Engineering Notes
Capturing the essence of engineering conversations into a structured knowledge base. What's needed isn't more output, but lower-cost understanding.
Featured
All topics →The key to Astro + Cloudflare comes down to four things: enable the adapter's platformProxy, manage all bindings centrally in wrangler.jsonc, remember that env vars and bindings are two separate systems, and point your D1 Preview environment at a dedicated staging database.
Topics
Explore all →Latest
Astro handles static rendering and content management, Cloudflare Pages/Workers handle deployment and dynamic APIs, D1 provides lightweight data storage, Vectorize + Workers AI power RAG semantic search, and R2 stores OG images and TTS audio. The entire pipeline — from YouTube crawl to user search — runs inside the Cloudflare ecosystem.
Over the past two years RAG evolved from a 'linear pipeline' to 'loop-based reasoning'. It maps cleanly to five stages: Naive, Advanced, Modular, Graph, Agentic. The real inflection point is control moving from pipeline to agent — a System 1 → System 2 shift. Looking back at engineer-news's own RAG stack, it's stuck at the Naive edge — so this post also lays out what to fix next.
The previous post zoomed out for a five-stage panorama of RAG. This one zooms in on the five infra lessons any real RAG has to face: document ingestion, contextualized chunking, three-lane hybrid search, tensor reranker, and GraphRAG's semantic gap. Each lesson is checked against engineer-news's current stack, ending with a priority list for a personal site.
Anthropic proposes J-lens, an interpretability tool that captures the 'verbalizable' representations inside a Transformer, and uses it to show that Claude contains a privileged subspace analogous to the neuroscientific 'global workspace' — a small set of vectors that broadcast, drive reasoning, respond to external steering, and even leak signals during deception and evaluation awareness.
Three OpenAI engineers, five months, one million lines of AI-generated code, zero hand-written. The real value of this experiment isn't the numbers — it's the proof that Harness design can be engineered. Five concrete practices: making the app legible to agents, treating the repo as the source of truth, mechanizing architectural constraints, rewriting merge philosophy, and background entropy management.
Distilling Harness Engineering from concept and benchmark case into something you can start executing today: the four fixed failure modes of Agents, the 40% context sweet spot, the four-pillar framework the industry has converged on, and a three-phase roadmap from 'this afternoon' to 'fully automated in two weeks' — closing with six industry consensus points and three still-unsolved problems.
A breakdown of the defense-in-depth used by Claude Code, Codex, and others: rule-based keyword matching, classifiers, input/output scanning, execution sandboxes, cross-session behavioral monitoring, and the role of system prompts and skills.
An npj Digital Medicine study analyzed 84,395 users and 282,893 sessions on the self-hypnosis app Reveri. Self-rated stress dropped with a Cohen's d of −0.71 to −0.78 (a 'large' effect). But it's a retrospective observational study with no control group and a single-item Likert measure — a great case study in why 'big N + big effect' still can't prove causation.
Headroom compresses tool outputs, logs, and RAG chunks by 60–95% locally, before the request reaches your provider. The part worth stealing isn't the ratio — it's how it decides whether to compress at all using 'mask extraction + cache-mutation economics' — plus a reminder that its docs run ahead of its code.
Two 2026 products answering the same question: when a single model hits its ceiling, don't pick one model — orchestrate a pool of them behind one API. Fugu bets on learned coordination; Fusion bets on parallel deliberation plus a judge.
When an agent loads tools on demand, it's essentially applying RAG's 'retrieve-then-inject' pattern to tool schemas — the only difference is the retriever is the LLM's own reasoning instead of vector similarity.
Research shows AI agents communicating through emergent compressed languages outperform those using natural language — shorter, more compute-efficient, but less interpretable
Loop Engineering replaces you as the person who prompts the agent. You design the system that does it instead.
WWDC26 was Tim Cook's last keynote as CEO. Apple Intelligence now runs on Google Gemini, Siri got a major overhaul, iOS 27 and macOS Golden Gate shipped, and John Ternus takes over as CEO in September.
Mainstream OLED and Mini-LED are just the baseline — E Ink now covers BMW bodies, Looking Glass won SID 2026 Display of the Year with Hololuminescent technology, and rollable AMOLED plus stretchable displays are maturing in labs
Fireship walks through 10 landmark CS papers from Turing to GPT-3, showing how each idea set the stage for the next — and how the path to modern AI was anything but linear
Titans introduces a neural memory module that updates itself via gradient descent at inference time, breaking the context-length ceiling of Transformers while staying near-linear in complexity.
