I Betrayed Myself

🎧️ Prefer to listen? I recorded this week’s newsletter for you.

Last Tuesday night, I went to a fireside chat with LuLu Cheng Meservy, the comms guru who previously led comms at Blizzard and now runs her own agency, Rostra. I was sitting on a bench, shoulder to shoulder with about 200 people, intently listening to her about what it means to “go direct.”

“When I talk to founders, they tell me their biggest regret is that they were too scared to say what they wanted to say. They were so worried about turning people off, they didn’t speak their truth.” she said.

By not speaking their truth, they’d hoped to appeal to a lot of people. Instead, their generic message appealed to no one.

I have the same problem! I’ll sit down to write this newsletter, and worry about what my audience will think–what if the topic is repetitive, or my conviction is too intense or someone unsubscribes if I write about motherhood again.

I think about what my audience wants before I think about what I want. It’s the people-pleasing Midwesterner in me that I can’t quite shake off.

Then I look at a Bjork quote I have saved: “If I try to please five people, I won’t please anyone. But if I only please myself, I might please ten people.”

I am reminded that worrying about everyone else is betraying myself. And my writing ends up sounding like a Hallmark card.

“You have inherent authority to tell your story. You do not need to get published in the New York Times or get a stamp of approval from an institution. You just need to tell it directly, and tell it in a way that inspires others to help you tell it.”

Ultimately you have to be selfish about your life in order to get someone else to care about it.

If you care about something, with passion, energy, and perspective–other people can tell. If you write a book report in order to fulfill an assignment or an obligation, people can also tell.

Be selfish and do what you want to do.

If you know someone who needs a reminder to be selfish, send them this newsletter.

Frontier Tech Roundup — SheFi Edition 🏇

• Democrats retook control of the U.S. Senate in the midterm elections, and the growing opposition to AI data centers could reshape America's AI infrastructure boom.

  The House was still being counted Tuesday night, but Democrats have secured the Senate. One issue drawing increasing political attention is the enormous demand for electricity from AI data centers.

  According to Data Center Watch, 32.4% of Democratic congressional candidates it tracked supported some form of moratorium on new data centers, compared with 3.4% of Republicans. Candidates from both parties have also called for stronger regulation.

  AI companies are investing hundreds of billions of dollars in infrastructure, but communities are increasingly concerned about electricity bills and environmental impacts. Congress could push for tougher rules, while state and local governments control many of the permits.

  The pace of AI development increasingly depends on how quickly the country can build the infrastructure to support it.

• Major AI companies are beginning to watermark the text their models generate, and xAI's Grok is the latest to join them.

  Starting November 7, Grok will watermark text from supported API models by default. Google and Anthropic already watermark text from supported models, while OpenAI is rolling out the technology for ChatGPT in Europe but hasn't enabled it by default elsewhere.

  A watermark is a statistical pattern embedded in the words an AI chooses. You can't see it, and it can survive copying and pasting the text into another document.

  Detection is imperfect, and access to official detection tools is currently restricted. A watermark also can't tell you how much of the thinking or writing came from a human.

  I think most people using AI to write have no idea that the words themselves can carry a detectable signature. This could have significant consequences for how employers, schools and publishers evaluate AI-assisted work.

• Hackers stole $6.3 billion from Binance in the largest crypto hack on record, using fake job interviews and a compromised internal AI tool.

  On October 29, attackers drained billions from Binance's hot wallets after spending months infiltrating the company's systems.

  According to Binance's investigation, one engineer downloaded malicious software disguised as a coding assignment during a fake job interview. The attackers then used compromised computers and internal AI tooling to reach the systems managing Binance's wallets.

  The FBI attributed the theft to North Korea, which denies involvement. Binance says approximately $2.85 billion has been frozen or intercepted and that it has replenished 98% of the affected token assets. The company has committed to covering customer losses.

  A compromised employee laptop eventually gave attackers access to billions of dollars in assets. It's a good reminder to review where you're storing your crypto and how much security you're delegating to an exchange.

• Anthropic says AI may soon dramatically accelerate its own development, and Claude already writes more than 90% of the code contributed to Anthropic's codebase.

  In its updated IPO filing from November 2, Anthropic disclosed that its internal estimates for future models suggest it may imminently cross its own safety threshold for AI dramatically speeding up AI research.

  The company hasn't designated any current model as having reached that threshold. It also reported that more than 90% of code contributions in September were written by Claude, and that its engineers are contributing far more code than they were two years ago.

  Anthropic's own engineers are increasingly directing and reviewing work done by AI. The company is now telling investors that AI could accelerate the development of its successors, while acknowledging that its ability to measure that acceleration remains limited.

