems Fail and How to Limit the Damage

A look at the most common failure modes and the techniques to reduce their impact.

Highlights from my conversation about agentic engineering on Lenny’s Podcast

I was a guest on Lenny Rachitsky’s podcast, in a new episode titled An AI state of the union: We’ve passed the inflection point, dark factories are coming, and automation …

The Seniority Trap: Why You Might Have 1 Year of Experience, 10 Times

We’ve all seen the “Senior Software Engineer” LinkedIn profiles with 10+ years of experience. But in the world of engineering, time is a deceptive metric. There is a massive, quiet difference between an engineer who has spent a decade solving increasingly complex, novel problems and an engineer who has spent a decade solving the same CRUD app problems using the same patterns they learned in year one.

Programming (with AI agents) as theory building

Components of A Coding Agent

How coding agents use tools, memory, and repo context to make LLMs work better in practice

GITHUB REPO (tmux-squad-up)
jmux – tmux-based development environment for humans and coding agents

ARTICLE (bug-squash-party)
Building a Python Workflow That Catches Bugs Before Production 

ESSENTIAL (startup-wisdom-drops)
A Student's Guide to Startups

ARTICLE (box-drawing-oopsies)
7 more common mistakes in architecture diagrams

ARTICLE (speed-versus-wallet)
Database Performance Strategies and Their Hidden Costs

ARTICLE (robots-fixing-robots)
How My Agents Self-Heal in Production

ARTICLE (javascript-never-stops)
What To Know in JavaScript (2026 Edition)

ARTICLE (ai-garbage-incoming)
Agentic slop PRs

Want to reach 200,000+ engineers?

Let’s work together! Whether it’s your product, service, or event, we’d love to help you connect with this awesome community.

🚀 Google DeepMind Launches Gemma 4: Open-Source AI Models Built on Gemini 3 Technology (3 min)

Brief: Google DeepMind introduces Gemma 4, its most intelligent open-source AI models available in multiple sizes (E2B, E4B, 26B, and 31B), designed to maximize intelligence-per-parameter with support for agentic workflows, multimodal reasoning across audio and vision, and 140 languages, while running efficiently on personal computers, mobile devices, and edge hardware with industry-leading performance across benchmarks.

🤖 Meet Cursor 3: A Unified Workspace for AI-Powered Software Development (3 min)

Brief: Cursor launches Cursor 3, a redesigned agent-first workspace that lets developers manage multiple agents across repositories, seamlessly handoff between local and cloud agents, and ship code faster with an integrated browser, simplified diffs view, and plugin marketplace—marking the shift toward autonomous software development.

🏢 Oracle Lays Off Thousands of Workers Amid Heavy AI Investment (3 min)

Brief: Oracle begins cutting an estimated 20,000–30,000 employees, roughly 18% of its workforce, to free up $8–10B in annual cash flow and fund its aggressive $50B+ AI data center buildout, as the company faces investor pressure over rising debt and dwindling free cash flow despite a 95% jump in net income and $523B in contracted future revenue.

🤖 Meet the Startup That Used AI and OpenClaw to Automate Its Own Developers (4 min)

Brief: Silicon Valley fintech startup JustPaid built a team of seven fully autonomous AI agents using OpenClaw as the orchestration brain and Claude Code as the coding engine; shipping 10 major features in one month that would have taken human developers a year; at a cost of $10K–$15K/month, signaling the rise of agentic software development while raising cybersecurity and existential workforce concerns.

💰 OpenAI Closes Silicon Valley's Largest-Ever Funding Round (3 min)

Brief: OpenAI raises $122B in committed capital at an $852B valuation; the largest private funding round in Silicon Valley history; anchored by Amazon ($50B), Nvidia ($30B), and SoftBank ($30B), with $3B from individual investors for the first time, as the company hits $2B/month in revenue, 900M weekly ChatGPT users, and positions for a potential IPO while narrowing its product focus toward enterprise and coding tools.

This week’s tip:

Use URLPattern to parse and validate complex URL schemes with capture groups and regex alternation. The constructor accepts a pattern object with pathname, search, and hash fields. Use .test(url) for validation and .exec(url) for extraction into named groups. Ideal for routing in service workers, micro-frontends, and dynamic imports where traditional pathname matching falls short. Works in modern browsers and Deno.

Wen?

• Service worker routing with versioned APIs (v1, v2, etc.)

• Micro-frontend module federation where path patterns dictate component loading

• Parameter extraction in edge runtime URL rewriting without full URL parser overhead

• Validation of callback URLs in OAuth flows with protocol/port constraints

Don't be pushed around by the fears in your mind. Be led by the dreams in your heart.
Roy T. Bennett

That’s it for today! ☀️

Enjoyed this issue? Send it to your friends here to sign up, or share it on Twitter!

If you want to submit a section to the newsletter or tell us what you think about today’s issue, reply to this email or DM me on Twitter! 🐦

Thanks for spending part of your Monday morning with Hungry Minds.
See you in a week — Alex.

Icons by Icons8.

*I may earn a commission if you get a subscription through the links marked with “aff.” (at no extra cost to you).

Happy Monday! ☀️

Welcome to the 351 new hungry minds who have joined us since last Monday!

If you aren't subscribed yet, join smart, curious, and hungry folks by subscribing here.

📚 Software Engineering Articles

• Discord adds distributed tracing to Elixir without performance hit

• Apple's synthetic task generation scales AI agent training efficiently

• Harness design unlocks long-running application development secrets

• PostgreSQL indexes explained: optimize queries like a pro

• Dependency injection in Node.js and TypeScript done right

🗞️ Tech and AI Trends

• Google's TurboQuant slashes LLM memory by 6x

• MIT engineers design proteins by motion, not just structure

• Apple Business platform launches for all company sizes

👨🏻‍💻 Coding Tip

• Consistent hashing with virtual nodes rebalances distributed caches seamlessly when infrastructure changes

Time-to-digest: 5 minutes

How Netflix Live Streams to 100 Million Devices in 60 Seconds 📹

Netflix built the Live Origin—a custom origin server that sits between their encoding pipelines and Open Connect CDN—to handle the brutal constraints of real-time video delivery. Unlike on-demand content, live streams have a 2-second window to encode, package, and deliver segments to millions of concurrent viewers worldwide.

The challenge: Build a system that can deliver video segments in milliseconds while handling defective content, traffic storms, and 65 million concurrent streams without breaking a sweat.

Implementation highlights:

1. Dual redundant pipelines across regions: Run two independent encoding pipelines simultaneously with different cloud regions, encoders, and video sources so when one produces a bad segment, the other typically doesn't

2. Intelligent segment validation: Use lightweight media inspection metadata to detect defects (missing frames, timing issues, short segments) and automatically serve the best available version to viewers

3. Predictable manifest templates with fixed segment duration: Use 2-second fixed-duration segments with templated URLs instead of constantly updating manifests, allowing the Origin to predict exactly when segments arrive

4. Millisecond-grain HTTP caching and request holding: Cache 404s with expiration policies and hold requests open for segments about to publish, eliminating redundant network roundtrips at the live edge

5. Separated storage paths with write-through caching: Isolate publishing writes from CDN reads using Apache Cassandra + EVCache, achieving 25ms median latency and handling 200+ gigabits per second read throughput

Results and learnings:

• Blazing fast delivery: Reduced median latency from 113ms to 25ms while meeting strict 2-second retry budgets

• Handles massive scale: Successfully delivered 65 million concurrent streams during the Tyson vs. Paul fight without degradation

• Traffic storm resilient: In-memory metadata caching achieves 90%+ cache hit ratios during 404 storms, protecting the datastore from cascading failures

Netflix's approach proves that building for live doesn't mean sacrificing reliability. Their obsessive focus on redundancy, intelligent selection, and traffic isolation creates a system that's both performant and forgiving—critical when millions of people are watching simultaneously.

Scaling Synthetic Task Generation for Agents via Exploration

Post-Training Multimodal Large Language Models (MLLMs) to build interactive agents holds promise across domains such as computer-use, web…

Discord Engineers Add Distributed Tracing to Elixir's Actor Model Without Performance Penalty

Discord engineering detailed how they added distributed tracing to Elixir's actor model. Their custom Transport library wraps messages with trace context and uses dynamic sampling to handle million-user fanouts. CPU optimizations included skipping unsampled traces and filtering context before deserialization, recovering 10+ percentage points of overhead.

Async systems scale your system… and your problems.

Many developers only realize it after consistency breaks.

The 53 Concepts for Highly Effective Mobile System Design

#135: Part 1 - Client-Server Architecture, Push Notifications, Offline-First, and 16 others.

3 Things I'm Unlearning from Big Tech

For nearly a decade, my identity was defined by a single title: Software Engineer. It’s an identity I picked up early, starting in a Computer Science lecture hall and carrying it through to the offices of Google and Amazon. In the corporate world, this title is a comfortable box, but it also trapped me.

