Back to news
Sunday report

Sunday Signal Report: September 13, 2026

Vibe editing is a step function. A union and a vendor made AI protections contractual. A year of phone-ban data shows enforcement decaying. Dario asked to pace the frontier. Sam matched the evaluator offer. Sacks told them to slow down without writing the law. The rule was never the outcome.

Rob here

A rule is not an outcome.

Video used to mean hours on a timeline after the idea was already finished. This week I watched someone describe the cut and let the machine execute.

That is a step function.

Alex posted a walkthrough of editing with GPT-6 Astra and a DaVinci Resolve MCP. He says the model cut about 95 percent of the video. He kept the last slice for whether the cut sounded right. That is vibe editing. You describe the cut. The machine executes. Taste stays with you.

I used to treat video as specialist craft. Timeline, audio, pacing, color. Hours after the idea was already clear. If a model can take most of that work, anyone who can record can ship a video they would have skipped. The scarce part is knowing what the piece is for.

Same shape as the rest of this week. Capability jumped. The interesting part is still what happens after the announcement.

On Sept. 9 the AFT, the UFT, and Microsoft announced a National AI Safety and Privacy Standard, ten principles a district can fold into its own Microsoft agreement and then enforce as contract language. That is a real lever. Read the definitions and you find the lever has a handle only on products built and marketed for education. General-purpose productivity, collaboration, search, and workplace assistance sit outside the fence unless a vendor agrees otherwise.

On Sept. 10 Silver, Rapaport, and Polikoff published a year of teen survey data on school cellphone policies. Nearly every school now has one. Reported phone use in core classes went from 28 percent in October to 42 percent in May. Forty-four percent of teens said their teachers got less strict as the year went on. Thirteen percent of adults noticed. That gap between the policy on paper and the policy in the room is the whole story.

On Sept. 12 Dario Amodei published We Must Pace the Frontier. Anthropic is giving third-party evaluators employee-level access. Sam Altman said OpenAI will match that. Elon Musk said Dario is right. David Sacks told them to slow down without asking Congress to bless it. Schools will still get the models. The scarce thing is still an adult who can tell a signed commitment from a verified one.

The primary lens is Implementation Reality. The secondary lens is Governance and Trust. Capability is cheap, and this week rules were cheap too. The scarce thing is the adult who owns enforcement in March, and the editor who still knows why the cut is there.

Two more in the same shape. MIT's faculty committee report on AI in teaching is a set of recommendations about assessment design, not an adopted institutional policy, and the district-transferable object is the test it implies: if a chatbot can finish the assignment tonight, the assignment is no longer measuring learning. And PaperCut finally replaced three emergency patches with maintenance releases that actually completed QA. Emergency builds stopped the bleeding. The QA'd release is the fix.

Strong signal

AI protections became contract language. Read the definitions before you celebrate.

What changed: on Sept. 9 AFT President Randi Weingarten, UFT President Michael Mulgrew, and Microsoft Vice Chair Brad Smith announced a National AI Safety and Privacy Standard, a memorandum of agreement with ten mandatory principles: no training on student data, no student tracking, no consequential decisions without human oversight, plain-language transparency for educators and parents, a two-year term, and revocation with public notice. Microsoft is the first signing provider and the remaining provider slots are blank. The clause that matters most to a district is the one that says any U.S. district can request the same substantive protections inside its own agreement, and that signers must stand up a findable, usable process for that within ninety days. Why it matters now: this is the first time the AI-in-schools argument has been moved from guidance documents into enforceable purchase terms, and it arrives while state and federal rules are still absent. The catch is the definition of an educational product: generative services primarily designed and marketed for students, educators, and administrators under an authenticated education agreement. General-purpose productivity, collaboration, communication, search, cloud, and workplace assistance are explicitly carved out, which puts a very large share of the AI surface actually running in schools outside the fence. Rob's take: this is the Governance Gap narrowing in exactly one lane and staying wide open in the others. It is a real procurement lever and a bad talking point. The sentence to never say out loud is that a vendor signed, therefore we are covered. The transferable skill here is not admiration, it is reading the definitions section first. Meanwhile Kansas is running the state version of the same question: KSDE made a draft instructional-technology policy public on Sept. 2, the state board received it Sept. 8 and spent most of Sept. 9 reworking the policy, the plan template, and the guidance, with action pushed to October or later. The draft names AI, privacy and security, grade-band device access, take-home rules, and parent acknowledgement before individual device access starting in 2027-28. Concrete implication for a district leader: pull your three largest AI-touching agreements this month and sort each one into educational product or general-purpose. For the general-purpose side, write down what protection you actually have today in plain language, because that is the column where a contractual standard will not help you.

