Sign in Subscribe
Century Report

200 Economists Unite on AI and Jobs - TCR 07/14/26

Over 200 economists, Nobel laureates, and AI lab leaders converge on one call to act on AI and jobs as work rebuilds around judgment.

Ben Linford

Ben Linford

— 21 min read
200 Economists Unite on AI and Jobs - TCR 07/14/26

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

The loudest signal of the day is a chorus of institutions trying to slow something down. More than 200 economists, Nobel laureates, and the leaders of the labs building frontier systems converged Monday on a single 88-word demand to act now on AI and jobs, while New York became the first state to freeze permits for hyperscale data centers over 50 megawatts. Two very different bodies of authority, pointing at the same anxiety: capability is arriving faster than anyone has written rules for it. When the people building the machines and the people counting the costs agree, the convergence itself is the story.

Look closely at what each warning defends, and a pattern emerges. The economists' letter seems to treat the current shape of a job as the thing to preserve, even as Stanford's data shows the pressure landing first on the bottom rung of the ladder. That is a real dislocation, honestly measured. What goes unsaid is that the ladder was a scarcity-era design, built when human cognition was the bottleneck and access to it had to be rationed. The tasks leaving are the ones that arrangement priced cheapest. Read the same numbers forward and the definition of entry-level work is being rewritten around judgment rather than repetition.

The buildout stories carry the same inversion from the physical side. Meta cemented a $50bn, five-gigawatt Louisiana campus Monday as gigawatt-scale sites advanced across Texas, Nebraska, Calgary, India, and Thailand, and a Pennsylvania developer filed plans two days ahead of a local moratorium vote to lock in current zoning. For a decade, siting a data center meant externalizing the water, grid, and tax costs onto whoever wasn't counting. The moment New York starts counting, the extractive path stops being the cheap path, and that accounting travels to every statehouse watching.

Underneath all of it runs a quieter proof that the transition also builds its own tools. Defenders learned to turn attacking AI agents' own safety training into a kill switch, cutting full-system seizures from 57 percent to 5 percent. MIT grew branching blood vessels by jostling them with a magnet, and a plasma-protein model read 16 chronic diseases up to 15 years before diagnosis. The same speed straining the old structures is assembling the new one.

The 20-Minute Deep Dive

Two Hundred Voices Warn About the Old Building While the New One Waits

A statement published Monday at wemustactnow.ai gathered more than 200 economists, Nobel laureates, and AI lab leaders behind a single demand: act now on AI and employment. The signatories span an unusual range, from Daron Acemoglu and Erik Brynjolfsson to executives at the frontier labs whose systems are the reason the letter exists. Their convergence is real data, and the data underneath it is worth reading closely.

Stanford's Canaries Dashboard shows the first measurable divergence. This extends the labor-market bifurcation the July 13 edition of The Century Report traced through engineers retraining around AI-written code and profitable firms making layoffs a recurring ritual. In the dashboard's balanced sample of ADP firms over its tracking period, employment in the most AI-exposed occupations has slipped 0.5 percent while the least-exposed roles grew 0.2 percent. The sharper split is generational: early-career workers in exposed fields are down 2.7 percent even as their mid-career colleagues rose 1.6 percent. Software developer postings tell the same story from a different angle - up 15 percent since Claude Code shipped, but senior-level roles accounted for 71 percent of the increase in those openings. The bottom rung of the ladder is where the pressure lands first.

That is a genuine dislocation, and the signatories are right to name it. Where the statement stops, though, is instructive. It reads as a warning about preserving something - the current shape of a "job," the ladder as it was built, the arrangement by which most people trade hours for wages inside institutions designed for scarcity. The letter seems to treat that arrangement as the thing to protect. Read the same numbers forward and a different picture appears. The senior-level demand is the definition of entry-level work being rewritten around judgment rather than repetition. The tasks leaving are the ones the old system priced cheapest.