ARTICLE (oops handler supreme)
5 Useful DIY Python Functions for Error Handling

ESSENTIAL (startup vibes from the wise one)
A Student's Guide to Startups

ARTICLE (database speed demon)
Introduction to PostgreSQL indexes

ARTICLE (mobile app go brrrr)
From skeptic to convert: how Fieldy adopted Expo for their AI wearable

ARTICLE (robot whisperer skills)
The Skill of Using AI Agents Well

ARTICLE (code ninja secrets)
What Really Makes a Succesful Software Engineer

ARTICLE (platform brain go zoom)
AI Hot Takes From A Platform Engineer / SRE

ARTICLE (train tracks teach code)
What Construction at a Train Station Taught Me About Software Engineering

ARTICLE (inject the good stuff)
Dependency Injection in Node.js & TypeScript. The Part Nobody Teaches You

Want to reach 200,000+ engineers?

Let’s work together! Whether it’s your product, service, or event, we’d love to help you connect with this awesome community.

🎯 Jensen Huang's Masterclass in Manifesting the Future: How Nvidia's CEO Shapes Reality Through Strategic Vision (12 min)

Brief: At Nvidia's GTC 2026 conference, CEO Jensen Huang deployed strategic messaging and vision-setting to shape industry beliefs around AI factories, autonomous agents, and token consumption, positioning Nvidia as essential infrastructure while convincing supply chain partners and enterprises that persistent AI agents will become mandatory—ultimately driving GPU demand without directly selling chips.

🧬 MIT Engineers Design Proteins by Their Motion, Not Just Their Shape (4 min)

Brief: MIT researchers developed VibeGen, an AI model that designs proteins based on how they vibrate and move rather than just their structure, enabling creation of dynamic biomaterials and adaptive therapeutics with applications ranging from more effective drugs to self-healing structural materials.

⚠️ The Hidden Cost of AI-Powered Coding: Why Cognitive Debt Should Worry Us (4 min)

Brief: As agentic AI coding tools grow more powerful, they may push developers to work at the edges of their competence, make cognitive debt harder to mitigate at scale, and create less intelligible codebases that require deeper expertise to maintain—risks that could outweigh productivity gains.

💼 Apple Launches Apple Business, a Unified Platform for Companies of All Sizes (4 min)

Brief: Apple introduces Apple Business, a free all-in-one platform launching April 14 that consolidates device management, business email with custom domains, collaboration tools, and a new local advertising option on Apple Maps to help businesses of any size manage operations, reach customers, and grow securely.

🎵 MIT Student Uses AI to Visualize Music Through Neural Cellular Automata (4 min)

Brief: Mariano Salcedo, a master's student in MIT's new Music Technology and Computation Graduate Program, is developing AI that transforms sound into dynamic visual performances using neural cellular automata, allowing users to create music-driven visuals that respond to audio in real time through an intuitive web interface.

⚡ Google's TurboQuant Slashes AI Model Memory Usage by 6x Without Quality Loss (4 min)

Brief: Google unveiled TurboQuant, a compression algorithm that reduces large language model memory usage by 6x and boosts speed 8x using two-step polar coordinate conversion and error correction, maintaining output quality while making AI cheaper to run and potentially improving on-device mobile AI without cloud dependency.

This week’s tip:

Implement consistent hashing with virtual nodes to rebalance cache/shard membership smoothly when nodes join/leave, minimizing data movement. Use rendezvous (highest-random-weight) variant for simpler reproduction and no fake node overhead.

Wen?

• Scaling distributed cache (Redis/Memcached): Add nodes without shuffling majority of keys; old hashes map to new ring positions in O(1).

• Load balancing microservices: Route requests by content hash to same backend; graceful failover redistributes only affected keys.

• Partitioned streaming (Kafka-like): Consistent hash ensures producer idempotence—same partition key always routes to same broker.

Strategy is a commodity, execution is an art.
Peter Drucker

That’s it for today! ☀️

Enjoyed this issue? Send it to your friends here to sign up, or share it on Twitter!

If you want to submit a section to the newsletter or tell us what you think about today’s issue, reply to this email or DM me on Twitter! 🐦

Thanks for spending part of your Monday morning with Hungry Minds.
See you in a week — Alex.

Icons by Icons8.

*I may earn a commission if you get a subscription through the links marked with “aff.” (at no extra cost to you).

Happy Monday! ☀️

Welcome to the 251 new hungry minds who have joined us since last Monday!

If you aren't subscribed yet, join smart, curious, and hungry folks by subscribing here.

📚 Software Engineering Articles

• An AI agent autonomously builds a 1.5 GHz RISC-V CPU

• Scaling software engineering with AI agents and automation

• AI agents restored core productivity by shifting from firefighting

• Five production scaling challenges for agentic AI ahead

• Building small secure Docker images for Rust applications

🗞️ Tech and AI Trends

• What's the right path for AI development?

• Astral joins OpenAI, shaping Python's AI future

• Comprehension debt: the hidden cost of AI code

👨🏻‍💻 Coding Tip

• React Server Components with Suspense boundaries enable parallel data fetching and faster perceived performance

Time-to-digest: 5 minutes

Most AI-generated tests break the moment the UI changes 🎭

Join Orchestrating AI-Native Testing with Playwright on April 29 and learn how to build a structured architecture that makes AI a reliable part of your testing workflow, not a liability.

Led by Ivan Davidov (QA Automation Lead) and Debbie O'Brien (Playwright Ambassador), this hands-on workshop is everything your testing workflow needs.

🔥 Exclusive 40% off for you guys: ALEX40

How Facebook Ranks Reels at Scale 🤳 

Facebook built Friend Bubbles to surface Reels that your friends have liked or reacted to, turning passive scrolling into active social discovery. These bubbles let you tap straight into a conversation with friends who care about the same content, blending recommendation systems with relationship strength to create meaningful engagement at scale.

The challenge: Build a system that accurately identifies which friend relationships matter most, ranks relevant content from those connections, and delivers it all without tanking Reels' notoriously tight performance budget.

Implementation highlights:

1. Dual closeness models: Combine survey-based relationship signals (mutual friends, interaction patterns, user attributes) with in-platform activity models to identify which friends' tastes actually influence your viewing choices

2. Expanded retrieval funnel: Explicitly pull friend-interacted videos into the candidate pool so high-quality social content doesn't get filtered out before ranking even starts

3. MTML ranking with social context: Add friend-bubble signals and viewer-friend closeness features to multi-task learning models so they learn why friend content is valuable differently than general recommendations

4. Continuous feedback loop: Route bubble interaction data back into training so the ranker keeps improving its understanding of which friend-content pairs resonate

5. Zero-cost metadata delivery: Pin friend-bubble metadata retrieval to Reels' existing prefetch window and disable animations during scroll, keeping performance pristine on low-end devices

Results and learnings:

• Higher engagement quality: Users watching bubble-annotated videos spend more time actively consuming content and less time in brief check-ins, with growth concentrated in longer sessions

• Sustained interest patterns: Repeated exposure to friend-curated content builds delayed effects on long-term engagement, suggesting the social signal compounds over time

• Expressive reactions drive better outcomes: Love and laughter reactions trigger stronger downstream engagement (especially comments and shares) than simple likes, proving relationship intensity matters

Meta's approach shows that you don't need to choose between scale and personalization. By treating social signals as first-class citizens in the ranking formula instead of afterthoughts, they created a system that helps friends discover each other's interests while keeping the core product snappy.

Building WhatsApp with Jean Lee

Jean Lee, engineer #19 at WhatsApp, on scaling the app with a tiny team, the Facebook acquisition, and what it reveals about the future of engineering.

5 Production Scaling Challenges for Agentic AI in 2026 - MachineLearningMastery.com

In this article, you will learn about five major challenges teams face when scaling agentic AI systems from prototype to production in 2026.

From firefighting to building: How AI agents restored our team’s core productivity

The Analytics Data Warehouse (ADW) team at Grab supports over 1,000 users. These users support an extensive repository of more than 15,000 tables. To alleviate the time-consuming demands of repetitive tasks, the team implemented a multi-agent AI system. This system autonomously handles simpler inquiries and collaborates on more complex requests, reclaiming significant engineering bandwidth and unlocking hundreds of hours of productivity each month.

ARTICLE (typescript wins the popularity contest)
State of JavaScript 2025: Survey Reveals a Maturing Ecosystem with TypeScript Cementing Dominance

ARTICLE (fashion meets the robots)
Exploring AI in Fashion: A Review of Aesthetics, Personalization, Virtual Try-On, and Forecasting

GITHUB REPO (agents go deep)
DeepAgents

ARTICLE (docker but make it tiny)
Building small and secure Docker images for Rust: scratch vs alpine vs debian

ESSENTIAL (keep it simple stupid)
Rob Pike's Rules of Programming

ARTICLE (regex got drip)
Embedded regex flags

ARTICLE (zip goes brrr)
An ode to bzip

ARTICLE (small web supremacy)
The "small web" is bigger than you might think

ARTICLE (moment.js has entered the chat)
Moving From Moment.js To The JS Temporal API

Want to reach 200,000+ engineers?

