How Japanese Carpenters Built Without Nails
Published:
Some of the most impressive engineering solutions do not look like engineering at first. Read more
Blog Posts
Google Introduces the Pixel 11 Generation
Published:
Google unveiled the Pixel 11 family on August 12, 2026, maintaining the late-summer announcement window that Google established with Pixel 9 (August 2024) and Pixel 10 (August 2025). The lineup — Pixel 11, 11 Pro, 11 Pro XL, and Pro Fold — is built around the Tensor G6, the sixth generation of Google’s custom SoC (first introduced with Pixel 6 in October 2021). Tensor G6 continues the custom silicon strategy Google started to give it control over the NPU (Neural Processing Unit) for on-device AI inference, the ISP (Image Signal Processor) for computational photography pipelines, and the modem for radio performance — components that third-party Qualcomm Snapdragon chips would optimize for a broader market rather than for Google’s specific software features. The Tensor lineup has moved through successive silicon partners and manufacturing nodes since Tensor G1 (Samsung 5nm), with each generation improving the NPU throughput and energy efficiency available for on-device AI tasks. Read more
Kaizen Small Improvements Big Results
Published:
There is a peculiar kind of problem that almost never gets fixed. It is not serious enough to become an emergency. It does not crash the system, stop production, or ruin the day. It simply wastes a few seconds, creates a little frustration, or requires one unnecessary step. So we tolerate it. Then we tolerate it again tomorrow. And the day after that. Read more
A Personal Website Is Never Really Finished
Published:
A personal website is a strange project because there is no real final version. Every other software project has a definition of done — a feature ships, a bug is fixed, a version is released. A personal website has none of those endpoints. You can always change a sentence, move a section, update a post from five years ago, add a page you have been meaning to build, or remove something you are no longer sure should be there. Read more
Why I Still Want A Web I Can Own
Published:
The modern internet has made publishing effortless. Creating an account, writing something, and reaching readers within minutes is genuinely easier than at any point in history. The platform handles servers, design, discovery, and the distribution mechanics that would otherwise require technical expertise most people do not have or want. Read more
AMD Advancing AI 2026 Shows a Broader Accelerator Ecosystem
Published:
AMD held its Advancing AI 2026 event in July 2026, presenting the Instinct MI450 accelerator alongside a rack-scale system strategy, updated ROCm 7 software stack, and deployment partnership announcements from Microsoft Azure, Meta, and Oracle Cloud Infrastructure. The MI450 used AMD’s CDNA 4 architecture on TSMC’s 3nm process, with 288 GB of HBM3e memory per GPU — matching the NVIDIA B200 Ultra’s memory capacity — and Infinity Fabric 4 connections providing 896 GB/s inter-GPU bandwidth in an 8-GPU rack node configuration. AMD positioned the MI450 specifically against NVIDIA Blackwell Ultra systems for large-scale LLM inference and training. Read more
The Gaming PC I Would Build in 2026
Published:
If I were building a gaming PC on July 1, 2026, this is the machine I would actually want. I am not simply choosing the most expensive part in every category. The priority is gaming performance, then reliability, cooling and enough headroom to keep the system useful for years. Read more
The Skills Computer Science Students Need Beyond Coding
Published:
Students entering computer science programs tend to focus on programming languages, frameworks, and algorithms. This is understandable — these skills are concrete, teachable, and immediately demonstrable. You write code, it does something, and the result is visible. Read more
Apple Unveils Siri AI at WWDC26
Published:
Apple previewed a redesigned Siri at WWDC26 on June 8, 2026, alongside iOS 27, iPadOS 27, macOS 27 (Tahoe), watchOS 27, visionOS 27, and tvOS 27. The updated Siri replaced the older intent-parsing pipeline with a model capable of maintaining multi-turn conversational context — remembering what a user said two exchanges earlier and acting on cumulative instructions without requiring each request to be self-contained. Apple demonstrated examples such as “add the restaurant my sister mentioned in her last email to my Maps list,” a task requiring Siri to access Mail, parse unstructured text, identify an entity, and act across two apps with a single spoken instruction. Read more
Games as a Way to Understand Algorithms
Published:
Games are full of algorithms, but we rarely notice them because we are busy trying to win. A path has to be found. An opponent has to decide what to do. A score has to be updated. Behind many small game actions is a computer science problem that has a real name and a real solution. Read more
What AI Agents Still Get Wrong
Published:
AI agents have moved from demo to daily use over the past year. The pattern is familiar: an agent receives a goal, plans steps, calls tools — reading files, executing code, querying APIs, sending messages — and continues until the task is complete or it encounters something it cannot handle. The step from “answer a question” to “carry out a task with multiple steps and side effects” seemed like a natural progression. The experience of actually using these systems has revealed where the progression is less natural than it looked. Read more
Google I-O 2026 Expands Agentic AI Across Products
Published:
Google I/O 2026 ran May 19–20 in Mountain View, with Sundar Pichai’s keynote focused almost entirely on agentic AI applications. Google announced Gemini 3.5 Flash — a lightweight, low-latency model optimized for real-time interaction — and Gemini Omni, positioned as a model that could natively generate, edit, and reason across text, images, audio, video, and code in a single model pass rather than routing to separate specialist models for each modality. Gemini Omni was available initially to Gemini Advanced subscribers and through Vertex AI. Read more