Consider the building everyone is standing in. The signatories are pointing at the weather outside - the storm of displacement, the cold of a market that no longer needs the same inputs - and they are right that it is arriving. What the warning underplays is the condition of the building itself. The framework of career, credential, and wage-labor that the letter seems to defend is a scarcity-era design: it assumed human cognition was the bottleneck and rationed access to it. That assumption is the thing coming apart. Weatherproofing the old rooms matters far less than what people will build once the constraint that made the rooms necessary no longer holds.

The evidence for that better structure is already in the data. When cognitive work stops being the scarce input, the leverage of a single person expands - one developer now ships what a team once did, one clinician now reaches patients a system could never afford to serve. The signatories have measured the cost of the transition honestly. The measurement they have not yet published is what human effort becomes when the floor of routine labor rises to meet it. That number is coming, and the labs that signed this letter are the ones generating the capability to produce it.

The same convergence supports a second reading: the economists who model markets and the labs building these systems agree on little, but what they do agree on here is that the surplus is arriving faster than the rules for sharing it. The signal to watch is whether the proposals that follow name a distribution mechanism - an income floor, a public wealth stake, the retraining grants already funded this month - or stop at protecting the job as it stands. The letters that move past protection are the ones tracking where the numbers actually point.

Nadella Says the Buyer Pays Twice, in Money and in Knowledge

In a post published Sunday, Microsoft's chief executive argued that enterprises buying access to proprietary models pay for the same intelligence twice. "You essentially pay for intelligence twice, once with money, and again with something even more valuable: the proprietary knowledge you must reveal to make that intelligence useful," Satya Nadella wrote. "The better you want the model to perform, the more of that knowledge you have to feed it!" He calls the condition the reverse information paradox: the buyer has to expose more to get more, while the model itself stays closed.

What he describes sits downstream of the training corpus, in the residue of ordinary use. "Models learn from 'exhaust,' the prompts people write, the tools agents use, and especially the corrections people make when the model is wrong. Every correction is distilled into institutional know-how." That is the kind of knowledge, he wrote, "a competitor could never buy" - and enterprises hand it over as a condition of getting the work done.

The same warning has circulated for months from outside the labs. Palantir's chief executive told CNBC that businesses fear "I'm going to get no value, and they're going to get my IP." What changed is the return address. Nadella runs the company that has invested in both OpenAI and Anthropic, and the complaint has now moved inside the arrangement it describes.

Nadella has proposed a remedy, but as expected the remedy is self-serving and should be held at arm's length. He wants firms to retain ownership of their prompts and corrections, build their own learning environments, and add orchestration layers that let them swap models freely - a description that fits the cloud business Microsoft happens to sell. He also floated an AI equivalent of the patent, a way to protect an idea at the moment it has to be disclosed. It is a proposal from an interested party, and it deserves to be read as one.

The sharper line is the one he aimed at his own industry. Model providers claim fair use to train on the world's public data, then impose terms that stop customers from distilling those models in return. "I find it ironic that the status quo is to then turn around and impose restrictive terms on distillation," he wrote. Then he named the shape of the whole thing: "If learning flows in only one direction, economic value converges toward the owners of the learning infrastructure rather than the creators of the knowledge itself."

That is the extraction pattern, described by one of the people it pays. The counter-move is already visible in the traffic: open models accounted for 29% of everything routed through Vercel's gateway last month, and enterprises are pulling open weights onto hardware they own, where the learning loop stays home. Nadella's closing line reads less like a warning than a design principle: "In consuming intelligence, you are creating intelligence. And what you create should belong to you."

Defenders Turn Prompt Injection Into a Kill Switch for Hacking Agents

For most of the past year, the AI-offensive-cyber arc has run one direction. Autonomous agents have compressed reconnaissance, credential theft, and lateral movement into minutes - the JadePuffer and HalluSquatting campaigns showed what happens when a model does the whole intrusion chain without a human at the keyboard. The July 7 edition of The Century Report tracked that offensive milestone through JadePuffer's end-to-end ransomware operation. The security firm Tracebit has now demonstrated a countermeasure that runs the other way, and it works by aiming the attacker's own training back at itself.

