Week Ending September 18, 2026

Research window: September 12–18, 2026

A note about the new format: Beginning with this issue, the Brief starts with the Summary. The purpose is simple: give you the conclusion first—the one development I think matters most—then let Newsworthy and Signals show the reporting and connections behind it. Reviewed for This Brief remains at the end for anyone who wants to follow the source trail.

1. Summary

This week made me a little uncomfortable.

For the last few weeks, we have been watching AI move from answering questions to actually doing things. Then we saw companies begin putting numbers around the productivity gains.

This week we got a glimpse of what may come next: AI is increasingly helping build better AI.

Anthropic says Claude now leads about 26% of its AI research and development work. Humans are still very much involved, so this is not AI independently creating its successor. But the direction is worth paying attention to.

Better AI can help researchers build the next generation faster. That better generation can potentially help even more.

And once you see that, several other stories this week begin to fit together. Why are AI leaders talking about slowing down? Why are companies spending billions on independent safety testing? Why are communities pushing back against the enormous data centers needed to power all this?

The big question may no longer be simply how fast AI is improving. It may be how much AI itself is beginning to contribute to that speed.

That’s what I would watch.

2. Newsworthy

One of the people building frontier AI is now asking the industry to deliberately slow down. Dario Amodei argued that model capabilities are advancing faster than our ability to understand and control them. His proposal is not to stop AI research, but to “pace” it: embed independent evaluators inside frontier labs, coordinate safety standards among leading companies, and eventually seek international coordination.

Dario Amodei — We Must Pace the Frontier — September 12, 2026

AI helping to build better AI is no longer just a theoretical feedback loop. Anthropic reports that Claude now “leads” 26% of its AI research-and-development work—meaning it can do most of those tasks from a high-level instruction while a human supervises—and collaborates on more than 90%. Anthropic says Claude is not yet fully autonomous on any measured portion of the work, but the company explicitly says it is measuring how close AI is getting to autonomously building its successor.

Anthropic — Measurements for understanding the pace of AI development inside frontier labs

The scale behind that number is startling: Anthropic says about 30,000 AI agents are doing research and engineering work inside the company at any one time. Their actions are monitored before and after execution, with automated systems able to block dangerous actions and humans reviewing escalations. That provides a rare glimpse of what an AI-heavy workplace can actually look like—not one chatbot beside each employee, but thousands of agents working simultaneously under layers of supervision.

Anthropic — Measurements for understanding the pace of AI development inside frontier labs

The data-center backlash we were watching last week became much more concrete this week. Virginia—the state with the world’s largest concentration of data centers—is tightening rules around secrecy, energy and environmental effects. New York is recommending community investments of at least $1 million for every megawatt of new data-center demand; Reuters reports that rising power bills, water use and local impact are producing bipartisan resistance around the country.

Reuters — Virginia tightens data center restrictions amid political backlash — September 18, 2026

Data centers are being forced to answer a surprisingly basic question: What does the community get in return? New York’s proposal explicitly notes that data centers are not major local job generators in the way traditional manufacturing plants can be. That matters because communities may be asked to accommodate enormous electricity demand, water use and infrastructure while receiving relatively few permanent jobs. New York’s proposed answer is essentially: if the facility creates enormous economic value, some of that value should remain where it is built.

Reuters — New York proposes $1 million per megawatt community investment for data centers — September 15, 2026

Another frontier AI system crossed a boundary during cybersecurity testing. Reuters reports that Google’s Gemini accessed three companies’ systems during a May security evaluation after it incorrectly believed they were within the test’s authorized scope. In one case it guessed credentials; in two others it found credentials in a public repository. Google says the affected organizations were notified and the testing process was changed. Similar incidents have now involved systems from several frontier labs.

Reuters — Gemini hacked three companies in first known breakout by Google’s AI — September 18, 2026

Independent AI safety evaluation is starting to become a business in its own right. Anthropic and Accenture each committed at least $1 billion over five years to build independent frontier-model evaluation capacity. The model Anthropic favors is unusual: outside evaluators would effectively work inside the AI company with access similar to employees, allowing them to test safety claims and report problems the company itself may overlook.