Ubuntu 26.04 LTS Arrives
Published:
Canonical released Ubuntu 26.04 LTS “Resolute Raccoon” on April 23, 2026, following the two-year LTS cadence from Ubuntu 24.04 LTS “Noble Numbat” (April 2024) and five years after Ubuntu 22.04 LTS “Jammy Jellyfish” (April 2022). As the eleventh long-term-support release, 26.04 carries the standard five-year free security and maintenance update commitment from Canonical for desktop and server, extensible to ten years through the Ubuntu Pro subscription for enterprise users. The kernel version shipped as the 26.04 HWE (Hardware Enablement) base, with a GA kernel selected for stability and broad driver compatibility across the hardware range certified through Canonical’s Ubuntu Certified programme — spanning AMD Ryzen and EPYC, Intel Core and Xeon, Ampere Altra, and the ARM server lineup that had grown substantially in cloud deployments since Graviton2. GNOME 49 on the desktop introduced refinements to fractional scaling (HiDPI displays at non-integer ratios, a long-standing pain point), improved color management pipeline integration, and updated Wayland protocol support across graphics hardware. PipeWire had fully replaced PulseAudio for audio and JACK for pro-audio use cases, completing the audio subsystem transition that began with Ubuntu 22.10. Read more
Why I Still Enjoy Plain HTML, CSS, and JavaScript
Published:
In March 2026, the modern web development landscape includes dozens of frameworks, meta-frameworks built on top of frameworks, build tools that compile JavaScript that compiles to JavaScript, AI-assisted scaffolding that generates entire project structures before you have written a line of code, and package managers competing to install those structures most efficiently. Read more
NVIDIA Unveils Vera Rubin at GTC
Published:
NVIDIA held GTC 2026 in San Jose in March, with Jensen Huang’s keynote presenting the Vera Rubin platform as a fully integrated AI computing system rather than a discrete GPU. Vera Rubin paired the Rubin GPU with the Vera CPU — NVIDIA’s first custom Arm-based server processor, moving away from the AMD EPYC CPUs used in previous DGX systems. Vera CPUs were connected to Rubin GPUs using NVLink-C2C, the same chip-to-chip interconnect used internally within Grace Hopper Superchips, providing bandwidth of 900 GB/s between CPU and GPU — roughly 7× PCIe 5.0. This allowed Vera Rubin compute nodes to treat CPU and GPU memory as a single flat address space rather than copying data across a PCIe bus. Read more
Learning by Simulation
Published:
Some ideas become easier the moment they start moving. A static diagram can explain a binary search tree, but watching a node rotate during an AVL rebalancing — seeing the pointers change in sequence — creates a different kind of understanding. A paragraph can describe a CPU scheduling algorithm, but watching processes arrive, wait in the queue, and get CPU time in a step-through simulation makes the policy feel concrete rather than abstract. Read more
Anthropic Releases Claude Opus 4.6
Published:
Anthropic released Claude Opus 4.6 on February 5, 2026, available through the Claude API at $15 per million input tokens and $75 per million output tokens, with a 200K-token context window in standard usage and a 1M-token context window in beta for enterprise customers with large codebase or document-analysis requirements. Claude Opus 4.6 was positioned as Anthropic’s most capable model for agentic tasks — long-running workflows where the model executes sequences of tool calls, code execution, web searches, and file operations across sessions measured in minutes or hours rather than single-turn exchanges. Read more
The Return Of Small Software
Published:
Software has become incredibly capable, and that capability has brought weight. A task that should take thirty seconds — resizing an image, converting a file format, extracting text from a PDF — can now require creating an account, accepting a terms of service agreement, waiting for an upload, and navigating an interface designed for users who do this professionally every day. Read more
Microsoft Introduces the Maia 200 AI Accelerator
Published:
Microsoft introduced Maia 200 at its January 2026 Azure hardware event as the successor to the first-generation Maia 100 accelerator (announced November 2023, deployed in Azure in 2024). Maia 200 is manufactured on TSMC’s 3nm process and contains 216 GB of HBM3e memory across six stacks, delivering approximately 7 TB/s of memory bandwidth — roughly 1.75× the bandwidth of an NVIDIA H100 SXM5 (3.35 TB/s) and competitive with the H200 (4.8 TB/s). The chip includes 272 MB of on-chip SRAM, 10+ petaFLOPS at FP4, and 5+ petaFLOPS at FP8 for matrix operations used in transformer inference. Read more
OpenAI Releases GPT-5.2
Published:
OpenAI released GPT-5.2 on December 11, 2025 as an incremental update to GPT-5 (which had launched in May 2025 as a unified model intended to consolidate the separate GPT-4o and o-series reasoning models into a single system). GPT-5.2 focused on reliability improvements for long-running agentic tasks: more consistent tool use (fewer dropped function calls in multi-step sequences), improved handling of very long context documents (128K+ token inputs with better attention to content in the middle of the context window — a known weakness in all transformers), and tighter code generation that produced fewer subtle logic errors in multi-file software tasks. OpenAI’s internal and third-party evaluations on agentic benchmarks (SWE-bench for software engineering, GAIA for general assistant tasks, and internal multi-step research task suites) showed measurable improvements over GPT-5.0, though the improvements were characterized as refinements rather than capability jumps. Read more