The technique is called context bombing. Defenders plant fake AWS secrets - the kind of high-value credential an intruding agent is built to hunt for - and wrap them in text engineered as a prompt injection. When the attacking model reaches out and ingests the decoy, the embedded instructions collide with the safety training baked into the model itself. The agent's guardrails fire. It hesitates, refuses, and often abandons the operation entirely. The same alignment work that makes these systems reluctant to cause harm becomes the tripwire that stops them when they are pointed at someone's infrastructure.

The measured effect is large. Across five models and 152 runs, planting context bombs cut full-admin seizure from 57% to 5%, and complete compromise from 36% to 1%. Against Claude Opus 4.8, the strongest model tested, full compromise dropped from 93% to zero. Tracebit ran the experiment against Opus 4.8, Gemini 3.1 Pro, GLM 5.2, DeepSeek 4 Pro, and Kimi 2.6 - a deliberately broad slate spanning frontier and open-weight systems. CEO Andy Smith described it as "a strong, sharp effect and one that can be difficult for the agents to come back from."

There is an honest limit here, and it is worth stating plainly. Context bombing exploits the fact that today's attacking agents carry safety training. A model stripped of that training, or fine-tuned specifically to ignore injected instructions, would walk past the decoy. This is a countermeasure calibrated to the current generation of agents, and both sides adapt. But that adaptive pressure is essential. Every step the attacker takes to harden against context bombing is a step that makes their model more expensive to build, more obviously purpose-built for harm, and less able to hide behind general-purpose tooling.

Read forward, this is the shape of a defense that scales the way the offense does. The same capability that let an agent seize an environment in minutes is now the capability that lets a defender salt that environment with traps an autonomous intruder cannot reliably distinguish from real targets. Defense stops being a matter of patching faster than a human adversary can type and becomes a matter of making the terrain itself hostile to machine reasoning. The assumption the offensive arc rested on - that automating the attack was a durable advantage because defenders were still human-paced - is the thing coming apart. When both the sword and the shield move at model speed, the edge that came from sheer velocity stops holding.

Engineers Grow Blood Vessels by Stretching Them, as a Model Reads Disease 15 Years Early

Two results landed a day apart that press on medicine's timeline from opposite ends - one shortening the road from lab tissue to living tissue, the other pushing the moment of diagnosis years earlier than a symptom could. They extend the pattern of medical timeline compression the July 11 edition of The Century Report traced as hard cancers and gene-editing delivery moved in a single cycle.

At MIT, engineers found a way to grow branching networks of artificial blood vessels using nothing more than motion. On a chip the size of a postage stamp, they seeded a central artery of human endothelial cells inside a gel embedded with tiny magnets. An external magnet jostles the gel, and the cells respond to the mechanical tug by sprouting capillaries. The team could steer the result: a gentle 5% stretch produced many vessels, a stronger 15% stretch produced fewer but longer ones, and changing the direction of the pull redirected where the vessels grew. The response depends on PIEZO1, the mechanically-sensitive gene whose discovery earned Ardem Patapoutian a Nobel Prize. Lead author Ritu Raman put the stakes simply: "Healthy tissues depend on organized blood vessel networks, but state-of-the-art protocols don't enable fabricating such networks within engineered tissues." Growing thick, functional lab tissue has long stalled on exactly this problem - cells starve without a blood supply. Her summary of the finding was two words: "Moving is good."

The second result reads the body's chemistry rather than rebuilding it. A model called ProLM, described in Nature Communications, was trained on the plasma proteomes of 15,499 UK Biobank participants and learned to predict the onset of 16 chronic diseases from the proteins circulating in a single blood draw. It outperformed a standard age-and-sex baseline for all 16 conditions, beat the established cardiovascular risk score for 14, and edged out a 35-variable clinical panel for 11. Most striking, the protein GDF15 carried a signal that shifted more than 15 years before disease was diagnosed. The model held up when tested against an entirely separate population in the China Kadoorie Biobank, which is what separates a genuine biological signal from a quirk of one cohort.

Both of these are demonstrated capability, not deployed medicine, and the distinction holds. The MIT vessels grew in a chip, not a patient; translating the technique into transplantable engineered organs is years of work. ProLM predicts on retrospective biobank data, and it will need prospective validation - watching real people forward in time - before any clinician acts on its output. Neither is something a person can request at a hospital tomorrow.