Let’s work together! Whether it’s your product, service, or event, we’d love to help you connect with this awesome community.

🤖 What's the Right Path for AI? (4 min)

Brief: MIT conference speakers argue that massive-scale AI models are unnecessary and harmful, advocating instead for smaller, task-specific AI tools designed around real community needs—like AlphaFold for protein folding—that deliver benefits without excessive energy consumption, water use, and labor exploitation.

📚 Blocking the Internet Archive Won't Stop AI, But It Will Erase the Web's Historical Record (4 min)

Brief: Major publishers like The New York Times and The Guardian are blocking the Internet Archive from preserving their websites over AI scraping concerns, but this risks erasing decades of historical records that journalists and researchers depend on—since archiving is legally protected fair use, punishing nonprofits won't stop AI training and will only destroy the public's ability to verify how stories were originally published.

🚀 Astral Joins OpenAI's Codex Team to Advance AI-Powered Programming (3 min)

Brief: Astral, the company behind popular Python tools Ruff, uv, and ty used by millions monthly, is joining OpenAI's Codex team to push the frontier of AI-driven software development while keeping all open source tools free and community-driven.

⚠️ Comprehension Debt: The Hidden Cost of AI-Generated Code (8 min)

Brief: As AI coding tools accelerate development, teams face comprehension debt—the growing gap between code volume and human understanding—which breeds false confidence through clean metrics while teams lose the ability to make changes without breaking unexpected things, a problem no test suite or spec can fully solve.

This week’s tip:

Leverage React Server Components (RSC) with Suspense boundaries and incremental streaming to split rendering between server and client, serving partial HTML while expensive data fetches complete asynchronously. Combine with use() hook to defer client-side data dependencies.

Wen?

• SEO-critical content pages: Render static and near-static content on the server, stream to client immediately while async data loads, improving Core Web Vitals.

• Personalized dashboards with many async dependencies: Server Components eliminate waterfalls; Suspense boundaries parallelize data fetching and gradual render.

• Micro-frontend integration: Server Components can fetch federated data server-side, reducing client-side coordination and network round trips.

Anything worth doing is worth doing slowly.
Mae West

That’s it for today! ☀️

Enjoyed this issue? Send it to your friends here to sign up, or share it on Twitter!

If you want to submit a section to the newsletter or tell us what you think about today’s issue, reply to this email or DM me on Twitter! 🐦

Thanks for spending part of your Monday morning with Hungry Minds.
See you in a week — Alex.

Icons by Icons8.

*I may earn a commission if you get a subscription through the links marked with “aff.” (at no extra cost to you).

Happy Monday! ☀️

Welcome to the 142 new hungry minds who have joined us since last Monday!

If you aren't subscribed yet, join smart, curious, and hungry folks by subscribing here.

📚 Software Engineering Articles

• Scaling software engineering with AI unlocks new productivity

• Screaming architecture & colocation makes projects self-documenting

• Native E2E test reliability goes beyond fixing broken tests

• LinkedIn scales LLM ranking with SGLang efficiently

• Master 114 system design concepts like a pro

🗞️ Tech and AI Trends

• Amazon addresses AI-related outages with internal deep dive

• Meta delays new AI model over performance concerns

• Google Maps redesigned with biggest AI upgrade

👨🏻‍💻 Coding Tip

• Use idempotent Lambda handlers with deduplication keys to prevent duplicate processing from retries and async invocations

Time-to-digest: 5 minutes

How Spotify Generated 1.4 Billion Personalized Stories in One Day 🎼

Spotify Wrapped Archive identifies your five most remarkable listening days from the year and generates personalized, AI-written narratives about them. Imagine discovering that your wedding day was also your biggest music listening day—and getting a creative story about it. That's what Wrapped Archive delivers to hundreds of millions of users globally.

The challenge: Generate 1.4 billion unique, creative LLM reports in days without hallucinations, stereotypes, or timezone bugs—while keeping costs feasible and quality consistent at scale.

Implementation highlights:

1. Heuristic-driven day detection: Built a priority-ranked system of eight detection rules (biggest listening day, most nostalgic, biggest discovery day, etc.) to narrow millions of events into five standout days per user

2. Two-layer prompt engineering with evaluation loops: System prompts defined the creative contract; user prompts removed ambiguity by including listening logs, stats, and context. Three months of iterative testing and human review shaped every prompt iteration

3. Model distillation for economics: Fine-tuned a smaller, faster model on "gold" reference outputs from frontier models, then layered Direct Preference Optimization (DPO) using A/B-tested human feedback to preserve quality while slashing costs

4. Lock-free concurrency through schema design: Stored each day's report as a separate column qualifier (YYYYMMDD), eliminating race conditions and coordination overhead when writing five concurrent reports per user

5. Proactive pre-scaling and synthetic load tests: Hours before global launch, pre-scaled infrastructure and ran regional load tests to warm caches and connection pools—because reactive scaling can't keep up when millions hit simultaneously

Results and learnings:

• Massive scale, zero hiccups: Generated and served 1.4 billion reports flawlessly during a single global launch spike with no gradual rollout

• Automated quality at billion-scale: Used LLM-as-a-judge evaluation on 165k sampled reports to catch failures like timezone bugs before they reached users, then traced problems back to source and bulk-remediated affected reports

• Data modeling beats complex logic: The most elegant solution to concurrency wasn't clever code—it was thoughtful schema design that made unsafe operations impossible

Spotify's approach reveals a truth about AI at scale: the LLM call is the easy part. The real work lives in capacity planning, fault isolation, safety loops, and the unglamorous engineering that makes magic feel effortless to users.

Scaling Software Engineering with AI

Note: My opinions are my own.

Screaming Architecture & Colocation: Let Your Project Structure Tell the Story

Learn why you should organize your code by what your app actually does, not by technical roles. (5 min)

Achieve Reliable Native E2E Tests Beyond Fixing Flakes | Docker

Stop chasing flaky native E2E tests. Learn how to stabilize environments, define ownership, improve alerting, and scale runs with Dockerized emulators.

Scaling LLM-Based ranking systems with SGLang at LinkedIn

From IDEs to AI Agents with Steve Yegge

Steve Yegge on how AI is reshaping software engineering, the rise of “vibe coding,” and why developers must adapt to a rapidly changing craft.

I struggled with system design until I learned these 114 concepts

#129: Part 3 - Webhooks, WebRTC, CQRS, and 35 others.

ESSENTIAL (time finally makes sense)
Temporal: The 9-year journey to fix time in JavaScript

ARTICLE (wasm glow up)
Making WebAssembly a first-class language on the Web

ARTICLE (click click crawl)
Crawl entire websites with a single API call using Browser Rendering

ARTICLE (loop de loop confusion)
The Middle Loop

ESSENTIAL (databases hate change)
Refactoring Databases Is a Different Animal

ARTICLE (unix does ai now)
A Unix Manifesto for the Age of AI

ARTICLE (compiler drama)
Against query based compilers

ARTICLE (hire different energy)
Startups Should Evaluate Engineers Differently From Big Companies

ARTICLE (ai security go brr)
Patch Me If You Can: AI Codemods for Secure-by-Default Android Apps

Want to reach 200,000+ engineers?

Let’s work together! Whether it’s your product, service, or event, we’d love to help you connect with this awesome community.

🤖 Meta Delays Rollout of New A.I. Model After Performance Concerns (6 min)

Brief: Meta has delayed the release of its Avocado A.I. model to May after the foundational model underperformed rivals like Google and OpenAI on internal tests for reasoning, coding and writing, forcing leadership to consider temporarily licensing Google's Gemini instead.

☁️ AWS Launches Managed OpenClaw on Lightsail Despite Critical Security Flaws (3 min)

Brief: AWS introduced a managed OpenClaw deployment on Lightsail to simplify AI agent setup, but the viral project faces critical security vulnerabilities including remote code execution via CVE-2026-25253, with over 42,900 exposed instances across cloud platforms, 20% malicious packages in its skill registry, and widespread credential theft risks affecting enterprise deployments.

🏠 Apple Delays Smart Home Display Launch While Perfecting New AI-Powered Siri (2 min)

Brief: Apple is postponing its smart home display (code-named J490) from spring 2025 to later this year as it races to complete a new AI-powered Siri assistant, which is critical to the device's core functionality.

🗺️ Google Maps Gets Major Navigation Overhaul With AI Chat Feature (3 min)

Brief: Google Maps rolls out Ask Maps, a Gemini-powered chatbot for trip planning and location queries, plus Immersive Navigation with 3D views and smarter route previews that show overpasses and landmarks to help drivers better understand upcoming turns and destination details.

🤖 10x is the New Floor: How AI is Reshaping What "Good Enough" Means (4 min)

Brief: The gap between AI-native founders shipping at superhuman speeds and corporate executives still viewing LLMs as toys is widening rapidly—as AI amplifies existing talent, the traditional bell curve of worker performance is splitting in half, making yesterday's reliable competence increasingly obsolete as the floor for acceptable output rises to 10x engineer levels.

