July 13, 2026
Hello and welcome to The Financial Diet's weekly newsletter!
A fun event for our NYC folks (or online!): our dear friend Erin Lowry (co-host of Asked & Answered and author of the Broke Millennial series) is hosting the second installment of her live show on July 18th at 4pm! Funny Money is part comedy show, part actionable financial advice. We went with a group of TFDers back in May, and it was so much fun! This month’s theme celebrates living a fun life while paying off debt effectively. The show culminates with one brave New Yorker asking our panel (Erin, Melissa Jean-Baptiste, and Alyce Chan) for help getting her financial life together. New Yorkers can come in person to Caveat on the Lower East Side, but anyone can livestream or watch it on demand! Get $10 off in-person tickets with the special code TFD10 or $3 off the livestream with LOL3. Hope to see some of you there!!
As always, make sure to read through the entire newsletter so you don't miss anything! Thank you for being a part of this special little corner of our community!
❤ TFD
By Zoe Lu
What We Should Worry About When We Worry About AI
If you haven’t been living under a rock in the past four years, you likely had a few nerves fried by AI. Data centers are destroying our planet. AI slop is rotting our brain. AI psychosis is fueling delusions. AI is trapping us in the “permanent underclass.” AI will be the end of humanity. Unsurprisingly, when Chelsea shared her opinions on AI, the audience did not hold back in their critiques and concerns.
I spent the past decade studying, researching, and working in the AI ecosystem, from infrastructure, training algorithms, and software tools to ethics and governance. I’m not here to convince you whether AI is good or bad, but rather to help you ground your stance in facts. We will start with an overview of how AI has evolved before myth-busting each part of the AI technology. I hope this explainer helps you separate the signal from the noise and walk away with a few concrete actions you can take for a better future.
A Brief History of AI
The concept of "artificial intelligence” and its implementation have evolved for 70+ years. We are demystifying the technical terms here, so you can be more discerning about buzzwords in news and advertisements.
While stories about thinking automatons began in antiquity, the scientific pursuit started in the 1940s with breakthroughs in neuroscience, electric networks, information theory, and theory of computation. The term “artificial intelligence” was coined in 1956 during a summer workshop at Dartmouth (fun fact: I did AI research under a student of the workshop’s co-organizer). The researchers at the workshop conjectured “every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it.” This definition is what people usually mean when talking about "artificial general intelligence” (AGI).
Early AI research focused on logic-based systems; however, these systems only worked in narrow settings and were hard to generalize. The field experienced a few “AI winters” with reduced funding and enthusiasm between 1970 and 2000. The revival came with the paradigm shift from rule-based to pattern-based learning (i.e., “machine learning”), using massive amounts of data and computing power. Machine learning is built on multiple layers of neural networks, where the nodes, like neurons, hold an internal representation of a piece of information (i.e., tokens, which can be a chunk of characters or a patch of image), and the edges, like synapses, let information flow from one node to another. Training (or “learning”) involves adjusting the weights for the nodes and edges to improve the accuracy of the next token prediction. Once trained, the models are judged by evaluations and get fine-tuned / adversarially tested before being released.
Prior to 2017, neural nets processed tokens sequentially. The invention of transformers allows each node to attend to different kinds of relationships in different layers of the neural nets, capturing much richer context. Transformers are at the core of all large language models (LLMs) we are familiar with today (e.g., ChatGPT, Claude) and have since been widely applied to other use cases such as gameplay, image recognition, and translation.
Myth Busting and Suggested Actions
For each aspect of AI technology below, I’ll examine the common beliefs (italicized) in greater detail and consider constructive actions to mitigate concerns. Given that midterm elections are coming up in the U.S., we specifically focus on policy and regulatory opportunities you can advocate and vote for.
AI Infrastructure
Only AI uses data centers. → Almost all digital consumer products (e.g., Google, Instagram, TikTok, Zoom) need data storage and compute, so data centers are not new. But the current scale of growth is unprecedented. McKinsey estimated that AI data centers accounted for about 10% of total data center demand in 2023, projected to grow to 70% by 2030.
Water is the primary method for cooling servers, but it’s not the only one. Older data centers used air cooling instead of liquid cooling, which is noisier and less energy-efficient. Most data centers today use direct-to-chip cooling with water flowing through the hot components, and researchers are experimenting with non-water liquids for even higher efficiency. The resulting warm liquid can be reused for heating in colder climates.
Other methods have their own trade-offs: immersion cooling (submerging servers in non-conductive liquid) uses zero water and is highly efficient, but harder to maintain; evaporative cooling works well in dry climates, but uses huge amounts of water; refrigerant-based cooling is energy-intensive and can pollute the environment. Cold climates, meanwhile, offer natural cooling that saves energy overall.
In short, data centers do use a lot of water and can cause pollution, but it's worth considering the baseline usage elsewhere and the pace of improvements.
Data centers must use a lot of electricity, which negatively impacts the local communities. → Data centers consumed 448 terawatt-hours of electricity globally in 2025, which is about three times the total energy use of New York State and 1.5% of global electricity consumption. When thinking about data center energy use, we need to differentiate the power usage effectiveness (PUE), total power, and the power source. AI companies are financially incentivized to maximize PUE, which mostly means making the cooling mechanism as efficient as possible. AI companies are also incentivized to scale up total power usage, as long as they can acquire more and newer GPUs. AI companies are agnostic of power sources. It’s possible for AI companies to pay for their fair share of power needs and bring in their own clean energy sources, rather than passing the negative externalities onto local communities.
