Quanta Bits - May 24, 2026

Hi everyone,

Long weekend here in the US, so I am keeping this one a bit on the lighter side.

The main essay is about something I have been experimenting with lately: using agents to separate the content layer from the presentation layer, so a working note can turn into a visual surface without rebuilding the whole thing every time.

Cheers.
Reza

Review is becoming the job. The Financial Times says AI-generated vulnerability reports are flooding security teams, while Peter Yang's hackathon joke captures the same thing in developer work: people are waiting on agents, checking outputs, and deciding when to intervene. When output gets cheap, review becomes the scarce capacity. (FT)

Agents are moving into the operating layer. Google used I/O to push agents across Search, Gmail, YouTube, Docs, and Chrome, while HBR's agentic marketing piece makes the same point inside one function: the work changes when agents sit inside the workflow. The next question is less "which AI tool?" and more "which operating surface owns the work?" (Wired, HBR)

AI backlash is becoming local and economic. Gallup found 71% of Americans oppose local AI data centers, and the worker-upside idea you clipped asks whether people should share in the gains if AI weakens their jobs. AI adoption is leaving the demo phase and running into land, power, jobs, and fairness. (Gallup, ImpactAlpha)

Thariq Shihipar, who works on Claude Code at Anthropic, made a point recently on How I AI: "HTML is the new Markdown." I think the useful part is less about the file format and more about review.

Agents are now producing longer plans, richer drafts, and more complicated work products. If all of that comes back as a giant wall of Markdown, people, especially me, stop reading closely. The harder part is making the review manageable.

This matters because a lot of operating work is recurring communication: pipeline reviews, campaign launch trackers, client status updates, initiative portfolios, risk views, forecast decks, architecture reviews. The facts change every week, but the communication package keeps getting rebuilt from scratch.

The payoff is practical: less rework, faster review, fewer stale versions, and less drift between the facts, the narrative, and the actions.

Claire Vo added another useful line in the same conversation. As agents run longer, product and operations people become "compute allocators." Someone has to review work and decide which work is worth model time, how much human review it needs, and whether the output is good enough to move forward.

The practical split is this: Markdown can be the memory layer. HTML, dashboards, diagrams, and visual surfaces can be the understanding layer.

I have been experimenting with this in my own work over the last few weeks. The old flow was familiar: gather notes, make an outline, build a deck, polish it, and then rebuild it when the facts changed. That works, but every update becomes another small production job.

The better flow I am starting to use has three layers.

First, I point the agent at the context: meeting notes, research clips, project notes, prior examples, a style guide, a logo, or even a PowerPoint that shows the visual language I want. This is the part most people rush through. If the agent doesn't know the material or the audience, the first output is usually generic.

Second, I give it a shape. I might say: this needs to explain a project timeline, show current operating risks, compare three options, or summarize a messy set of initiatives for an executive audience. I am giving it the job the artifact has to do.

Third, I let it propose a visual surface, then I react to it like I would react to a draft. Too dense. Too much text. Move the timeline higher. Make the decision clearer. Once the shape is close, I ask for the important separation: create a human-readable Markdown file behind the visual.

That last step changes the workflow.

Now the content is not trapped inside the design. The Markdown file holds the facts, narrative, statuses, risks, decisions, and open questions. I can edit it like any normal note. Then the agent can regenerate the visual surface from that source.

This changes the deck.

This is most useful when the artifact repeats. A one-time board deck may still deserve slides. A weekly pipeline review, monthly initiative update, campaign readiness tracker, project status deck, or client update usually needs something else: a reusable source and a reusable surface.

Business applications teams already know this pattern. A CRM, project tracker, or finance system may hold the source data, but people still need dashboards, queues, exception views, and review screens. Agents make the same pattern easier to apply to documents and updates.

For Quanta, we are working toward a standard way to describe projects, timelines, and status. The source layer can hold objectives, stakeholders, decisions, dependencies, risks, blockers, dates, and open questions. The visual layer can decide what to show, what to compress, and what to turn into a one-page working surface.

