Daily Digest — 2026-09-12

The core of the video revolves around the Navier-Stokes Millennium Prize Problem, a long-standing mathematical challenge focused on describing the motion of fluids and gases. The problem, originating in the 1800s with initial contributions from Claude-Louis Navier and refined by George Stokes, asks whether solutions to the Navier-Stokes equations always exist and remain smooth (i.e., don't 'break down' or become infinite) in three dimensions. The Clay Mathematics Institute offered a $1 million prize in 2000 for a solution (either a proof of smoothness or a counterexample proving breakdown).

Tristan Buckmaster, a math professor at NYU and a recipient of the 2019 Clay Research Award for his work on Navier-Stokes and Euler equations, had been collaborating with Levent Alpöge, a mathematician at Anthropic, on this problem for the past year. Their research, which involved using a 'smooth external force to construct a finite-time blowup' for the Euler equations (a simpler version of Navier-Stokes without the viscosity term), was making slow but steady progress. Critically, their work was being documented and shared in OpenAI's Codex sessions.

In mid-August 2026, Alpöge realized the potential of LLMs for their research and they began leaning more heavily on Anthropic's Claude, OpenAI's Codex, and GPT-5.6 Sol/Astra for assistance, leading to a significant acceleration in their progress. By August 15, 2026, they had successfully demonstrated a blowup for the Euler equations with smooth forcing, marking a significant step, though not the full Navier-Stokes solution. This success was shared within their private Codex sessions.

Shortly after, OpenAI announced that they had made progress on the Navier-Stokes problem. Their internal models, specifically GPT-6 Astra and an internal model, were claimed to have established 'statement C' and 'statement D' of the Millennium Prize Problem formulation, essentially showing a breakdown of Navier-Stokes solutions. This was achieved using 10,000 concurrent agents and reportedly cost $20 million in compute. Notably, OpenAI's solution utilized an approach (the 'smooth external force' method) that was identical to the novel approach developed and shared by Buckmaster and Alpöge in their Codex sessions.

On September 3, 2026, Buckmaster emailed OpenAI for clarification and had a phone call with Sebastien Bubeck, an OpenAI researcher leading the project. During this call, Buckmaster was initially told that OpenAI's model was simply given the problem statement with 'very little human input.' However, as the conversation progressed, it became clear to Buckmaster that OpenAI's team had actively worked on the problem, including using their own research on Euler equations, and that the prompts for the model were written by Codex. When Buckmaster inquired if their model had been trained on or had access to his Codex sessions, he was initially told it 'did not look up user data' and was given no direct answer when he pressed on the training aspect.

Buckmaster claims OpenAI then offered two proposals: 1) Buckmaster and Alpöge publish their Euler results first, and then OpenAI publishes their Navier-Stokes results the next day, crediting Buckmaster for getting 'closest' to the solution. 2) Buckmaster could write the Navier-Stokes paper himself, but without crediting Alpöge, due to Alpöge's affiliation with Anthropic (a competitor). Buckmaster refused both, threatening to go public. Bubeck allegedly responded, 'Why would you ruin your career?' and stated, 'If you don't want me to be nice, then I don't have to be nice.'

Sam Altman, OpenAI's CEO, later publicly stated that his team acted with 'integrity and generosity' and that they believed the other team had also solved the problem, wanting a 'joint release.' Bubeck also tweeted clarifications, denying he ever asked for Alpöge to be removed from authorship and claiming discussions about authorship were in the context of a rewrite of OpenAI's proof.

On Tuesday, September 8, 2026, Buckmaster and Alpöge posted their papers, along with a detailed four-page statement explaining their version of events. Later that day, OpenAI posted their solution, along with a blog post claiming that their researchers and agents 'did not see any of their work through any means until they released it publicly' and that 'no specific user data was accessed.' They also explicitly stated that an investigation confirmed that Buckmaster's Codex prompts over the preceding two months could not have influenced their system, including through training.

