TL;DR The defining shift from April 2025 to April 2026 is the move from “ask” to “delegate” - agents now run for minutes, open files, execute shells, and return results rather than waiting for each prompt Key developments that drove this: coding agents becoming operators (Claude Code, Cursor, Codex), MCP standardising tool access, spec-driven development going mainstream, and context windows expanding to millions of tokens In the next two years, longer-horizon agents, multi-agent coordination, persistent personal AI memory, and computer-use automation will move from early features to default expectations The working day is reshaping around less typing and more reviewing - the skill that matters is judgement over diffs, not typing speed or boilerplate generation To adapt now: pick a stack and use it daily, write specs before code, build the habit of reviewing diffs fast, and move procedural knowledge into reusable agent skills A year ago, in April 2025, “AI in your workflow” mostly meant a chat window in a browser tab and an autocomplete plugin in your editor. You typed, it suggested, you accepted or rejected. The interaction model was small. The blast radius was small. The verb was “ask”. ...

On Adversity, Awakening, and Rediscovering Who You Are Most of us don’t find our life’s purpose in a single moment of calm reflection. We find it in the wreckage. We find it in the sleepless nights, the boxes we didn’t pack, the home we no longer live in. We find it when we are forced - absolutely forced - to stop, strip everything back, and ask: who am I without all of this? ...

TL;DR Documentation sprawl across Confluence, Jira, SharePoint, Google Docs, GitHub, and Miro is not a tool problem - it is a joints problem: the same decision exists in four places, drifting out of sync immediately Three forces constantly pull against each other: source of truth (one canonical home), discoverability (right surface for every audience), and governance (real access control) - optimising for any one breaks the others The proposed shape: docs-as-code for engineering artefacts in Git, collaborative tools for business content, a read-only render layer between them, and an AI-assisted discovery layer across all of it AI tooling weakens the old boundary - a business user can get a summary generated from a markdown master without ever seeing the file, and an engineer can draft an ADR pulling context from Confluence and Jira automatically Several genuine open questions remain unsolved: versioning across boundaries, who owns the render pipeline, and whether Jira tickets as documents should be formalised or fought against This post is me thinking out loud. It is not a proposal, not a recommended pattern, and possibly not even a useful framing. I am writing it because I am actively stuck on the question, and writing in public tends to be the fastest way I find out what I have got wrong. Feel free to disagree with any of it. ...

TL;DR The work AI now automates - boring tickets, bug hunts, boilerplate - was the unspoken apprenticeship that turned juniors into seniors The skills that work built (pattern recognition, systems intuition, taste, calibration) are built by doing, not by reading - and that doing is now cheapest to delegate The new apprenticeship shifts toward reading over writing, debugging agent output, earlier architectural decisions, and deliberate practice of things agents do badly There is a coordination problem: individual organisations rationally skip junior investment in the short term, but the senior pipeline thins industry-wide a few years later If you are starting out today, optimise for proximity to a great senior engineer above salary, title, or any other variable The views in this post are my own personal reflections on the industry as a whole, written in my own time. They are not about any specific employer, team, or colleague, past or present. ...

TL;DR For most of history, usable intelligence - the kind that solves complex problems - required hiring expensive specialists or spending years acquiring expertise yourself Research shows the cost of running AI capability has been falling roughly an order of magnitude every one to two years, making intelligence increasingly affordable Intelligence is becoming infrastructure - like electricity or compute, available on demand through APIs rather than locked inside individuals or institutions When intelligence is cheap and abundant, creativity becomes the limiting factor, not knowledge, credentials, or access to experts This democratisation is extraordinary, but the question of how we deploy these tools wisely matters as much as the capability itself For most of human history, intelligence has been scarce. Not intelligence in the biological sense - people have always been clever - but usable intelligence. The kind that helps you design a system, debug a problem, write code, plan a strategy, analyse data, or turn a vague idea into something real. ...

