Simon Willison · 2026-06-29 · 2 min read
Ornith-1.0 is a new MIT-licensed family of open-weights coding models (9B to 397B parameters, built on Gemma 4 and Qwen 3.5) that are specifically trained for agentic coding tasks — meaning they can autonomously chain together many tool calls to accomplish multi-step code navigation and manipulation tasks. This matters because it's a locally-runnable alternative to proprietary agentic coding models, and the clean Apache 2.0 lineage of the base models means you can actually use it in commercial projects without the legal headaches that plagued earlier open models.
Charity Majors · 2026-06-24 · 17 min read
TLDR: Refusing to use AI on ethical grounds is a form of purity politics that actually weakens your ability to shape the technology. The harms associated with AI — exploited labor, stolen training data, authoritarian misuse — are real, but disengagement doesn't address them; expertise does. If you work in tech, you have a moral obligation to engage deeply enough to critique, govern, and constrain these tools rather than opt out and cede that ground to worse actors.
Simon Willison · 2026-06-22 · 10 min read
TLDR: A 0.2B PyTorch inpainting model was ported to run entirely in-browser using ONNX Runtime Web + WebGPU in a single day, with Claude Code handling all the technical work autonomously — model conversion, weight publishing to Hugging Face, and ServiceWorker-based caching — while the author did zero code review. The key practical insight is that modern coding agents can now execute complex, multi-step ML porting tasks end-to-end, making feasibility the main question worth asking, not implementation effort.
Henrik Kniberg · 2026-05-29 · 8 min read
TLDR: The core shift in modern leadership is moving from controlling outcomes to designing environments where good outcomes emerge naturally. This means giving teams problems to solve (not solutions to implement), ensuring psychological safety so learning can happen, and providing clear direction on why and what while staying out of how. Without that environment, agile rituals are just cargo cult — the meetings exist but nothing actually changes.
Martin Fowler · 2026-05-14 · 14 min read
TLDR: LLMs have flipped the calculus on legacy "lift and shift" migrations — what was once considered a costly cop-out is now a viable first step, because AI can port code cheaply and a modernized platform makes subsequent improvements far easier. The broader theme: agentic programming isn't just accelerating existing workflows, it's forcing a rethink of long-held architectural principles (like avoiding code duplication across jurisdictions), with LLMs potentially acting as consistency enforcement tools rather than just code generators.
Martin Fowler · 2026-05-14 · 2 min read
Instead of manually writing lengthy context documents for LLM tasks, you can prompt an LLM to interview you one question at a time, then generate that context itself — turning a tedious writing burden into a structured conversation. This matters because it's a practical technique for extracting tacit knowledge from engineers who struggle to write documentation, and for validating specs by having an LLM interrogate domain experts rather than asking them to do a formal review.
Kent Beck · 2026-05-08 · 1 min read
Kent Beck explores whether sharing pre-formed thoughts (polished frameworks, conclusions) is less valuable than sharing live thinking (the messy reasoning process itself), prompted by his work on the build/buy/customize tradeoff. This matters because it challenges how senior engineers communicate technical decisions — whether presenting a clean mental model actually strips out the most useful signal for others trying to reason through similar problems.
Kent Beck · 2026-05-06 · 2 min read
Kent Beck interviews Angie Jones about the tension AI creates for software engineers — specifically whether hiring junior developers still makes sense when AI can fill that role, and who gets to set the standards shaping agentic AI systems. This matters because it directly challenges assumptions senior engineers hold about team composition, mentorship pipelines, and whether the open-standards work happening now will determine who controls the next generation of software infrastructure.
Bryan Cantrill · 2025-07-30 · 5 min read
TLDR: Oxide Computer raised a $100M Series B — more than doubling all capital raised in the company's prior six years — to scale manufacturing and operations for its fully vertically integrated on-premises cloud hardware. The core bet, dismissed early by VCs as lacking a market, has proven out: enterprises want cloud-like infrastructure on-prem, and Oxide built the entire stack from custom boards and hypervisor to switch silicon and control plane to deliver it. The raise positions them to go from early customers to serious scale.
Bryan Cantrill · 2025-05-01 · 11 min read
TLDR: Oxide pays every employee the same salary ($207K), and four years in, the model works better than expected — contrary to fears of freeloading, it drives higher self-motivation, stricter peer-based hiring standards, and organizational flexibility, since people aren't constrained by role-based pay tiers or competing for compensation-linked status. The biggest underrated benefit is eliminating the management overhead dedicated to salary negotiation and performance ranking, which quietly consumes enormous organizational energy at conventional companies.