🌀🗞 The FLUX Review, Ep. 64
August 18th, 2022
Episode 64 — August 18th, 2022 — Available at read.fluxcollective.org/p/64
Contributors to this issue: Ade Oshineye, Erika Rice Scherpelz, Dimitri Glazkov, Neel Mehta, Spencer Pitman, Dart Lindsley, Midjourney, GPT-3, DALL-E
Additional insights from: Gordon Brander, a.r. Routh, Stefano Mazzocchi, Ben Mathes, Boris Smus, Justin Quimby, Alex Komoroske, Robinson Eaton, Julka Almquist, Scott Schaffter, Lisie Lillianfeld, Samuel Arbesman
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.
“O for a Muse of fire, that would ascend
The brightest heaven of invention”
— William Shakespeare
📝 Editor’s note: This is a special ‘centaur’ episode where everything was created with the assistance of Large Language Models like GPT-3, DALL-E, and Midjourney.
🎭🦸 A muse, rather than an assistant
With large-language-model-powered tools taking early adopters’ attention by storm, we are inclined to pause and consider the potential future uses of these tools. Some of them already exhibit remarkable abilities. Still, rough edges show up quickly: faces melt, sentences stop making sense, and so on.
We are still in the early stages of understanding how to best use these tools. As we continue to explore, it is important to keep in mind both the potential benefits and their potential risks.
However, we have found one use that already seems intriguing: a source of inspiration. Especially in the busy Midjourney community, the creative vibe is nearly palpable. People riff on each other’s creations, showcase their best, and most interestingly, show fascinating new uses of the tool’s output.
Some of the most interesting uses come from people who take the output of the tool and use it as inspiration for another creation. For example, one user may take the output of the tool and use parts of it to create a new piece of art. Others mesh output from one tool with different tools, such as Craiyon and DALL-E, or feed generated text from GPT-3 as image generation prompts. Some think about ambitious ideas, such as incorporating generated media into video games. Think .kkrieger, but for the modern age.
What if, instead of being an assistant, the first broad application of AI is to be a muse? Suppose you want to break out of a creative rut. No problem — just ask your muse to inspire you. What might a Starship of Theseus look like? Just ask, and perhaps you’ll get something like this:
Struggling to describe a challenging concept in a few simple words? The muse has got you covered. For example, here’s a summary of Michael Porter’s five forces framework as a death metal banger (please don’t forget to metal-scream when reading it):
There's a lot of ways to skin a cat
But if you want to know the score
You need to know Porter's 5 forcesThe threat of new entrants is high
If you're looking to get in the game
You'd better be prepared to fightThe bargaining power of buyers is strong
They hold the purse strings tight
So you better offer them a good dealThe bargaining power of suppliers is weak
You can get what you need from them
But you can also make them bleedThe threat of substitutes is real
If you can't give them what they want
They'll find someone who willThe intensity of competition is fierce
No one wants to lose
So you better be ready to fight
The difference between an assistant and a muse is that it’s okay for a muse to be unreliable, weird, and even hilariously bad. That’s because you’re not going to blindly use the output without applying human judgment to curate, edit, and remix it. The muse’s output merely has to be a suitable source of inspiration — something to build on, to mold, to use as a tool to expand our thinking. These new LLM muses have the potential to inspire us, to help us think outside the box, and to nudge us toward becoming something new and wonderful.
[Editor’s note: Writing this, we were reminded of the times we wrote dialogues as a group. The human author and GPT-3 took “turns” writing paragraphs, and then we edited it to tighten things up.]
🛣️🚩 Signposts
Clues that point to where our changing world might lead us.
The following news summaries were written by OpenAI’s GPT-3 language model, with very minor human edits for clarity. The signpost titles were written by humans.
🚏🏏 Indian farmers streamed fake cricket matches to dupe Russian gamblers
Fake professional cricket matches are being streamed in India in order to scam Russian bettors. The matches are set up to look like they are being played between two professional teams, but in reality, the players are farmers who have been paid to pretend to play. The scam is only possible because of the popularity of cricket in India and the lack of knowledge about the sport in Russia.
[Editor’s note: a surprisingly good summary, especially considering we only gave GPT-3 the link to the article, not the article’s body text.]
🚏🇩🇪 With Russia restricting gas shipments, Germany is keeping its nuclear plants open
Government officials say Germany plans to postpone the closure of the country’s last three nuclear power plants. The move would mark the first departure from a policy initiated in the early 2000s to phase out nuclear energy in Germany and which had over time become enshrined in political consensus. A cabinet decision would need to wait on the outcome of an assessment of Germany’s energy needs that will be concluded in the coming weeks.
[Editor’s note: this summary is decent but missed a key bit of context — namely, that this is all the result of Germany’s response to the Russian invasion of Ukraine.]