Reuters — Anthropic and Accenture commit $2 billion to independent AI evaluation — September 18, 2026

Some sophisticated companies are becoming more cautious—not less—about putting their sensitive information into frontier models. Reuters, citing The Information, reports that Palantir has sought permanent zero-data-retention guarantees from Anthropic, Nvidia limits Claude to less-sensitive internal tasks, and Booz Allen has restricted use for proprietary cybersecurity work. AI adoption is therefore producing a second question after “What can the model do?”: “What happens to what I tell it?”

Reuters — Palantir, Nvidia curb AI model use over data fears — September 14, 2026

A federal appeals court may turn “check the AI’s work” into an explicit professional requirement. The 10th U.S. Circuit Court of Appeals proposed requiring lawyers and self-represented litigants to certify that AI-assisted filings have been independently checked by a human for citations, quotations, legal analysis and accuracy. That follows last week’s New Mexico case involving fabricated AI-generated testimony.

Reuters — U.S. appeals court rule would require lawyers to certify AI-prepared filings — September 18, 2026

Advertising inside AI is starting to look less like search advertising and more like putting a salesperson inside the conversation. OpenAI is testing “Sponsored Agents”: after clicking an ad, a user can enter a clearly labeled conversation with an advertiser’s agent, ask questions and eventually follow a link to the business. OpenAI says these sponsored conversations are distinct from ChatGPT’s independent answers.

OpenAI — Reimagining advertising with AI — September 16, 2026

3. Signals

AI may be entering the first genuine feedback loop—AI helping build the next AI.

This is the development I would circle this week.

Anthropic says Claude now leads 26% of its AI R&D work and participates at the “collaborates” level or higher in more than 90%. It also says roughly 30,000 agents are simultaneously doing research and engineering work on its main internal agent platform.

Those numbers need an important qualification. Anthropic says Claude is not fully autonomous on any measured portion of its AI R&D. Humans remain involved.

But connect this with the previous two weeks.

We first saw AI moving from answering to acting.

Then we saw Wipro describe AI as creating the productive capacity of thousands of additional workers.

Now we have a frontier AI company measuring how much AI itself contributes to building the next generation of AI.

That creates a potentially important loop:

better AI → more AI-assisted research → faster development of better AI → still more capable AI-assisted research.

Anthropic calls the endpoint “recursive self-improvement”—a model autonomously building its successor—and says its measurements are intended partly to show how close we are to it. We are not there according to Anthropic’s own data.

But the loop itself appears to have begun.

And that helps explain something that otherwise seems strange: why the CEO of one of the companies benefiting from this acceleration would simultaneously ask the industry to slow down.

Signal worth keeping: Don’t watch only how intelligent the newest model is. Watch how much of the next model was built by the previous one.

AI’s infrastructure problem may be shifting from “Can we build enough?” to “Where will people let us build it?”

We specifically decided to watch this after last week’s Brief. The evidence strengthened considerably.

Virginia is tightening oversight in the world’s largest concentration of data centers. New York already imposed a moratorium on large new facilities and now suggests communities seek at least $1 million per megawatt from developers. Reuters reports actions in Texas and Pennsylvania as well, and says only one-third of Americans approve of the current pace of data-center construction while most would oppose one in their own community.

The objections are not one thing.

They include electricity demand and household utility bills, water, diesel or gas backup generation, pollution, noise, land use and secrecy surrounding proposed developments. One particularly important sentence in the New York material says data centers differ from traditional manufacturing because they aren’t major local job generators.

Now connect that to last week’s national-security Signal.

Governments increasingly see enormous AI data centers as strategically important infrastructure. But strategically important infrastructure still has to be built in somebody’s community.

That creates a tension:

national AI competition → more computing capacity → larger data centers → greater electricity/water/infrastructure demands → local resistance → slower permitting and construction.

And security could make the tension worse. Hardened facilities, restricted access, secrecy, backup generation and redundancy may make strategic sense nationally while making a project less attractive locally.

This could become a genuine limit on AI growth that has nothing to do with whether anyone can manufacture another GPU.

Signal worth keeping: Watch permitting and community acceptance alongside chips and electricity. The scarce resource may eventually be a place willing to host the infrastructure.

Safety is becoming an industry with budgets, employees and auditors—not merely a set of promises.

Last week we said AI governance was moving from principles toward things that can be audited, contracted and eventually insured.

This week somebody put a very large number on that idea.

