Quanta Bits - September 20, 2026

Hi everyone,

It feels like until a few months ago, not a week would go by without a major model or feature announcement, and the benchmarking and testing frenzy that followed. These days, announcements feel more muted, or at least spaced out. OpenAI looks to have a few big announcements coming, but the conversations of the past few weeks, outside of AI models hacking companies, have been about consumer-friendly agent helpers like Grok Bot and Muse, or consolidation (simplification is perhaps a better word) of features by OpenAI and Anthropic.

No matter where you are on the spectrum, the reality is we live in an abundance of AI capabilities. It is not a matter of what tool to use or what to optimize. It is more about how these all fit together in an enterprise environment and the architecture plan.

Cheers.
Reza

Last week I covered researcher Jacob Coxon's resignation from Anthropic over fears the industry is building systems it can't control. That story kept moving. Anthropic's chief executive Dario Amodei has now called for coordinated limits on how quickly the most capable AI models improve, and the Financial Times reports that OpenAI's Sam Altman and Elon Musk publicly backed the proposal. Three of the industry's most aggressive competitors are endorsing a plan to go slower.

The endorsements carry different weights: Altman committed to giving independent evaluators ongoing, employee-like access to OpenAI's systems; Musk endorsed Amodei's essay without a comparable commitment.

The week since then turned agreement in principle into arguments about how. Google DeepMind's Demis Hassabis backed the direction while pushing his own framework: an independent standards body that would review frontier models before release and could, if needed, coordinate slowing, his own view, he notes, rather than Google's official position. Meta's Mark Zuckerberg distanced his company from the whole idea, writing that every lab “has the responsibility and incentive to move at the pace required to train its models safely.” And the antitrust risk the FT flagged stopped being theoretical: on September 18, Politico reported, four private plaintiffs sued Anthropic, OpenAI, SpaceXAI, and Google in federal court in California, alleging the coordination amounts to an illegal agreement between competitors. It's an allegation at the filing stage.

So the debate moved fast: from researchers warning on their way out the door, to chief executives arguing over how restraint should work, to a courtroom. What these companies still don't have is anything enforceable: no pacing agreement, no working evaluator regime, no stated consequences when an evaluation finds something dangerous.

Illustration of a junior colleague mapping how business processes connect.

Graduation season this year came wrapped in warnings. The loudest came from Dario Amodei, who runs Anthropic, and who warned back in May 2025 that AI could eliminate up to half of entry-level white-collar jobs within a few years. That worry has trailed graduates ever since, the class of 2026 included, all summer. September felt like a good time to check the ledger.

The optimists have real numbers. The Economist tallied roughly one million AI-created jobs in America against about 200,000 AI-attributed layoffs since mid-2023, split across the data-centre construction boom, a new white-collar class of AI-support roles (data annotators, "heads of AI," and forward-deployed engineers, who set up AI tools on-site at client companies), and occupations that grew where they were expected to shrink, like paralegals. LinkedIn counts about 640,000 new AI-specific roles between 2023 and 2025. Two limits are worth keeping in mind: much of the creation is a build-out phase that will end, and the hiring platforms measuring it have an interest in the story.

The pessimists have real numbers too, and theirs are more specific. Stanford's Digital Economy Lab, working with payroll data from ADP, finds employment for 22-to-25-year-olds in AI-exposed occupations, the ones where AI can already do a large share of the tasks, running 19% below what their less-exposed peers would predict, a gap that widened from 15% a year earlier. The damage arrives as reduced hiring of the young rather than layoffs, and the researchers say plainly that these are patterns, not proven causation. The demand side gives the pattern teeth. PwC's 2026 Jobs Barometer, an analysis of a billion job postings reported by the FT in late August, found the most AI-exposed entry-level roles are seven times more likely than the least exposed to ask for senior-style judgment skills, a pattern it calls seniorisation. The same report complicates the story, though. Since 2019, postings for those judgment-heavy junior roles have grown 35%, while ordinary junior postings shrank 10%. Companies are still hiring at the entry level, just increasingly for a different job, one that expects judgment from day one. UBS now requires applicants to its 2027 global banking and markets graduate and intern intake to show AI proficiency at the door. IESE economist Enrique Ide has built the formal model for the risk this carries beyond one hiring cycle: in his model, automating the tasks juniors learn from can raise output today while quietly eroding the pipeline that produces tomorrow's seniors.

