Life can only be understood backwards but it must be lived forwards.
-
The open-source “andrej-karpathy-skills” repo
Four rules that make AI coding agents behave more like a careful senior engineer. The open-source “andrej-karpathy-skills” repo is a single CLAUDE.md file built from Andrej Karpathy’s public observations on where LLMs go wrong when they write code: overcomplicated code and APIs, bloated abstractions, dead code left behind. It boils down to four principles: 1.…
-
Graft
We keep teaching people to use fewer tokens without losing quality. Now that effort finally has a real assistant. Graft (now part of Trail-code) builds a map of your codebase: files, functions and who calls what. Coding agents such as Claude Code, Cursor, Codex and GitHub Copilot use that map instead of searching the repository…
-
OpenBot by CopilotKit
AI coworkers are coming. The question is who sets the rules. CopilotKit released OpenBot, an open-source, self-hosted platform for AI coworkers. Each bot gets its own browser, files and tools. Every action goes through a gateway that checks policy and writes an audit log before anything runs, and the policy engine denies by default. A…
-
Where will tomorrow’s senior engineers come from?
If AI takes over much of the coding, design and debugging that junior engineers used to learn from, where will tomorrow’s senior engineers come from? Microsoft and Carnegie Mellon University are answering with PRAISE – Preceptorship for AI-accelerated Software Engineering – presented on 15 October. It borrows the preceptorship model from nursing: – Junior engineers…
-
Alibaba’s OpenSandbox
Letting AI agents run code is easy. Letting them run code safely is the real work. Alibaba’s OpenSandbox is an open-source sandbox runtime for AI agents. It gives each agent an isolated place to run code, commands, files and even a browser or desktop, on your own infrastructure: Docker on a laptop, Kubernetes in production.…
-
Atlassian AI-native SDLC on PROD
At Team ’26 Europe on 7 October, Atlassian moved its AI-native SDLC from playbook to production. What I find interesting is that it targets concrete business pain points across the whole delivery lifecycle, not just coding: – Planning: Planner turns a rough idea into a shared Confluence spec with options, risks, owners and review questions,…
-
Jarvis on EvenRealities G2
I have been using AI for more than two years. Over time, I’ve automated much of my daily work. But something was always missing.That something was a digital assistant. Because what am I without my suit of armor?Genius, Father, Believer, Mentor, Idealist, Human. So I decided to build my own digital assistant, and only one…
-
Anthropic’s AI-native SDLC playbook
Anthropic’s AI-native SDLC playbook makes a point I keep coming back to: code is no longer the bottleneck. Agents can write code faster than organisations can plan, review, test and ship it – security teams, for example, are sized for human output. The playbook’s answer is to turn the linear SDLC into a loop, where…
-
Cybersecurity is now the number one risk to growth for CEOs
The KPMG 2026 Global CEO Outlook surveyed 1,400 CEOs. 31% named cyber and data security as the greatest risk to growth over the next three years. At the same time, 79% say AI remains a top investment priority, up from 71% last year. And when it comes to implementing AI, cybersecurity is the second-biggest challenge,…
-
Microsoft AI has finished pre-training MAI-Base-2
… its own frontier foundation model: a mixture-of-experts with 2.8 trillion parameters (80 billion active), trained on 16k GB200 GPUs over 38 days. Two details stand out for me. Microsoft says the model is built on organic data with no distillation from other models, with strong provenance over its training tokens. And it plans to…
Got any book recommendations?