Terrence Tao, a renowned mathematician, offered a nuanced perspective, congratulating Alpöge and Buckmaster on their significant achievement in fluid dynamics. However, he warned that 'prematurely solving the problem by purely AI-powered methods — particularly without full transparency into the solution process — can contaminate this process to the point where it actually becomes a net negative for the progress of mathematics as a whole.' Tao emphasized the 'substantial opportunity cost' in converting historically productive mathematical problems into mere viral social media benchmarks, advocating for deeper advancement and fostering future research rather than a race to 'solve' problems. The video concludes by highlighting the privacy risks of using AI chatbots and implicitly urging caution in trusting AI companies with sensitive work, while promoting Railway as a trustworthy alternative for project hosting.

The core premise of the video is that every operating system is a collection of "frozen decisions" made by developers years ago. Omarchy, a Linux desktop by David Heinemeier Hansson (DHH), aims to "time travel" these decisions, allowing users to alter and customize their OS using AI agents. The speaker, Nate B Jones, emphasizes that this concept isn't limited to Linux and can be applied to existing Mac and Windows systems.

Jones illustrates this with a compelling example: a father named Jankees wanting different screen time rules for his two children. One child needs a system that locks and rewards more screen time for math, while the other needs a system based on agreements with no locks. Such diverse needs are hard to accommodate in a single, fixed OS design, highlighting the need for highly personalized computing environments. The shift is towards an "audience for a change can be just you," meaning individual customization is paramount.

Traditionally, modifying a computer required several steps: 1. Find The Setting: Locating the specific configuration or code. 2. Understand The Format: Deciphering how the setting is structured. 3. Apply The Change: Implementing the modification. 4. Know If It Worked: Testing the change and verifying it didn't break anything else.

This process is often arduous and error-prone. AI agents, however, can automate these steps. Omarchy is built with this in mind, offering a Quickshell Unified Desktop with plugins and a configurable agent (Omarchy 4 "Quattro," released August 14, 2026). This allows for a coherent starting point and mechanisms to change it. The Omarchy homepage offers Mac and Windows trials, enabling users to explore this agent-driven customization without altering their main setup.

Jones outlines a "Try One Window Rule" workflow for safe experimentation: 1. Keep A Copy of the original setting. 2. Ask For A Change from the agent, focusing on one specific rule. 3. Test The Result to see what changed. 4. Undo The Change to return to the original state.

This iterative process allows users to experiment and learn without fear of irreversible damage. The example of adjusting window tiling and application placement (e.g., browser on the right, notes on the left, a specific app on a second screen) demonstrates how agents can manage such tasks by finding settings, understanding formats, applying changes, and verifying success. Omarchy's AI Manual provides explicit guidance for the agent on config changes, top bar adjustments, diagnosis, and activity logging.

Critically, Jones stresses the importance of scoping requests. Vague commands like "redesign my desktop and make it magical" are ineffective. Instead, users should give specific jobs to the agent: "Name the setting, keep the original, make one change, check, and undo." This approach mirrors good engineering practices of small, verifiable changes.

For smaller applications, the "wrapper pattern" is introduced. Ludovic Frank, a web developer, built "Omaport" on Omarchy using `rclone` (an existing maintained tool) to achieve his desired file transfer experience. This demonstrates how agents can leverage existing tools and wrap a custom interface around them, rather than inventing solutions from scratch. This allows users to adapt an application's behavior to their preference without taking full responsibility for its underlying functionality.

Security and access control are also vital considerations. Different jobs require different levels of access: inspecting a config file might need "inspect only" access, editing might need "scoped change" access, installing a new package needs wider impact, and administrative tasks require "broad access." Users must be cautious if an agent requests more access than necessary. Omarchy addresses this by offering an opt-in passwordless sudo for 15 minutes, but Jones warns that this applies to every program running as the user, including the agent itself. He advises asking "why" if an agent asks for broad access, emphasizing that models can be over-optimistic.

