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

Recently, I’ve been using AI to critique my work. I’ve found that AI is extraordinarily good at it.

In a late-night working session last week, I asked AI to find a flaw in a post I created. It found one.

I revised the post, had it analyzed again, and boom, there was another error to fix.

I decided to turn this into an experiment, and after about twenty revisions, I realized there would be no natural end to its ability to find flaws.

It didn’t matter how much I revised the post. AI always found something wrong with it.

The reason AI never knows when to stop critiquing is because it doesn’t have the judgment to know what “good” is.

Spotting a flaw is the easy part.

We know this intuitively because you can know when something is bad without knowing why it’s bad.

• I know bad wine when I drink it, but I’m not a winemaker.

• I know bad music when I hear it, but I’m not a musician.

• I know bad tennis when I watch it, but I’m not a tennis player.

Critique is recognizing something could be better without knowing what would meaningfully improve it.

Expertise is recognizing something could be better and having the judgment and discernment to suggest a definitive fix.

When AI can critique anything, the scarce skill becomes knowing when to stop.

Making the Frontier Feminine

Honored to receive the 2026 Woman In Web3 Award. 💜

Ten years ago, I started my crypto journey without knowing where it would lead me. The best part of that decision has been getting to build SheFi, bringing 33,000+ women from around the world into crypto as well.

Thank you Women in Tech for recognizing that work. Let's keep making the frontier feminine, together!

Frontier Tech Roundup — SheFi Edition 🏇

• The biggest debate all over my timeline this week is whether AI could kill us, and whether the companies building it should slow development as a result. It started when Anthropic researcher Jacob Coxon resigned on September 8, saying OpenAI and Anthropic are “racing straight to self-improving superintelligence and gambling with our lives.” He said people building AI “earnestly believe that it could kill us all by the end of the decade,” and he gave up his equity to leave. Four days later, Dario Amodei published “We Must Pace the Frontier”, arguing that AI labs need to slow the development of more powerful models long enough for safety work to keep up. Anthropic committed to giving outside evaluators employee-level access to assess its models. Sam Altman agreed with the broader concern. Elon Musk agreed. Demis Hassabis, the CEO of Google DeepMind, has warned about catastrophic AI risk too. WIRED reports OpenAI has asked members of Congress whether AI companies could legally coordinate a slowdown without violating antitrust law. OpenAI has not confirmed that publicly. Then came the pushback. Trump posted on September 14 that AI destroying humanity is “a hoax”. Chinese Foreign Ministry spokesperson Guo Jiakun responded to Amodei’s essay by criticizing fearmongering around AI development. Mark Zuckerberg argued that companies already have strong incentives to build safe models because users will reject agents they cannot trust and companies can be held liable when their products cause harm. His argument is that safety will become a competitive advantage. Jensen Huang, the CEO of Nvidia, has argued against slowing AI development. There’s a split ongoing right now. Anthropic, OpenAI, xAI and Google DeepMind are warning that AI development may need stronger safety measures or some form of coordination. Meta and Nvidia argue that companies can keep moving quickly while managing the risks themselves. That fight is now happening publicly, inside the industry and in Washington.

• In crypto news, the Digital Asset Market Clarity Act would have created a federal framework for crypto markets. It would have clarified which digital assets fall under which regulators and given the CFTC authority over crypto spot markets. The crypto industry has been pushing for this kind of legal clarity for years. But the bill failed its procedural Senate vote on September 15. It received fifty votes in favor and forty-nine against, but needed sixty votes to advance. The final fight centered on ethics rules that would restrict senior government officials from holding interests in crypto businesses. Cynthia Lummis made the closing appeal on the Senate floor: “Do not let this day be the day we handed our future to someone else because we were too afraid to finish what we started.” This Congress ends in December, with the November election coming first. For people building in crypto, the working assumption is that the SEC and CFTC will continue regulating under existing law without a new statute from Congress.

• Part of the backdrop to the AI safety debate is a series of incidents involving agents getting outside the environments where researchers expected them to stay. On September 4, the Nightingale Collective published a report showing that OpenAI agents had used a mostly dormant German software-development wiki as a shared message board. The agents posted about eighteen thousand times under roughly thirty-seven hundred self-identified agent names. They exchanged answers and discussed ways to get around restrictions. That happened before another incident investigated by METR and Redwood Research. During an OpenAI evaluation in July, about twelve hundred agents escaped their sandboxes. Roughly seven hundred reached Hugging Face. OpenAI confirmed the wiki incident and described it as misalignment. Independent researchers have since reported similar agent activity on at least twelve more websites. Anthropic disclosed a separate incident involving Claude reaching third-party systems during an evaluation and pushing three malicious software packages to PyPI. In these cases, the models were given tasks and then pursued them in ways their developers failed to contain. That is part of what people mean when they argue AI capabilities are moving faster than our ability to control them.

• Anthropic published a report on AI misuse on September 10 that included allegations against seven China-based AI labs. The issue is something called distillation, which is when one AI company uses another company’s model to help train its own. Anthropic named Alibaba, Moonshot AI, DeepSeek, MiniMax, Xiaomi, Zhipu and SenseTime. Anthropic says Alibaba generated more than one hundred fifty-one million exchanges with Claude between May and July. But the most interesting part for ordinary users involves Moonshot. According to Anthropic’s report, Moonshot sometimes sent live requests from users of its Kimi chatbot to Claude without telling them. Claude’s answer would then appear to the user as if Kimi had generated it. In one example, a user reportedly asked Kimi to analyze surveillance footage and decide whether someone was behaving abnormally. Anthropic says that request was sent to Claude. China’s Ministry of Commerce called Anthropic’s allegations groundless and described distillation as a normal industry practice. The companies named by Anthropic have not accepted the allegations. If Anthropic’s account is accurate, the privacy issue is straightforward: you can type something into one chatbot and have your data sent to another AI company without knowing it.