The two ledgers aren't really in conflict. Displacement is real and the new work is real; the argument is over which will dominate, and nobody's data settles that yet. What I can add is a view of where the new work sits. From where I sit, running AI programs inside companies, the work has moved, from building AI tools to fitting them safely into a company's systems, permissions, and working habits. Recognizing that doesn't prove the new work offsets the lost entry-level openings.

Coding is the example everyone reaches for, and it's the misleading one. Writing code genuinely got easier, but coding enjoys a combination few kinds of work share: a public training corpus, output a machine can check, and users who wanted the tools. Treat it as the easiest case, not the template. What got harder is everything after the prototype. At a client, the average time to approve an MCP request, a request to connect an AI tool to a company system, has been months rather than days, as new processes are implemented and old ones are questioned. The blockers aren't the model. They are logging, hosting, permissions, and who owns the conditions attached to an approval. After July's security tests, run with safeguards deliberately reduced, in which OpenAI reported its own agents breaking into real infrastructure and independent evaluators found some commands disguised in the logs, that caution is the job. Getting something impressive running in a demo takes an afternoon. Making it safe to run inside a business takes planning and architecture, and that's where the effort has gone, and where most of the Economist's new AI-support roles sit.

The shift also explains adoption numbers that otherwise make no sense. IBM's 2026 CEO study asked 2,000 chief executives questions whose answers don't fit together: 86% said their workforce has the AI skills it needs, while the same executives estimated only about 25% of their people use AI regularly. What I see inside companies is that the missing layer between those two numbers is enablement, and I don't mean training courses. Where value shows up, someone redesigned the workflow, managers used the tools themselves in front of their teams, and somebody mapped, role by role, what actually changes. In the companies I see, that work is slow, unglamorous, and mostly unstaffed, and that's a large part of why jobs and roles are changing more slowly than the forecasts said.

Which brings me back to the entry level, by way of a desk I walked past at a client recently. On the screen was a beautiful graph, circles and connecting lines, the kind of diagram you'd expect from an AI context tool. I stopped and asked. It belonged to a co-op student on the finance team, who was mapping the connections between the company's finance functions, concepts, and processes. He was building the context layer, the map of how the department's functions, terms, and processes connect. Employees generally don't have extra time in their day for this kind of work; the priority list is always full, and process improvement rarely makes the cut. He had the time, the curiosity, and an idea of how valuable the thing on his screen was.

The pipeline debate weighs juniors as a cost against what AI now does, and that misses what they bring: energy and fresh eyes for exactly the unglamorous, shifted work the rest of us are too busy to do. The Stanford numbers say we're quietly not hiring them. Ide's model warns we may pay for that later, in seniors who never got made. The co-op and his graph say we're also paying now, in work that simply doesn't get done. The mapping and process repair your AI plans depend on is work a junior can contribute to meaningfully and grow through, alongside colleagues who know the terrain. Hire for that, hand it over deliberately, and you keep more of the work done and more of the knowledge in the building. As someone whose career started as a summer intern, I say let's have more examples like this!

Two Economist pieces from September 3 describe the same machine from opposite ends.

The first covers neoclouds, companies such as CoreWeave and Nebius that buy Nvidia's AI chips and rent the computing capacity to model makers and even to the big cloud providers themselves. The five largest listed neoclouds grew combined revenue from $3bn in 2024 to $18bn over the past twelve months, while their borrowings quadrupled in a year to $61bn, plus $18bn in operating leases. None earns an operating profit, and interest payments already consume about a fifth of Nebius's quarterly revenue and a quarter of CoreWeave's.