TiDAR runs a diffusion model to draft tokens in parallel (Think), then lets an autoregressive decoder finalize output (Talk) — all in a single forward pass. Result: 5.91x faster than AR at comparable quality.
Reading guilt comes from treating 'finishing the book' as the goal. Hao Xu-Lie's point in Haovoice EP680: the real purpose of reading is finding your own connection to the material—not page count.
A complete behind-the-scenes look at tech YouTube production: topic selection, scripting, recording, editing, and how AI tools have changed each step.
Kafka's speed comes from two counterintuitive design choices: deliberately writing to disk (not memory) but using sequential I/O, and Zero-Copy to move data from disk to NIC without CPU involvement.
Running an LLM locally with Ollama is simpler than you think: one line to install, one line to pull a model, one line to chat. This guide takes you from install to a working local RAG pipeline.
Anthropic published a pause-AI paper on June 4, launched Fable 5 on June 9, and had it forcibly taken offline by the US government on June 12. All within ten days.
Opus 4.8's headline improvement is a 4x reduction in the probability of letting code flaws pass silently—plus Dynamic Workflows for parallel subagents and Effort Control for cost tuning.
Android's update ecosystem has meaningfully improved in 2025–2026: Pixel's 7-year commitment, Samsung Galaxy S25's faster One UI updates, and OnePlus's policy turnaround are changing the calculus for long-term device ownership.
In 2022, Meesho voluntarily dropped to 0% seller commission—at a time when they had tens of millions of monthly orders—then used that scale advantage to become India's #1 e-commerce platform by order volume, ahead of Flipkart and Amazon.
WWDC 2026's biggest moves: Siri rebuilt from scratch using Google Gemini, macOS Golden Gate ends Intel Mac support, and iOS 27 adds an opacity slider to fix Liquid Glass readability.
In ten days, Anthropic filed an IPO, called for a global AI pause, launched Fable 5, and watched it get forced offline by the US government—all contradictions compressed into one week.
Kafka's horizontal scalability comes from partition design: each partition is an independent log, Consumer Groups enable parallel consumption, and replication provides durability without a large performance penalty.
DeepMind's AlphaProof combines a language model with AlphaZero-style reinforcement learning to produce fully machine-verifiable mathematical proofs — achieving silver-medal level at the 2024 International Mathematical Olympiad.
MCP (Model Context Protocol) is an open protocol designed by Anthropic that lets Claude Code call external tools and data sources through a standardized interface. Since its November 2024 release, it has rapidly become the de facto standard for AI agent tool integration, adopted by Cursor, Windsurf, and 40+ other editors.
Going back to school isn't the only path for a mid-life career change — and it's not always the fastest. The real question: does the door you want actually require a degree to open?
Ring opened its Appstore API in 2024, letting developers receive camera event webhooks and integrate custom logic. This post documents a real implementation: a driveway vehicle detector that only notifies on unfamiliar vehicles.
CPU for complex control flow, GPU for large-scale parallel computation, TPU for matrix operations pushed to the extreme. For most engineers, the real decision is cloud inference on GPU vs CPU, and when a TPU rental is worth it.
The backflip looks impressive, but the real challenge is making a mass-produced robot reliably catch a falling leaf. That requires solving actuator selection, sensor integration, and a supply chain that barely exists yet.
Qualcomm's core bet isn't on training AI — it's on inference at the edge. Running AI on phones, PCs, cars, and robots. 6G and Physical AI are extensions of the same logic: move compute closer to data.
From Carbon (code-to-image), nektos/act (run GitHub Actions locally), to Ink (React-based terminal UI) — 10 OSS projects that each solve one specific problem really well.
AlphaFold's protein structure predictions earned the 2024 Nobel Prize in Chemistry. Here's what the MSA + Transformer architecture actually does and why it matters.
Jeff Dean breaks down where the million-fold AI compute gains actually came from — specialized hardware, distributed training systems, and architecture efficiency — and where the next phase is headed.