K-12 leadership

Phone bans are nearly universal. Enforcement is not.

What changed: on Sept. 10 Dan Silver, Amie Rapaport, and Morgan Polikoff published year-over-year results from a nationally representative teen panel surveyed in October 2025 and again in May 2026. Virtually every student now attends a school with some cellphone restriction. Reported phone use during core classes rose from 28 percent to 42 percent across the year. Forty-four percent of teens said teachers became less strict about enforcement, while only 13 percent of adults in the same households noticed any loosening and 60 percent perceived no change. Bell-to-bell policies held up better but still slipped: a 10-point increase in core-class phone use under bell-to-bell rules versus an 18-point increase under partial restrictions. Why it matters now: the adoption fight is over and the enforcement fight is the one nobody staffed. The adult blind spot is the finding with teeth. If leaders are surveying parents about how the policy is going, they are sampling the group that cannot see the decay. Rob's take: this is Attention Is a Leadership Resource meeting The Implementation Layer, and it is the cleanest illustration I have seen this year of why an announcement is not a system. A policy that depends on a few hundred teachers making an unsupported judgment call forty times a day will erode by March, not because anyone rebelled, but because early-year routines gave way to covering content. Also worth holding: teens got less negative, not more positive. The share reporting a negative effect on their happiness fell from 48 percent to 27 percent, and most of that movement landed on no impact. Kids adapting to a rule is not the same as kids learning something from it. Restriction buys attention. It does not spend it. Concrete implication for a district leader: build a mid-year enforcement check that asks students, not parents, and do it in January rather than August. Then decide in advance what the answer changes, because a survey with no attached decision is another artifact.

Frontier governance

Dario asked to pace the frontier. Sam matched the evaluator offer. Sacks told them to slow down without writing the law.

What changed: on Sept. 12 Anthropic CEO Dario Amodei published We Must Pace the Frontier. The argument is not a pause. It is a three-part plan for slowing how fast frontier companies improve model capabilities so safety work can keep up. Step one is Embedded Evaluators: ongoing, employee-like access for a third-party team such as METR, so they can verify safety practices, report incidents, and assess alignment in the training pipeline, not only in a finished model. Anthropic committed to that step unilaterally. Sam Altman replied the same day: he agrees the frontier needs pacing, OpenAI has been discussing it for weeks, independent evaluators with employee-like access is a great idea, and OpenAI will do the same, with more to share soon. Elon Musk replied in three words: Dario is right. On Sept. 13 David Sacks told them to go ahead, and to stop asking for permission. His post says Anthropic and OpenAI already have a duopoly on frontier intelligence by market share, revenue growth, and model capability; that they should slow down if the unreleased models are scary; and that they should stop pretending they need an antitrust waiver, a regulatory process that supersedes product liability, or METR to police companies that are not at the frontier. He calls METR intertwined with Anthropic investors and staff, which is his claim, not a finding I independently verified. He also says the Hugging Face episode already makes reliability good business, that China is unlikely to join a global agreement, and that demanding their preferred regulatory framework as the price of slowing down will look like blackmail. If they slow down, they buy goodwill. If they do not, he says we will know it was regulatory capture or an election-season psyop. Why it matters now: the first step now has two labs on the record, not one. The argument moved in 24 hours from whether to pace to who writes the rules. Rob's take: this is Governance and Trust at the lab layer, and it is not a district planning document. Schools do not pace the frontier. They inherit whatever ships. Sam matching employee-like evaluator access is the first industry match. Elon's three words are an endorsement, not a program. Sacks is the implementation read you should keep in the room: if a lab needs a new law to slow down, it was never only a safety decision. The transferable skill is the same one in the AFT-Microsoft standard: read the verification clause before you celebrate the headline. Concrete implication for a district leader: do not wait for a paced frontier. Assume the models keep arriving, keep the AFT-style contract language on the tools you actually buy, and treat any vendor's safety commitment the way you would treat a policy with no named owner in month seven.

