🌀🗞 The FLUX Review, Ep. 90

March 9th, 2023

Ladybug. // Photo: Stefano Mazzocchi, FLUX

Episode 90 — March 9th, 2023 — Available at read.fluxcollective.org/p/90

Contributors to this issue: Erika Rice Scherpelz, Dimitri Glazkov, Neel Mehta, Boris Smus, Ade Oshineye

Additional insights from: Gordon Brander, a.r. Routh, Stefano Mazzocchi, Ben Mathes, Justin Quimby, Alex Komoroske, Robinson Eaton, Spencer Pitman, Julka Almquist, Scott Schaffter, Lisie Lillianfeld, Samuel Arbesman, Dart Lindsley, Jon Lebensold

---

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.

“Every successive stage of thinking is a conclusion in which the meaning of what has produced it is condensed; and it is no sooner stated than it is a light radiating to other things — unless it be a fog which obscures them.”

— John Dewey, Experience and Nature

⚔️🦾 Iron sharpens iron

“Us vs. them” is a classic broken lens. Even more insidious is the “us vs. us” variation. “Us vs. us” is a classic movie trope that underlies many thrilling murder mysteries and exciting spy novels: “The call is coming from inside the house.” However, when this dynamic develops within an organization, it’s a lot less fun. Let’s look at the script.

Our story begins with a belief that someone in our tight-knit, well-performing team is actually working against us. Whether they are a sneaky free-rider or an active saboteur doesn’t really matter. There’s no longer a single “us.” Now, there’s “us”... and an unknown “other” hiding in our midst. To protect ourselves from the harm they could bring to us, we create policies that assume that anyone could be a bad actor. 

This seems like the path to safety, but when we assume that anyone at all can be a bad actor, we harm our team culture. We start to act on the assumption that people in our team cannot be trusted. We add friction. We remove permissions. We expect people to prove that they deserve trust before we start to trust them. The long-term effects of these mitigations are more damaging than they initially seem. An organization that tries to to protect itself by assuming that no one is to be trusted will, ironically, feel less safe over time.

A more productive way of looking at these situations is “iron sharpens iron.” When we encounter danger, instead of looking at our teammates as potential threats, we can see them as fellow competitors in an iterative zero-sum game with bounded stakes and clear safeguards designed to achieve shared goals. Instead of focusing on preventing danger, we should focus on resilience in the face of danger. We are all competing against the same enemy that would infiltrate our systems, leak our data, etc. This is the mindset that yields ideas like orange team security exercises. 

Striking this balance of vigilance and camaraderie can be challenging. If we do it right, we’ll be safer and maintain the benefits of team trust. It’s up to us and our leaders to design organizations that channel behavior in positive directions rather than trying to lock down everything in order to prevent the possibility of bad behavior. 

🛣️🚩 Signposts 

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

🚏🦙 Meta’s LLaMA language model leaked online, and anyone can download it

Most large language models (LLMs) are locked behind an API, meaning that companies can gate-keep who can use them and for what. But when Meta shared its raw LLaMA model with a limited set of researchers and companies, someone leaked it on 4chan, and it quickly became freely available on GitHub — so anyone can now have their own personal copy of the LLM. Commentators speculate that Meta spent several million dollars in hardware costs alone to train the model (to say nothing of engineering salaries).

🚏👛 Consumer debt has hit its highest level since 2009

The number of US residents who are behind on their credit cards, auto loans, or personal loans has soared to nearly 25 million — the highest point since 2009, and nearly double the figure from about 2 years ago. One journalist attributed the spike in debt to “high inflation on basics,” the “end of pandemic government aid,” and “overspending.” Plus, some subprime borrowers saw their credit scores rise after paying off debts in 2020 and ‘21, leading them to take out too much new debt.

🚏🧬 LLM-like AI models can generate proteins never before seen in nature

Biotechnologists have long been using machine learning to predict how a given chain of amino acids would fold up, but scientists think that AI can now help generate the biological “code” for all-new proteins. They can input a structure they want, and the AI model (which behaves similarly to a large language model) will output an amino acid sequence that’s likely to fold in the right way. The problem is analogous to text generation: amino acid sequences are like text, but they must obey certain rules of biological “grammar,” “syntax,” and “semantics.”