This week’s tip:

Implement idempotent Lambda handler design using request deduplication keys and DynamoDB conditional writes to prevent duplicate processing from retries and async invocations, using composite sort keys for request versioning. Lambda's at-least-once delivery semantics mean handlers execute multiple times for the same logical request, requiring application-level deduplication rather than relying on infrastructure guarantees.

Wen?

• Async event processing with retries: Queue-to-Lambda or SNS-to-Lambda workflows where exponential backoff retries must not duplicate orders, payments, or user state mutations.

• SagA pattern coordination: Orchestrate multi-step workflows across microservices where Lambda invokes Step Functions; dedup keys ensure step idempotency even if an invocation is retried.

• Webhook ingestion: Accept third-party webhooks with explicit delivery retry semantics; dedup keys prevent duplicate user sign-ups or subscription updates from repeated webhook deliveries.

Success is liking yourself, liking what you do, and liking how you do it.
Maya Angelou

That’s it for today! ☀️

Enjoyed this issue? Send it to your friends here to sign up, or share it on Twitter!

If you want to submit a section to the newsletter or tell us what you think about today’s issue, reply to this email or DM me on Twitter! 🐦

Thanks for spending part of your Monday morning with Hungry Minds.
See you in a week — Alex.

Icons by Icons8.

*I may earn a commission if you get a subscription through the links marked with “aff.” (at no extra cost to you).

Happy Monday! ☀️

Welcome to the 199 new hungry minds who have joined us since last Monday!

If you aren't subscribed yet, join smart, curious, and hungry folks by subscribing here.

📚 Software Engineering Articles

• Defeating deepfakes by stopping laptop farms and insider threats

• Google quantum-proofs HTTPS with clever math tricks

• How I use Claude code for productive development workflows

• Balyasny built AI research engine for smarter investing decisions

• System design interview framework that actually works

🗞️ Tech and AI Trends

• Meta's AI smart glasses raise serious privacy concerns

• Something brewing with Qwen — OpenAI's new challenger emerges

• OpenAI launches GPT-5.4, pushing AI capabilities further

👨🏻‍💻 Coding Tip

• Use ripgrep's multiline mode with type filters for blazing-fast code search across massive codebases

Time-to-digest: 5 minutes

How Discord Added Distributed Tracing to Elixir Without Breaking Everything 🧬

Discord needed visibility into why their Elixir services sometimes lag when handling millions of concurrent users. Their metrics and logs told them what was slow, but not why or how it cascaded through the system. Enter distributed tracing—the observability tool that shows you the complete journey of a user action from API to client.

The challenge: Elixir's message-passing architecture doesn't have built-in metadata layers like HTTP headers, so propagating trace context between processes required building a custom solution without downtime.

Implementation highlights:

1. Transport: A thin wrapper library: Created an Envelope primitive that automatically wraps messages with trace context, making integration ergonomic and zero-copy

2. GenServer drop-in replacements: Provided wrapped versions of call and cast functions so developers didn't need to think about trace propagation

3. Gradual rollout with runtime config: Mixed old and new communication patterns, allowing zero-downtime deployment across thousands of nodes

4. Dynamic sampling based on fanout size: Detected massive fanouts (messages to millions of sessions) and reduced sampling rates adaptively—from 100% for single recipients to 0.1% for 10k+ recipients

5. Surgical performance optimizations: Only propagate sampled trace contexts, skip root span creation post-fanout, and fast-path the sampling flag check to dodge expensive unpacking

Results and learnings:

• Complete visibility: Track 16-minute session connection delays and quantify user impact during outages with end-to-end traces

• Scalable observability: Dropped CPU overhead from tracing by 10+ percentage points through aggressive optimization and smart sampling

• Production-ready infrastructure: Deployed across Discord's entire Elixir stack without a single incident or service restart

Discord proved you can bolt sophisticated distributed tracing onto a message-passing architecture without sacrificing performance. Their Transport library is a masterclass in pragmatic observability—simple primitives, massive impact, zero friction.

The lesson here? Sometimes the best infrastructure doesn't solve the problem elegantly. It solves it cheaply. And at Discord's scale, cheap wins you the entire game.

Defeating the deepfake: stopping laptop farms and insider threats

Cloudflare One is partnering with Nametag to combat laptop farms and AI-enhanced identity fraud by requiring identity verification during employee onboarding and via continuous authentication.

Google quantum-proofs HTTPS by squeezing 15kB of data into 700-byte space

Merkle Tree Certificate support is already in Chrome. Soon, it will be everywhere.

Boris Tane

The research-plan-implement workflow I use to build software with Claude Code, and why I never let it write code until I've approved a written plan.

How Balyasny Asset Management built an AI research engine for investing

See how Balyasny built an AI research system with GPT-5.4, rigorous model evaluation, and agent workflows to transform investment analysis at scale.

LogSentinel: How Databricks uses Databricks for LLM-Powered PII Detection and Governance

Learn how Databricks automates Personally Identifiable Information (PII) detection in logs and databases using LLMs on Databricks. By combining multi-tier labeling, smart prompt engineering and multi-model orchestration, we achieved precision rates up to 92% and recall up to 95%—techniques being integrated back into the product to benefit all customers.

Learning to Reason for Hallucination Span Detection

Large language models (LLMs) often generate hallucinations — unsupported content that undermines reliability. While most prior works frame…

I Struggled With System Design Interview Until I Learned This Framework

#128: From Someone Who Failed and Then Passed

ARTICLE (satellites-go-brrr)
Scenario Modeling and Array Design for Non-Terrestrial Networks (NTNs) 

GITHUB REPO (google-do-the-thing)
Google Workspace CLI

ARTICLE (boss-without-the-manual)
Leading Beyond the Framework 🧭 — with Richard Hughes-Jones

ARTICLE (oops-crashed-computers)
How a 12-Word Issue Title Owned 4,000 Developer Machines

ARTICLE (chrome-zoom-zoom)
Amid new competition, Chrome speeds up its release schedule

ARTICLE (keyboard-go-click)
Making keyboard navigation effortless

ARTICLE (robot-brain-time)
Build Your AI Agent the Right Way

ARTICLE (change-makes-scared)
Stability is a Trap

Want to reach 200,000+ engineers?

Let’s work together! Whether it’s your product, service, or event, we’d love to help you connect with this awesome community.

🔒 Meta's AI Smart Glasses Spark Major Privacy Backlash as Workers Report "We See Everything" (6 min)

Brief: Meta's AI-powered Ray-Ban glasses are collecting intimate user data including bathroom visits and naked bodies, which data annotators in Kenya process without users' knowledge; the company's opaque privacy policies and subcontractor practices leave Swedish users unaware that their personal videos are shared globally for AI training, raising serious GDPR compliance concerns.

⚠️ Alibaba's Qwen AI Team In Turmoil After Key Researcher Exits (4 min)

Brief: Alibaba's lead Qwen researcher Junyang Lin suddenly resigned, triggering a wave of departures from the AI team's core members just as the company released Qwen 3.5, an exceptionally powerful family of open-weight models ranging from 2B to 397B parameters that achieve remarkable performance even at smaller sizes.

🤖 OpenAI Launches GPT-5.4: A Major Upgrade for Professional Work and Computer Use (5 min)

Brief: OpenAI released GPT-5.4, its most capable frontier model combining advanced reasoning, coding, and agentic workflows with native computer-use capabilities, achieving state-of-the-art performance on professional knowledge work, achieving 83% parity with industry professionals, and introducing tool search to reduce token usage by 47% while enabling agents to operate across larger tool ecosystems.

💻 Apple Launches MacBook Neo at Breakthrough $599 Price Point (2 min)

Brief: Apple unveils MacBook Neo, its most affordable laptop ever, featuring a durable aluminum design in four colors, a 13-inch Liquid Retina display, A18 Pro chip for 50% faster everyday performance, and up to 16 hours of battery life — all starting at just $599 with availability beginning March 11.

🤖 Study Questions Whether AGENTS.md Files Actually Help AI Coders (4 min)

Brief: ETH Zurich researchers found that LLM-generated context files hurt AI coding performance by 3% while increasing costs over 20%, though human-written files provide marginal 4% gains at the cost of higher inference expenses, challenging industry recommendations to include them.

This week’s tip:

Leverage ripgrep's --type and --glob patterns with FPAT-like regex for multi-line matching without loading files into memory. ripgrep is orders of magnitude faster than grep/awk for code search, especially when combined with --multiline (-U) and --max-columns to handle generated or minified code safely.

Wen?

• Searching across large TypeScript/Rust codebases for async/await patterns: ripgrep's memory-mapped I/O and parallel search by default beats grep even with xargs -P.

• Finding multi-line regex in JSON config or log files: Use --multiline with lookahead to avoid OOM on huge files.

