🌀🗞 The FLUX Review, Ep. 235
May 14th, 2026
Episode 235 — May 14th, 2026 — Available at read.fluxcollective.org/p/235
Contributors to this issue: Justin Quimby, Erika Rice Scherpelz, Neel Mehta, Boris Smus, MK,
Additional insights from: Ade Oshineye, Ben Mathes, Dart Lindsley, Jasen Robillard, Jon Lebensold, Lisie Lillianfeld, Robinson Eaton, Spencer Pitman, Stefano Mazzocchi, Wesley Beary, and the rest of the FLUX Collective
We’re 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 we’ve noticed in recent weeks.
“Part of the appeal of the fantastic is taking ridiculous ideas very seriously and pretending they’re not absurd.”
— China Mieville
🏓🫂 Playing for keeps
When you play a game, you’re immersed (hopefully). Perhaps you’re a quack doctor filling your cauldron one chip at a time. You push your luck. Your pot explodes. The joy, the regret, the triumph, they’re all real. Nine rounds later, the person with the most points wins. The chips go back in the bags, and you move on to the next game. You don’t mourn the potion that blew up in round six. The goals of the game were real while you were playing and irrelevant now that you aren’t.
Philosopher C. Thi Nguyen posits in Games: Agency as Art that games inherently involve adopted ends: ends you take on voluntarily for the duration of play and set down when you’re done. A chess player doesn’t want checkmate the way they want food or love or recognition. They want it in a temporary, bracketed way.
Nguyen contrasts this with enduring ends: the things you care about in a more stable way, such as relationships, craft, or your sense of self. These don’t switch off when the session ends.
We can use the idea that some ends are disposable to think about LLMs. An LLM agent receives a task. It pursues the task with apparent conviction. But when the session ends, the agent moves on with no residue of having invested in that task at all.
Sometimes we say that AI agents are like a genius with amnesia, but Nguyen’s framework suggests another way to frame this. AI agents act within the constraints of disposable ends. The prompt defines a goal, the same way the rules define a game. When the session is over, the end is disposed of because it is no longer relevant.
Being able to dispose of ends is valuable, but the agent operates entirely within disposable ends. There is no outside the game where it goes home and reflects on what mattered. There are no enduring ends that the various disposable ends accumulate into. Jailbreaking an LLM is less like shallowly fooling it and more like convincing it to change the rules it is acting under (human social engineering often works the same way).
For humans, our disposable ends are dominated by our enduring ends. We might be willing to be ruthless playing against our friends, but we don’t really want to damage our relationship with them. At the same time, our enduring ends are not static. Slowly, our goals and values do shift. This lack of intermediate state between disposable ends and the broad alignment encoded in training makes it hard for an AI agent to have an enduring sense of direction.
What would it mean to give an AI agent enduring ends? The memory systems and long-running agents being built right now are attempts at exactly that. But so far, it feels like they’re merely scaffolding enduring ends onto a substrate that, by its nature, only knows how to take on disposable ends. It seems like what’s missing is a distillation layer. Perhaps we’ll learn that the distillation of experience and reasoning into opaque emotional response is load-bearing, not some irrational, vestigial quirk of human nature.
🛣️🚩 Signposts
Clues that point to where our changing world might lead us.
🚏🥣 Japanese snack packaging is turning grayscale amid ink shortages
Calbee, a large Tokyo-based manufacturer of potato chips and cereal, is turning its chip bags and cereal boxes grayscale, with greatly reduced decoration. Japan, which imports almost all of its oil, is facing a shortage of naphtha, a petroleum-based liquid that’s used in ink—so Calbee’s move helps it reduce ink usage and thus “respond flexibly to changing geopolitical conditions.”
🚏🥇 Big tech workers are “tokenmaxxing” to rise up AI usage leaderboards
Amazon is reportedly tracking how many LLM tokens employees are using (and some believe the stats will be used in performance evaluations), while Meta reportedly had an unofficial leaderboard tracking who used the most tokens (the high scorer had burned 281 billion tokens). This has led to the rise of “tokenmaxxing,” a slangy term for intentionally burning tokens to look better to one’s boss. For instance, many Amazon staffers reportedly ran their company’s internal AI-powered automation tool on “trivial tasks” to “inflate their token counts.” (Goodhart’s law, anyone?)