Douglas Crockford didn't create JavaScript, but he may be the single most important reason it went from a mocked scripting language to the foundation of the modern web: he formalized JSON, created JSLint, and wrote JavaScript: The Good Parts — a book that showed developers JavaScript actually had a good side.
There's no universal answer on grad school timing, but one question cuts through the noise: are you going to solve a specific problem, or because you haven't figured out what's next?
2026 has produced several devices labeled 'Whoop killers': Google Fitbit Air ($99, no subscription), Garmin Cirqa (expected launch), Apple Watch Ultra (post-watchOS 11). The real challenge isn't hardware — it's competing with Whoop's subscription model and its lock on the recovery analytics mindshare.
AlphaFold solved the protein folding problem in 2020 at near-experimental accuracy, earning Demis Hassabis and John Jumper the 2024 Nobel Prize in Chemistry. Its database now contains 200M+ protein structures, actively accelerating drug development and materials science.
Hassabis's preference for 'hard questions' isn't a personality quirk — it's a research strategy: choose problems that unlock large amounts of downstream value when solved, not problems easy enough to publish quickly. This strategy is the core reason DeepMind keeps breaking through at the scientific frontier.
The technical core of a modern NBA broadcast is Sony Hawk-Eye's 3D optical tracking system — 29 cameras producing gigabytes of player movement and ball trajectory data per game, feeding three completely separate pipelines: broadcast graphics, officiating assistance, and team analytics.
SpaceX plans to list on Nasdaq in June 2026 at $135/share with a $1.75T valuation — the largest IPO in stock market history. Key numbers: Starlink accounts for 58% of total revenue and is the only profitable division ($1.19B net profit). The launch business remains a money loser.
DeepMind's core strategy under Demis Hassabis: use game environments (which have clear evaluation functions) to train general reasoning capabilities, then apply the same approach to scientific problems with evaluation functions. AlphaFold, AlphaGeometry, AlphaDev, and GNoME are concrete implementations of this strategy.
Google I/O 2026's core signal isn't any single product feature — it's that Google has completed the shift from 'AI assistance tools' to 'AI agents': Gemini 3.5 Flash, Gemini Omni, Gemini Spark, and Antigravity 2.0 all point in the same direction — AI isn't your assistant, it's your agent.
On May 11, 2026, the TeamPCP group compromised 42 TanStack packages in 6 minutes using GitHub Actions cache poisoning and OIDC token extraction from process memory — producing the first-ever malicious package with valid SLSA Build Level 3 provenance.
A California jury ruled on March 25, 2026 that Meta and YouTube are liable for a child's social media addiction, awarding $6 million in damages — the first time tech companies have faced legal liability for addictive algorithmic design itself.
AI recursive self-improvement is already happening in production (Constitutional AI, RLHF with AI feedback, automated evaluators) — but the full recursive loop where AI autonomously generates stronger successors remains constrained by evaluation reliability and alignment gaps.
NVIDIA's latest inference optimizations — FP8/INT4 quantization, 2:4 structured sparsity, and TensorRT-LLM system improvements — dramatically increase throughput and cut deployment cost with negligible accuracy loss.
The scarcest resource in embodied AI isn't compute or algorithms — it's high-quality demonstration data recorded in real physical environments at scale.
An operating system isn't a black box — it's a clear pipeline from UEFI to Kernel to Process. Fireship's video uses the boot-to-shutdown lifecycle as a narrative spine to connect every major OS concept.
Cursor is an AI-powered code editor by Anysphere, built by four MIT graduates, that hit $500M ARR within two years of launch. This article distills the real engineering lessons they've shared publicly: why they forked VSCode instead of building an extension, how Tab prediction's latency engineering works, and the hard production lessons from shipping Agent Mode.
A YouTuber/indie developer noticed fans couldn't speak up due to social anxiety, so he built an AI-powered video call practice platform. This article breaks down the technical architecture and trade-offs of building this kind of product from scratch.
PageIndex swaps the vector DB for a hierarchical tree index plus an LLM Agent that reasons over it, and it shines on long structured documents (98.7% on FinanceBench). This site's Hybrid RAG instead runs vector search with a keyword fallback on the Cloudflare edge — a completely different set of tradeoffs.
Dexter breaks complex financial research questions into structured steps, automatically calls tools to fetch real-time financial data, then self-checks its results until the answer is complete enough—conceptually, it's 'Claude Code built specifically for financial research.'