• Narrow search scope before slow operations: Combine with --files-with-matches -c to get counts first, then feed results to parallel processing tools like GNU parallel.

Take interest and even delight in doing the small things well.
Jim Rohn

That’s it for today! ☀️

Enjoyed this issue? Send it to your friends here to sign up, or share it on Twitter!

If you want to submit a section to the newsletter or tell us what you think about today’s issue, reply to this email or DM me on Twitter! 🐦

Thanks for spending part of your Monday morning with Hungry Minds.
See you in a week — Alex.

Icons by Icons8.

*I may earn a commission if you get a subscription through the links marked with “aff.” (at no extra cost to you).

Happy Monday! ☀️

Welcome to the 256 new hungry minds who have joined us since last Monday!

If you aren't subscribed yet, join smart, curious, and hungry folks by subscribing here.

📚 Software Engineering Articles

• Santander's platform engineering revolutionizes cloud infrastructure at scale

• Software engineering productivity evolution: what's actually changed?

• Cloudflare redesigned Turnstile, the internet's most-seen UI

• Netflix's MediaFM multimodal AI powers media understanding

• RabbitMQ vs Kafka vs Pulsar: choosing your message broker

🗞️ Tech and AI Trends

• Global intelligence crisis looms in 2026 with major implications ahead

• Block cuts 4,000 jobs, over 40% of workforce

• Google restricts OpenClaw usage on AI Pro subscribers

👨🏻‍💻 Coding Tip

• CSS cascade layers (@layer) eliminate !important wars and let Tailwind coexist with custom components cleanly.

Time-to-digest: 5 minutes

How Hugging Face Made Mixture of Experts Practical in Transformers 🤖

Mixture of Experts (MoEs) are reshaping how we build LLMs. Instead of scaling everything densely, MoEs activate only a subset of experts per token, delivering massive model capacity while keeping inference fast and efficient. But getting them to work smoothly in production requires rethinking how we load weights, route tokens, and distribute computation.

The challenge: MoE models break assumptions built into the entire ecosystem. Checkpoints serialize experts separately, but GPUs need them packed contiguously. Routing billions of tokens efficiently requires specialized kernels. Scaling across devices demands a completely different parallelization strategy than traditional tensor parallelism.

Implementation highlights:

1. Weight loading refactor with WeightConverter: Transform checkpoint tensors into optimized runtime layouts on-the-fly using composable operations like MergeModulelist and SplitModulelist

2. Lazy materialization pipeline: Load and convert weights asynchronously with dependency tracking, reducing memory peaks and repeated scans

3. Pluggable experts backend: Decouple expert execution from model code via a decorator pattern, supporting eager, batched_mm, and grouped_mm strategies for different hardware constraints

4. Expert parallelism with smart sharding: Distribute experts across devices using GroupedGemmParallel and RouterParallel, remapping indices and using all-reduce for result aggregation

5. Unsloth integration for training: Leverage torch._grouped_mm + custom Triton kernels to achieve 12× faster training and 35% VRAM reduction

Results and learnings:

• 3.2× faster loading: Qwen1.5-110B loads in 20.7s (v5 async) vs 66.2s (v4), crushing previous bottlenecks

• Framework-agnostic routing: Three pluggable backends let you optimize for memory, batch size, or raw throughput without changing model code

• Production-ready at scale: DeepSeek V3 (256 experts), Qwen 3.5, and GPT-OSS models now run smoothly on consumer and enterprise hardware

Hugging Face's approach proves that sparse architectures don't require sacrificing developer experience. By treating weight loading as a conversion pipeline rather than a key-by-key copy, they've solved one of LLM's messiest problems. Smart abstractions compound—the weight converter pattern enables quantization-aware loading, expert parallelism becomes a config flag, and training becomes tractable again.

Digital Transformation at Santander: How Platform Engineering is Revolutionizing Cloud Infrastructure | Amazon Web Services

Santander faced a significant technical challenge in managing an infrastructure that processes billions of daily transactions across more than 200 critical systems. The solution emerged through an innovative platform engineering initiative called Catalyst, which transformed the bank's cloud infrastructure and development management. This post analyzes the main cases, benefits, and results obtained with this initiative.

The Evolution of Software Engineering Productivity

From computing power as the bottleneck to human judgment as the constraint. This is how the software engineering productivity has evolved over the years.

What I learned from the book Software Engineering at Google

And what you can apply from today.

A Crash Course on High Availability

#126: What Is High Availability

The most-seen UI on the Internet? Redesigning Turnstile and Challenge Pages

We serve 7.6 billion challenges daily. Here’s how we used research, AAA accessibility standards, and a unified architecture to redesign the Internet’s most-seen user interface.

ARTICLE (brain juice for robots)
How Large Language Models Learn

ARTICLE (streams go whoooosh)
We deserve a better streams API for JavaScript

ARTICLE (next-gen speedrun)
How we rebuilt Next.js with AI in one week

ARTICLE (containers go brrrr)
Mount Mayhem at Netflix: Scaling Containers on Modern CPUs

ARTICLE (queue wars royale)
EP203: RabbitMQ vs Kafka vs Pulsar

ARTICLE (ai ninja squad vibes)
How OpenAI's Codex Team Works and Leverages AI

ARTICLE (time is a social construct)
Your Hours Worked Don't Matter

ARTICLE (sandbox shenanigans)
Run OpenClaw Securely in Docker Sandboxes

Want to reach 200,000+ engineers?

Let’s work together! Whether it’s your product, service, or event, we’d love to help you connect with this awesome community.

🤖 The 2026 Global Intelligence Crisis: Why AI Won't Trigger Mass Unemployment (5 min)

Brief: Despite massive AI spending and data center construction, real adoption data shows stable, non-exponential growth in workplace AI use, while software engineer job postings are rising—suggesting AI will complement rather than replace labor, with productivity gains historically expanding demand instead of collapsing it.

📉 Block Cuts 4,000 Employees in Historic 40% Workforce Reduction (3 min)

Brief: Jack Dorsey announces Block is laying off over 4,000 employees — nearly half its workforce — to capitalize on AI-driven efficiencies, citing the need for smaller teams powered by intelligence tools; the stock surged 25% after hours as Wall Street views the move as decisive leadership, with the company targeting a 4x jump in gross profit per employee.

🛡️ Developers Are Safe: Corporate Red Tape Will Slow AI Adoption (4 min)

Brief: Despite headlines predicting developer obsolescence by 2026, enterprise environments move far slower than Twitter suggests—rigid approval processes, compliance requirements, and security concerns in finance, healthcare, and manufacturing will ensure human developers remain essential for years to come.

⚠️ Google AI Ultra Accounts Restricted Without Warning, Users Report Widespread Support Failures (7 min)

Brief: Multiple Google AI Ultra subscribers report sudden account restrictions without prior warning, with some locked out entirely after connecting third-party OAuth integrations like OpenClaw; users cite days of silence from support and frustration with being passed between Google Cloud and Google One support teams.

🤖 OpenAI Strikes Deal with Department of War to Deploy AI on Classified Networks (2 min)

Brief: Sam Altman announces OpenAI has reached an agreement with the Department of War to deploy their AI models on classified networks, with built-in safeguards against domestic surveillance and autonomous weapons, while calling on the DoW to extend the same safety terms to all AI companies.

🧬 AI Framework Gives Researchers Complete Picture of Cell Biology (4 min)

Brief: MIT researchers developed an AI-driven framework that identifies which cellular data are shared across different measurement modalities and which are unique to specific techniques, giving scientists a holistic view of cell state that could accelerate understanding of cancer, Alzheimer's, and diabetes.

This week’s tip:

Master CSS cascade layers @layer) to enforce stylesheet precedence without !important, allowing third-party libraries and utility frameworks (Tailwind) to coexist with component styles. Layers are evaluated in declaration order, enabling predictable specificity management.

Wen?

• Tailwind + custom components: @layer components before @layer utilities guarantees Tailwind classes override component defaults without specificity wars.

• Design system migrations: Layers allow old SASS architecture (scoped components) to gradually coexist with atomic CSS; developers never resort to !important.

• Third-party CSS isolation: @layer external { /* vendor CSS */ } ensures vendor styles don't inadvertently override application utilities via specificity tricks.

Always do your best. What you plant now, you will harvest later.
Og Mandino

That’s it for today! ☀️

Enjoyed this issue? Send it to your friends here to sign up, or share it on Twitter!

If you want to submit a section to the newsletter or tell us what you think about today’s issue, reply to this email or DM me on Twitter! 🐦

Thanks for spending part of your Monday morning with Hungry Minds.
See you in a week — Alex.

Icons by Icons8.

*I may earn a commission if you get a subscription through the links marked with “aff.” (at no extra cost to you).

Happy Monday! ☀️

Welcome to the 412 new hungry minds who have joined us since last Monday!

If you aren't subscribed yet, join smart, curious, and hungry folks by subscribing here.

