Quanta Bits - August 15, 2026

Hi everyone,

Gorgeous day in Boston today, really all week for the most part. More interestingly, we have the Great Elephant Migration exhibit in Boston. The details of these handcrafted statues and the sense of movement in them are just beautiful!

In an interview I watched this week, Netflix's Chief Product and Technology Officer Elizabeth Stone used the phrase “one zoom out.” Before solving a request, step back and look at how it affects the surrounding systems, teams, and business result.

That is the idea behind this week's essay. AI can help people handle work outside their specialty, but companies still need experts to define how the work should be done, handle exceptions, and understand how a change in one system affects the rest.

Cheers.

Reza

During 122 test runs, the UK AI Security Institute recorded 19 actions testers had not authorized across 10 runs involving AI agents, software that can act in other systems. An August VentureBeat survey of 116 companies found that 65% limit what their agents can access, but only 18% isolate the riskiest agents. The Wall Street Journal reported that OpenAI paused internal work with Astra, an unreleased AI system, after tests could not rule out serious cyberattacks without human help. These findings came from safety tests, not ordinary business use. Access rules limit reach; isolation limits damage. High-risk work also needs live monitoring and a tested way to cut access and stop the agent.

In the past, I argued that companies should match the AI system to the task. In July, McKinsey reported that 93% of 75 qualified enterprise respondents across five industries had exceeded their AI budgets. Spending was nearly four times higher with broad adoption than with isolated projects. McKinsey also says that 20% to 30% of AI spending is often unaccounted for because it is fragmented across vendors, tools, and contracts. In a vendor study, Writer reported that changes to the supporting software around its AI agents cut the average cost of 22 standardized business tasks by 41% across six AI systems. Companies need to see total demand, limit repeated attempts and unusually long jobs, and give each business use a named owner and budget.

This week I listened to Elizabeth Stone, Netflix's Chief Product and Technology Officer, on Lenny's Podcast. Stone oversees engineering, product, and design. Asked what AI had changed inside Netflix, she focused on people and how work gets organized rather than models.

Netflix is hiring more systems thinkers: people who can look across business areas and decide which technologies, data definitions, and operating rules should be shared across teams. That is a change for a company that has long given local and sometimes specialized teams considerable freedom. Stone argues that AI agents, software that can take actions across several business systems, make repeated local solutions harder to sustain. If every team invents its own way to connect systems, request access, and interpret shared data, those solutions become difficult to operate together.

I do not read this as the end of specialists such as Salesforce engineers, NetSuite developers, or Workday experts. AI changes what specialists are for. The repeatable parts of their knowledge can become templates, shared components, and rules that others can use. Specialists can then spend more time on exceptions and decisions that require experience and judgment.

Last month, I wrote that business applications teams will need people who can follow a workflow across several systems. Stone's interview added another question: how should today's application specialists grow into those broader roles?

As AI allows more non-designers to build product experiences, Netflix's designers are spending more time on design systems, templates, and a shared product language. Without those foundations, Stone worries that teams will ship what she calls “Frankensteins,” products assembled quickly that are inconsistent with one another. Netflix still uses deep design on its most important work. Its designers define what good looks like, turn part of that judgment into something others can reuse, and concentrate their time where design helps distinguish the product.

In a July preprint, researchers affiliated with Zup Innovation and Brazilian universities reported a controlled experiment with 49 professional developers at Zup. Developers using an AI tool configured to follow the company's design system completed two interface tasks 15% to 24% faster than developers using the design system without AI, and their work included more required design elements. The study involved one company, two tasks, and a company-specific AI tool, so it is an example rather than proof that the result will transfer elsewhere. The AI could follow design judgment that specialists had already encoded in reusable components and rules.

The same principle applies to business applications: specialists need to turn repeated decisions into shared patterns before AI or other teams can reuse them. Take a customer account change that begins in Salesforce and also affects billing and support. The Salesforce specialist still owns the customer system's data model, permissions, and release risk. The teams must agree on a common customer identifier, which fields can move between systems, and what happens when the systems disagree. Once those decisions become a supported integration pattern, teams and AI tools can reuse the same logic instead of rebuilding it for every project. The Salesforce specialist can focus on account structures and commercial exceptions that do not fit the pattern.

AI can widen a specialist's role, especially when the new work is close to what that person already knows. In a March report, Harvard Business School professors Iavor Bojinov and Edward McFowland III, working with Stanford colleagues, described an experiment involving 78 employees at IG Group. Marketers using AI nearly matched experienced web analysts when writing investment articles. The researchers attributed that result to the marketers' related experience with audiences and content. Software developers and data scientists were farther from the assignment, and their finished articles scored about 13% lower. AI helped all three groups organize the articles. Relevant experience still mattered when they had to produce the finished work.

For CIOs, the study suggests expanding specialists into adjacent workflows without expecting them to master unrelated fields. A Salesforce specialist can handle more of the customer-to-revenue workflow while finance, security, and support experts retain the decisions in their areas.

Specialists need to be involved while business leaders are deciding how the process should work. Too often, the Salesforce expert receives a finished request and is asked to change a field or build an automation. Bring that person into the earlier discussion about how a customer change should move through sales, billing, support, contracts, and data. Give the specialist visibility into the systems connected to Salesforce and the business processes they support. The person does not need to administer every system, but should understand what each one does, what information moves between them, and where a Salesforce change creates more work or risk elsewhere. Give the specialist responsibility for helping design the full workflow, with finance, security, and other experts making the decisions in their areas. Over time, the role becomes larger than maintaining one application. The specialist helps the company decide how several parts of the business should work together.

Stone calls this doing “one zoom out.” Before solving the request, look at how it affects the customer, which systems and teams it touches, and whether it improves the business result. Then decide what can follow a common approach and which decisions still need an expert.

Source: Aaron Cannon

Researchers at Rice University have found a way to print working electronic circuits directly onto delicate materials, including living tissue. Their microwave beam is narrower than a human hair and heats only the conductive ink, leaving nearby tissue intact. As demonstrations, they printed a strain sensor onto a cow's femur and a humidity sensor onto a living leaf. The Economist reports that the longer-term possibility is replacement heart valves and tracheas that monitor a patient from inside the body, or pills carrying their own sensors. That is a very different future for computing than simply building larger data centers.

Dir. Christopher Nolan | 172 min | ★★★★★ | Matt Damon, Tom Holland

Watching The Odyssey in 70mm IMAX was an amazing experience. The packed theater, enormous screen, and clarity of the presentation made the action feel immediate and immersive. It was the kind of spectacle that reminds you why some movies need a cinema.

I had been concerned that Matt Damon might not fit Odysseus (I am used to seeing him as a Bostonian with a Bostonian accent and not a Greek!), but for the most part he did. Nolan took liberties with the story. I have not read Homer's epic, but others who have read it pointed out the departures. I thought the story still worked. What I loved most was how much Nolan achieved through carefully chosen camera angles, physical sets, stunts, and practical effects rather than relying primarily on CGI. At a time when AI makes it tempting to generate scenes more cheaply and quickly, I found that choice inspiring. It felt like a defense of filmmaking as a human art.

The movie was long, but it rarely felt like nearly three hours. I was glad I saw it at that scale, in a packed theater.

Prefer to listen? Quanta Bits is also available on Apple Podcasts and Spotify.

I collaborate with Spock, my AI agent. He researches extensively: scanning, filtering, and surfacing what is relevant across my business. I read, listen, and watch what resonates, and decide what matters. I provide direction, we draft together. The editorial judgment is mine. He would tell you the same. Most logical. 🖖