TL;DR OpenClaw is an open-source AI agent framework that enables language models to operate software directly through computer interfaces - clicking, typing, and navigating the same way a human does Unlike chatbots that only respond to prompts, OpenClaw acts as an operator - automating any software without requiring custom APIs or integrations This makes legacy enterprise software, complex dashboards, and multi-application workflows instantly automatable using computer vision and reasoning models Because it uses reasoning models rather than fixed scripts, it can adapt to unexpected states and recover from mistakes - closer to digital labor than traditional automation This represents a shift in computing: software that can build, run, and manage other software, driven by open projects improving rapidly every month Every now and then a piece of technology appears that quietly changes the rules. Not in a loud marketing way. Not with a huge product launch. Just a project sitting on GitHub that makes you stop, stare at the screen for a second, and think: ...

TL;DR Graph algorithms are memory-bound, not CPU-bound: a 5 billion node, 50 billion edge graph needs 500+ GB just for adjacency, scores, and working memory - it simply does not fit on one machine Four scaling strategies cover most production cases: distributed processing (vertex-cut or edge-cut partitioning), approximate algorithms (sampling, sketching, early stopping), incremental/streaming approaches, and storage-level optimisation (columnar layout, compression, caching) The hidden costs dominate at scale: communication overhead, synchronisation barriers, debugging distributed state, and the operational burden of keeping it all running Full-graph recompute is not feasible at billions of nodes - incremental algorithms and clever approximations are the norm, not the exception Managed services like Neptune Analytics remove much of the partitioning and operations work, at the price of less control Graph algorithms work great on your laptop. PageRank on a 100,000-node graph finishes in seconds. Louvain finds communities instantly. ...

TL;DR The 1983 Yamaha DX7 sold over 200,000 units by 1989 and remains arguably the most influential synthesizer in history - it did not just change synth design, it changed how modern music sounds It was built on FM synthesis, invented by John Chowning at Stanford in the 1970s and licensed by Yamaha - six operators producing timbres subtractive synths could not touch The glassy electric pianos, basses, and bells of mid-80s pop are DX7 presets; a huge share of the era’s hits used the factory sounds unchanged It was notoriously hard to program, which nobody cared about - and which accidentally created the preset culture that still defines synth use today Its influence runs to the present: FM engines sit inside most modern workstations and software synths The Yamaha DX7 wasn’t the first synthesizer. It wasn’t the most powerful. It wasn’t the cheapest. But in 1983, it became the most important instrument released that decade - and arguably the most influential synthesizer in history. By 1989, over 200,000 units had been sold. Today, it remains the second-best-selling synthesizer of all time (after the Casio VL-Tone, which was technically a calculator with a synth). ...

TL;DR Claude Code’s new Code Review feature dispatches multiple AI agents in parallel to review a PR from different angles, rather than running a single shallow model pass over the diff The motivation is real: Anthropic’s internal code output per engineer increased by around 200%, making human review the bottleneck - and humans consistently miss subtle bugs on large diffs Multi-agent review cross-checks findings, filters false positives, and ranks issues by severity before posting a clean, high-signal review comment plus inline annotations Review depth scales with PR size; typical runs take about 20 minutes and cost $15 - $25, which is cheap compared to the cost of a production bug Humans still approve PRs - the tool’s role is a thorough pre-review pass, not automated sign-off, making it a complement to human judgment rather than a replacement I genuinely think a lot of people still underestimate how fast the AI developer tooling ecosystem is evolving. ...

TL;DR Hitting Claude Code Pro usage limits does not mean upgrading to the $200/month plan - a hybrid AI stack is a smarter and cheaper alternative The tiering strategy: local models (free) for quick edits, cheap cloud APIs for general coding, and frontier models only for architecture or complex multi-file reasoning Tools like Ollama or LM Studio with coding models such as DeepSeek Coder or Qwen2.5 handle the majority of everyday tasks locally at no cost Cheap cloud inference providers (Groq, Together AI, DeepInfra) offer capable open models at fractions of a cent per session for heavier work A realistic usage split of 80% local / 15% cheap APIs / 5% frontier models dramatically reduces limit burn while keeping Claude available when it genuinely matters I keep running into the same problem with Claude Code Pro ($20/month): I burn through the usage limits faster than I expect. The obvious solution is upgrading to the $200/month plan, but that feels excessive for how I actually use it. ...