Finally, Jones highlights the distinction between changing the application (control on your computer) and processing data (which may use the cloud). While agents can control local applications, they might still send data to hosted models for processing. Users should be aware of data privacy implications. He encourages users to explore the companion examples guide for Mac and Windows to discover what their existing OS can do differently, leveraging tools like AeroSpace for Mac window management or PowerToys Workspaces for Windows desktop arrangements. These tools offer "useful handles" for agents to interact with, enabling fine-grained control and automation. The ultimate goal is to bridge the "99% vs. 100% dependable" gap, allowing users to make their computing environments both reliable and perfectly tailored to their needs. There's "more room in your OS than you know about" for customization when agents are involved.

Thais Castello Branco, CEO of Taste Labs, kicked off her talk at the AI Engineer World's Fair by emphasizing Taste Labs' mission to "end AI slop," which she identifies as her personal enemy. She explains that achieving this requires decoding subjective domains. While significant effort has been invested in making AI models excellent at objective tasks like coding and math, it’s time to extend this rigor to subjective areas like design and writing.

Taste Labs operates in two primary ways: First, they collaborate with frontier labs to evaluate models, pinpoint their weaknesses, and then construct targeted post-training data or reinforcement learning (RL) environments to enhance their taste, vision, and design capabilities. Second, they develop application-layer tools to help agents and apps produce better outputs. The key challenge in subjective domains is defining "greatness." For math, correctness is objective; great is correct. But for writing or design, defining "great" is harder. It often involves uniqueness, care, attention to detail, craft, and authenticity—qualities currently lacking in much AI-generated content. Conversely, AI slop is easier to define: it’s characterized by repetition, lack of fit, and low intent.

Branco notes that "slop is not a new phenomenon" but has been accelerated by AI. The ease of generating content with AI means that individuals without honed taste or skills in specific domains (e.g., design) can instantly create things, leading to homogenized, low-quality output. To combat this, she believes in measuring, understanding, and then affecting the problem. Taste Labs analyzed over 2.4 million websites from the past 10 years and thousands of AI-generated sites to identify patterns and trends in design. They found that even before AI, the internet was becoming more homogeneous, with similar color palettes and layouts spreading rapidly. With AI, this repetition is amplified, often generating designs that are almost identical regardless of the context.

To quantitatively identify slop, Taste Labs developed "probes" (or "baby classifiers"). These are designed to detect specific features that indicate AI-generated slop, such as: weight-only type pairing, neutral/light palettes, Bootstrap/jQuery styling, fade-in entrances, single font families, card surfaces, rounded buttons, purple/violet accents, sticky navs, and neo-grotesque (inter-class) fonts. The probes effectively separate designer-approved sites from designer-rejected sites with an AUC of 0.74, a higher accuracy than most Large Language Models (LLMs) used as judge methodologies. This indicates that overused patterns equate to slop.

The critical difference, Branco states, is that "the cost of production is now going to zero, but not the cost of judgment." While improving models at the base layer is crucial, she argues that addressing the problem at inference time—when the end-user interacts with the AI—is equally, if not more, important. Taste Labs' solution involves building systems to fight slop by creating mechanisms for: 1. Repetition: Forcing intentional distribution of outcomes through a Creativity API that acts as an "inspiration machine" to guide agents toward diverse, out-of-distribution outputs. This isn't about randomness but strategic divergence from norms. 2. Fit: Ensuring brand adherence through a Brand API. This API takes an existing brand URL, extracts its specific design components (colors, typography, layout), feeds this structured context to an agent, and then scores/verifies the improved output against brand guidelines. This helps agents maintain brand consistency and avoid generic outputs. Moreover, for users without existing brands, Taste Labs is building a repository of pre-created, cohesive brand systems that users can retrieve and apply, thereby democratizing good design. 3. Intent: Improving intent mapping through prompt enhancement. 4. Judgment: Implementing a "slop gate" via a baby-classifier that filters out low-quality AI outputs in production.

Branco concludes by reiterating that to achieve "great" AI, we must first "raise the floor" by learning how to measure, decompose, and affect the quality of AI-generated content. This systematic approach will ensure that AI not only produces content efficiently but also with care, originality, and contextual relevance.