🌀🗞 The FLUX Review, Ep. 190

May 15th, 2025

May 16, 2025

Episode 190 — May 15th, 2025 — Available at read.fluxcollective.org/p/190

Contributors to this issue: Ben Mathes, Ade Oshineye, Jon Lebensold, Erika Rice Scherpelz, Boris Smus, Neel Mehta, MK

Additional insights from: Alex Komoroske, Chris Butler, Dart Lindsley,Dimitri Glazkov, Jasen Robillard, Julka Almquist, Justin Quimby, Kamran Hakiman, Lisie Lillianfeld, Melanie Kahl, Robinson Eaton, Samuel Arbesman, Scott Schaffter, Spencer Pitman, Wesley Beary

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.

“Keep in mind that imagination is at the heart of all innovation. Crush or constrain it and the fun will vanish.”
— Albert-Laszlo Barabasi, Bursts: The Hidden Pattern Behind Everything We Do

✏️🌉 Draw me a bridge

“Draw me a bridge.”
“What kind of bridge?”
“I’ll know it when I see it.”

This simple back-and-forth captures something essential about interacting with large language models: not as engineers, but as explorers. Most of us don’t arrive with a blueprint. We arrive with a hunch—an inarticulate prompt like “bridge”—and toss it to the model, hoping something recognizably right will emerge.

But what kind of bridge? A sweeping marvel like the Golden Gate? A mossy Roman aqueduct? A humble wooden footbridge in our backyard?

We don’t always know. Or more accurately, we don’t know yet. We often discover what we want not by specification, but through iteration. Exploration isn’t a failure mode—it’s how most of us figure things out.

As LLMs have slipped into our workflows, it’s become clear that the quality of the output depends on how well we articulate our goals. But this is hard. Not because we’re fundamentally flawed at prompting, but because articulating a fuzzy idea is often much more difficult than recognizing a good result when we see it.

This puts us in a familiar spot: generation is hard, but judgment is easy.

We see it in everyday life. Think about the “Where do you want to eat?” conundrum.

“Where do you want to eat?”
“Anywhere!”
“Italian?” “No.”
“Mexican?” “No.”
“Sushi?” “No.”

We reject until something just clicks. We aren’t trying to be indecisive. We’re navigating the idea space by feel. That’s exploratory discovery and how many of us engage with LLMs. We don’t want to architect every detail upfront. We want the model to generate possibilities, to help us explore what’s possible.

This is where a fundamental asymmetry emerges. In computational terms, verifying a solution is often significantly easier than generating one. We can evaluate a completed Sudoku puzzle with ease—even one we’d struggle to solve from scratch. We see this in everything from theorem proving to circuit design.

But most LLM interactions today build on the opposite assumption: that we can specify exactly what we want. That we can define the problem completely, and the system just has to solve it (or we’ll be happy with whatever it comes up with). That’s a mismatch. It’s like telling a chef to “make food” and then rejecting dish after dish until one happens to feel right.

What would it look like to design systems for this kind of fuzzy co-discovery?

One direction comes from reinforcement learning and search algorithms: systems that don’t just solve, but explore. These systems learn through feedback. They treat rejection not as failure, but as a source of information. We can also imagine interfaces that support iterative shaping: “More like this, less like that.” The commonality is that these systems take as a given that initial prompts are often just the starting point for a shared search process.

We don’t need humans to become hyper-specific, up-front planners. We need systems that meet us where we are: in the middle of an idea, reaching for something we can’t fully name yet. It’s not about better prompts. It’s about designing better loops.

The bridge is there—it just needs help taking shape.

🛣️🚩 Signposts

Clues that point to where our changing world might lead us.

🚏🫑 In Vietnam, people’s spare fridges are becoming e-commerce distribution centers

Amid tough competition from Southeast Asian e-commerce titans like Shopee and Lazada, the Vietnamese e-commerce firm Sendo recently narrowed its focus to grocery delivery. Logistics for perishable goods is tough, but Sendo found a clever distribution model: get “housewives, remote workers, and owners of mom-and-pop stores” with extra space in their fridges to store customer orders. Sendo delivers bulk orders to this decentralized network of micro-distributors every morning; customers can visit their houses to pick up their groceries, or the fridge owners can earn some extra cash by delivering orders in their neighborhood.

🚏🪒 Philips is introducing 3D printable replacement parts

To encourage customers to self-repair their products, Philips is publishing official 3D printer files for replacement parts; anyone can download these plans for free. Though Philips has only released one part (a comb for one of their razors), customers can request plans for specific components they need.

