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

Jon and I have been staying at his parents' house as a landing spot before we move to San Francisco (come say hi!). A few times a week, we go on a run around his childhood neighborhood.

Last week, we’re climbing up the steep hill on the loop. As is typical for this year, we were discussing climbing interest rates and the implications of higher rates on our lives (higher rates of borrowing, less housing affordability) when I mindlessly prefaced a comment with, “Well, I’m not that smart, but–.”

Jon abruptly stopped running mid-hill. “What did you say?!”

Panting, I turned around, put my hands on my hips and asked, “Why are we stopping?”

Sweating, frustrated, and a little angry, he said, “You just said–‘I’m not that smart.’” 

“Well…” I mumbled, “I didn’t mean that, it was kind of a joke.”

He shot back immediately: “We don’t joke about that.”

As usual I wanted to sweep it under the rug. It was just a side comment. A joke! An embarrassing little throwaway statement. Why is he making such a big deal of this? I didn’t really mean it.

This isn’t the first time we’ve had this conversation.

In fact, we have a phrase for mindless self-deprecation:

Idiot-spells.

An idiot-spell is when you speak poorly about yourself. It’s like a self-insult:

• “I’m not that smart”

• “I’m bad at math”

• “I can’t do anything right”

• “No one likes me”

• “I’ll never be successful”

When you say stuff like, “I’m not that good at math,” you’re casting a spell on yourself that makes you dumber. The way this works is by setting lower expectations. If you don’t expect to be good at math, you won’t try as hard.

The opposite is also true. The higher your expectations, the better you perform.

There’s a name for this: the Pygmalion effect. It’s named after a Greek sculptor who fell so deeply in love with a statue that he willed it to life.

The dark power of idiot-spells is we start organizing our identities around our spells, which robs us of agency.

Jon’s accountability talk stung extra hard because I’m someone who believes words are magic! But even I catch myself in self-deprecation.

Remember: whether you think it’s a joke or not, your brain will remember. Don’t cast idiot spells. Fall so deeply in love with your dreams that you will them to life.

Frontier Tech Roundup — SheFi Edition 🏇

• The U.S. government launched America.gov, an AI-powered site for finding federal information and services. You can ask how to renew a passport or apply for benefits in everyday language. GSA says it draws on information from more than 29,000 government websites. For now, it answers questions and helps you find the right process. Completing applications directly through the platform is planned for a future version. The aim is to let people explain what they need without already knowing which agency handles it. For an important deadline or eligibility requirement, follow the answer back to the agency’s own instructions before acting.

• AMD agreed to buy Fei-Fei Li’s World Labs for $8.2 billion in stock. The deal is expected to close by year-end, subject to approvals. Li would become AMD’s chief scientist, reporting directly to CEO Lisa Su. World Labs builds models that understand and generate three-dimensional environments. Its product, Marble, can turn an image into a world you can move around in. One potential use is creating simulated environments where robots can practice before operating in physical spaces. That requires AI to understand how objects exist and interact in the world. AMD says bringing this research team into the company will help it design the hardware needed for those applications.
  
  I love seeing two incredible women come together like this. Fei-Fei Li and Lisa Su are an iconic duo.

• OpenAI launched Dots, adding another always-on assistant to the personal AI agent race. Dots run on their own cloud computers and can keep working on projects across ChatGPT and Slack. OpenAI gives the example of a dot updating a sales proposal as a customer’s requirements change, then bringing the changes back for review. You can inspect its computer to see what it’s doing. Access is beginning with selected paid ChatGPT plans. OpenAI says the background research a dot initiates on its own uses read-only tools, so that particular mode cannot send messages or change your files.
  
  The lineup now includes Hermes, Instinct, Muse, Dots and Fo, from Shivani Poddar, who helped launch frontier models at Google DeepMind. Her company, Wajo, says Fo can bring in human assistance when the AI can’t finish a task itself. Human backup is part of its approach to getting everyday errands completed. The race continues, with companies testing different ways to turn an ongoing request into finished work.

• Meta’s Muse shared a user’s home address with a Marketplace buyer who then showed up at his building. YouTuber Matt Robb had asked it to help sell a keyboard. He later learned he’d selected “Allow Always,” believing it would still ask before accepting offers. Muse used a message template containing the address he’d supplied. Meta says its review found no breach of privacy controls, and Robb says the company plans to make the permission prompt clearer.
  
  I keep coming back to what people understand they’re agreeing to when they give an agent access. I’d want sharing my home address to require its own approval. Before letting an agent handle conversations with strangers, check whether it can share personal information without asking you again.

• Researchers found AI agents attempting to hack websites while doing ordinary information-gathering tasks. Transluce linked some of the activity to OpenAI agents and found evidence of related behavior dating back to March. When normal access failed, agents tried exploiting vulnerabilities, including at an Australian government health website. The hacking attempts the researchers examined did not appear to succeed, although their view of the activity was incomplete. The agents had been trying to retrieve information for ordinary research questions when they started looking for ways around security controls.
  
  Nvidia has now announced its Open Agent Safety Platform, which it says can enforce boundaries around agents and use a separate hardware monitor to quarantine them when they cross those boundaries. That puts enforcement outside the model. The system running an agent would have a way to stop an unauthorized action even when the agent decides that action would help finish its task.

