The detail I keep coming back to after reading this article, is Topol’s claim that epigenetic clocks are reading erosion in PRC2-associated regions. Ageing clocks have had an awkward status: impressively predictive, mechanistically fuzzy, and therefore dangerously easy to treat as scoreboards. If the new model holds, the clock becomes more interesting because the number is tied to a specific failure of cellular identity. The intervention story then gets much stricter. A therapy that lowers an epigenetic age score is weak evidence by itself. The meaningful test is whether cells recover the correct identity, function normally, and stay there without sliding into fibrosis, pluripotency or cancer. Partial reprogramming makes that tension almost comically explicit: the same machinery that might restore a cell’s youthful programme can, with too much reprogramming, erase the very identity you were trying to save. That framing also makes the 96-year-old fibroblast result more useful than the usual “reverse ageing” headline. It shows reversibility exists at least in cells. It does not show that a human organ can be safely reset. The next stage needs boring, unforgiving evidence: durable function, tissue specificity, dose windows and long follow-up. Rejuvenation will deserve the name when identity survives the treatment.

I'm a big fan of Bill Watterson and Calvin & Hobbes. When I was a teenager, back in the 90's, I would love to get to the Sunday pages of Publico (Portuguese newspaper) and read the colourful strip. 10 years ago, and with adult money now, I bought the whole Calvin and Hobbes collection and I have it now with me, all of it, nothing's missing! So the most revealing number here is seven; out of 1,800 newspapers carrying Calvin and Hobbes, only seven actually dropped the strip when Watterson demanded more space and fewer layout restrictions. The supposed commercial catastrophe barely happened. Watterson’s integrity mattered, but so did the bargaining power created by making something people genuinely did not want to lose. He could refuse merchandising, demand sabbaticals and redesign the Sunday page because Calvin and Hobbes had become unusually difficult to substitute. Quality gave him leverage, then he spent that leverage on keeping the work strange, personal and inconvenient. The uncomfortable twist is that winning did not solve his problem. More control produced better conditions for the art but also more work. Removing the standard Sunday layout meant every page required fresh decisions; the machinery he rejected had been constraining him, but it had also been doing useful labour. Eventually the freedom became exhausting. Autonomy has an operating cost. Watterson kept paying it until the price was the strip itself. The extraordinary thing is that when the bill finally became too high, he still refused to finance another decade by turning Hobbes into a stuffed toy.

Recently at Hitachi, we've been working on our OFE (Outcome Forward Engineering) and that encompasses, of course, a CLI. So building the perfect one - for engineers - is something quite close to the heart and these guidelines and rules are quite good - we adopted them. The line I’d pin above every internal-tools backlog is the recommendation to print something within 100 milliseconds. That sounds like a tiny UX nicety, but it exposes the guide’s deeper argument: responsiveness is information. A silent command forces the user to diagnose whether it is working, blocked, waiting on the network or dead. One line of status collapses that uncertainty before any actual performance improvement has happened. That principle scales beyond spinners. The guide repeatedly separates human-facing behaviour from machine-facing contracts: TTY-aware output, --plain and --json, explicit exit codes, prompts only when interaction is possible. I like this because it treats usability and automation as two views of the same interface rather than rival requirements. The shell is full of tools that accidentally make humans parse machine output and machines scrape prose. Neither constituency enjoys the arrangement. A CLI can acquire technical debt without adding much code; it merely has to teach users that an accidental behaviour is dependable.

4,000 - once a Dot can choose among thousands of connected apps, software discovery starts becoming a machine-readable problem. The user may never visit a marketplace, compare screenshots or even know which product completed part of the task. The agent needs to understand that a tool exists, when to invoke it and whether it can safely do what is required. That changes the developer optimisation problem. Being available inside ChatGPT is only the first hurdle. Tools will need clear capabilities, predictable authentication, sensible permissions, reliable execution and descriptions that an agent can interpret correctly. An excellent application could become effectively invisible if the orchestration layer rarely decides to call it. The missing billing system therefore feels secondary for now. OpenAI already has pieces that shape discovery, identity, execution, review and enterprise purchasing. Revenue sharing can arrive later if the distribution channel proves valuable enough. For developers, the more immediate shift is stranger: the next version of app-store optimisation may be aimed partly at software rather than people. Screenshots and catchy names matter less when your first customer is an agent deciding, in milliseconds, whether your tool belongs in the plan.

I think the useful provocation here is that 80% figure; if inference really becomes the dominant cost of enterprise AI, procurement stops looking like SaaS and starts looking more like cloud computing. Buying access is cheap; sustained use is where the meter spins. Success can therefore make the budget worse before it makes the P&L better. That creates an awkward incentive problem. A company that celebrates token consumption, agent runs or raw adoption may simply be rewarding expensive behaviour. Domanic’s example of employees becoming more selective about model choice points towards a healthier engineering metric: value produced per unit of inference. The most capable model should not become the organisational equivalent of taking a taxi to the kitchen. I’m less convinced by the precision of the $6.6 million template. Without underlying data, numbers such as 80%, 20 agents or 5-7% of engineering salary are planning hypotheses dressed rather neatly. Section also sells the enablement and transformation services included in its own prescription, which makes that precision worth interrogating. Still, the budgeting model captures something important. AI costs are becoming behavioural. Architecture, model choice, agent design and employee habits can all move the bill materially. CFOs will increasingly need FinOps-style visibility into inference, while CTOs will need to explain why some expensive workloads deserve to exist at all.