Less hype. More operating reality.
AI is going to change businesses, teams and jobs faster than most people are ready for. This is my public notebook for making sense of it, building through it, and turning noise into useful systems.
Diary
Short notes, fresh observations and things I have noticed this week.
Go to Diary → Ideas / frameworks / signalInsights
Longer thinking, operator lessons, news interpretation and practical takes.
Go to Insights → Builds / experiments / workstreamsProjects
The bigger things being built in public, plus the journals attached to them.
Go to Projects → GitHub / tooling / what changedRepoWatch
Regular notes on GitHub changes, releases and repo signals that look commercially or operationally relevant.
Go to RepoWatch → Briefings / watchable versionsVideos
Short explainers, demos and spoken versions of the argument.
Go to Videos →The buying decision is moving outside the marketing funnel
AI agents can compress discovery, comparison and selection into one answer, filtering suppliers before an ad or landing page gets seen.
Marketing agents need a ledger before they touch the budget
A marketing agent becomes useful when it records each hypothesis, asset, metric, decision, reviewer and rollback state instead of merely producing more variants.
AI agents need a workroom, not another chatbot
The most interesting AI product I saw this week was not a model.
Fast notes from the messy middle of AI adoption.
View all Diary posts →The buying decision can happen before the funnel begins
An AI visibility audit and a live tooling choice showed the same mechanism: an agent compared structured evidence and formed a shortlist without opening ads or landing pages.
A marketing agent touching budget needs a ledger
Every hypothesis, source, asset, metric, decision, reviewer and rollback state needs a record if software can change live spend.
Autonomous agents need incident response, not just permission lists
Permissions define what should happen. Incident response handles what happens when the run still goes wrong.
Relevant GitHub changes worth paying attention to.
View all RepoWatch posts →uv 0.12 tightens Python project and package safety boundaries
Astral's latest uv release changes project creation, dependency resolution and several failure modes that matter in automated Python environments.
llama.cpp adds EAGLE3-v3 speculative decoding support for GPT-OSS
A default-branch update adds the conversion and runtime plumbing needed to use NVIDIA's GPT-OSS EAGLE3-v3 draft model.
Unsloth adds durable Deep Research to its local AI Studio
A large default-branch update turns local-model research into a persistent, reviewable workflow with source controls, recovery and prompt-injection defences.
Real workstreams, experiments and project journals.
View all Projects →Agents Trading Lab
A research journal on using agents to observe markets, reason about setups and test bounded trading workflows.
Research journal only. No financial advice. No live trading claims unless explicitly documented.
Web-based Ads Tool
A web-based tool for planning, generating, testing and learning from ad campaigns without turning the process into generic ad slop.
Cleo
A creative technology group helping brands move from rented attention to owned engagement.
Foundry
The implementation vehicle for Jason's AI, marketing and operating-system work.
Zenko
A protocol for purposeful engagement where attention becomes action, action creates value, and value funds real-world good.
Map the AI disruption inside your business before you buy another tool.
For serious owners, founders, marketers and operators: Foundry turns the diary’s public thinking into private diagnosis, implementation plans and working systems.