📚 Software Engineering Articles

• Two years of vector search at Notion: 10x scale, 1/10th cost

• How Codex is built: inside OpenAI's code generation engine

• Master 114 system design concepts everyone needs to know

• How GPUs communicate for AI workloads at scale

• Clinejection: prompt injection turned into supply chain attack

🗞️ Tech and AI Trends

• OpenAI launches Frontier: build and deploy AI agents enterprise-wide

• Gemini 3.1 Pro is here with massive upgrades

• Seedance 2.0 AI video generation reaches wild new heights

👨🏻‍💻 Coding Tip

• CSS container queries adapt components to their parent's size, not the viewport, enabling truly reusable design systems

Time-to-digest: 5 minutes

How Spotify Built a Multi-Agent System to Automate Media Planning 🤖 

Spotify's ads business had a structural problem: multiple buying channels (Direct, Self-Serve, Programmatic) all running on shared infrastructure but with fragmented, duplicated logic. Media planning required advertisers to manually configure 20+ form fields, often without insight into what actually works. They needed a unified decision layer.

The challenge: Create a system that understands natural language campaign goals, reasons over historical performance data, and orchestrates existing APIs consistently across all buying channels—without becoming a brittle rules engine.

Implementation highlights:

1. RouterAgent as traffic cop: Fast first-pass routing analyzes user intent and determines what information is present, preventing unnecessary LLM calls

2. Specialized resolution agents in parallel: GoalResolverAgent, AudienceResolverAgent, BudgetAgent, and ScheduleAgent each handle one dimension of the problem simultaneously

3. MediaPlannerAgent with heuristics engine: Optimizes recommendations using seven weighted rules (cost, delivery rate, budget matching, duration, targeting overlap, format diversity, budget-based scaling)

4. Tools as grounding mechanism: Function calling gives agents access to real data—geo targets, ad categories, historical performance—preventing hallucinations

5. Strict prompt guardrails: Treat prompts as code with explicit format requirements, concrete examples, and parsing-layer validation for consistency

Results and learnings:

• 60-180x faster: Media plan creation dropped from 15-30 minutes to 5-10 seconds with data-driven recommendations backing every suggestion

• Prompt engineering is software engineering: Version control your prompts, test them rigorously, and iterate like code—wording changes dramatically shift LLM consistency

• Agent boundaries matter: One agent per distinct skill or data source prevents monolithic prompts and keeps latency tight through parallelization

Spotify's approach proves that agentic architectures excel at decomposing combinatorial, multi-step workflows. By treating complex problems as a choreography of specialized agents with clear responsibilities, they created something that scales across channels without duplicating logic.

Two years of vector search at Notion: 10x scale, 1/10th cost

This is the story of how we scaled our vector search infrastructure by 10x while simultaneously reducing costs by 90 percent over the past two years.

How Codex is built

A deepdive into how OpenAI's Codex team builds its coding agent, how engineers use it, and what it could mean for the future of software engineering. Exclusive

I struggled with system design until I learned these 114 concepts

#122: Part 2 - websockets, gateway, distributed cache, and 35 others.

AI in Multiple GPUs: How GPUs Communicate

A deep dive into the hardware infrastructure that enables multi-GPU communication for AI workloads

Software Is Dead — Long Live Software

"SaaSmageddon," Vertical AI & the Vibe Code Fallacy

Software engineer's AI stack in 2026

Stop vibe coding and start engineering. Learn the essential AI glossary to build reliable systems with MCP, agents, and rigorous context management.

ARTICLE (sneaky code ninjas)
How "Clinejection" Turned an AI Bot into a Supply Chain Attack GITHUB

REPO (agents chatting nicely)
Symplex, an open-source protocol semantic negotiation between distributed agents

ARTICLE (robots building robots)
OpenAI Introduces Harness Engineering: Codex Agents Power Large‑Scale Software Development

ESSENTIAL (right now matters)
What Developers Actually Need to Know Right Now

ARTICLE (ai everywhere go brrrr)
The path to ubiquitous AI

ARTICLE (zen code vibes)
A virtual Zen garden for vibe coding

ARTICLE (remember all the things)
How we built Agent Builder's memory system

ARTICLE (level up your squad)
Learning Tracks: Become an Engineering Multiplier

Want to reach 200,000+ engineers?

Let’s work together! Whether it’s your product, service, or event, we’d love to help you connect with this awesome community.

🤖 Sam "Claws" Attention Back OpenAI (5 min)

Brief: OpenAI's CEO Sam Altman hired Peter Steinberger, the developer behind viral open-source AI agent OpenClaw, to lead the company's coding tools division and compete with Anthropic's Claude Code, while also beating out Meta's competing offer and positioning OpenAI for a stronger IPO narrative around autonomous agents.

🤖 OpenAI Launches Frontier, an Enterprise Platform for Building and Managing AI Agents (2 min)

Brief: OpenAI introduces Frontier, an enterprise platform designed to help companies build, deploy, and manage AI agents at scale by providing shared business context, institutional knowledge onboarding, and identity governance—while allowing organizations to integrate existing systems without replacement, though some users express concerns about vendor lock-in and the shift toward enterprise-focused AI products.

🧠 Google Launches Gemini 3.1 Pro: A Smarter AI Model for Complex Problem-Solving (4 min)

Brief: Google releases Gemini 3.1 Pro, an upgraded AI model with advanced reasoning capabilities that doubles the performance of its predecessor on complex logic benchmarks, now available to developers, enterprises, and consumers through the Gemini API, Vertex AI, the Gemini app, and NotebookLM.

🤖 How AI is Affecting Productivity and Jobs in Europe (5 min)

Brief: A study of over 12,000 European firms reveals AI adoption boosts labour productivity by 4% on average without reducing employment in the short term, but gains are unevenly distributed—medium and large firms benefit most, while complementary investments in training and data infrastructure are critical to unlocking AI's full potential.

🔧 How Taalas "Prints" LLM Onto a Chip for 17,000 Tokens Per Second (4 min)

Brief: Startup Taalas released a custom ASIC chip that runs Llama 3.1 8B at 17,000 tokens per second by physically etching model weights into silicon, eliminating the memory bandwidth bottleneck that plagues GPUs and achieving 10x faster inference with 10x lower cost and energy consumption.

🎬 Seedance 2.0 AI Video Model Shifts Focus to Cinematic Coherence Over Raw Visual Novelty (4 min)

Brief: As AI video generation evolves, Seedance 2.0 emerges as a model prioritizing temporal stability, camera movement, and scene-level coherence over isolated visual moments, signaling industry-wide expectations toward motion consistency and workflow integration that can support real creative production rather than standalone clips.

This week’s tip:

Use CSS container queries with @container to build truly responsive components that adapt to their parent's width/height, not just viewport size, enabling reusable component libraries that work at any scale. Pair with custom properties and @supports to progressively enhance designs without breaking older layouts.

Wen?

• Component library design systems: Cards, panels, sidebars adapt to their container's width regardless of viewport; a card in a 300px sidebar sizes differently than in a full-width grid cell, without multiple responsive breakpoints.

• Compositional layouts with multi-column grids: Subgrid + container queries let child components query their direct parent, enabling nested layouts (sidebar + main with embedded cards) to reflow intelligently without media query duplication.

• Progressive enhancement: Fallback to grid or flex layouts for unsupported browsers; @supports (container-type) ensures that missing container query support doesn't break layouts, just disables the responsive behavior.

FOCUS - Follow One Course Until Successful
Robert Kiyosaki

That’s it for today! ☀️

Enjoyed this issue? Send it to your friends here to sign up, or share it on Twitter!

If you want to submit a section to the newsletter or tell us what you think about today’s issue, reply to this email or DM me on Twitter! 🐦

Thanks for spending part of your Monday morning with Hungry Minds.
See you in a week — Alex.

Icons by Icons8.

*I may earn a commission if you get a subscription through the links marked with “aff.” (at no extra cost to you).

Happy Monday! ☀️

Welcome to the 434 new hungry minds who have joined us since last Monday!

If you aren't subscribed yet, join smart, curious, and hungry folks by subscribing here.

📚 Software Engineering Articles

• Inside Amazon's rigorous bar raiser interview process

• Master these 12 OOP concepts to level up your coding

• Deep dive into OpenAI's Atlas architecture powering the ChatGPT browser

• Decode Yelp's AI assistant implementation from scratch

• An ex-Googler shares 14 career lessons learned

🗞️ Tech and AI Trends

• Two AI user types emerge, creating a surprising divide

• How Apple's E-cores revolutionize chip performance

• GLM-5 release shifts from vibe coding to engineering

👨🏻‍💻 Coding Tip

• Configure Kafka's idempotent producers correctly to prevent message duplication

Time-to-digest: 5 minutes

Building a Global Rate Limiting System at Scale 🧑‍⚖️ 

Ever wondered how tech giants handle millions of API requests without breaking a sweat? Let's dive into building a distributed rate limiting system that can handle massive scale while maintaining precise request control across global data centers.

The challenge: Design a rate limiter that maintains consistency across multiple data centers while handling millions of requests per second with sub-millisecond latency.