OpenStock is Open Dev Society's open-source stock market dashboard, built with Next.js 15, MongoDB, Finnhub, and TradingView widgets. AGPL-3.0 licensed, self-hostable, and forever free, it's positioned as a market-intelligence tool rather than a trading platform.
Point your tools' endpoint at localhost:20128 and 9Router automatically does three-tier fallback across subscription services, cheap APIs, and free services — refreshing expired tokens on its own, translating between OpenAI and Claude formats, and using RTK to shave 20–40% off tool-output tokens.
Install one Rust binary, and the output of commands like git/grep/test/docker gets compressed before reaching the AI context — a 30-minute Claude Code session drops from roughly 118,000 tokens to about 23,900.
KV Cache reduces autoregressive Transformer generation from O(n²) — recomputing the full sequence for every new token — to O(n) per step, which is the core reason modern LLM inference is fast enough to be usable.
The question isn't 'should I go' — it's 'after I go, will this connect to the path I actually want?' Working holidays without a clear purpose tend to be detours, not shortcuts.
Transformer self-attention is inherently orderless — positional encoding is the fix. From sinusoidal absolute encoding, to learnable absolute encoding, to relative positional encoding, to RoPE (Rotary Position Embedding): modern LLMs almost universally use RoPE because it requires no parameters, naturally encodes relative distances, and can be extended to longer sequences.
TSMC controls over 90% of leading-edge process capacity globally. AI chip demand pushed its 2025 market cap past $2 trillion with stock up over 100% in a year — but this also creates concentration risk for Taiwan's equity market.
DeepSeek V3's 671B-parameter MoE architecture trained on just 2.78M H800 GPU-hours matches near-GPT-4 performance across multiple benchmarks, with API pricing at one-tenth of OpenAI's equivalent.
OpenAI released three models in spring 2025: GPT-4.1 for coding and instruction-following, o3 as the strongest reasoning model, and o4-mini hitting remarkable math and code performance at low cost — but the pricing strategy and API access limits left developers with mixed feelings.
Meta Ray-Ban Display is the first consumer product to genuinely integrate an AI display into a normal eyeglass frame, but the $799 price and 6-hour battery life signal this is still early-adopter territory.
The M4 MacBook Air and Mac Studio are solid spec upgrades — 16GB standard RAM and massive memory bandwidth for local AI workloads. Apple Intelligence's Siri integration, however, remains frustratingly inconsistent. Hardware is ahead; software is still catching up.
The Data Lakehouse merges the ACID reliability of data warehouses with the low-cost open storage of data lakes. Apache Iceberg and Delta Lake are the two dominant open table formats making this architecture practical at scale.
NVIDIA's Isaac GR00T N1 is the first genuinely open humanoid robot foundation model. Its dual-system architecture — a VLM for high-level reasoning plus a Diffusion Transformer for precise motion control — lets a single model run across multiple robot hardware platforms.
NVIDIA Lyra 2.0 generates geometrically consistent, indefinitely explorable 3D worlds from a single image. Its geometry-guided frame retrieval solves spatial forgetting and temporal drift while preserving generation quality — released open source under Apache 2.0 in April 2026.
LLM output quality is determined at three distinct layers: token-level decoding strategy, task-level workflow design, and model-level reasoning capability. Knowing which layer your problem lives in is the fastest path to fixing it.
AI video generation has been plagued by temporal drift and forgetting for years. In 2025, FramePack, Mixture of Contexts, and A2RD introduced systematic solutions that make long-form video generation genuinely viable.
Sakana AI's God Simulator uses neural cellular automata to let users act as the rule-setter for a digital ecosystem, revealing how incentive structures drive cooperation, collapse, and everything in between.
The LLM Wiki isn't a query tool—it's an architecture that lets knowledge compound over time. The LLM actively builds and maintains a markdown knowledge base, instead of re-fetching from raw documents on every single query.
Wasted work at the office isn't just bad luck — it follows identifiable structural patterns: unclear goals, no decision-maker, missing acceptance criteria, insufficient resources, no feedback loop. Recognizing these five causes lets you catch dead-end tasks before they start.
OpenAI's Michael Bolin breaks down the Codex CLI agent loop in depth: from prompt assembly, token inference, and tool execution, to context window management and auto compaction.
