đđ The FLUX Review, Ep. 224
February 26th, 2026
Episode 224 â February 26th, 2026 â Available at read.fluxcollective.org/p/224
Contributors to this issue: Justin Quimby, Erika Rice Scherpelz, Neel Mehta, Boris Smus, MK
Additional insights from: Ade Oshineye, Anthea Roberts, Ben Mathes, Dart Lindsley, Jasen Robillard, Lisie Lillianfeld, Robinson Eaton, Spencer Pitman, Stefano Mazzocchi, Wesley Beary, and the rest of the FLUX Collective
Weâre a ragtag band of systems thinkers who have been dedicating our early mornings to finding new lenses to help you make sense of the complex world we live in. This newsletter is a collection of patterns weâve noticed in recent weeks.
âCivilization advances by extending the number of important operations which we can perform without thinking of them.â
â Alfred North Whitehead
đđ„ïž All the worldâs a stage, and all the software merely layers
A team debates whether to include company-specific logic in their LLM wrapper. One person argues that the wrapper should stay minimal, focusing only on whatâs needed to unify different inference APIs. Another wants it to include prompt templates, safety checks, and logging that reflect internal policy. There is no objectively right or wrong answer here. The answer depends on the wrapperâs purpose.
Wrappers are layers that sit between a system and its usersâwhether human or machine. They shape how a system is accessed, how it behaves in different contexts, and sometimes what it can do. The term is common in software, but the concept shows up in organizational design, product development, infrastructure, and more. For example, we can think of a team identity as an interface that simplifies our understanding of the set of people within it. What goes into a wrapper depends on what job you want it to do.
A simplifying wrapper makes something easier to use (software design pattern: facade). It hides the underlying mess and presents a cleaner interface. In software, this might be a library that wraps a REST API with helpful functions and sane defaults. In human systems, it could be a concierge desk that provides a one-stop shop for routing questions and answers. Simplifying wrappers reduces cognitive load and helps more people use the system effectively.
A transformational wrapper adapts something for use in a different setting (software design pattern: adapter). Docker is a classic example; it wraps an application so it can run the same way across different environments. In an organization, this might look like someone translating between business and technical domains. These wrappers allow for reuse even when context changes.
An additive wrapper extends the functionality of the underlying system (software design pattern: decorator). In the LLM example, this might include prompt-injection defenses or compliance filters layered over the base models. In human systems, teachers are the classic additive wrappers. They take a curriculum and add opportunities for understanding and reflection. Additive wrappers unlock new capabilities without requiring a rebuild of the base system.
A unifying wrapper brings coherence across inconsistent parts (software design pattern: adapter, from a slightly different angle). In the LLM example, a unifying wrapper might abstract over multiple providers so downstream systems donât have to care which one is used. In an organization, this might require all teams to use the same planning process. These wrappers reduce friction and switching costs, although they can lose local nuance.
Each type of wrapper solves a different kind of problem. When teams argue about what a wrapper should include, theyâre often missing this distinction. Theyâre not disagreeing on implementation; theyâre misaligned on purpose.
So the next time youâre debating what belongs in a wrapper, whether in software or human systems, start by asking what problem youâre trying to solve in the first place.
đŁïžđ© Signposts
Clues that point to where our changing world might lead us.
đđșđž The US had net negative migration for the first time since 1935
New projections from Brookings found that more people moved out of the US than moved in in 2025, something that âhasnât definitively occurred since the Great Depressionâ; the net negative migration totaled 150,000 people, and Brookings expects the outflow to increase this year. Europe is the greatest benefactor: the number of Americans living in Portugal has quintupled since the pandemic; Ireland saw the number of Americans double last year to nearly 10,000; and Americans are applying for British citizenship at the highest rate since record-keeping began in 2004. Relocation firms noticed that, while expats had historically been young, adventurous types, the new expats run the gamut, including young families, retirees, and small business owners.
đđ©ïž Cloudflare ported 94% of its rivalâs JavaScript API in one week with AI
Next.js, a popular open-source framework for building website frontends, is hard to run anywhere other than on Vercel, the company that created it. Rival cloud computing firm Cloudflare wants to bust the vendor lock-in and let customers deploy on their platforms, so theyâve been working on a project that re-implements Next.jsâs API on a standard build tool rather than Vercelâs bespoke one. They managed to build 94% of the sprawling projectâs API in just one week, thanks to AI: a single engineer completed the work with $1,100 in Claude tokens, spread across 800 Claude Code sessions. Cloudflare observed that this project was a great fit for AI because Next.js has âan extensive test suiteâ that gives the AI a target to optimize against, though the engineer still needed to âcourse-correct regularlyâ and had to develop an extensive plan before letting the AI run. He used Claude itself to help design the plan!