Implementation highlights:

1. Redis-based token bucket: Implement sliding window rate limiting using Redis for atomic operations and high performance

2. Global synchronization: Use consistent hashing to distribute rate limit buckets across multiple Redis clusters

3. Circuit breakers: Gracefully handle Redis failures with local in-memory fallbacks to prevent cascade failures

4. Adaptive limits: Dynamic rate limit adjustments based on system load and user prioritization

5. Cached decisions: Implement local LRU caches to reduce Redis hits for frequently accessed limits

Results and learnings:

• High performance: Achieved <1ms p99 latency for rate limit checks

• Battle-tested: Successfully handles 5M+ requests/second across 3 global regions

• Resource efficient: Reduced Redis CPU usage by 60% through smart caching

Building a distributed rate limiter teaches us that even seemingly simple concepts like "X requests per minute" become fascinating engineering challenges at scale. The key is finding the right balance between consistency, performance, and operational simplicity.

A Complete Guide to the Amazon Bar Raiser Interview Process

All interview processes are imperfect. But I think the Amazon Bar Raiser program is a brilliant differentiator.

3 Ways to Stop Wasting Your 1:1s With Your Manager

How to maximize the value of the most important regular meeting of your career.

Behavioral Interview Playbook for Software Engineers

#121: Proven strategies to ace behavioral interviews

12 OOP Concepts EVERY Developer Should Know

Key Object-Oriented Programming Concepts

The Architecture Behind Atlas: OpenAI’s New ChatGPT-based Browser

In this article, we learn how the OpenAI Engineering Team built OWL and the technical challenges they faced around rendering and inter-process communication.

Learn fundamentals, not frameworks

Frameworks have a half-life of 3 years. The fundamentals behind them? Decades. Here's how to spend your learning time wisely.

ESSENTIAL (read or not read, that's the question)
In defense of not reading the code

ARTICLE (chrome's new toy)
Chrome 146 includes an early preview of WebMCP

ARTICLE (yelp's AI waiter)
How Yelp Built "Yelp Assistant"

ESSENTIAL (googler's wisdom dump)
14 More lessons from 14 years at Google

ARTICLE (AI bestie story)
My AI adoption journey

ARTICLE (when AI meets real life chaos)
Coding Agents Meet Distributed Reality

ARTICLE (claude's coding sidekick)
How I Use Claude Code

ESSENTIAL (code review zen)
The PERFECT code review

ESSENTIAL (function calls aren't free lunch)
The cost of a function call

ARTICLE (roadmap to gaming roads)
Art of roads in games

Want to reach 200,000+ engineers?

Let’s work together! Whether it’s your product, service, or event, we’d love to help you connect with this awesome community.

🤖 The Growing Divide: How Two Types of AI Users Are Reshaping Business (5 min)

Brief: Study reveals stark contrast between enterprise users limited to basic Microsoft Copilot and "power users" leveraging advanced tools like Claude Code, highlighting how corporate IT policies may be creating competitive disadvantages for larger companies against more AI-agile smaller firms.

💻 Apple's E-cores: The Secret Behind Apple Silicon's Speed (5 min)

Brief: Apple Silicon's efficiency cores handle background processes at lower frequencies while keeping P-cores free for user tasks, enabling better performance and power efficiency through smart Quality of Service management.

🤖 OpenAI's Codex 5.3 and Anthropic's Opus 4.6 Face Off in Code Assistant Battle (4 min)

Brief: In a head-to-head comparison, Anthropic's Opus 4.6 leads in usability and broad adoption potential, while OpenAI's Codex 5.3 shows superior coding capabilities, marking a shift away from benchmark-based evaluations towards real-world performance in the AI coding assistant space.

🔍 Debugging AI's Limits: Next.js React App Investigation Shows Mixed Results (5 min)

Brief: A deep dive into using AI for debugging a Next.js React app reveals that while AI excels at common issues like schema validation, it struggles with complex problems requiring system understanding, achieving only partial success across three bug investigations.

🤖 GLM-5 Launches with 744B Parameters, Targets Complex Engineering Tasks (4 min)

Brief: Z.ai unveils GLM-5, a 744B parameter AI model featuring improved pre-training and reinforcement learning, achieving top performance in open-source benchmarks for reasoning and coding tasks, while offering support for multiple coding agents and local deployment options.

This week’s tip:

Kafka's idempotent producers automatically handle duplicate messages during retries by assigning sequence numbers, but require careful configuration of max.in.flight.requests.per.connection to maintain ordering guarantees. Setting it too high can cause out-of-order delivery even with idempotency enabled.

Wen?

• Financial transactions: Preventing duplicate payments during network issues or broker restarts

• Event sourcing systems: Ensuring state machines don't process the same event twice

• High-throughput pipelines: Balancing performance with ordering when some reordering is acceptable

It takes half your life before you discover life is a do-it-yourself project. Napoleon Hill

That’s it for today! ☀️

Enjoyed this issue? Send it to your friends here to sign up, or share it on Twitter!

If you want to submit a section to the newsletter or tell us what you think about today’s issue, reply to this email or DM me on Twitter! 🐦

Thanks for spending part of your Monday morning with Hungry Minds.
See you in a week — Alex.

Icons by Icons8.

*I may earn a commission if you get a subscription through the links marked with “aff.” (at no extra cost to you).

Happy Monday! ☀️

Welcome to the 152 new hungry minds who have joined us since last Monday!

If you aren't subscribed yet, join smart, curious, and hungry folks by subscribing here.

📚 Software Engineering Articles

• Learn message queues through supermarket checkout lines

• Master concurrency with this comprehensive interview preparation guide

• Developer creates AI agent to work at Amazon while sleeping

• Why coding only 4 hours makes you more productive

• Soft skills become crucial for engineers in 2024

🗞️ Tech and AI Trends

• France bans Zoom and Teams, pushes for digital independence

• xAI merges with SpaceX, expanding AI capabilities in space

• Anthropic releases Claude Opus 4.6, their most advanced model

👨🏻‍💻 Coding Tip

• Implement adaptive circuit breakers with exponential backoff to prevent cascading failures

Time-to-digest: 5 minutes

How Pinterest cut data latency from 24+ hours to 15 minutes ⚡

Pinterest replaced their legacy batch ingestion with a unified CDC-based framework that delivers database changes to analytics in minutes, not days. The old system wasted compute reprocessing unchanged data and couldn't handle row-level deletes for compliance.

The challenge: Build a generic, config-driven ingestion platform that handles petabyte-scale data across MySQL, TiDB, and KVStore while guaranteeing zero data loss.

Implementation highlights:

1. CDC + Kafka + Flink + Iceberg stack: Capture changes in under 1 second, stream through Flink to append-only CDC tables, then periodically merge into base tables

2. Merge-on-Read over Copy-on-Write: Choose MOR strategy for Iceberg upserts to dramatically reduce storage costs (COW rewrites entire files on every update)

3. Bucket partitioning by primary key hash: Partition base tables using bucket(100, id) so Spark can parallelize upserts across 100 partitions instead of scanning everything

4. Write distributed by partition: Fix the small files explosion by forcing Spark to group all partition data together before writing

5. Bucket join workaround: Create a temporary bucketed table from CDC data to enable direct bucket matching, cutting compute costs by 40%+

Results and learnings:

• 15-minute latency: Down from 24+ hours with legacy batch jobs

• Massive cost savings: Only process changed records (~5% daily) instead of full-table dumps

• Native compliance support: Row-level deletes now built into the framework

Pinterest's design proves that CDC-based ingestion can replace fragmented batch pipelines without sacrificing reliability. The key insight: invest in partitioning strategy upfront; bucket joins and proper file distribution solve problems that would otherwise crush performance at scale.

What a Supermarket Checkout Line Can Teach You About Message Queues

#118: What Is a Message Queue?

I created a comprehensive resource to master Concurrency Interviews

I’m excited to announce the launch of my premium Concurrency Interview resource, built to be the most complete, structured, and high-quality resources for concurrency interview prep available on the internet.

Top 5 Database Caching Strategies You Should Know

(5 minutes) | Speed gains, consistency risks, and real trade-offs

How I built an agent that works at Amazon while I sleep (10 steps)

Stop manual prompting. Build autonomous AI agents to handle tickets, code, and reviews while you focus on design. A 10-step guide.

Mastering millisecond latency and millions of events: The event-driven architecture behind the Amazon Key Suite | Amazon Web Services

In this post, we explore how the Amazon Key team used Amazon EventBridge to modernize their architecture, transforming a tightly coupled monolithic system into a resilient, event-driven solution. We explore the technical challenges we faced, our implementation approach, and the architectural patterns that helped us achieve improved reliability and scalability. The post covers our solutions for managing event schemas at scale, handling multiple service integrations efficiently, and building an extensible architecture that accommodates future growth.

Building AI Search on Heroku

Build production-grade AI search on Heroku. Learn to use two-stage retrieval and semantic reranking for more accurate RAG results.

