2026-09-22: AI Agents Meet Their Real-World Limits

Welcome

Welcome to the NextWith.ai Daily podcast, an AI-narrated retrospective look at the reporting from September 22. Today’s theme is simple: AI products may be moving fast, but the real-world tests are getting more specific. We start with consumer adoption, then move through merchant permission, data-center infrastructure, state regulation, and finally a new governance step for mathematical claims.

Meta’s Muse and the retention test

Now, to Meta’s Muse. The new personal AI agent got a strong early signal on acquisition, not yet on staying power. According to the reporting here, Apptopia estimated 2.8 million mobile installs in Muse’s first 12 days. In a narrower U.S. and Canada iOS comparison, Muse was estimated at 1.8 million downloads versus 1.3 million for ChatGPT in its first 12 mobile days. Apptopia also estimated 642,000 daily active Muse users on U.S. mobile, compared with 231,000 for ChatGPT at the same point in its mobile launch. Those are third-party estimates, and the article is careful not to treat them as a verdict on durable demand.

The limitation is important. Muse launched across iOS and Android in a narrower region, while ChatGPT launched globally on iOS, so the products did not arrive under the same conditions. Appfigures added a different data point, estimating 1.1 million devices by September 18 and a climb to the top of the U.S. free-app charts. The direction is clear: people were willing to try Muse quickly. But a download is not the same as a repeated task, a weekly habit, or a user granting deeper permissions and keeping them in place.

The takeaway is straightforward: launch rank measures attention, not retention. For an agent, the real question begins after installation, when users decide whether the result is useful enough to repeat.

Amazon blocks Muse from shopping

Now, to Amazon’s block on Muse. This story moves from adoption to permission. Amazon said it blocked Meta’s Muse personal AI agent from shopping on Amazon.com after trying, and failing, to get Meta to leave Amazon out of the experience. In the reporting here, Amazon called Muse an unauthorized AI agent and said it had not been told Muse would access Amazon. Amazon also said the agent does not identify itself while browsing and could raise privacy and security risks if it handles customer credentials. Meta did not immediately respond in the source material, so those objections remain Amazon’s account.

What matters operationally is the boundary. A shopping agent can search, compare and move toward checkout, but if a merchant refuses the arrangement, the transaction may stop there. The article’s recommendation is to treat that as a visible outcome, not a quiet success. In other words, “found an item” is not the same thing as “placed an order.”

That distinction is useful for anyone building or buying an AI shopping assistant. A demo can look polished and still fail at the point where the merchant decides whether the agent is welcome. So the takeaway here is that merchant permission is part of the product contract. If the store says no, the agent has to stop cleanly and make the failed handoff clear to the user.

NVIDIA’s DSX Ready qualification

Now, to NVIDIA’s DSX Ready program. This is not about chat or shopping at all; it is about the infrastructure that makes AI factories possible. NVIDIA introduced DSX Ready as a qualification program for partner power and cooling products that meet selected requirements in its DSX reference designs. The first categories are battery energy storage systems and cooling distribution units. The article lists initial qualified partners including Hitachi Energy, LG Energy Solution and Tesla for BESS, and LG Electronics, LiquidStack and Vertiv for CDUs.

The practical point is that power and cooling are no longer background facilities work. A fast accelerator does not help if a site cannot absorb load changes, reject heat, or integrate equipment safely. But the reporting also stresses the limit: passing DSX Ready does not replace site-level engineering, and it does not imply site-level stability. A qualified component can still be wrong for a particular facility if water temperatures, redundancy plans, controls, or maintenance models do not match.

The takeaway is that NVIDIA is offering a narrower buying signal, not a full answer. For infrastructure buyers, the badge may shorten the first pass at vendor selection, but the real work still happens in commissioning, integration and acceptance testing.

California’s data-center rules

Now, to California’s new data-center laws. Governor Gavin Newsom signed seven bills on September 21 that tighten how data centers are disclosed, reviewed and charged for water- and power-related impacts. The state’s stated goal is to give communities more information about energy, water, workforce and land use before a project is built, not after. The package adds reporting and disclosure requirements for proposed data centers, including information on water use, supply, efficiency and drought planning.

The reporting also says the laws are meant to stop costs from being shifted onto ordinary customers. According to the governor’s office and The Verge’s coverage, the package is intended to require data-center operators to help fund grid upgrades and new clean-energy supply where needed, and the California Public Utilities Commission may create a new rate class for data centers. The article is careful not to overclaim: the evidence shows expanded disclosure and cost accounting, but not a blanket local veto over projects.

The takeaway is that California is treating AI infrastructure as a regulated utility issue as much as an economic-development issue. For builders and operators, that means water, power and rate evidence will matter more in permitting conversations.

OpenAI’s math advisory group

Now, to OpenAI’s new math advisory group. OpenAI said it is creating an independent advisory group on mathematics and artificial intelligence after claiming that a new internal model, trained starting August 28, solved the Navier–Stokes Millennium Prize problem and more than 100 other open problems. Those are OpenAI’s claims, not independently verified in the material here. The news value is the governance response: OpenAI is adding outside mathematical judgment after a public controversy over how to evaluate frontier results.

The group is described as independent, with members not paid by the company and able to make their advice public. TechCrunch reported that it is hosted at the Institute for Advanced Study and includes nine prominent mathematicians. The group’s role is advisory, though, not a veto. OpenAI says it will help review results, judge significance, coordinate release, and align the work with academic and professional standards, but it will not control how quickly OpenAI continues its research.

The takeaway is that the proof burden is expanding. For math-capable AI, the question is no longer just what the model can say, but how those claims get reviewed and communicated before anyone treats them as results.

Closing synthesis

Taken together, today’s reporting points in one direction: AI is running into the real world. Muse shows that early downloads are easy to count, but habits are harder. Amazon’s block shows that merchant permission can still stop an agent at checkout. NVIDIA’s qualification program and California’s new laws show that power, cooling, water and cost allocation are becoming central to AI deployment. And OpenAI’s advisory group shows that the more ambitious the claim, the more the review process matters.

The common thread is not that AI is stalling. It is that success now depends on evidence at each step: user return, permission boundaries, site-level engineering, regulatory disclosure and outside scrutiny. That is a better way to read the field than treating any single headline as the whole story.

That’s our retrospective for today. For the transcript and sources, visit NextWith.ai.