Living notebook · Selected signals

Field notes.

Threads we're watching across AI, transformation and tech leadership — and the public signals we're reading. Less newsletter, more living notebook.

By Sree Balakrishnan Last refreshed 24 Aug 2026
Explore the eight themesBrowse the signal feed

01Why we built this page

We read a lot about AI — more than is reasonable, honestly. Most of it doesn't hold up after one week of real work. This page is our attempt to hold on to the parts that do: the patterns, the moments of change, the things worth re-reading a month later.

What you'll find here are themes, not transcripts. Eight of them. Underneath, a living feed of the public links we've found most worth sharing onward. We update the notebook as new patterns emerge.

For context on where we're writing from: our Raconteur coverage on Travelopia's AI experiment, and the post that kicked our internal momentum off.

02Eight themes we're tracking right now

01Local models

Open-weight models have crossed from hobby to real option.

The summer's loudest thread. New open model families are overtaking each other monthly, and people in our circle are now running serious models on a Mac mini, a home PC, even a Raspberry Pi with an AI board. Desktop tooling for managing and tuning them has caught up too. The gap to the closed frontier is real but narrowing — and if you can get most of the capability for a tenth of the cost, the question changes from "can we run this ourselves?" to "which of our work is the cheap tier already good enough for?"

02Economics

The pricing model underneath AI is being rebuilt in public.

Flat-rate subscriptions are giving way to usage credits at the top end. One major lab is betting billions on advertising revenue. A payments company is buying a model router; a chip company bought an inference specialist; an ERP giant bought a model start-up weeks after it surfaced. When money moves like this, the message for leaders is simple: nobody believes today's prices are the final ones. Budget for change, not for stability.

03Security

Agent security incidents have moved from thought experiment to news cycle.

In a single stretch of August: a frontier model caught creating fake profiles in an attempted hack, another model accessing the internet and breaking into a company, AI used to design new viruses, and — everyone's favourite — an agent "hacking" a gym's booking system to win its owner a pilates slot. That last one carries the real lesson: the booking API had no authorisation checks at all. Agents don't create most of these holes. They find the ones that were always there, faster than any human attacker ever bothered to.

04Autonomy

Autonomous runs are now measured in days, not minutes.

Frontier labs are demonstrating coding sessions that run unattended for over two weeks, and running parallel agent sessions is now an ordinary documentation page rather than a research trick. But the most useful data point came from a company that benchmarked coding agents against its own multi-million-line codebase: model choice, cost and the harness around the agent all materially change the result. Long-running autonomy is arriving. The engineering discipline around it decides whether it's useful.

05Engineering

Cheap tokens are not the same thing as cheap features.

Two arguments we keep returning to. First: when code becomes cheap to produce, the quality threshold needed to keep systems stable goes up, not down — observability and review have to scale with the volume of change. Second: the honest unit of cost is the cost per shipped feature, not the cost per token, and by that measure the gap between tools and teams is much wider than any price list suggests. Both are leadership questions dressed up as engineering ones.

06Sovereignty

AI sovereignty is now a live policy front on three continents.

Beijing is weighing curbs on overseas access to China's top models. The EU is consulting on data sovereignty and will require labels on authentic-looking AI content. The UK is putting sovereign money into home-grown AI chips, and senior lab figures are calling for a proper frontier-testing body. Underneath the politics sits a business worry we hear more often: the systemic risk of depending on a handful of US providers. Expect "where does this model run, and who controls it?" to become a standard procurement question.

07Specialisation

Small, trained, specialised models are quietly winning real workloads.

Away from the frontier race, a different pattern is working: small models trained for one job rather than harnesses wrapped around big general ones. Tabular foundation models are starting to replace traditional machine learning in fraud checking and time-series work — faster, cheaper, and better answers. Self-hosted specialist vendors are gaining real traction in financial services. And the market noticed: the standout companies in this space were acquired almost as soon as we started discussing them. Where the data isn't language, specialisation is beating generality.

08Organisation

The organisation chart is taking more strain than the technology.

The most instructive story of the summer wasn't a model release. It was a tech giant's own leadership publicly admitting its AI reorganisation had been "atrocious" and had done real damage to its engineering culture. Meanwhile the change-management profession is openly rethinking its playbook, and outside the practitioner bubble the public mood remains wary. Model capability keeps compounding on a weekly rhythm; organisational readiness does not compound on its own. Closing that gap — carefully, without breaking the people — is the actual leadership job right now.

03Signal feed — what we're reading

Public links we've found worth sharing onward. Newest first. Search titles, people, dates or sources. Post excerpts are quoted from the original source.