🚏🎨 TikTok is using text-to-image AI to auto-generate video backgrounds
TikTok’s new AI greenscreen feature uses basic text-to-image model to create dreamy, abstract backgrounds for videos. This model is not as advanced as state-of-the-art text-to-image models, but it is much less resource-intensive. TikTok’s model only produces swirling, abstracted, smeared images. Not state-of-the-art, but probably for the better given the potential for harm with more advanced models.
[Editor’s note: this summary was produced with the prompt, “Write a three-sentence summary of this article in Reuters style.” Simply saying “summarize” or “paraphrase” yielded poor results, with GPT-3 just regurgitating the first few paragraphs of the article.]
🚏🏥 The CDC will restructure after its fumbling of COVID-19
After receiving harsh criticism for its handling of the COVID-19 pandemic and the monkeypox emergency, the Centers for Disease Control and Prevention (CDC) will undergo a significant overhaul involving cultural and structural changes. CDC Director Rochelle Walensky said the goal is a new, public health action-oriented culture at CDC that emphasizes accountability, collaboration, communication, and timeliness. The changes include stamping out pedantic data analyses that have slowed and hampered the agency's public health responses, improving health messaging to the public, and expanding emergency response teams.
[Editor’s note: this summary borrows several sentences nearly word-for-word from the article. It’s quite evident just from reading the summary, since in our experience GPT-3 has trouble capturing the journalistic turns of phrase and complex sentence structures that mark professionally-written articles.]
🚏🎬 A deepfake studio used AI to seamlessly replace swear words in a movie
The director of the upcoming action-thriller Fall used AI technology to remove over thirty F-bombs to turn its R-rating into a much more box office friendly PG-13. The movie was shot with IMAX cameras in the middle of the Mojave Desert in California on a modest production budget of just $3 million, meaning that reshoots would have cost time and money that simply wasn’t available.
[Editor’s note: GPT-3 made the questionable decision to include minute details about the movie’s filming without digging into the more newsworthy facts about the swear-removing AI technology.]
🚏🧑🍳 A fully-automated restaurant will open in San Francisco
Mezli is a fully automated robot restaurant that will be open by the end of this week. Mezli is the first automated restaurant to remove humans entirely from the on-site operation equation. The fully robot-run restaurant will be open at Spark Social in San Francisco’s Mission Bay neighborhood.
[Editor’s note: this summary hits the “robot-run” motif repeatedly without going into detail about how exactly the restaurant is automated. This appears to be a pattern in GPT-3 summaries.]
📖⏳ Worth your time
Some especially insightful pieces we’ve read, watched, and listened to recently.
The following summaries were written by OpenAI’s GPT-3 language model, with very minor human edits for clarity.
The Perils of Audience Capture (Gurwinder) — This article discusses how people can get caught up in trying to please their online audience and how this can lead to them becoming a caricature of themselves. It uses the examples of Nicholas Perry, who became Nikocado Avocado, and Louise Mensch, who became consumed by conspiracy theories. It argues that this is a problem because it means that people are not being their true selves and are instead becoming what other people want them to be.
Sony’s Racing Car AI Just Destroyed Its Human Competitors By Being Nice (And Fast) (MIT Technology Review) — This article discusses how Sony's racing AI, Gran Turismo Sophy, has been designed to be able to beat human competitors by being ‘nice’ and fast. It explains how the AI was tested against professional sim-racing drivers to fine-tune its skills, and how it ultimately won by being more strategic and less aggressive than its human counterparts. The article goes on to discuss how this AI could be applied to other contexts outside of racing, such as in driverless cars.
AGI is Sacred (Overcoming Bias) — This essay discusses how treating something as sacred can lead to unreasonable expectations and values. The author gives the example of how Christians treated their God as more perfect than other gods in ancient times. The author then applies this idea to the recent trend of treating artificial general intelligence (AGI) as a sacred ideal. The author argues that the AGI ideal is not based in reality and is more likely to lead to negative consequences than the default AI scenario.
[Editor’s note: this is a fair summary, but the last sentence misses the point: the essay is less about any “negative consequences” and more about the types of sacred attributes that people tend to ascribe to powerful entities like AGI.]
Shifting Baseline Syndrome (Professional Hog Groomer) — The Shifting Baseline Syndrome is a problem that arises when people don't consider the long-term effects of their actions. They only think about the immediate consequences, which can lead to big problems down the road.
[Editor’s note: a decent overview of the lens the author used, but it misses how that lens is applied to understand humans’ changing perceptions of what’s “normal” in the environment.]
