Unfixed Newsletter — September 3–16, 2026
Editors’ Note: We are returning to our once-every-two-week cadence for the newsletter and just in time. There were three gigantic stories this week adjacent to higher ed and one interesting campus based approach. We want to flag the story about the labs collectively discussing slowing down because it might be the first time we can catch our breath since ChatGPT was released in Fall 2022. All the stories this week matter, but this one is the sort of sea change many of us have been hoping for.
The stories
1) AI rivals are starting to agree that the race may need to slow down
Anthropic CEO Dario Amodei’s call to slow frontier AI development is drawing support from some of his biggest competitors. OpenAI CEO Sam Altman has signaled support for coordinated action and suggested private discussions among major labs are already happening. Elon Musk has also backed Amodei’s position. OpenAI had separately argued that future development may need to be deliberately paced so safety work can keep up. The unusual part here is the emerging consensus: companies competing aggressively to build more capable systems are increasingly acknowledging that competition itself may be creating unacceptable risks.
Why this matters: I (Zach) am going to take a little moment of editorial discretion here. For years Nik and I have worked with faculty who are deeply worried about AI and X-risks, or environmental impacts, or the existence of the technology at all. We have repeatedly told folks we don’t have control over any of those things, let's stay focused on our classrooms. Now, for one of the first times, we see rival labs recognizing risks and there is hope they might collectively slow down. This might go nowhere, might be a short reprieve, or it might end up being a time for us to reflect on the moment we are in. We are thankful the conversation has at least started.
Links:
https://openai.com/index/research-acceleration-view-inside-openai/
https://www.wired.com/story/openai-wants-to-know-if-an-ai-industry-slowdown-would-even-be-legal/
https://www.theguardian.com/technology/2026/sep/13/openai-sam-altman-elon-musk-back-anthropic-calls-brakes-ai-development
2) New York and Los Angeles put the brakes on student AI use
Two of the country’s largest school systems opened the year with new restrictions on generative AI. New York City announced a one-year moratorium on student-facing generative AI for students through eighth grade, while allowing teacher use and limited high-school pilots. Los Angeles Unified went further, blocking generative AI for all students on district-issued devices while it reviews instructional uses and safeguards. Both moves cut against the assumption that schools would steadily expand student AI access.
Why this matters: There is always cross-over between K-12 and higher ed so we should expect the “ban it” group to catch some momentum on all fronts. We covered this in a blog post, but even in K-12 environments where we have more control “ban it” does not itself constitute a policy. Students have access to the internet outside the classroom and on their own devices often at school. It is worth paying attention to the implementation details of this story to see if districts try to implement parental education about AI and/or surveillance technology on student work.
Links:
https://www.nyc.gov/mayors-office/news/2026/09/mayor-mamdani-and-chancellor-samuels-put-students-first-with-nat
https://laist.com/news/education/lausd-students-barred-artificial-intelligence-tools
3) The Millennium Problem breakthrough became a fight about AI research norms
OpenAI announced September 8 that an internal AI system had produced a formally verified solution related to the Navier-Stokes Millennium Prize Problem. The mathematical claim quickly became tangled in a conflict over who actually solved the problem. NYU mathematician Tristan Buckmaster and Anthropic researcher Levent Alpöge had been pursuing related work, and Buckmaster questioned whether prior interactions with OpenAI’s systems could have influenced the company’s effort. OpenAI says an investigation found no such influence and argues the groups solved materially different problems. The dust-up puts familiar questions about, attribution, data provenance, and research competition into a much stranger setting.
Why this matters: There is a lot here. This is a story about data security with the question of whether OpenAI mined usage for hints to solutions. This is a story about the place of elite researchers. The solution to this problem was the combination of research and LLM or mainly just LLM either way elite math is now being engaged differently. Somehow lost in the drama of this story is the fact that this was a long standing problem people have been working on for decades and now we have a solution. We have been waiting for the investments in AI to payoff with scientific discovery. Solving Navier-Stokes does not immediately impact my life, but it is tough to look at this and not see we have arrived at a moment of scientific breakthrough.
Links:
https://openai.com/index/navier-stokes-solution/
https://www.theverge.com/ai-artificial-intelligence/994255/openai-millennium-prize-problem-tristan-buckmaster-competition
https://www.nature.com/articles/d41586-026-02842-5
4) Penn State puts AI-use expectations directly into Canvas
Penn State has launched an optional Canvas tool that lets faculty mark courses and assignments as AI allowed, limited, or prohibited using a common set of icons and editable language. The idea is straightforward: move AI expectations out of lengthy syllabus policies and put them where students encounter individual assignments. Faculty retain control over the rules, but students get a consistent visual vocabulary across courses. It is a modest intervention compared with campuswide AI mandates, but it addresses one of the most persistent practical problems of the last few years: students encountering different AI rules in every class and sometimes every assignment.
Why this matters: This story is a bit more direct than the others. Penn State has put forward a systematic approach that increases transparency for students. The part we like the most is what this communicates to faculty–we need to tell students what our expectations are for each task. Maybe we are being naive, but we don’t really see a downside to this as an approach and we would like to see other institutions experiment with it.
From Our Work
Ep. 39: Beyond Pro-AI and Anti-AI with Andy Masley
September 7, 2026
Andy Masley joins Zach and Nik to discuss why AI debates collapse so quickly into pro- and anti-AI camps, along with environmental claims, technological tradeoffs, and the growing data-center backlash.
https://www.meltsintoair.org/unfixedpodcast
Ban AI in NYC and LA Schools
September 14, 2026
This post is the full featured analysis of the AI bans in NYC and LA. There are real challenges here and the overly simplistic “Ban AI” approach papers over the reality of what it means to actually get AI out of education.
https://www.meltsintoair.org/chatgpt/8a8n3ns5w2hk3xese4032o3l4ntutn
The AI Retrofit™ was featured in the most recent ChatGPT for education subtack
September 15, 2026
Zach was interviewed for this post which also featured real examples of adaptation from faculty at Chico State.
https://edunewsletter.openai.com/p/the-ai-retrofit-chico-state