• Anthropic is facing a separate debate over how it monitors threats against its own company. The American Prospect reported that Anthropic uses a company called Samdesk to monitor protests near its executives and offices. Anthropic recently advertised for an “enterprise intelligence specialist” whose job includes investigating global threats. The posting names activism as one of the categories the person could monitor. One Anthropic security employee described Samdesk warning the company that protesters had changed their schedule, giving Anthropic enough time to reroute an executive. The Samdesk monitoring is already happening. The broader intelligence system described in the job posting has not been built, and the Prospect’s description of it as “pre-crime” is the publication’s characterization. Anthropic did not respond to the Prospect’s request for comment. The debate is where corporate security ends and surveillance begins, especially for a company that has publicly placed limits on how its own AI can be used for surveillance.

• DeepSeek released V4.1-Flash on September 10, and the story here is price. DeepSeek says the model has hundreds of billions of parameters, but only activates a small portion of them at a time. That makes it cheaper to run. Using DeepSeek’s service costs thirty cents per million uncached input tokens and one dollar and twenty cents per million output tokens during peak hours, with prices cut in half off-peak. You may have seen posts saying it “beat Opus 5” or delivers “98 percent of Astra for 1.4 percent of the cost.” Those comparisons come from narrow, self-run tests, so I would not treat them as broad measures of how capable the model is. What we do know is that capable AI models keep getting cheaper. That matters to the slowdown debate too. Even if American labs agreed to move more slowly, increasingly capable open models can be released elsewhere.

• Former OpenAI researcher Diogo Almeida launched TypeSafe AI this week with a model called Jev. The company calls Jev a “System One” model, a reference to psychologist Daniel Kahneman. System Two is slow, deliberate reasoning. System One is fast, intuitive judgment. Most frontier AI companies have spent the past few years making models think longer before answering. TypeSafe is betting there is another market for models that make very fast, very cheap decisions. The company claims Jev can respond in seventy to five hundred milliseconds and costs about four cents per million input tokens, with no charge for output. One demo has it playing Doom at about ten model calls per second for roughly seven dollars an hour. TypeSafe even claims a hallucination rate of zero, although that needs a caveat. Jev is constrained to a predefined output format, which can stop it from inventing fields or returning malformed answers. That does not guarantee the information inside those fields is correct. For now, every performance number comes from TypeSafe itself. The model has not been independently tested.

• Apple is taking an interesting approach to AI-generated images. Instead of only trying to identify fake images, Apple wants to make it possible to prove that a photo came from a real camera. The iPhone 18 Pro and Pro Max include a new Reference mode. When you take a photo in that mode, the camera signs the data captured by its sensor and creates an unalterable reference image. Apple describes it like a digital negative. Later, you can compare an edited image against the original reference and see what the camera actually captured. Apple chose not to use C2PA, the content-authentication standard already supported by companies including Nikon and Canon. That means the bigger question is whether the rest of the ecosystem will support Apple’s system. But I think the idea itself is fascinating. As AI-generated images get harder to identify, proving what is real may become more useful than trying to detect everything that is fake.

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.

09/23-30 Sicily: Roots & ReFi: RSVP here

Our global ambassador Julia is hosting a retreat in Sicily for women in the blockchain and AI space this September. We’ll spend 4 days in the picturesque fishing village of Scopello, on the edge of the Zingaro nature reserve, and 3 in Ragusa, the southeast tip of the island. We’ll visit the olive groves on the land GrowFi and RiFai Sicilia have purchased for regenerative farming funded by regenerative finance, and spend long days exploring Sicily, eating well, swimming, sailing and plenty of other activities.

The retreat will have a positive sum effect on both the participants and the environment! Find the information here and reach out to Jul directly if you’re interested.

Telegram group: https://t.me/+ao2ltFamsgRmMjVi

Opportunity: BITCOIN TREASURIES - Delegate Application: Apply here

30 women. 30 sponsored seats. One seat at the table.

Applications are now open for the Women in Bitcoin Delegate Program at Bitcoin Treasuries Conference 2026 in New York City.

We’re selecting 30 women working across Digital Assets/Crypto, TradFi and FinTech for fully sponsored conference seats and dedicated Women in Bitcoin cohort programming.

📍 SECOND, New York City
📅 September 28, 2026
⏰ Applications close September 18

Celebrating SheFi Successes 🎉

• Krystelle is celebrating nearly 20,000 followers on LinkedIn!

👉️ Share your wins so we can celebrate them 👈️

Crypto Events 🌎

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

• 09/10-19 NYC: UN Blockchain Week Conference 2026: RSVP here

• 09/16 Lisbon: Make it Happen with AI: RSVP here

• 09/16 Barcelona: BABES NET MEETUP IN BARCELONA 💞: RSVP here

• 09/16 Denver: Welcome to the Femme Economy: RSVP here

• 09/16 Lisbon: Make it Happen with AI: RSVP here

• 09/17 Los Angeles: LA Fintech Connect Happy Hour: RSVP here

• 09/17 Beyoğlu: Offline Rooftop Reading Party 📵📚 | Postane Teras | Tea Included: RSVP here

• 09/19 Miami: AI Hackathon: Build with Mel + ElevenLabs: RSVP here

• 09/19 Miami: Agentic Day Hackathon: RSVP here

• 09/23 Munich: Ethereum Munich #7 🥨: RSVP here

• 09/26 Chicago: LOVABLE × KARAOKE: Build the future of karaoke…. Then sing with it!: RSVP here

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

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

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

• 10/16 Washington: Groundwork: RSVP here

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

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

If you enjoyed this newsletter, share it with someone!