🌀🗞 The FLUX Review, Ep. 178
February 20th, 2025
Episode 178 — February 20th, 2025 — Available at read.fluxcollective.org/p/178
Contributors to this issue: Ben Mathes, Erika Rice Scherpelz, Neel Mehta, Boris Smus, MK
Additional insights from: Ade Oshineye, Alex Komoroske, Chris Butler, Dart Lindsley,Dimitri Glazkov, Jasen Robillard, Jon Lebensold, 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.
“Only he who has no use for the empire is fit to be entrusted with it.”
― Zhuangzi
😵💫🍞 Gell-Mann amnesia for LLMs
Alright, picture this: you’re scrolling through the internet, right? You hit an article on, say, artisanal sourdough starters. You’re a sourdough wizard; you know your hydration ratios, the whole shebang. And this article? It’s wrong. Like, hilariously, offensively wrong. But, you chuckle, maybe roll your eyes, and then you’re off to the next article, something about, I dunno, quantum staplers. And suddenly, you're believing everything you read.
That's the Gell-Mann Amnesia effect. This happens to all of us, and it’s basically how we get duped by LLMs. They screw up your area of expertise, you write ‘em off. But then they spit out something about the optimal temperature for growing space kale, and suddenly, you’re taking notes. What do you know about space kale? And the LLM sure sounds like an authority.
See, this isn’t new. LLMs are like those newspapers from back in the day, except they are much faster and more confident. They’ll write code that looks slick but turns into a maintenance nightmare. They’ll tell you about the history of cheese graters with the authority of a thousand librarians, even if they’re making half of it up. And because you don’t know cheese graters, you assume they’re right. You’re like, “Wow, this LLM is smart!” forgetting that it just told you sourdough is made with motor oil. That’s the trap. We see the errors where we know… but blindly trust where we don’t. It’s like LLMs are great at “good enough,” but “good enough” ain’t “expert.” We keep forgetting the difference.
Imagine if LLM outputs looked like Balsamiq wireframes — sketchy, hand-drawn, clearly a draft. The visual cues would scream, “Hey, this is an idea, not gospel!” We need that kind of visual communication of uncertainty baked into LLMs so we don't mistake their confident-sounding BS for actual knowledge. Right now, the style of LLM output is way more authoritative than the content itself.
So, here’s the thing: we gotta remember these LLMs are just pattern-matching machines. They’re like (no, not exactly, please don’t come after us with pitchforks, researchers) really, really good parrots. They can mimic expertise, but it’s not clear they actually have it. They’re like your friend who knows how to sound authoritative on everything, even when they aren’t an expert. That’s especially true of the stuff that sounds super impressive, but you don’t actually understand. Because that's where the LLM is probably making up the best BS.
Most LLM outputs are sketches and drafts. You should treat them as such. The tone should probably match… at least until the accuracy becomes truly expert.
🛣️🚩 Signposts
Clues that point to where our changing world might lead us.
🚏🇨🇦 Canadian cafés are replacing “Americanos” with “Canadianos,” and brands are advertising “0% American” products
As US–Canada tensions continue to heat up, a number of Canadian coffee houses and roasting companies are renaming the Americanos on their menus “Canadianos.” Elsewhere in Canada, food brands have been putting up billboards saying things like “Cheestrings: Made with 0% American cheese”. At the same time, one Canadian grocer app rolled out a button that would let users substitute products in their cart with Canadian-made alternatives.
🚏🛰️ Nokia is putting a 4G cellular network on the moon
Intuitive Machines, the company behind the first lunar landing by a private company, is mounting a second mission this year — and they’re working with Nokia to deploy a cellular network to help the lander and two rovers communicate. Nokia had to design a “network in a box” with components that are “robust against radiation, extreme temperatures, and the sorts of vibrations that will be experienced during the launch, flight, and landing.” This technology could be helpful for Artemis IV, the planned manned mission to the Moon slated to launch later this decade.
🚏🎨 Christie’s is running an auction of AI-generated art
The noted auction house Christie’s is running an online auction called “Augmented Intelligence,” featuring several pictures and videos generated with AI. (Some of these pieces were also minted as NFTs.) This likely makes Christie’s the first major auction house selling AI art. The move has drawn backlash from artists, who criticize AI art models as being trained on copyrighted work from human artists; one open letter urging Christie’s to cancel the auction has gotten over 5,000 signatures.