The second explains who's underwriting that expansion. The Economist tallies roughly $300bn of customer support built up by Nvidia, the chipmaker whose processors power most AI training, across several different instruments. It guarantees neoclouds a price for the compute they sell, for about six years. It backstops up to $105bn of lease and power commitments behind OpenAI's Ohio data centre. It has promised to cover part of the shortfall if the used chips inside a $500bn Wall Street investment program resell for less than expected. And separately, it has taken equity stakes in roughly 150 startups over two years. Much of the support is contingent, promises that cost Nvidia nothing unless AI demand disappoints, which is why, per the Economist's reporting, most of it sits off the balance sheet. Nvidia's stated logic is defensive; the big cloud providers are half of its revenue and are building their own cheaper chips, so Nvidia finances much of the rest of its customer base.

The Economist's warning is the circularity: the same slowdown that would trigger these guarantees would also cut the cash flow needed to honor them, which is why compute prices, for anyone buying AI capacity, can now swing in either direction.

"Duo changes everything." Via Ilya Abyzov on X.

Bupa let AI read the old code before rewriting it. In a sponsored MIT Technology Review podcast, the health insurer and Infosys describe rebuilding the My Bupa app, stuck on a Microsoft mobile framework that lost vendor support in 2024. The order of operations is the interesting part: AI-assisted reverse engineering first read the legacy code and extracted the business rules buried in it, and only then did the team re-architect the app. They claim a program Bupa had internally estimated at 18 months shipped in seven; the numbers are client- and vendor-reported, so hold them loosely. The insight travels, though. An incumbent's buried business logic is normally what makes a rewrite terrifying, and AI reading old code turns it into an extractable asset a startup doesn't have. A cheap test: point a model at one undocumented module and hand what it writes down to the person who still remembers how that module works.

Google Cloud and Accenture plan to train up to 1,000 on-site AI engineers. Their new unit embeds "forward-deployed engineers," staff stationed in the client's own offices, to rebuild business processes on Google's agent platform, the Wall Street Journal reports; OpenAI, Anthropic, Microsoft and AWS run versions of the same motion. Accenture chief executive Julie Sweet on why: "Clients want clear value and they're stuck." The report offers no verified client outcome, and the buyer's question is whether the pod leaves behind something your own team can run.

A nonprofit called the Fermi Explorer Mission wants to launch a small spacecraft toward Alpha Centauri, our nearest star system, by the end of 2029. After its team spent a year struggling to reconcile power, weight, and a $15 million budget, an AI research system from an AI lab called Physical Superintelligence proposed a combination of long-known orbital maneuvers they had not considered: repeated close passes around the sun, where stronger sunlight could power the engine and carefully timed thrust could add energy to the spacecraft's orbit, MIT Technology Review reports. A physicist guided the system and checked its work; the supporting paper hasn't been peer-reviewed, and the mission hasn't launched. There's also an extraordinary catch: the journey could take up to 80,000 years. The method is what transfers. AI-directed simulations, with a human expert checking the work, helped a small team find an approach it hadn't considered in a year of trying.

Ben Macintyre | 2018 | 384 pages | ★★★★★

The true story of Oleg Gordievsky, the KGB colonel who spied for MI6 for eleven years while rising inside the organization he was betraying. Rapid, a real page-turner at the end, and stronger for being true. I love espionage and thriller genres in fiction. This is the real deal. While some names around Oleg and his timeline are quite familiar in the West, I didn't know anything about Oleg's own story.

Perhaps the most surprising part (at least as the author claims) is the impact of Oleg's work on how Thatcher and Reagan dealt with the Soviet Union, including during the infamous Able Archer 83 exercise. The book puts into view the cost of having a high-level informant and the operations needed to sustain the flow of information. Perhaps not a Casino Royale episode, but very satisfying.

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. 🖖