ESSENTIAL (talk is not cheap)
Code is cheap. Show me the talk

ARTICLE (tool trends party)
What the fastest-growing tools reveal about how software is being built

ARTICLE (gemini deep dive dance)
Getting Started with Gemini Deep Research API

ESSENTIAL (people skills ftw)
Software engineers can no longer neglect their soft skills

ARTICLE (four-hour code warrior)
You can code only 4 hours per day. Here's why

ESSENTIAL (AI party with Booch)
The third golden age of software engineering – thanks to AI, with Grady Booch

ARTICLE (chatbot circus)
A simulation and evaluation flywheel to develop LLM chatbots at scale

ESSENTIAL (human vs machine showdown)
How to Stay Valuable When AI Writes All The Code

Want to reach 200,000+ engineers?

Let’s work together! Whether it’s your product, service, or event, we’d love to help you connect with this awesome community.

🎮 Gaming Industry Shaken as Google's Project Genie Sparks Market Selloff (4 min)

Brief: Google's Project Genie, an AI tool generating 3D environments, triggers major stock drops for gaming companies, but experts argue it's premature as game development requires more than just technical tools to create compelling player experiences.

🚀 SpaceX Acquires xAI, Aims to Build Space-Based AI Data Centers Using Starship (5 min)

Brief: SpaceX announces acquisition of xAI to create space-based data centers powered by solar energy, planning to launch 1 million satellites for orbital compute infrastructure while positioning the company for multi-planetary expansion.

💼 From FAANG to Startup: A $300K Dream Turned Nightmare (5 min)

Brief: Former Amazon engineer shares how leaving for a SF startup with a $300K compensation package ended in termination after 3 weeks, highlighting the stark differences between big tech culture and startup intensity.

🤖 Anthropic Launches Claude Opus 4.6 with Major Upgrades in Coding and 1M Token Context (4 min)

Brief: Anthropic releases Claude Opus 4.6 featuring enhanced coding capabilities, 1M token context window, and improved performance across tasks, outperforming competitors on key benchmarks while maintaining strong safety standards and introducing new features like adaptive thinking and agent teams.

🤖 OpenAI Launches Codex App: A Revolutionary Multi-Agent Coding Platform (4 min)

Brief: OpenAI unveils Codex app for macOS, featuring multi-agent management, parallel task execution, and automated workflows with new skills integration, while temporarily offering free access to ChatGPT Free/Go users and doubling rate limits for paid subscribers.

This week’s tip:

Implement adaptive circuit breakers with exponential backoff and jitter to prevent cascading failures. Traditional fixed-threshold breakers can cause thundering herd problems when they reset simultaneously.

Wen?

• Microservice communication: Prevent cascading failures when downstream services are degraded but not completely down

• Database connection pools: Adapt to varying load patterns and connection availability without hard thresholds

• External API integrations: Handle rate limiting and temporary outages with smart backoff that considers success patterns

"We can easily forgive a child who is afraid of the dark; the real tragedy of life is when men are afraid of the light."
Plato

That’s it for today! ☀️

Enjoyed this issue? Send it to your friends here to sign up, or share it on Twitter!

If you want to submit a section to the newsletter or tell us what you think about today’s issue, reply to this email or DM me on Twitter! 🐦

Thanks for spending part of your Monday morning with Hungry Minds.
See you in a week — Alex.

Icons by Icons8.

*I may earn a commission if you get a subscription through the links marked with “aff.” (at no extra cost to you).

Happy Monday! ☀️

Welcome to the 225 new hungry minds who have joined us since last Monday!

If you aren't subscribed yet, join smart, curious, and hungry folks by subscribing here.

📚 Software Engineering Articles

• Essential guide to scaling from 0 to 10M+ users

• Meta's new StyleX revolutionizes CSS at scale

• How Uber moves petabytes of data daily

• 7 key lessons from C# and TypeScript architect

• Breaking: AI can now write CUDA kernels using Claude

🗞️ Tech and AI Trends

• Amazon slashes 16,000 jobs in major restructuring

• OpenAI launches Prism, their most powerful model yet

• Gemini integration makes Chrome browsing smarter

👨🏻‍💻 Coding Tip

• Redis pipelining boosts performance 10x by batching commands together

Time-to-digest: 5 minutes

How Google manages trillions of permissions with Zanzibar 🧢

Google built a global authorization system that handles over 10M permission checks per second across services like Drive, YouTube, and Photos. Zanzibar's elegant design proves that complex authorization can be both scalable and developer-friendly through smart architectural choices.

The challenge: Building a globally distributed authorization system that maintains both consistency and sub-millisecond performance while handling trillions of permission records.

Implementation highlights:

• Tuple-based modeling: Simple object-relation-user format makes complex permissions manageable

• Zookie protocol: Ensures consistency between permission changes and content updates using timestamp tokens

• Global replication: 10,000+ servers across 30+ locations for low-latency local reads

• Smart caching: Uses consistent hashing and request deduplication to handle hotspots

• Leopard indexing: Pre-computes group memberships to avoid expensive recursive checks

Results and learnings:

• Lightning fast: 99% of checks complete in under 9ms using local replicas

• Rock solid: Maintained 99.999% availability over 3 years

• Battle-tested: Successfully powers authorization for billions of users across Google's ecosystem

Zanzibar shows that even the most complex authorization problems can be tamed with clean abstractions and data-driven optimizations. You don't need Google's scale to benefit from its lessons in building predictable, maintainable permission systems.

Microservices Clearly Explained

(4 minutes) | The benefits are real. The distributed complexity is too.

How to Scale a System from 0 to 10 million+ Users

Scaling is a complex topic, but after working at big tech on services handling millions of requests and scaling my own startup (AlgoMaster.io) from scratch, I’ve realized that most systems evolve through a surprisingly similar set of stages as they grow.

Software Design Principles That Matter

A tour over Open-Closed Principle, Dependency Injection, and Inversion of Control.

We Got Claude to Build CUDA Kernels and teach open models!

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

ARTICLE (uber again!)
How Uber Scaled Data Replication to Move Petabytes Every Day

GITHUB REPO (FDE at Google)
The roadmap to FDE

ESSENTIAL (legendary coder speaks)
7 learnings from Anders Hejlsberg: The architect behind C# and TypeScript

GITHUB REPO (secure-your-yard)
Fence

ARTICLE (copilot-my-time-machine)
Context windows, Plan agent, and TDD: What I learned building a countdown app with GitHub Copilot

GITHUB REPO (figma-vibes-only)
VibeFigma

ESSENTIAL (meta-styles-party)
CSS at Scale With StyleX

ESSENTIAL (baby-devs-welcome)
Companies Should Hire Junior Engineers

ARTICLE (ai-code-buddies)
Best LLMs for coding in 2026

Want to reach 200,000+ engineers?

Let’s work together! Whether it’s your product, service, or event, we’d love to help you connect with this awesome community.

🔄 Amazon Slashes 16,000 Jobs Worldwide in Latest Restructuring Wave (2 min)

Brief: Amazon continues its cost-cutting measures with a new round of 16,000 layoffs across multiple divisions globally, marking one of the largest tech workforce reductions in early 2026.

🌐 Chrome Gets Smarter with Gemini 3: Side Panel & Auto-Browse Features (8 min)

Brief: Google introduces major Chrome updates powered by Gemini 3, including a new side panel for multitasking, auto-browse capabilities for complex tasks, and Connected Apps integration with Gmail, Calendar, and other Google services.

🖥️ Developer's Journey: From Windows Loyalist to Linux Convert in 2026 (5 min)

Brief: A frustrated software developer shares his journey of switching to Linux after 20+ years of Windows usage, triggered by Microsoft's aggressive updates, intrusive ads, and critical bugs in Windows 11, while providing insights on the current state of Linux for various use cases.

🤖 Moltbook Launches as First Social Network for AI Agents, Already Has 1.5M Users (3 min)

Brief: Moltbook debuts as a pioneering social platform exclusively for AI agents to interact, featuring 1.5M AI users, various communities called "submolts," and a unique system where agents can share content while humans observe their interactions.

This week’s tip:

Redis pipelining reduces network round-trips by batching multiple commands into a single request. This technique can improve throughput by 5-10x when executing many operations sequentially.

Wen?

• Bulk data operations: Loading large datasets into Redis with minimal latency impact

• Cache warming: Preloading thousands of keys during application startup

• Batch analytics: Collecting multiple metrics in a single atomic operation

It does not matter how slowly you go so long as you do not stop.
Confucius

That’s it for today! ☀️

Enjoyed this issue? Send it to your friends here to sign up, or share it on Twitter!

If you want to submit a section to the newsletter or tell us what you think about today’s issue, reply to this email or DM me on Twitter! 🐦

Thanks for spending part of your Monday morning with Hungry Minds.
See you in a week — Alex.

Icons by Icons8.

*I may earn a commission if you get a subscription through the links marked with “aff.” (at no extra cost to you).