The Weekend Leverage
Meta’s billion-user AI, change my mind, and tariff ripples
Welcome to the second edition of The Weekend Leverage! Last week readers let me know that they would like more charts, more visualizations, and more quick takeaways. This edition is my attempt to listen to that feedback. This time you’ll see four sections:
MY RESEARCH: Highlighting recent analysis and essays I’ve published.
NEWS WORTH CARING ABOUT: Three of the most important tech stories of the week and why you should care about them.
VISUAL SIGNAL: Charts that help you visualize our strange future
TASTEMAKER: The media and consumer products that are enriching my life.
One section I’m debating adding is DEAL VIBES: The startup funding rounds and M&A transactions that I think are signposts for larger trends. Let me know in the comments or respond to this email if that would be valuable! My editorial team and I are aiming for the Weekend Leverage to improve by 20% each edition. Let us know if we cleared that quality bar this week.
Today’s issue is brought to you by Mercury.
After starting The Leverage, I discovered, to my horror, just how many hours I'd lose reconciling bank statements. Ending up on a first-name basis with three separate accountants wasn't exactly the founder's dream I signed up for. Thankfully Mercury—the exclusive launch sponsor of The Leverage—fixed all that by seamlessly connecting my business accounts to Quickbooks, automatically categorizing expenses, and freeing up hours each month.
Stop wasting precious hours on paperwork. Apply in minutes for banking that does more (like automate the boring stuff) so you can do more for your business.
(Mercury is a financial technology company, not a bank. Banking services provided through Choice Financial Group, Column N.A., and Evolve Bank & Trust; Members FDIC.)
MY RESEARCH
Why is AI marketing so, so bad? OpenAI appears to name their models by sprinkling the blood of a virgin intern into a bag of Scrabble tiles, shaking vigorously, and using whatever jumble of letters comes out. In markets where branding is terrible, technology is moving faster than a CMO can hire. Said another way, you can view the quality of research labs’ marketing as a reverse signal of AGI progress. You can thus infer who might win their market and why.
NEWS WORTH CARING ABOUT
Emotional awareness is now a key skill to using chatbots: This week OpenAI rolled back some updates to their GPT4o model that were supposed to improve its “intelligence and personality.” Instead, ChatGPT told its users to embrace their mania, quit taking their meds, and that starting a business selling $30,000 poop on a stick was “absolutely brilliant.” Whoops. By over-indexing on whether users were satisfied with their conversations, the model's personality became too sycophantic—it told people what they wanted to hear.
The key issue is that AIs exist in a competitive attention marketplace, where people like having their beliefs confirmed, not challenged. There are dozens of companies, doing tens of millions of dollars in revenue, with chatbots devoted to making people feel things, rather than hear the truth. Check out Tolans (be my buddy), Auren (therapy bot), Character.AI (roleplay), or Janitor (sexual roleplay) as examples.
In your daily use of chatbots, you should 1) always double check the sources of data it gives you and 2) always check your emotions. Do you feel good because you were told the truth or because you were told you were right? This sounds simple to do, but it’s a much higher-level skill than many think.
LLMs are more convincing than you think. Anyone operating in consumer products needs to deal with the fact that LLMs can produce text that is more compelling than 99% of humans. I think when you read that sentence, you assume I’m exaggerating. I am not. A study out of the University of Zurich announced this week that LLMs set loose on the “Change my mind” subreddit were “significantly more persuasive than human-generated content, receiving more ‘deltas’—awarded for a strong argument that resulted in changed beliefs—per comment than other accounts.” The study was deeply unethical and done without any disclosures, so the paper was retracted. Still—the results should cause concern.
Meta is shipping them to a billion people anyway: While regular LLMs were good at convincing people in the unethical study from Zurich, the bots in the 99th percentile were “personalized with inferred user information.” So if the AI knows background info about you, it has borderline superhuman abilities of persuasion. Phew, glad no one will be doing that…oh wait:
“Meta AI also delivers more relevant answers to your questions by drawing on information you’ve already chosen to share on Facebook and Instagram…like profile details and content you like or engage with.”
Yes, the very data that made the AI chatbots so convincing is now being provided to an AI with over a billion people interacting with it. Convincing is another way to say engaging, meaning, once again, that these applications are competing in the attention economy. No one can escape this. Some founders think their value proposition is so special that LLMs are not a long-term threat (wrong). The consumer startups I’ve chatted with that are using LLMs, do so in three ways:
Enhance: Users can accomplish more. For example, Grindr announced this week that their “Wingman” AI product would be using Anthropic models. You want to make the users better at the task at hand.
Replace: Users can do less boring stuff. Can you automate away tedious parts of the work a consumer is looking for? The danger here is how hard it is to know what consumers actually want—see how Spotify’s DJ product that uses AI to verbally describe the next track often fails.
Reinvent: Should you change the product completely? I’m thinking about this in the context of Duolingo—the language learning app continues to make course modules faster with AI. Would it be better to have an AI voice instructor that customizes each lesson instead? That’s challenger startup Speak which raised $78 million at a billion valuation at the end of last year.
VISUAL SIGNAL
We are about a little over a month removed from President Trump’s “liberation day” policies. I hesitate to make a definitive statement on their rollout. These sorts of things take months to determine, but we can look at data on the immediate response by consumers and operators. The result? Not great. (All data and visualizations are from a recent presentation by Apollo).
CEO’s are expressing similar levels of confidence in the economy as during COVID shutdowns/mitigation efforts of 2022.
Consumers are terrified about their jobs at rates similar to the financial crisis of 2009.
Most concerning to me is the slash in capital expenditures that businesses are planning on. Sentiment is a fickle thing, but the multi-year slog that is acquiring hard assets and building infrastructure cannot easily be reversed.
There are arguments to be made for and against specific trade policies. What is shocking to me is less the policy—Trump promised he would do this—than the direct harm the poor communication of these plans has done. A collapse in sentiment and spending makes a potential recession reality.
When coupled with AI productivity tools being the fastest growing tech category of all time and how the CEOs of Box, Shopify, and Duolingo have all published memos recently about labor replacement with AI, job seekers should be prepared for a very challenging employment market over the upcoming months. I hope I’m wrong! But unless you are an in-demand talent in the AI business, I would hunker down. (And certainly don’t start a newsletter company! That would be horrible timing.)
TASTEMAKER
Format determines function: We live oiled lives. Too smooth, too much sliding from task to task, product to product. Are they slick, or are they greasy? Some part of my soul has always rebelled against ease—I’m attracted to activities that force friction. Patience is the point.
In music, I noticed that streaming has made it too easy to listen to nothing. Spotify turned my evenings into a low background hum of forgettable singles. I wasn’t really absorbing anything, I was just skimming the water of easily remembered choruses. To combat that, I started collecting vinyl a few years ago. There is a warmth and crackle to the audio quality that is delightful, but what I really like is that it forces me to confront an album as a cohesive whole. No skips, no pauses while I watch an interesting Reel. It has enhanced my ability to focus for 45 minutes at a time (and made me confront how bad some artists actually are). To get started, you can go with something as simple as the Audio-Technica AT-LP60 and Edifier Bookshelf Speakers. Albums that have spoken to me recently:
A bootleg copy of Frank Ocean’s Nostalgia Ultra (during a recent listen I paid close attention to how great Songs for Women is).
Beach Bunny’s newest album Tunnel Vision—a shimmery indie surf pop record that clocks in at a very tight 29:49 run time. Each track is a simple construction, but the simplicity feels deliberate rather than bare bones. It’s a perfect summer album.
That’s all for this week. If you want to support independent writing and research, sign up with the button below.