Hey folks,

One of our favourite recurring themes on the podcast has been the idea that intelligence is no longer just a fixed block you hire, but something that can be broken into small, right-sized pieces and applied exactly where it’s needed.

Traditionally, if a business needed intelligence, it used to be person-shaped. A salaried, fixed-size, usually-on, block of cognition you pay for all year round.

But AI allows us to apply small, right-sized portions of machine judgement exactly where they’re needed, while leaving the rest of the workflow to software, automation or people. Instead of wrapping an entire process in intelligence, we can insert snack sized units of intelligence where it is genuinely needed.

But now it’s more than just one of our recurring themes. We’re pleased to introduce you to our all new media brand and format for the podcast and newsletter - Intelligence Snacks.

Our old newsletter format was an essay built around one idea from the podcast that week, but there were often 10-15 themes that we thought were worth sharing.

So, we’ll be sending out a weekly mixtape of bite-sized distillations of the big themes and ideas we’re exploring from the world of AI.

We also have a new website where we’ll be compiling our latest episodes, contributors and snacks all in one place.

Check it out at intelligencesnacks.com

This week we were joined by Anthony - an independent researcher who left the corporate world in 2024 to work in Bitcoin full time. He’s also an alumni of Chaincode’s 2025 ₿OSS program and holds a PhD in computer science from the University of South Australia.

We dug into the shift toward local AI models and what that means for privacy, sovereignty and the future of AI adoption, and why open-weight models are now capable enough for everyday knowledge work.

We also explored why the next phase of AI will be defined by the orchestration layer that coordinates intelligence, rather than the models themselves, and how we're using Wingman in this way to run our business and help other SME's do the same.

For small manual tasks, hosted frontier AI is currently much easier for consumers to access than capable local models. People can point a hosted model at everyday jobs immediately, while running a capable model at home still means buying hardware powerful enough to support it.

That creates a meaningful hardware threshold. Anthony’s local setup calls for roughly 32 GB of GPU memory, turning private AI into an investment before it delivers any value. He contrasted that with spending around $5,000 on local AI hardware versus about $20 a month for an Anthropic or ChatGPT subscription.

Awareness widens the gap even further. Most people don’t realise capable local models are even an option now, so they continue using hosted frontier models for tasks that don’t necessarily require frontier intelligence. Until local AI becomes simpler and more affordable, hosted models remain the easiest path for most consumers.

Once you’re working with multiple AI agents, a common communications and orchestration layer becomes really useful. One component is a kanban board. When an agent delegates work, it creates a task on the board, giving both people and agents a shared view of what has been assigned, what is in progress and what needs review.

The board is paired with an agent that starts every hour and does catch-up work, looking for anything that’s been dropped or left in progress. The business also defines its core scopes, giving work a clear structure around products, teams or domains. Together, delegation, task tracking, scheduled follow-up and scoped workspaces provide a practical foundation for coordinating both people and AI agents across different areas of the business.

Communication across this layer doesn't have to be equally open in every direction. For example, a chief-of-staff agent might be allowed to contact anyone, while other agents have much narrower permissions. This preserves a coordinating role without giving every agent unrestricted access. The target is also deliberately modest. The goal isn’t to support thousands of users, but to work well for a team of around twenty people.

AI automation works best when routine operations stay in code and agents appear only where the work calls for judgement.

Simply replacing parts of an existing workflow with an agent often leaves the underlying process unchanged, even though it may be poorly suited to AI. Pete described early systems where agents spent time running scripts that could just as easily have been executed directly, consuming context and tokens on work that never required an agent in the first place.

Andy saw the same pattern while building Pliny, a personalised AI-powered research briefing. Initial pipeline runs took 15 to 20 minutes because too much of the workflow had been placed inside the agent runtime. Reworking the pipeline meant identifying which steps could simply be code, leaving the agent to appear only where judgement was actually required.

Deterministic work doesn’t belong inside an agent runtime. It belongs in ordinary code, with the agent entering only at the points that require intelligence.

If this resonated, we’d love for you to forward this newsletter to someone who might enjoy exploring these ideas too.

See you next week!

Cheers,
Pete & Andy