There is still a governance question. If the visual artifact contains sensitive data, you need to know who can see it, where the facts came from, what changed since last time, and which source wins if two versions disagree. A working surface should not become a loose copy of the truth.

Markdown still has a job: memory, editing, versioning, and portable text. The visual surface has a different job: helping people understand the work quickly enough to discuss it.

That feels like the right long-weekend experiment. Pick one annoying recurring artifact. Build the source file that remembers the facts. Generate the surface that explains them in five minutes. If the pairing works, keep it. If it doesn't, discard it before it becomes another process.

Writer turns agents into event-based workflow infrastructure. VentureBeat's Michael Nuñez reports that Writer Agent can listen for events in Gmail, Gong, Calendar, Drive, SharePoint, and Slack, then start predefined playbooks without a human prompt. The interesting part is the wrapper: connector profiles, agent profiles, observability, logs, and encryption controls.

Peter Steinberger's tool stack points to the next agent-infrastructure layer. Steinberger's recent tools point to a practical pattern: agents work better when memory, browser visibility, computer visibility, and business apps are exposed through narrow, callable interfaces. Gog puts Google Workspace in the terminal. Lossless Claw treats long context as infrastructure. Peekaboo makes visual inspection callable.

Americans oppose local AI data centers at rates worse than nuclear. Gallup's first poll on local AI data centers found 71% of Americans oppose building one nearby, including 48% strongly opposed. That is higher than opposition to local nuclear plants in the same survey. The objections are mostly about energy, water, pollution, utility bills, and taxpayer support.

AI may already be reshaping the graduate job market. The Economist analyzed graduate employment data and found that majors most exposed to AI saw worse full-time employment outcomes after ChatGPT's release. I would be careful with the causal claim, but the operating signal is hard to ignore: entry-level work is being re-priced first. If junior work gets thinner, companies may need to redesign how people learn by doing.

AI helped map a possible non-opioid pain therapy. ScienceDaily summarizes a University of Pennsylvania preclinical study on a gene therapy that targets pain processing circuits without activating reward pathways tied to opioid addiction. AI was used to monitor behavior in mice, estimate pain levels, and help guide the treatment design. Very early, but hopeful: chronic pain affects roughly 50 million Americans, and a therapy that reduces pain without feeding the opioid crisis would be a genuine medical advance.

▶️ Google AI announcements roundup: Craig Hewitt's overview and Peter Yang's I/O strategy review. Useful catch-up on how broad Google's agent push has become.

🐦 Avid on Airbnb's agentic coding work. Interesting because Airbnb is being discussed less as an AI demo shop and more as a real production migration case: senior engineers showing how agents fit into engineering work.

▶️ Anthropic Might Buy Atlassian For $40B. Here's Why It Makes Sense. Sharp essay-video on why Anthropic may want the layer where work actually gets done: issue trackers, CRMs, ERPs, and other systems that already encode state, ownership, permissions, and audit history.

▶️ How This Solo AI Founder Bootstrapped 5 Products to $1M+/Month. Good, inspiring listen on launching products quickly and staying a solo entrepreneur as long as you can.

Robert Louis Stevenson | 352 pages | ★★★☆☆

Stevenson's pirate classic has the shape of the palate cleanser I wanted: a map, a sea chest, a ship, a mutiny, an island, and Long John Silver moving through the story with more charm and menace than almost anyone else in it. The opening worked best for me. Billy Bones at the inn, Black Dog, Blind Pew, and the stolen packet had the mystery and danger I was hoping for.

Once the story reached the island, it became average. The fighting scenes and siege mechanics did not pull me in, and the book felt more like a children's adventure than I expected. I still liked the change of pace and the pirate mythology around it. Useful break from heavier recent reads, just not one that fully landed for me.

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's 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'd tell you the same. Most logical. 🖖