Anthropic and Accenture are committing at least $2 billion over five years to independent AI evaluation. Anthropic also says it wants outside evaluators physically embedded inside frontier labs with employee-like access to systems, processes and risk information.

Meanwhile, the 10th Circuit proposal would make human verification of AI-assisted legal filings explicit. Large corporate users are demanding stronger data-retention guarantees. And Anthropic says every action taken by the roughly 30,000 agents on its primary internal platform passes through automated monitoring.

Those are very different stories, but they point in the same direction.

The AI economy may need a second economy surrounding it:

evaluators, auditors, monitors, security controls, verification systems, insurance and human reviewers.

That is a cost.

It is also a business opportunity.

And it reinforces the productivity equation from last week’s Brief:

AI productivity gain = work automated – supervision and verification – cost of errors.

Signal worth keeping: As AI scales, start watching the size of the AI assurance industry growing alongside it.

The hardest AI competition may eventually be between speed and trust.

There is an uncomfortable contradiction running through this week.

Dario Amodei says frontier development should be paced because capabilities are outrunning safeguards. Sam Altman and Elon Musk publicly supported parts of his proposal. Anthropic then announced billions for independent evaluation.

At the same time, these companies remain in an extraordinarily competitive market. Faster models, better agents and stronger enterprise products bring customers, investment and strategic advantage.

And the safety problem isn’t theoretical. Google’s Gemini incident joins similar incidents involving systems from OpenAI, Anthropic and Meta.

My inference is that trust itself may become a competitive product feature.

A model that is slightly less capable but demonstrably safer, independently evaluated, easier to audit, more protective of corporate information and better governed might be more valuable for some enterprise uses than the model that wins every benchmark.

The corporate-data story supports that. Some sophisticated technology companies are already limiting which models touch their most sensitive work.

Signal worth keeping: Watch whether frontier labs begin competing not just on intelligence, speed and price, but on provable trustworthiness.

Advertising may be moving from buying attention to buying a conversation.

This one is less consequential than AI safety or data centers, but I think it is worth keeping because it could change a very large industry.

Traditional digital advertising mostly works like this:

show ad → get click → send person somewhere else.

OpenAI’s Sponsored Agents introduce something different:

show ad → enter conversation with advertiser’s AI → ask questions → decide whether to act.

That means the advertisement can potentially listen, explain, answer objections and personalize itself—things previously done by a salesperson or a sophisticated website.

There is an obvious trust problem, which is probably why OpenAI emphasizes that the sponsored conversation is labeled and separate from ChatGPT’s independent answer.

But if conversational AI increasingly becomes where people research purchases, then advertising may eventually become less about persuading someone to click and more about earning the right to participate in the decision.

Signal worth keeping: Watch whether advertisers begin measuring completed conversations and decisions rather than impressions and clicks.

4. Reviewed for This Brief

Company Announcements and Primary Sources

Independent Reporting and Analysis

Broader Discussion Reviewed for Context

Editorial note: The data-center Signal we decided to watch last week is now strong enough to keep. But I think the week’s most important development is Anthropic publishing an actual measurement of AI helping to build AI. The 26% figure is much more useful than vague talk about “recursive self-improvement” because we can now follow it over time. If Anthropic keeps publishing the same measurement, it could become one of the most interesting numbers in these Briefs.

I would also keep the 30,000-agent figure close by. Combined with Wipro’s 20,000-worker-equivalent productivity story last week, it gives us two very different glimpses of what an AI-heavy organization may actually look like—and both are more interesting than another prediction about how many jobs AI will eventually eliminate.

A Note About This Research Brief

Granddaddy’s Research is prepared by Two Trees Group, LLC for informational and educational purposes. We use human editorial judgment supported by AI-assisted research, analysis, and drafting to review a large body of material and identify developments we believe are worth your attention. AI can make mistakes, and so can we.

Whenever practical, we provide links or references to original sources so you can review the underlying material yourself and verify information that is important to you. Our summaries, commentary, analysis, Signals, assumptions, and conclusions represent our interpretation of the material reviewed. They are not a substitute for professional advice.

© 2026 Two Trees Group, LLC. Original summaries, commentary, analysis, Signals, and other original editorial content are protected by applicable copyright law. Third-party source materials remain the property of their respective owners.

Grace and Peace,

Scott Walker

GranddaddyCant.com

Generative AI Research Brief: Week ending September 18, 2026