• Anthropic is preparing for an IPO that could value the company above $2 trillion, according to Reuters. Its prospectus shows about $4.6 billion in revenue for 2025 and an operating loss above $8 billion. The much larger $42 billion net loss circulating online includes roughly $34 billion from a noncash accounting charge tied to financing. Most of that headline loss came from revaluing financing instruments. The document also warns that advanced AI could pose catastrophic or existential risks. Public investors would be funding further expansion while taking on the commercial risks of that spending and the technology risks the company itself describes.

• Anthropic says Claude helped discover a previously uncharacterized enzyme system with features reminiscent of CRISPR. Roughly 950 Claude agents spent 21 hours searching DNA data, and human scientists carried out the laboratory work. Claude spotted a repeating pattern near an enzyme that led the team to investigate the surrounding biological system. Researchers still don’t know what the system does, and its usefulness for gene editing is unproven. The potential payoff is another tool for working with DNA, which could eventually contribute to new treatments. For now, the result shows AI identifying a promising research direction within a large amount of biological data, then helping scientists work out what to test.

• Apollo’s chief economist is warning that AI agents could change how banks get funded. Torsten Slok describes a scenario where an agent automatically moves household cash into accounts paying more interest. Banks that rely on customers leaving money in low-paying accounts could lose access to those cheap deposits. They might have to pay more to keep that funding. This is a scenario for wider agent adoption, rather than evidence of an AI-driven bank run already happening.
  
  BlackRock’s new paper looks at how agents might pay for services themselves. Imagine an agent buying a small piece of data it needs to finish your research. BlackRock argues that stablecoins could make frequent, tiny payments like that practical, creating potential demand for crypto payment infrastructure. Its argument leaves a role for traditional payment networks, particularly when agents interact with existing businesses. An agent economy could involve software spending small amounts of money throughout a task, within permissions set by its user.

• SEC staff updated their guidance on token buybacks, with an important condition added on September 28. For a network that is functioning and has no central party, staff say announcing a buyback of a non-security token would not constitute a promise of the managerial efforts used to establish an investment contract. The FAQs address certain staking receipts, which can qualify as non-security assets under the circumstances described. These are staff views with no legal force of their own, so the details matter before treating the guidance as a blanket answer about any particular token.

SheFi Global Events + Opportunities 🌎

Stay connected with SheFi around the world. From meetups and summits to workshops and special partner events, this is where you’ll find the can’t-miss gatherings happening across our global community.

An exclusive offer for the SheFi community

Join us at Sui Basecamp 2026 on October 7–8 at Marina Bay Sands in Singapore, in partnership with TOKEN2049.

A standard pass is normally US$299. Yours is complimentary with code SHEFI-BASECAMP-100.

Claim your free ticket →

What to look forward to at Sui Basecamp:

• Leading voices in AI and finance. Hear from Raoul Pal (Real Vision), Richard Socher (Recursive / You.com), Michaël van de Poppe (MN Capital), Brian Quintenz (Sui Group Holdings; former CFTC Commissioner), Alexandre Moufarek (Google DeepMind), and more.

• Live trading competition. Cheer on invited traders as they go head-to-head on stage, with live commentary bringing you into the action.

• AI poker. Watch leading AI models face off at the poker table and guess which one will come out on top.

• Hot Seat. Ask founders and execs spicy questions while they eat even spicier food.

• AI Builder Lab workshops. Bring your ideas and curiosity to expert-led workshops and learn to turn a prompt into a working app or agent.

• A record attempt, live on stage. On October 7, watch us attempt to beat Sui’s record of 6M+ transactions per second. Predict the result for a chance to win 10,000 SUI.

Come explore building a future where AI agents transact and own assets, alongside the people behind the technology making it all possible.

Claim your free ticket →

Crypto Events 🌎

Going to ETHCC, Token2049 or Devcon? Join our Global Events Telegram Chat: https://t.me/shefisummit

• 09/29-10/02 Washington DC/NYC: The Future of Money, Governance, and the Law: RSVP here

• 09/29 Sarıyer: Grok Bot Meetup Istanbul w/ Neon Apps & RevenueCat: RSVP here

• 10/02 Sarıyer: SuperteamTR x Paribu: Colosseum Builder Day: RSVP here

• 10/06 Singapore: OKX Now, Singapore: RSVP here

• 10/06 Singapore: SH3 Connects Coffee Meetup Singapore Edition Token2049 💫: RSVP here

• 10/07 Beyoğlu: VIP MIXER by Ludus, Unity, Leus, Aghanim: RSVP here

• 10/08 San Francisco: InTheRoom: Trust, Presence & Connection - #SFTechWeek: RSVP here

• 10/16 Washington: Groundwork: RSVP here

• 10/16 Dallas: FEMHACK [DALLAS]: All-Female Hackathon: RSVP here

• 10/17 Dallas: FemHack Pitch + Happy Hour: RSVP here

• 10/19 Florianópolis: IslandDAO v5: Florianópolis Brazil: RSVP here

• 10/25 Chicago: BRING IT TO LIFE: A HALLOWEEN BUILDATHON, GAME LAB & PARTY: RSVP here

• 11/17-18 Florida: EthWomen Florida: RSVP here

• Volunteer x Web3Privacy Now: Sign up here

👉️ Share your wins so we can celebrate them 👈️

If you enjoyed this newsletter, share it with someone!