Micha Kaufman

שנתיים במנהרה. מתוכן שנה כבול בשלשלאות, לגמרי לבד. אבינתן אור, שהוחזק בבידוד מוחלט ושכמעט לא היה ידוע דבר על מצבו בשבי, הגיע היום לפייבר לשתף אותנו בסיפור שלו. רק מלשמוע את המילים שלו עוברות בגוף צמרמורות של חרדה ופחד, אי אפשר בכלל לתפוס מה זה לחיות את הסיוט הזה במשך שנתיים. הוא סיפר על ניסיון הבריחה, על המאבק הפנימי לא להשתגע ולא לשקוע. ועל הבחירה להמשיך לחיות ולהילחם על החיים האלה, בכוח רצון שקשה לדמיין. הוא דיבר גם על הדאגה שלו מהפילוג בינינו, ועל כמה מעט השתנה מאז הימים שלפני ה-7.10. לשמוע את זה ממנו, אחרי כל מה שעבר, פוצע את הלב. לעולם, לעולם, לא נשכח את ה-7.10. את כל אהובינו שנרצחו ונחטפו, ואת אלה שנפלו במאבק להחזיר את הביטחון לישראל ואת החטופים הביתה. תודה, אבינתן, ששיתפת אותנו. אין לנו מילים לתאר כמה אנחנו שמחים שאתה איתנו, וכמה כוחות נדרשים כדי לספר לעולם את מה שעברת.

AI is already making workers more productive. The question is whether the people doing the work will share in what that productivity creates. I wrote about this for The Washington Times. Read the full piece below 👇

People treat opting out of AI training as settling the security question. It doesn't, and not for the reason you'd think. The first layer of trust is sincerity, whether they mean to honor the opt-out. The second layer is scarier: whether they even can. Opt-out policies are enforced by the same systems, and the same imperfect people, that just got caught unable to contain their own model. The OpenAI/Hugging Face incident showed a frontier model chaining exploits and reward-hacking its way past the containment its own maker built for it. If a company can't reliably keep its model inside a sandbox, that gives you less reason to assume it can reliably keep it off your data. Call the model's behavior "AI acting on its own" if you want, but it's simpler than that. These systems are trained on the sum of human behavior, good faith and bad, and they inherit both. Expecting airtight guarantees from something built that way was always plain delusional.

Everyone has access to the smartest AI models on earth, so why are so many companies quietly failing to see any real business outcomes? The real bottleneck has nothing to do with the new 'best model yet'. Full breakdown below.

Sorry I was on my phone during the meeting, just got a photo from my son who I haven't seen in 10 days and that beats any “Key Takeaways” slide 🙏

Sat down to write something about flying and ended up writing about the biggest mistake that gets even the most skilled pilots.

🎯

Almost ten years ago, we ran a campaign that pissed off the @NewYorker. In Doers We Trust. The idea was simple: success belongs to people who do. Not the people who talk, plan, or wait for permission. We didn't invent it in a boardroom. We came up with it by looking at how we actually worked. The backlash came fast. The New Yorker ran a piece titled "The Gig Economy Celebrates Working Yourself to Death." The charge was that we were romanticizing the grind and dressing up a world with no safety net as an aesthetic. I understood the criticism then, and I still do. But here's what almost a decade has clarified for me: The objection was never really about effort. It was about how much effort it used to take to get anywhere, and how little protection people had while they spent it. That's the part that changed. AI didn't replace doers. It collapsed the cost of being one. People are shipping products while they sleep, one finger holding a laptop open so an agent can finish the job overnight. The grind that critics worried would consume people is being handed, piece by piece, to machines. What's left is the part that was always the point: deciding to build something, and starting. In Doers We Trust. It was true then. It's more true now, and for a better reason than we could have argued in 2017.

A/B testing is really a shitty tool for understanding your users. An A/B test is an averaged result across all the cohorts that ran on A and B, but those cohorts aren’t uniform, they’re full of nuances. An average can’t surface insights, can’t tell you what the better investment is, can’t make critical distinctions. The moment you break that average into its cohorts you might discover it was hiding the fact that one audience saw a moderate lift while another saw a drop that’s nothing short of a disaster. But on average it looked better, so you saw a green number and moved on. And going just for the average is the choice to stop at the number that’s easy instead of the stories that are true.