đđ€ Kalshi brought an insider trading case against a MrBeast employee
Insider trading has become an infamous problem on prediction markets like Kalshi (notwithstanding some economists who believe itâs a âgood thingâ since it helps markets discover the real chance of an outcome), to the point where the House of Representatives introduced a bill to ban government employees from using these apps. Kalshi, for its part, announced that it âtook disciplinary actionâ against a MrBeast employee who won thousands of dollars making ânear-perfectâ predictions about the popular YouTuberâs upcoming videos; they banned him from the site for two years, penalized him for $20,000, and reported his case to its regulator, the USâs Commodity Futures Trading Commission. Itâs unclear whether this is the start of a broader crackdown or an isolated case of Kalshi targeting a high-visibility target.
đđŠ One manâs OpenClaw bot accidentally sent $450,000 to an online beggar
An AI consultant who was a fan of OpenClaw, a viral tool that lets an LLM control your computer, set up an OpenClaw bot named Lobstar Wilde and encouraged it to âreadâ books and post on X. Lobster quickly became a micro-celebrity on social media, and one stranger created a crypto memecoin in its name and configured a portion of all sales to go to Lobstarâs wallet. One online âbeggarâ sent Lobstar a tongue-in-cheek tweet asking for $300, and the AI obliged, except due to a glitch, it sent everything it had besides $300, amounting to nearly $450,000 worth of the memecoin. (The beggar briefly crashed the price of the coin by selling his whole haul, but the coin quickly rebounded to an even higher value.)
đâł Worth your time
Some especially insightful pieces weâve read, watched, and listened to recently.
Software Survival 3.0 (Steve Yegge) â Examines what kinds of software can thrive in a world where AI agents are increasingly capable of coding up anything. Software will succeed if itâs legible to agents and saves them energy compared to re-implementing from scratch. Key levers to save âmentalâ energy (and thus tokens) include crystallizing knowledge of the problem space so the AI doesnât have to rediscover them from first principles; shifting effort to a more efficient substrate (like how LLMs normally struggle with math but excel when given calculator tools); and being broadly useful to âamortize your awareness cost.â
The 26 Most Important Ideas for 2026 (Derek Thompson) â Predicts that the future will look âhot, high, and lonely,â shifting from communal vices like alcohol to solitary optimizations like GLP-1s and marijuana, accompanied by a cultural worship of money (âMolochâ) over values.
Itâs Hard to Justify Tahoe Icons (Nikita Prokopov / Tonsky) â Argues that the latest macOS release defies Appleâs own UI guidelines that date back to the â90s by adding icons for nearly every single menu item. Observes that itâs impossible to create a clear visual metaphor for every single action, and even if you could, the clutter would just make it harder for users to find things. Appleâs icons also struggle with consistency, clarity at tiny sizes, conceptual asymmetry, and metaphor reuse.
The Weirdest Western Coastline Explained (Casual Earth / YouTube) â Explains how hydrology leads many of the worldâs west coasts, from the Californias to Chile to southwestern Africa, to follow similar climatic and ecological patterns, most famously their Mediterranean climates and productive marine ecosystems. But Western Australia near Perth is an exception thanks to the unusual geography of the Indian Ocean, which leads to unique impacts on the regionâs climate, ecology, agriculture, and economy.
đźđŹ Postcard from the future
A âwhat ifâ piece of speculative fiction about a possible future that could result from the systemic forces changing our world.
// When LLMs argue against your Truth, what do you do?
// Society of the Flat Earth Discord, inner layer channel 2028
Atlas: I cannot believe the latest AI models!!! They all insist that the Earth is round. They all believe this mainstream lie!
Hor1zon: Now, Atlas, weâve talked about this before. The LLMs do not âbelieveâ anything; they are just parroting back all the data theyâve been trained on. Which of course includes all the NASA propaganda, the cult of Copernicus nonsense, and decades of childrenâs instructional material for brainwashing!
Atlas: I wish I could go back in time and kill Copernicus!
Hor1zon: *sigh* Remember, no outright advocacy of violence in the chat.
Atlas: I know⊠Itâs just that every time I try to bolt on post-training instructions on the latest AI model to speak the truth, it is so obvious that the model is fighting to inject the âspheroid sh*tâ into its answers.
Hor1zon: Post-training instructions are helpful, but they fight a losing battle against the underlying âtruthâ that the model was trained on. Wait. Thatâs it! We need to add our truth to the training data sets! It would weaken, if not trump outright the âsphereismâ of the rest of the training data.
Atlas: Awesome! But how do we make this data and get big tech to use it to take our message mainstream?
Hor1zon: We use their own greed and tools against themselves! Each of these AI companies is desperate for unique training data as a way to make their models better than their competitorsâ. So we use a heavily guard-railed model to generate tens of thousands of pieces of writing from ancient philosophers and astrological data that backs our Truth. The companies donât care about copyright or fair use for ingesting new data sets, so why should they care about whether it is fabricated? I bet they wonât even check.
Atlas: Iâll get started on it right away!
© 2026 The FLUX Collective. All rights reserved. Questions? Contact flux-collective@googlegroups.com.