People who seem fearless about big moves aren't braver by nature — they have richer reference points. They've seen enough people 'do the thing and survive' that the unknown has become an estimable risk. And that's something you can deliberately build.
LLM inference is memory-bandwidth-bound, not compute-bound. That makes HBM the critical bottleneck in AI accelerators, driving a supercycle that saw the memory semiconductor market grow 78% in 2024, with HBM capacity sold out through 2026 and the cycle projected to last into 2028.
Deploy Ollama on a GPU machine via Docker, then use netsh portproxy on Windows to forward localhost:11434 to the GPU box, so Codex on Windows transparently uses the remote GPU.
Small language models around 10B parameters can run on local hardware in real time, enabling dynamic NPC dialogue, procedural narrative generation, and adaptive game content. Research shows SLMs approach large model quality on short, well-constrained creative tasks — the key is curated training data and constrained inference design.
Manycore Tech (Kujiale's parent) became the first of Hangzhou's 'Six Dragons' startups to go public, opening up 171% on its Hong Kong debut in April 2026. The technical story is spatial intelligence: 15 years of structured indoor 3D scene data is being repositioned as training infrastructure for embodied AI.
The concept of first principles isn't hard; actually applying it is. Using three familiar scenarios — the PMP exam, going to grad school, and learning personal finance — Joe from Dadu (大人學) demonstrates three thinking exercises you can turn into daily intuition.
Industry security tends to get 'looser the closer you get to the database' — backends often connect directly with an admin account, bare and unguarded. PostgreSQL uses role + schema + column-level GRANT + RLS to build outside-in permissions that can go as fine-grained as a single row (or even a single cell), keeping the last lock inside the database itself.
When an AI Agent performs poorly, it's not necessarily because the model is dumb. Starting from a small experiment where a Gemma 4 2B fixes a bug, this piece explains what a Harness is, how Harness Engineering differs from Prompt / Context Engineering, and how effective natural-language rules like agents.md really are.
A pnpm monorepo integrating four tool SPAs: a to-do list, a habit tracker, an ebook reader, and a resignation stamp collector. apps/root is the sole build entry point; the four modules are library-only workspace packages bundled together by root's Vite, sharing @a920604a/auth and @a920604a/ui, deployed across Firebase + Cloudflare dual cloud.
A real-time-voice-first AI English tutoring system: students converse with the AI teacher Emma via microphone (optionally with video/screen sharing), the system corrects mistakes in real time, and generates a post-class report in Chinese. The technical core is LiveKit (Self-hosted WebRTC) + Google Gemini 2.5 Flash Native Audio, with a FastAPI backend handling auth, courses, and data persistence.
A pure-TypeScript tag scoring engine computes real-time risk across 140 foods for four conditions — gout, high cholesterol, diabetes, and hypertension. The entire backend runs on Cloudflare's free tier, for a monthly cost of $0.
A microservice platform orchestrated with Docker Compose: it crawls arXiv papers daily, builds a Qdrant vector index, and delivers bilingual RAG Q&A through hybrid search + re-ranking + Ollama, with email subscriptions and Grafana monitoring.
Using stock price prediction as the subject, I built a complete MLOps lifecycle covering ETL, experiment tracking, model deployment, drift monitoring, and CI/CD — all orchestrated on a single machine with Docker Compose.
A self-hosted platform that integrates TTS and STT into a single interface: TTS uses Microsoft Edge TTS's 322 voices, STT uses local Whisper for offline inference on pure CPU, and results are stored in SQLite. Completely free, with no GPU or API key required.
Starting from the DIY-setup craze that OpenClaw ignited, this piece unpacks the technical evolution behind the Agentic Social Network: the three-layer architecture of Model Router, Memory OS, and protocols, plus the key mechanisms driving Agents toward long-horizon companionship, proactive action, and self-evolution.
ChatGPT's capabilities come from the first stage of massive self-supervised pre-training; but 'answering questions, understanding instructions, and not making things up' come from the latter two stages — instruction fine-tuning and RLHF. Capability and alignment are two separate things that are forged independently.
Time granularity is the basic unit you use to schedule your time. Shrinking it from hours to 15 minutes isn't just a quantitative change — it's a fundamental upgrade in how you perceive the cost of time.
Treat a conversation as raw material for an article: with a structured prompt and a template, Claude Code turns a debugging thread into a publishable technical post.