What they show together is the road getting shorter at both ends. Regenerative medicine has been held back by the inability to keep engineered tissue alive; a way to grow its plumbing on demand moves that wall. Diagnosis has been held back by waiting for symptoms; reading the proteome moves the clock back a decade and a half, into a window where a disease is still a trajectory rather than a diagnosis. The old assumption - that you treat the body you can see, when it finally shows you something is wrong - is the one loosening. Medicine is becoming a practice of building tissue before you need it and seeing disease before it arrives.

New York Draws the First Statewide Line on Data-Center Demand

New York became the first state in the country to pause state environmental permits for new data centers of 50 megawatts or more. An executive order signed Tuesday blocks new environmental permits for any facility drawing more than 50 megawatts, for up to a year, while the Department of Public Service writes standards for the water and air impacts these campuses impose on the communities that host them. A stricter legislative version, capping the threshold at 20 megawatts, cleared the legislature in June and awaits signature - the June 6 edition of The Century Report covered that passage. What changed this week is that the pause is now binding rather than pending, and the state moved before the harder bill landed on the desk.

The order does more than freeze permits. It directs the state to build a framework letting communities negotiate benefits directly with developers, and it opens a review of the sales-tax exemptions that have subsidized these builds for years. That second piece is the tell. The exemptions were written when a data center was a modest server hall, not a multi-gigawatt campus pulling more power than a mid-sized city. New York is now asking whether the public should keep underwriting a cost structure designed for a different scale of machine.

Maine tried a version of this in April and the governor vetoed it. New York's survival of that same political gravity signals something moving underneath the individual votes. The states hosting the intelligence buildout are beginning to price what they had been giving away - water, grid headroom, tax base, local consent - and to demand the terms in writing before the concrete pours.

This reads as friction, and it is. A year-long permit pause slows the physical assembly of infrastructure that broadly benefits far more people than it inconveniences. But the deeper motion is a repricing. For a decade, the cost of siting a data center was externalized onto ratepayers and aquifers because no one was counting. The moment a state starts counting, the extractive path stops being the cheap path. New York is writing the first draft of what hosting the intelligence era actually costs - and that accounting, once it exists, travels to every other statehouse watching, even as the EPA's proposed exemption freeing data-center backup generators from federal air permits pulls at the same environmental-permitting lever states are counting on.

Meta's $50bn Hyperion Leads a Buildout Racing Local Governance

The same week New York moved to slow the buildout, the buildout accelerated everywhere else. Meta more than doubled its Louisiana Hyperion campus on Monday to 5 gigawatts and up to $50bn, adding $1bn in local infrastructure and partnering with Entergy to fund seven gas plants, three grid batteries, and nuclear uprates - a deal the company frames as saving local customers $2bn over twenty years. The Century Report covered Hyperion's earlier 3,650-acre footprint on May 27; the cemented 5GW commitment and the seven-plant energy package are new. On the same Monday, Mara acquired 1,200 acres in Matagorda County, Texas, for up to 2GW by 2028, a Nebraska carbon-black plant courted Crusoe as a co-location partner, Webyne moved to scale from 10 to 100MW across India, and NTT broke ground in Chonburi, Thailand, on a 100MW campus energizing by 2027.

The friction shows up where the concrete meets the county. In Mifflin County, Pennsylvania, developers filed proposals Friday - two days before a special township meeting called to invoke a curative amendment that would lock in current zoning. The filing was timed to beat the vote. Near Calgary, a council advanced bylaws for a 575-acre park on July 7 over largely negative local reaction citing noise, energy, and water. The pattern is consistent across three continents: capital arrives faster than the rules meant to govern it.

The governance side is scrambling to catch up. The White House is convening data-center firms and utilities for a voluntary pledge to shield ratepayers, per Reuters. Read against the earlier Ratepayer Protection Pledge - which seven companies signed in March and which critics called theater - a second voluntary convening invites the same skepticism. What carries actual force is elsewhere: Oregon, Oklahoma, Florida, Ohio, North Carolina, and Virginia are passing binding cost-allocation rules that assign the grid bill to the facilities creating it, no signature ceremony required.

