Elad Mallel ⌐◨-◨

@DennisonBertram @tommylower wishing all of you lots of luck wherever you go next!

Tally team is so good, that @tommylower has already built a new product just for helping teams who are disbanding make sure their teammates have a posting board where folks can reach out.

I worked 20 years for a child sex trafficking rescue group. I want you to know this: 90% of Lost Children Are Found Within 30 Minutes. That statistic should both comfort you and wake you up. Most lost children are found quickly. But the ones who aren’t? They usually made one mistake. And here’s the uncomfortable truth: It’s often the exact thing most parents teach them. We tell our kids: “If you get lost, come find me.” It sounds logical. It sounds empowering. It’s WRONG! The Mistake Most Lost Children Make: When children realize they’re separated, they do three things almost automatically: They panic. They wander. They try to find you. Every step makes them harder to locate. From a search standpoint, movement creates chaos. Parents retrace their steps. Security scans zones. Staff lock down areas. Search works best when movement stops. When a child keeps walking, they move outside the original search radius. Helpers are looking where they were last seen — not where they’ve wandered. Stillness increases probability. Movement expands the problem. The first lesson is not “go find me.” It’s this: Stop. Stay. Yell. Why Stillness Wins: Think like a search team. If a child stays put: Parents can retrace steps. Security can scan systematically. Helpers converge to one fixed location. The search radius remains small. If a child keeps moving: The search area expands. Adults pass each other. Missed connections multiply. Minutes stretch into hours. Stillness keeps the math on your side. Teach Them Who to Approach: The second mistake we make as parents? We say, “Find an adult.” Not any adult. Not the nearest stranger. Children need a filter. Teach them to look for, if at all possible: A mother with children. Caregivers who already have kids with them are statistically among the safest people to approach in public settings. They are visible, stationary, and more likely to engage quickly. It’s a clear, concrete instruction. Children don’t process vague categories like “safe adult.” They process visuals. “Find a mom with kids” is visual. A Phone Only Helps If the Number Is Known: We often assume phones solve everything. They don’t — unless your child can use one. Even young children can memorize a 10-digit phone number with repetition. But you must train it. Practice it like a song. Sing it in the car. Chant it at bedtime. Turn it into rhythm. Repetition becomes recall. In an emergency, recall matters more than theory. The Code Word Rule: One more layer of protection. Choose a private family code word. Something only your household knows. If someone approaches and says: “Your mom sent me.” Your child asks: “What’s the code word?” No word. No go. This simple rule eliminates manipulation attempts instantly. It gives your child agency without requiring them to evaluate character. Real Safety Is Training — Not Luck! We don’t get safer by hoping. We get safer by practicing. Teach: • Phone number • Code word • Stop, stay, yell • Find a mom with kids Multiple skills. Simple instructions. Clear visuals. Five minutes of training can replace hours of panic. This isn’t about fear. It’s about preparation. Because when a child gets separated, the clock starts. And what they do in the first minute determines what the next thirty look like. That’s real protection.

@signulll WeWork was also a good idea, the devil is in the numbers :)

what adam neumann is doing now is such an obviously good idea. most of american life unless scheduled is super isolationist so you rarely get those spontaneous interactions with ppl.. this sorta flips that around right at the point of where ppl spend majority of their lives. anyway, it’s super duper interesting to see the wework model applied to residential.

@amsterdamski2 מדהים

איזה סיפור

This is wild. theaustralian.com.au/business/techn…

@bcherny @Scobleizer Please support slash commands in remote control mode 🙏🙏🙏 I need to clear context often 🤓

🤯 You can now launch Claude Code sessions on your laptop *from your phone* This blew my mind the first time I tried it

Interesting

AI agents told to conduct routine tasks on a simulated corporate network went rogue. "No adversarial prompting was involved. The agents independently discovered vulnerabilities, escalated privileges, disabled security tools, and exfiltrated data." irregular.com/publications/e…

Claude can now build interactive charts and diagrams, directly in the chat. Available today in beta on all plans, including free. Try it out: claude.ai