• Hackers remotely raised the temperature of supermarket refrigerators and freezers across Europe, destroying food and disrupting stores.

  On November 2, refrigeration systems malfunctioned in more than 100 supermarkets across several European countries, including Belgium, Sweden and Luxembourg.

  In some cases, temperatures were reportedly raised to 50 degrees Celsius, or 122 degrees Fahrenheit. Stores had to remove spoiled products, while technicians worked to restore the equipment.

  The attacks resemble an earlier incident affecting hundreds of French supermarkets, although investigators haven't established whether the same people were responsible.

  Refrigeration systems are increasingly connected to networks so operators can monitor and adjust them remotely. That convenience also creates an entry point for attackers.

  Software security is becoming a physical safety issue when compromised systems control the equipment people depend on every day.

• The best AI model scored just 12% on a new benchmark measuring difficult, high-value professional work.

  Research firm SemiAnalysis launched GDPx, which evaluates AI systems on 206 complex tasks across 11 fields, including chip design, biology and finance.

  Claude Opus 5.5 led with 12%, followed by Google's Gemini 4 Carbon at 10.9% and OpenAI's GPT-6 Astra at 9.6%.

  The benchmark deliberately focuses on exceptionally challenging tasks, and some scores were affected by models refusing certain cybersecurity and biological assignments. So 12% should not be treated as a universal AI success rate.

  But the results show how much complex professional work still defeats even the best models.

  This is useful context for anyone building AI agents. A model can produce something that looks complete while missing requirements that an experienced professional would immediately recognize.

• More than 80% of Roblox's most popular games were built with help from AI tools last month.

  In its third-quarter shareholder letter, Roblox reported that AI usage among its top creators has tripled since the second quarter. Two games created almost entirely with AI even broke into the platform's top 20.

  Roblox now lets creators connect external AI models directly to its development environment, making it easier to turn an idea into a playable game.

  Its new prompt-to-game creation tool has already been used by more than 500,000 people in an early release across four countries.

  These are company-reported figures, but the adoption is striking.

  I love seeing technology make it possible for more people, including young creators, to build something they previously wouldn't have had the technical skills to produce.

• A study of more than 9,000 workers found that people using AI had workdays up to 90 minutes longer than those who weren't.

  ActivTrak analyzed computer activity from 9,140 workers across 467 organizations and found that heavy AI users switched between applications more frequently and spent less time in uninterrupted focus.

  The study is observational, so it doesn't prove that AI caused the longer workdays. People with demanding jobs may be more likely to adopt the tools in the first place.

  But I've been teaching people how to use AI to save time at work, and I think we need to pay attention to what happens to those savings.

  If AI makes us more productive, we have to decide how to use the extra capacity. Otherwise, it's easy to fill every hour we've saved with another task.

• The FDA removed certain prominent safety warnings from three more menopause hormone therapy patches, potentially changing how women access treatment.

  On October 30, the agency approved updated labeling for Minivelle, Vivelle and Vivelle-Dot, all estradiol patches used to treat menopausal symptoms.

  The changes remove boxed-warning language related to cardiovascular disease, breast cancer and probable dementia.

  Relevant risks remain elsewhere in the prescribing information, including the boxed warning about endometrial cancer for systemic estrogen-only products.

  For years, prominent warnings have influenced how women and their doctors think about menopause treatment.

  Updating the labels to reflect the agency's current assessment of evidence could change those conversations. I'm interested in whether this ultimately helps more women get appropriate, evidence-based care.

Share The Newsletter, Unlock Frontier Is Feminine Merch (and a Small-Group Session with Me) ✨

Over 80% of SheFi’s growth has always come from one place: you telling a friend, a coworker, or your group chat about what you’re learning here.

Starting this week, we’re officially turning on the SheFi Newsletter Referral Program so we can reward the readers who help us grow.

Every time someone subscribes to the SheFi Newsletter using your personal referral link below, it counts toward 5 reward milestones:

• 10 referrals — SheFi Insider Badge + NFT: A digital SheFi Insider badge plus a commemorative NFT marking you as an early referrer.

• 25 referrals — Newsletter Shoutout: A shoutout to you (and your work or project) in an upcoming edition of the SheFi Newsletter.

• 50 referrals — Frontier Is Feminine Hat: The embroidered black SheFi “The Frontier Is Feminine” baseball cap from the SheFi Shop, shipped to your door.

• 100 referrals — Frontier Is Feminine Shirt: The SheFi “The Frontier Is Feminine” tee, shipped to you.

• 250 referrals — Small Group Session with Maggie: A private small-group session with me and other top SheFi insiders.

Or copy and paste your personal link anywhere:
https://newsletter.shefi.org/subscribe?ref=PLACEHOLDER