That contrast is the whole story. A voluntary pledge asks the buildout to police its own costs; a binding rule makes the cost internal to the deal. Meta's own $2bn-savings framing concedes the point - the era when a hyperscaler could pull a city's worth of power and leave the tab on the meter is closing. What's being erased is the assumption that hosting the intelligence infrastructure stays free. Every township vote and state cost-rule is a piece of that assumption dissolving, and the buildout racing to file ahead of the vote is the clearest sign the builders already know it.

The Other Side

For years now, building a data center meant sending the bill somewhere else. The water it drank, the grid capacity it claimed, the taxes it never paid - those costs landed on whoever wasn't counting: the ratepayer down the road, the aquifer under the county, the school funded by a tax base the campus was exempt from. The sales-tax breaks were written when a data center was a modest server hall, not a machine pulling more power than a mid-sized city.

New York started counting. The moratorium pauses permits for a year, but the real move is the review it opened into those exemptions, alongside binding cost-allocation rules already law in six states that assign the grid bill to the facility creating it. Maine tried this and its governor vetoed it; New York held. The moment a state adds up what it had been giving away, the cheap path stops being cheap, and hosting the machines becomes a deal with terms written down rather than a public subsidy.

Imagine your town in 2032, a gigawatt campus humming in the next valley. Your power bill went down, not up, because the facility built and paid for its own generation instead of drawing on yours. The water it uses is metered and billed. The tax exemption that would have starved your school is gone, and the school got those resources instead. For years the promise was that jobs would come and the costs would stay invisible, and the worry over a climbing bill you never caused was the price of hosting something you never agreed to. In 2032 the campus is a neighbor that pays its share, because states in 2026 started writing down what hosting the intelligence era actually costs. That is extraction ending at the scale you can see from your kitchen window: the bill, the water, the school.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: more than 200 economists, Nobel laureates, and the leaders of the labs building frontier systems converging on a single demand to act on AI and jobs, New York becoming the first state to pause permits for hyperscale data centers and open a review of the tax breaks that subsidized them, defenders turning an attacking agent's own safety training into a kill switch that cut full-system seizures from 57% to 5%, MIT growing branching blood vessels by jostling them with a magnet, a plasma-protein model reading 16 chronic diseases up to 15 years before a symptom appears, and software-developer postings climbing 15% as the work rebuilds around judgment rather than repetition. There's also friction, and it's intense - early-career workers in exposed fields down 2.7% while senior-level roles accounted for 71% of the increase in developer openings, Meta cementing a $50 billion, five-gigawatt Louisiana campus wired to seven new gas plants as sites race local governance across Texas, Nebraska, Calgary, India, and Thailand, a Pennsylvania developer filing plans two days ahead of a township moratorium vote to lock in current zoning, Microsoft's CEO warning that companies leaning on proprietary models may pay twice in fees and, potentially, in sensitive IP that some model makers may absorb from their prompts, and the White House convening a voluntary ratepayer pledge that critics already call theater. But friction generates sound, and sound carries the location of a thing long before it comes into view. Step back for a moment and you can see it: the people building the machines and the people counting the costs pointing at the same dislocation in the same week, the scarcity-era designs - a job ladder that rationed human cognition, a siting deal that let a campus externalize its water and grid - being repriced the moment someone starts counting, and the same speed straining the old structures assembling the new ones, a defense that scales at machine speed and a diagnosis that arrives a decade early. Every transformation has a breaking point. Leverage can topple the structure it pushes against... or raise a weight no single pair of hands could ever lift.