Agency and assessment

If a chatbot can finish it tonight, it was never measuring learning

What changed: two assessment documents moved into the district conversation this week, though both were written earlier and should be read as recent rather than breaking. MIT's Ad Hoc Committee on AI Use in Teaching, Learning, and Research Training, co-chaired by Eric Klopfer and Sam Madden, dated its final report Aug. 13 and released it publicly in late August. It says generative AI can already produce credible answers to nearly any undergraduate written assignment, that detection tools are unreliable, and that policing student use damages the teacher-student relationship. Its recommendations are oral exams, semester portfolios, out-of-class work paired with in-class conversation, a regular in-person social component in every subject, more teaching assistants rather than fewer, and an explicit warning against replacing undergraduate research apprenticeships with cheaper agents. Separately, a preregistered randomized trial from OpenAI with Bocconi, Duke, and Berkeley researchers, run with 1,053 first-year undergraduates, tested two treatments at once: teaching causal reasoning, and providing GPT-4o. Causal training moved how students thought, lifting mechanism identification by roughly 0.55 standard deviations and falsifiability by roughly 0.85. Model access moved the score, adding 0.86 points on a five-point rubric from a 2.09 baseline. Combined, the model still won the scoreboard. Why it matters now: put those two side by side and you get an uncomfortable finding. The thinking treatment worked and the rubric did not pay for it. Rob's take: this is What Should Stay Stubbornly Human, stated as a measurement problem rather than a values problem. We keep asking whether students should have the tool. The sharper question is whether our assessments can still tell the difference between a well-formed answer and a formed mind. They increasingly cannot, and that is our design failure, not the model's. MIT is a residential research university and its brand does not transfer, but its test does. Concrete implication for a district leader: pick one graded assignment per department this semester and run it through a frontier model. If the output would earn full credit, that assignment is now a writing sample, not an assessment, and the department owns the redesign. Do not start with a detection tool.

Cybersecurity and privacy

The emergency patch stopped the bleeding. The maintenance release is the fix.

What changed: PaperCut updated its security bulletin on Sept. 10 to publish official NG/MF security maintenance releases, builds 26.0.5, 25.0.13, and 24.1.10, which replace Emergency Patch Releases 1 through 3 and are now the recommended install for every customer whether or not the server faces the internet. Unlike the emergency builds, these completed full QA and carry additional hardening. CISA's known-exploited catalog still lists the two PaperCut CVEs with a Sept. 14 federal remediation date. In the same window ConnectWise updated its Trust Center advisory for a ScreenConnect vulnerability fixed in 26.6.5, with cloud instances already patched and on-premises deployments needing the upgrade plus a host-client reinstall and access-agent update. CISA added four more known-exploited vulnerabilities on Sept. 8, including a Citrix NetScaler authentication bypass and a Fortinet heap overflow, both carrying Sept. 12 federal due dates and forensic triage requirements. Why it matters now: three of these four live in the boring middle of a district stack, print management, remote support tooling, and edge appliances, which is exactly where a school network gets pivoted through. Rob's take: the interesting item is not the CVE, it is the sequence. A vendor shipped three emergency builds in two weeks, each one a partial answer, and only now has a release that went through real QA. If your team applied Release 3 and marked the ticket closed, the work is not done, it is deferred. Federal KEV due dates do not bind a school board, and treating them as a private deadline anyway is the cheapest maturity upgrade available. Concrete implication for a district leader: ask for a one-page confirmation this week that print servers, site servers, and secondary print servers are on a maintenance build rather than an emergency patch, and that remote-support tooling is at or above the fixed version with agents actually reinstalled. Ask for build numbers, not a yes.

Agentic AI checkpoint

Agentic AI checkpoint: two patch releases and a fast stable cadence, no new authority

Hermes Agent published two stable tags inside the window: v0.21.1 on Sept. 7, described in its canonical notes as a rollup of main since v0.21.0 with full curated notes deferred to v0.22.0, and v0.21.2 on Sept. 11, a targeted reliability patch. The v0.21.2 notes are the honest part and the useful part: the large session-store rewrite that shipped in v0.21.0 made the local state database fragile on some installs, with second writers cancelling each other's locks, healthy databases reported as corrupt, and a single bad row breaking a session listing. Six pull requests closed that class of failure by removing the second writers rather than retrying harder. That is a maturity signal about recovery and data integrity, not a capability announcement. OpenClaw shipped stable 2026.9.2 on Sept. 5, 2026.9.3 on Sept. 8, and 2026.9.4 on Sept. 11, plus 2026.6.35 on Sept. 10 as the final June extended-stable release. The 2026.9.3 notes emphasize rehearsing core and plugin changes in isolated candidate state before activation and recovering abandoned update records without stopping a healthy gateway. All of these are stable tags with prerelease false, not betas, and the June LTS line closing is the item to note if anything in a building is pinned to it. Nothing in this window grants an agent new operational authority. A weekly stable cadence and a state-integrity patch are evidence that these projects are being maintained, not evidence that an agent should hold credentials to a student information system. Concrete implication for a district leader: if an agent platform is under evaluation, ask which tag is stable today, what the last regression was and how it was found, whether the extended-stable line you would pin to is still receiving fixes, and whether a restore from a corrupted local state store has actually been tested rather than assumed.