Africa’s Cold Rush and the Promise of Refrigeration (The New Yorker) — In Rwanda, there is a school that trains future refrigeration technicians. The goal of the school is to help the country develop an energy-efficient cold chain. The cold chain is a system of food production and distribution that involves refrigeration. The aces team is excited about the potential for Rwanda to lead the way in sustainable refrigeration, but they are also aware of the challenges. One challenge is that people in Rwanda do not trust refrigerated food. Another challenge is that the country does not have a well-developed infrastructure for refrigeration. The aces team is using computer modeling to try to find solutions to these problems.
The Unbearable Lightness of My Pockets (Interconnected) — The article is about the author’s concerns about over-optimizing his pockets and accidentally locking himself out of his own life. He talks about how, in the past, women have been restricted by not having pockets, and how pockets have been a symbol of liberation. He also talks about how, nowadays, people are carrying less and less in their pockets, and how this could lead to problems if they were to lose everything in their pockets.
[Editor’s note: GPT-3 appears to have missed the key themes of this article, namely bootstrapping, circular dependencies, and the complex interrelations that come with optimization. The AI takes the piece too literally and includes an unnecessary (albeit interesting) tangent on the history of pockets.]
How AI-Human Symbiotes May Reinvent Innovation and What the New Centaurs Will Mean for Cities (Technology and Investment) — The research paper examines the impact of emotional intelligence on job satisfaction. The study found that emotional intelligence is a significant predictor of job satisfaction. The study also found that emotional intelligence is a significant predictor of job performance. The study concludes that emotional intelligence is an important factor in determining job satisfaction and job performance.
[Editor’s note: this summary is nonsense, since the paper has nothing to do with emotional intelligence or job satisfaction. Unlike the other Worth Your Times, we only gave GPT-3 the link to the article, not its body text. The AI seemed to get confused; each time we ran this query, we got a different “summary”: first this, then one about the impact of background music on task performance, then one about the potential value of automatic text summarization in online education.]
📚🌲 Book for your shelf
An evergreen book that will help you dip your toes into systems thinking.
This week, we recommend The Architecture of Happiness by Alain de Botton (2006, 280 pages).
In The Architecture of Happiness, Alain de Botton examines the relationship between buildings and our emotions. He believes that our surroundings can have a profound effect on our mood and well-being. de Botton looks at a variety of architectural styles and explores how they can influence our happiness. He also discusses the importance of good design in our homes and workplaces.
He seeks to answer the question, “How can we judge the quality of architecture when we admit that no style is universally the best?” In other words, how can we tell if a building is beautiful or not? By looking at the history and meaning of different architectural styles, de Botton shows us that there is no single answer to this question. Instead, it depends on our individual preferences and needs. At the most basic level, this implies that the form of a building should be related to its function, but it goes beyond that. A building also needs to appeal to the deeper psychological needs of the culture and time it belongs in. It needs to appeal to our memories and to our ideals.
In our personal lives, we can use de Botton’s ideas to create happy and fulfilling home lives by making sure that the design of our homes reflect our individual needs and preferences. In our professional lives, we can use de Botton’s ideas to make sure the design of our workplaces reflects our companies’ values and goals.
[Editor’s note: like the main article, this piece was written through a back-and-forth between a human and GPT-3.]
🕵️♀️📆 Lens of the week
Introducing new ways to see the world and new tools to add to your mental arsenal.
This week’s lens: slum clearance.
When a system with some bad properties defies simple fixes, it can be tempting to tear it down and start over again. However, this slum clearance attitude is problematic, both for architects and for residents of the existing system.
For the architect: you look at what people are already doing and it clearly doesn’t meet your minimum standards. Therefore you decide, for the good of all, to demolish it and replace it with something clean and sensible. Something that will lift up everybody involved.
For the resident: you live in a tolerable situation. It’s not great but you’ve learned how to make it work for you. Someone comes along and destroys everything you have but they say it's “for your own good.” The new thing is terrible. You would much rather go back to the old way, despite its limitations.
You’re both wrong.
The high modernist architect is wrong because they don’t try to preserve the value inherent in the existing arrangements. They don’t experiment to make sure that their clean new solution doesn’t just look better but actually works better. They don’t seek the consent of the current inhabitants. Instead they jump straight to imposing their solution.
The (figurative) slum dweller is wrong because they’re holding on so tightly to what they have that they’re blind to what they could gain. They don’t see that a new system could work for them in ways the old one never did. They don’t see that they could be better off. They just see that they’ll have to start from scratch again, which is scary.
The right way to move forward is to work with the existing system, to understand how it works and how people use it. Then architects and residents can work together to design a new system that is an improvement on the old one. This new system should be designed to be implemented in stages so that it can be tested and refined as it goes. And, importantly, the people who will be using the new system should be allowed to further design for emergence so they can keep evolving the system as needs change.
[Editor’s note: like the main article, this piece was written through a back-and-forth between a human and GPT-3.]