🚏🇮🇷 Iran says it found the world’s 2nd-largest lithium deposit

Lithium is in high demand thanks to the rise of electric vehicle batteries, so analysts took note when an Iranian official said that a large lithium reserve had been found in the country. The deposit apparently holds 8.5 million tons of the rare metal (good for 10% of the world’s total reserves); it would give Iran more lithium reserves than any other country besides Chile. Experts say that this could push down the price of lithium and revive Iran’s economy, which has been badly battered by sanctions.

🚏🗃 4% of employees have put sensitive business data into ChatGPT

A data security firm that monitors ChatGPT inputs reported that 4.2% of employees at its client companies had tried to put confidential or sensitive data into the chatbot. One executive pasted their firm’s strategy document into ChatGPT to create a slide deck; one doctor put his patient’s name and medical info into ChatGPT to generate a letter to the insurance company. Companies worry that AI services could integrate this proprietary data into their models and potentially surface it at a later date.

📖⏳ Worth your time

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

• A Climate Risk Assessment for the United States (Risky Business) — Explores how climate change will impact each region of the US by 2100, which industries could be the most affected in each region, and the various risks to life and property in each area. Includes some great maps that show projected temperature change, sea level rise, flooding risks, and more.

• How Teams Change at Different Sizes (Jonathan Maltz) — Describes how individual contributors and managers need to work differently at different team sizes: from 2 or 3 people (“an ideal small squad”), to 6–9 (“a good balance of size and redundancy”), to 13+ (“not even a team anymore… a bunch of sub-teams that secretly need to break off”).

• Reddit’s Donut Crypto Experiment Turns Sour (Decrypt) — Tells the story of how one user on a crypto-focused subreddit built a tool to let users tokenize their upvotes as “donut” tokens and trade them for money on decentralized exchanges. Within days, people started buying and selling influence (having more donuts gave your votes more weight in community polls) and gaming the system to earn more donuts (such as by making fake accounts to upvote one’s posts).

• Paved Paradises (Scope of Work) — Shows the invisible but wide-ranging effects that mandatory parking minimums have had on North American cities, and highlights the work that advocates (starting with Donald Shoup, author of The High Cost of Free Parking) have been doing to cut down on these requirements.

• Arabic and Islamic Themes in Frank Herbert's “Dune” (The Baheyeldin Dynasty) — Shows how, from the transparent analogy between spice and crude oil to the stylings of the Fremen, Herbert's masterpiece was clearly inspired by Middle Eastern themes. The author digs deeper into an etymological tour of the series, revealing more Arabic influences than a typical Westerner could glean.

🔍📆 Lens of the week

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

This week, we’re bringing you a lens portmanteau: the revealed specialist.

Building a team naturally brings up questions of team composition. One dimension that often comes up is the mix of generalists and specialists. Generalists are like a multi-tool; they can tackle various problems with a consistent level of effectiveness. Specialists are best applied to problems that they… well, specialize in. Otherwise, their expertise is wasted.

However, it turns out that there is no such thing as a pure generalist. Even those who consider themselves generalists make choices that, when added up over time, reveal their particular way of specializing.

This is where we run into our next lens: revealed preferences, first articulated by Paul Anthony Samuelson, who studied consumer behavior in the early 20th century. The key insight of revealed preferences is that our preferences are often better assessed from what we do rather than what we say. 

If our hiring process prefers generalists, people are more likely to present themselves as generalists. It is only after some time that we learn, for example, that the person we hired is a huge Rust + monorepo fan. They may be a generalist in theory — they can use multiple repositories and program in the languages we use — but given the opportunity, they will act like a specialist. Our team conversations start gravitating toward switching to Rust, moving all of our source to a monorepo, and so on. 

There is nothing wrong with being a revealed specialist. Most people will become one as they go through their career. It is only when we believe too firmly in generalists that revealed specialists seem flummoxing.  Some organizations pride themselves on only hiring generalists, touting their ability to adjust to a changing environment. This may backfire. Such an organization may believe it is highly adaptable and flexible, but when a need for change comes, it keeps acting in the same familiar way. Perhaps our generalists aren’t actually that general. Perhaps they are revealed specialists.

The idea of generalists and specialists can still be useful if we allow for the limitations of revealed preferences. Instead of assuming that stated generalist preferences are representative, we can expect that the particular way an individual prefers to specialize will be revealed over time. And knowing that it’s coming, we can be ready to make the effort to integrate that specialist into the team structure.

---

© 2023 The FLUX Collective. All rights reserved. Questions? Contact flux-collective@googlegroups.com.