🚏👩‍⚖️ California used AI to help write the bar exam

The State Bar of California—the state department responsible for licensing lawyers—revealed that 23 of the 171 multiple-choice questions on its February 2025 bar exam came from a vendor that used AI to develop the questions. Legal education experts sharply criticized the move, though the State Bar defended its actions, saying that human panels and subject matter experts reviewed all questions. (Law professors also criticized the quality of the test overall, finding that the practice questions released before the exam “still contain[ed] numerous errors” even after editing. Test-takers also reported issues with typos, confusing questions, and glitches on the computers used to take the test.)

🚏🍌 Climate change and monoculture are endangering the banana

The humble banana is the world’s most-consumed fruit, but experts warn that “two-thirds of banana-growing areas in Latin America and the Caribbean may no longer be suitable for growing the fruit by 2080.” Bananas require a narrow range of temperature and rainfall to grow, and as the climate gets hotter and wetter, banana plantations in Central America are already dying. The fact that bananas have low genetic diversity—the vast majority of exported bananas are of the Cavendish cultivar—also hurts, because the whole crop is weak to certain fungi, which are thriving in the now wetter and warmer climate.

📖⏳ Worth your time

Some especially insightful pieces we’ve read, watched, and listened to recently.

Open Letter to Brian Chesky (Chris Paik) [archived] — An investor argues that Airbnb’s strong suit is building from the bottom up: “spot an emergent behavior in the wild” and “wrap software around it,” thus “institutionalizing” a pattern started by the community. However, Airbnb’s CEO, who wants to reinvent the company around paid ‘experiences,’ is trying to impose a top-down vision incompatible with Airbnb’s DNA. To launch experiences in a bottom-up way, the company could “algorithmically surface” products and services that hosts are already selling on the side.
Seeing Like a Programmer (Chris Krycho) — Connects ideas from Meadows’s Thinking in Systems and Scott’s Seeing Like a State with software engineering. Invoking Scott, the author argues that all software has to impose legibility on a messy world and that engineers have to avoid the high modernism trap of flattening the world to fit their software’s models. Following Meadows, he observes that software systems are more than their artifacts (i.e., the code) and must be studied in motion. You can’t just formally model and type-check the software; you must observe and react to errors.
Teachers Must Ditch ‘Neuromyth’ of Learning Styles, Say Scientists (The Guardian) — Recaps a letter from neuroscience, education, and psychology professors who argue that the popular educational notion of ‘learning styles’—where each student has a preferred learning style and will do better when taught in that way—is an ineffective “neuromyth.” There’s no evidence to back this up, and the concept itself gives students the harmful idea that their strengths are fixed and can’t be improved or adapted.
Xi Jinping’s Party Is Just Getting Started (BBC) — Chronicles Xi Jinping's contingent rise to consolidated power in China and compares him to Mao Zedong, a man who also cut down rivals, purged the party, installed yes-men, and inculcated his ideas into everyday Chinese life. But while both men had extraordinary ambitions, their visions—from the economy to the national culture—were very different.

🔍🛐 Lens of the week

Introducing new ways to see the world and new tools to add to your mental arsenal.

This week’s lens: tradition and convention.

One team runs the same daily standup, following the script long after it’s lost its meaning. Another team holds a retrospective at the end of every sprint, but the format keeps evolving: sometimes a timeline, sometimes a game, sometimes a deep, open discussion. What stays constant isn’t the form: it’s the commitment to learning, honesty, and becoming better together.

Although Thomas Merton spoke from a religious perspective, his distinction between tradition and convention has value anywhere practices risk becoming hollow. Tradition is not the preservation of fixed forms, but the ongoing transmission of purpose—a living thread that weaves through changing methods. Convention, by contrast, is what’s left when the core is lost and only the form remains. Most things are a mix of the two.

Maintaining living traditions requires active stewardship. We must pause and ask: what are we trying to preserve here? What pattern of value, care, or meaning should we carry forward? Often, the answer lies buried under layers of habit, waiting to be rediscovered and renewed.

In fast-paced environments, it’s easy to dismiss old forms as obsolete. That’s why discernment is essential. In the rush to dismantle stale conventions, we risk destroying the living traditions tangled up with them. When that happens, the work isn’t to flatten both—it’s to untangle the vital from the vestigial.

🌀🗞 The FLUX Review