🚏🚗 There’s a new $10,000 EV in China
Electric cars in the USA sell for an average of $55,000 (with a floor price of about $30,000), but a new bestselling Chinese EV, the MG4, starts at just $10,000. It’s already sold 100,000 units in eight months, which set “a new record for the fastest-selling pure electric hatchback.” The MG4 is also the world’s first electric car with semi-solid-state batteries; these are more fire-resistant than the standard liquid lithium-ion batteries but have better performance than solid-state batteries (which haven’t been productionized yet in EVs).
🚏💱 South Korea is considering an “AI dividend” for citizens
South Korea’s presidential policy chief wrote on social media that some of the profits and tax revenue from the AI boom “should be structurally returned to all citizens,” because the gains from AI were driven by Korea’s decades-long infrastructure investments whereas “excess profits in the AI era are, by nature, concentrated.” The comments came just a month after tens of thousands of protesters gathered at a Samsung chip factory to demand that employees get a cut of AI-derived profits.
📖⏳ Worth your time
Some especially insightful pieces we’ve read, watched, and listened to recently.
Dawn of the Electric World Order (Phenomenal World) — Argues that the Iran War has been a boon for the renewable energy transition, because only “geopolitical catastrophe… creates the supply-side shocks necessary for big changes in the energy complex to unfold.” Unlike the ‘70s oil shock or even the 2022 war in Ukraine, “this is the first energy shock with a superior alternative,” as adoption of EVs, batteries, and solar panels has hyper-accelerated in the last four years.
On Seeing Through Unseeing: The Hacker Mindset (Gwern) — Examines how computer hacking, gaming speedruns, heist movies, linguistic puns, and many other diverse techniques share a common pattern: “seeing through” the apparent rules and abstractions of a complex system and instead dropping down to the atomic elements, which you can manipulate in unexpected ways to achieve creative ends.
AI, the Shadow Prince (M. Alejandra Parra-Orlandoni) — Argues that, whether or not it becomes sentient, AI already operates as a Machiavellian “Shadow Prince,” an indispensable but unconscious minister to whom we willingly cede agency out of convenience. Framing AI as a novel “fourth actor” alongside government, business, and citizens provides a highly useful lens for rethinking alignment, demanding new checks and balances to protect our “cognitive security” and prevent this shadow minister from quietly concentrating absolute power in the corporate labs that control it.
Is There Actually a Word With a Longer Etymology Chain Than Pog? (Unyes / YouTube) — Following up on a popular linguistics video about the long etymology of “pog” (which went from fruit drink to gamer slang), this video explores more words with etymologies that hop around the globe or reach into the deep past. Examples include “apricot” (which looped around the Mediterranean from Latin to Syriac to Arabic to Catalan to English) and “divan” (which came all the way from Sumerian to the Ottomans to Hong Kong, where it now means ‘drug den’). Also mentions some fun examples of semantic shifts, like how “dinner” used to mean lunch (and before that, breakfast), and how the financial term “check” ultimately comes from chess.
🔍📆 Lens of the week
Introducing new ways to see the world and new tools to add to your mental arsenal.
This week’s lens: the context-free reader.
A colleague pings you: “I read the strategy doc, but I’m still not sure what I should do.” You read it again. It seems clear... but then, it seemed clear when you wrote it, too.
What we need when we write is the context-free reader. They don’t know the team’s unwritten norms. They haven’t read all the Slack threads. They don’t have years of historical background that makes certain assumptions obvious. They can only work with what is on the page. (Here, “context-free” is a stand-in for the minimal shared context of our target audience.)
While frustrating for a writer, context-free readers do provide value in that they do not suffer from the curse of knowledge: once you know something, it becomes nearly impossible to imagine not knowing it. These readers can see assumptions that have grown stale and processes that were built for a different environment.
Finding a context-free reader used to be a challenge. Where do we find someone who doesn’t have the context but cares enough to read our drafts? Today, we all have one at our fingertips. The positive flipside of the LLM’s disposable ends is exactly that lack of enduring context. Feed your favorite LLM the doc plus a list of questions you want any reader of the doc to be able to answer. If the LLM hedges, contradicts itself, or confidently lands on the wrong answer, the doc isn’t clear. If it comes away with the answers you were hoping for, then you have some hope!
© 2026 The FLUX Collective. All rights reserved. Questions? Contact flux-collective@googlegroups.com.