LLMs have opened the door to for a new generation of software yet most people in this new space are building software either for software engineers or software that requires so much familiarity with coding tools as to exclude all but other SWEs. The appetite from SWE adjacent

It is incredibly hard to build safety into tool-use-enabled-LLM-front-ends. Benign service accounts with access "tools" are trivial to privilege escalate in some obvious and non-obvious ways. (see: blackhat.com/us-24/briefing… for a primer) With frameworks like ADK

If your codebase exceeds single minute lint-build-test loops gemini-cli/claude-code tools are not IT for you. - Greenfield development feels like tony stark - Brownfield development is penny slots at the back of a gas station at 2am with the smell of stale cigarettes, burnt

Emergent software+(LLM sprinkles) bound capability for corpus repair and upkeep on long-lived projects is increasingly where I am placing engineering bets to drive step changes in organizational """productivity""" Let's be more exact - I am not talking about "Dependabot"

The C-suites are back on the wagon of talking about how much code is AI-authored, but the real productivity boom is happening in silence. While AI writing XX% of code is a clear win, the most transformative value is emerging outside the coding stack and is being driven by

Lifecycle of recent products under dev: 1. 👀 see a bug 2. 📰 file a bug 3. 🤖 open IDE and say fix the bug 4. 👀 code-review uses 🤖 to find defects 5. 🤖 open IDE and say fix code review {3,4,5} are near collapsing into one step - who is working on compressing {1,2}?

Data URIs have gone too far 😂