🚏💉 Scientists made an “injection” that could make lithium-ion batteries last 6x longer
Lithium-ion batteries wear down over time as their ions become less efficient at moving between terminals. To counter this, a team of Chinese researchers developed a chemical solution that can be “injected” into lithium batteries to revive the lithium ions. In their tests, an electric-car battery that got the injections only lost 4% of its health after almost 12,000 charging cycles, compared to its usual lifespan of only 2,000 cycles. This “lithium carrier molecule” tech is also remarkably cheap and compatible with various batteries.
📖⏳ Worth your time
Some especially insightful pieces we’ve read, watched, and listened to recently.
Another Tale of Two Islands (Hoser / YouTube) — Uses two similar African island nations (Cabo Verde and São Tomé and Príncipe) to explore how being rich in natural resources often curses countries (like São Tomé, which has a lot of oil) to slow growth. Some reasons: extracting resources generates little spillover knowledge or productivity growth; the volatile price of commodities “throws a wrench in any sort of long-term [economic] planning”; exporting a lot makes your currency strong, which makes other exports less competitive (Dutch disease); and resource extraction encourages rent-seeking, which in turn encourages corruption and autocracy.
50 Years of Travel Tips (Kevin Kelley) — Shares some delightful tips for engaging more deeply with the world when traveling. Our favorites: pay your taxi driver to take you to visit their mother (you’ll get to see a local’s home and often taste some home cooking); organize trips around themes (e.g. “obscure cheeses or naval history”) rather than destinations; follow “gradients” of crowds on the street and you’ll usually end up at something interesting like a festival or market.
The MTG Color Wheel (Duncan Sabien) — A novel take on personality taxonomies that categorizes personalities using the five colors of magic from the card game Magic: The Gathering, arranged along the vertices of a pentagon. Each color has a dominant value and a mean of attaining it (e.g. White is Peace Through Order), two natural allies which share a value (e.g. White–Green is Community) and two natural converses which form a continuum (e.g. White–Red is Group–Individual).
Tariffs, Lobbying, and the Spoils System of Exemptions (Clare Brock) — Observes that trade wars (including tariff fights) are especially lucrative for lobbyists, whom companies hire to “petition [the] government for loopholes and exceptions.” This, in turn, preferentially benefits larger companies, who have more money to spend on lobbyists.
🔍📆 Lens of the week
Introducing new ways to see the world and new tools to add to your mental arsenal.
This week’s lens: causal graph.
Amazon popularized the idea of leading indicators as controllable input metrics. In their book Platform Engineering, authors Camille Fournier and Ian Nowland enrich this idea by connecting it to the idea of a causal graph:
Imagine constructing a causal graph, where nodes have different inputs you might measure, like storage or experiment throughput, and outputs that are connected to other nodes. Node outputs may be differently sensitive to inputs: maybe the number of experiments you can run increases with throughput at some rate, until you hit diminishing returns because of some other factor. At the end of your graph should be some KPI that’s important to the business, like revenue.
You should collect impact metrics at the interesting nodes in that graph. … Metrics let you check whether your impact theory is sound and adjust when it isn’t.
By thinking of your theory of change as a network of cause-and-effect relationships, where nodes are tied to measurable metrics, you create a map linking the controllable to the consequential. This graph illuminates the pathways between inputs (what you can influence) and outputs (what you aim to achieve). It strengthens your ability to choose KPIs with intention, ensuring they reflect the underlying dynamics of your system.
A well-crafted causal graph also acts as a guardrail for metric design. By surfacing the connections and sensitivities between nodes, it helps maintain discipline. It prevents a myopic view of metrics for their own sake by forcing you to trace a metric’s relevance back to core business outcomes. Conversely, a causal graph encourages you to experiment and adapt when reality deviates from the model. When teams understand their metrics’ position in the causal graph, they gain agency to innovate within boundaries aligned with larger goals.
Nothing can completely prevent the flattening of metrics caused by Goodhart’s law. Still, by explicitly connecting metrics to a theory of change, the causal graph provides a way of thinking that, hopefully, helps metrics retain at least some value even in the face of becoming a target.
© 2025 The FLUX Collective. All rights reserved. Questions? Contact flux-collective@googlegroups.com.