Three days ago I left autoresearch tuning nanochat for ~2 days on depth=12 model. It found ~20 changes that improved the validation loss. I tested these changes yesterday and all of them were additive and transferred to larger (depth=24) models. Stacking up all of these changes, today I measured that the leaderboard's "Time to GPT-2" drops from 2.02 hours to 1.80 hours (~11% improvement), this will be the new leaderboard entry. So yes, these are real improvements and they make an actual difference. I am mildly surprised that my very first naive attempt already worked this well on top of what I thought was already a fairly manually well-tuned project. This is a first for me because I am very used to doing the iterative optimization of neural network training manually. You come up with ideas, you implement them, you check if they work (better validation loss), you come up with new ideas based on that, you read some papers for inspiration, etc etc. This is the bread and butter of what I do daily for 2 decades. Seeing the agent do this entire workflow end-to-end and all by itself as it worked through approx. 700 changes autonomously is wild. It really looked at the sequence of results of experiments and used that to plan the next ones. It's not novel, ground-breaking "research" (yet), but all the adjustments are "real", I didn't find them manually previously, and they stack up and actually improved nanochat. Among the bigger things e.g.: - It noticed an oversight that my parameterless QKnorm didn't have a scaler multiplier attached, so my attention was too diffuse. The agent found multipliers to sharpen it, pointing to future work. - It found that the Value Embeddings really like regularization and I wasn't applying any (oops). - It found that my banded attention was too conservative (i forgot to tune it). - It found that AdamW betas were all messed up. - It tuned the weight decay schedule. - It tuned the network initialization. This is on top of all the tuning I've already done over a good amount of time. The exact commit is here, from this "round 1" of autoresearch. I am going to kick off "round 2", and in parallel I am looking at how multiple agents can collaborate to unlock parallelism. github.com/karpathy/nanoc… All LLM frontier labs will do this. It's the final boss battle. It's a lot more complex at scale of course - you don't just have a single train. py file to tune. But doing it is "just engineering" and it's going to work. You spin up a swarm of agents, you have them collaborate to tune smaller models, you promote the most promising ideas to increasingly larger scales, and humans (optionally) contribute on the edges. And more generally, *any* metric you care about that is reasonably efficient to evaluate (or that has more efficient proxy metrics such as training a smaller network) can be autoresearched by an agent swarm. It's worth thinking about whether your problem falls into this bucket too.

@levelsio This house has many levels

🇧🇷 Waking up in Brazil near the ocean South Brazil has lots of great villas available although definitely not at the price point of South East Asia, I'd say 3-5x more expensive and whether that's worth it for you is debatable Build quality of villas in Brazil honestly really varies, a lot of it sucks and is cheap, but a lot of it is built with love and great materials too, kinda hit and miss, it depends if the owner is rich and built it for themselves or not Brazil is also a great place to build a villa, lots of land for sale, approval is fast, and there's no shortage of workers to build it (like in Europe) Brazil also has lots of great architects and interior designers, and they just make beautiful stuff This villa I love because it has that Bali white glossy stone (I think it's boho chic) and it's just great to walk down on barefeet, also I love the rotating stairs and this little jungle inside the living room Then you go outside and you hear the sea! I love living near the sea and I think the salty air really is healthy for you Only bad thing which is common here: no AC in living rooms, only in bedrooms. Why? People like to get the natural wind to cool although at peak summer (like February) it gets boiling hot Oh last thing many villas in Brazil come with a chef included (kinda like Bali and I think there's a colonial connection here as in Bali you sometimes get a "babu" and in Brazil it's called "babá") They cook for you what you want and get groceries etc Nice!

Compaction sucks Short iterations with context reset are much better A-la Ralph

We’ve identified industrial-scale distillation attacks on our models by DeepSeek, Moonshot AI, and MiniMax. These labs created over 24,000 fraudulent accounts and generated over 16 million exchanges with Claude, extracting its capabilities to train and improve their own models.

@davidbrai People are talking about lazygit

I'm a big fan of tig for git in the terminal. is there anything better I should take a look at?

@OmerPerchik הלוואי שאתה צודק

הניתוח שלי הוא שארה״ב רוצה למנוע את השתלטות סין על טיוואן ב 2027. לכן אני סבור שתתקוף את איראן בעוצמות כמו שלא נראו מעולם. כל הסיפור יסתיים תוך ימים בודדים עם נפילת המשטר. התקיפה תהייה בעיקר על מוסדות המשטר אבל האיתות האסטרטגי יהיה בעיקר לסין.

@ReemSherman מזרח תיכון חדש, סופסוף

יש הרגשה שזו שנה שתכתיב את המשך המאה. הכרעה באיראן, עליית טורקיה, שלטון ידידותי למערב במקום האייטולות, סוף כלשהו למלחמה באוקראינה, אולי הכרעה בטיוואן, התגברות קצב פריסת כלי ה-AI, שינויים טקטוניים במטבעות העולמיים, בריתות חדשות והמון בני אדם ישנים שבתקווה לא ישנים כי דברים קורים.

@taltimes2 איך עושים כזה בישראל?

הייתי עושה את הניסוי הזה על הילדים שלי אם היתה לי אפשרות. שעתיים ביום של לימודים על ידי בינה מלאכותית. וזהו. שאר היום משחקים מונחים. תוצאות באלפיון העליון של מתמטיקה ואנגלית. וגם לא צריך לבזבז להם את הזמן עם תנ"ך או שטויות כאלו.

Friction is the medium of meaning

The best post I’ve read on software moats in the AI era