Actions you can take:
While it’s tempting to protest away data centers in your neighborhood, this approach usually shifts the buildout elsewhere, likely to places with more marginal voices (e.g., black communities, the Global South). Demand holistic, systematic, and institutional changes.
Advocate and vote for policies/regulations that make data centers use clean power, cooperate with local communities’ needs, and adopt energy- and water-efficient cooling.
If you’re concerned about water conservation, on a personal level, reducing meat and dairy consumption saves a lot more water than not using LLMs. Similarly, fighting animal agriculture is more effective than banning data centers at the institutional level.
AI Training
LLMs violate my data privacy. → You can run open-source models locally. This approach requires some technical setup but keeps your data completely private. Otherwise, you can still turn off training in the web UIs (ChatGPT example). It’s also good practice to avoid uploading sensitive data and delete sensitive information periodically from an LLMs’ memory (ChatGPT example).
AI training is biased against vulnerable communities. → Unfortunately, this is true, not because AI companies are intentionally discriminating against certain groups, but because the existing training data is not maximally diverse, inclusive, and equitable. For instance, some LLMs are trained on Reddit posts. Since young males from English-speaking countries are the dominant demographic group on this forum, their values are over-represented in this training data.
AI companies can and should instill better values or add more guardrails to LLMs. → AI alignment is the type of work to ensure AI acts in accordance with human values, but it’s notoriously hard due to the sheer size of neural nets used in training. Even frontier AI companies don’t understand exactly what is happening within neural nets, so they can’t guarantee that a specific value will be honored or that a guardrail will be enforced.
Actions you can take:
If you or your loved ones choose to use LLMs, learn about how to protect your privacy.
Advocate and vote for policies/regulations that make training data public, inclusive, and equitable.
Advocate and vote for policies/regulations that require AI companies to invest in alignment and interpretability, even at the expense of development and growth.
AI Tools
LLMs are sycophantic and addictive. → They used to be more sycophantic, but this is improving. You can personalize how LLMs speak to you (e.g., Claude) to guard against sycophancy.
LLMs hallucinate. → This is true but also improving. Personalized instructions can encourage LLMs to admit when they don’t know something to reduce hallucinations.
We can’t distinguish between AI slop and human work. → Detection algorithms exist, but they're struggling to keep pace with generative AI. Creators and institutions need to disclose AI use to ensure that end users can distinguish AI outputs from human work.
Remember AI encompasses far more than chatbots and image generation: agents are rapidly improving at coding, machine learning powers backend systems like Google Search, and beyond civilian use, AI is deployed for everything from surveillance and military applications to disease detection and drug synthesis.
Actions you can take:
Avoid AI tools if you prefer, but treat those who use them with compassion (see Dr. Fatima’s video on “how to (anti) AI better”). If you or your loved ones choose to use them, beware of their limitations.
Advocate and vote for policies/regulations that require creators and institutions to disclose AI usage.
Advocate and vote for policies/regulations that limit AI use for negative applications.
The economic, social, and ethical questions that AI raises are beyond this explainer, but these conversations will be more productive with a clear picture of the technology itself. Whatever side of those debates you land on, you're now more equipped to engage with them honestly.
Currently an advocate for animal welfare, AI safety, and everything in between, Zoe previously worked as an R&D Technical Product Manager in quantitative trading, scaling high-performance computing infrastructure and building internal tools. She double-majored in Computer Science and Mechanical Engineering and minored in Philosophy at MIT.
If you want to do more for the environment and the animals, Zoe encourages you to support animal advocacy organizations in any capacity you’re comfortable with. She personally donates to and supports Faunalytics (advocacy research), The Humane League (advocacy for farmed animals), and the Good Food Institute (advocacy for a more sustainable food system).
The TFD Toolkit
Think of these as your financial basics: essentials for getting a handle on your finances, no matter where you’re starting from. (Some of these will be affiliate links, which help us keep the lights on — but fear not! We will not endorse a product we wouldn’t personally recommend.)
A budgeting app that makes it easier for you to keep track of your spending, not more annoying. We love Monarch’s user-friendly dashboard and rule-setting functionality that allows you to minimize how much time you need to check in on your budget.
A HYSA (high-yield savings account) for storing short-term savings, such as your emergency fund or emergency savings. Here’s one we recommend, which has a much higher interest rate than you’d likely get keeping extra funds in your checking account. (affiliate, terms apply)
If you’re just starting out investing, a robo-advisor like Betterment can make it painless to start actively (but also very, very passively) work towards your long-term money goals.
If you’ve moved beyond the basics and are looking for expert insights for your ongoing financial decisions, now might be a great time to explore a a fee-only advisory service, like our partners at Advisor.com.
The Society at TFD
Become a Society member!
The Society at TFD is our members-only community with access available on both YouTube and Patreon. Joining The Society is the best way to directly support TFD! The Society offers the exact same things on both platforms, so choose whichever one you prefer!
The Society at TFD Lite: $2.99/month
Monthly office hours with Chelsea to chat and get your personal questions answered
Access to our monthly book club hosted by TFD Creative Director, Holly
Illustrated tech backgrounds every month
Access to Society Discord
The Society at TFD: $4.99/month — includes everything in the $2.99 tier plus:
Monthly ad-free videos
Exclusive members-only events and workshops
Complete post archive (including exclusive members-only videos of Chelsea ranting on different topics)
The Society Premium: $9.99/month — includes everything from our $2.99 & $4.99 tiers, plus:
Weekly newsletter from Chelsea
Monthly multi-page workbook/guidebook on a different topic each month
Members-only podcast episodes