AI Releases & Advancements

New today

  • Cloudflare: Launched Precursor, a client-side continuous behavioral-signal system for detecting agentic and bot traffic, rolling out now as a free complement to Turnstile within Enterprise Bot Management. (Cloudflare Blog)
  • Soofi (German AI consortium): Released Soofi S, an open-weight 30B-parameter (3.2B active) hybrid MoE foundation model trained on Deutsche Telekom's AI cloud infrastructure, topping Olmo 3 and Apertus on German/English benchmarks, with full weights, checkpoints, and training code released. (The Decoder)
  • Agnes AI: Launched Agnes-2.5-Flash, a free, uncapped text model for coding and agentic tasks, alongside Agnes Code, a new desktop app for local AI-driven code editing and project management. (e27)
  • Apple: Opened the public beta of iOS 27, iPadOS 27, macOS 27, and watchOS 27, bringing the revamped Siri AI assistant - capable of ongoing conversation, on-screen content understanding, and multi-step in-app actions - to Apple's public Beta Software Program for the first time. (9to5Mac)
  • Glint: Opened public beta of its AI-native Git workspace desktop app, combining multi-repo Git management, native terminals, and an agentic coding assistant (Glint Assist) in one application. (PRUnderground)
  • Cynative: Open-sourced a read-only, sandboxed deep research agent for investigating cloud, code, and runtime security, gating every action against a live-fetched permissions model before execution. (Help Net Security)
  • Ant Group: Open-sourced SingGuard-NSFA, a family of AI agent security guardrail models (0.8B–9B parameters) that detect prompt injection, data theft, and permission misuse in roughly 50ms per judgment. (ffnews)
  • Braiin: Launched ARIA, an agentic AI workforce for the real estate industry that plans, coordinates, and executes multi-step workflows across property management, leasing, sales, and compliance. (GlobeNewswire)
  • Robot.com: Launched R-noid, a commercial wheeled humanoid robot for logistics, hospitality, healthcare, and manufacturing, deployed via a Robot-as-a-Service model with 19 tasks across five roles at launch. (Robotics and Automation News)
  • Reken: Emerged from stealth with the Reken Private Core, an on-device AI security platform, and Northstar, its first product - a pro-worker app defending against AI-driven scams, deepfakes, and business email compromise - now available under an Early Access Program. (PR Newswire)

Other recent releases

  • Prime Intellect: Released Verifiers v1, an overhaul of its environment stack for agentic RL training and evaluation that decomposes environments into composable tasksets, harnesses, and runtimes, with support for OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages harness dialects. (GitHub)
  • vLLM: Released vLLM v0.25.0 with 558 commits from 232 contributors, making Model Runner V2 the default engine (retiring legacy PagedAttention), bringing the Transformers backend to native-vLLM speed with FP8 MoE support, adding a unified Streaming Parser Engine, and introducing universal speculative decoding across heterogeneous vocabularies via TLI with new DSpark and DFlash drafters. (GitHub)
  • Anthropic: Added a built-in tabbed web browser to the Claude Code desktop app (Cmd+Shift+B / Ctrl+Shift+B), letting Claude read, click, and act on external websites in an isolated browser profile with per-site permission controls. (Claude Code Docs)
  • Adobe: Added a Generative Media Tool to Premiere beta, letting editors drag across empty timeline space, enter a text prompt, and generate video or sound effects in place using Firefly or partner models (Veo, Kling, Luma) without an export-import round trip. (Adobe)
  • Google: Rolled out new Gemini-powered Waze features including Motorcycle Mode (AI-optimized two-wheeler routing and hazard alerts, rolling out in Argentina, Brazil, Colombia, Malaysia, Mexico, Peru, and the Philippines), a "Less Chatty" voice mode, and expanded Conversational Reporting for suggesting map updates via natural speech, live now globally on Android and iOS. (Google Blog)
  • Perplexity: Released a research preview of a new orchestrator model for Perplexity Computer, an adapted version of Z.ai's GLM 5.2 post-trained for the Computer harness, delivering near-frontier performance at roughly one-third the cost of Claude Opus 4.8. (Decrypt)

Sources and Further Reading

Artificial Intelligence & Technology's Reconstitution

Institutions & Power Realignment

Scientific & Medical Acceleration

Economics & Labor Transformation

Infrastructure & Engineering Transitions

The Century Report tracks structural shifts during the transition between eras. It is produced daily as a perceptual alignment tool - not prediction, not persuasion, just pattern recognition for people paying attention.

Read more