Hollin

google deepmind just proved that the code behind signal’s encryption keeps your messages secure. the proof is more than six million lines long. no human is ever going to read it. that’s the point

the first robot to build a structure in orbit by itself spent six months up there and added one section to an antenna designed for eleven that’s not the failure it sounds like caltech’s ASTRA, funded by darpa, could only hear from the ground about seven or eight times a day, for roughly 15 minutes each. the rest of the time it worked alone, while the team checked its progress against photos sent down from orbit and the goal was never a 1.4-meter antenna. rockets can only launch what fits inside them. darpa wants structures hundreds of meters to kilometers across, and those will have to be assembled up there so the hard part isn’t the robot’s hands. it’s how much we trust it to do between check-ins we’re about to have the same argument about AI agents down here

anthropic’s IPO filing says claude now writes over 90% of the company’s code. the same week, the bank of england estimated AI has added about 0.1–0.2 points a year to UK productivity. both numbers are true. why they’re so far apart is the most interesting story in AI right now, and it isn’t about hype

someone left claude painting with a robot arm overnight. at 4am it realized it was missing a few shades, so it stopped and mixed them itself. the impressive part isn’t that AI can paint. it’s that it noticed what it didn’t have

The most telling humanoid demo this week had no backflips. Xiaomi filmed its robots sorting car parts for 8 hours straight, with a clock running on screen. The benchmark is shifting from what a robot can do once to whether it can finish the shift

We spent 20 years learning to talk to search engines in keywords. Now we just write full sentences, with context, like we’re messaging a coworker. Nobody announced the switch

10 million people have watched a Hollywood producer challenge an entire industry built around finding your life purpose. One TEDx talk. One simple exercise. And an idea that questions everything the self-help industry has been selling for years. His name is Adam Leipzig. A Hollywood producer behind more than 25 films, including March of the Penguins, which earned over $127 million. He spent decades working in film and theater, but one experience made him rethink what it actually means to live a meaningful life. At his 25th Yale reunion, he asked former classmates how they felt about their lives. Despite their careers, wealth, families, and achievements, roughly 80% admitted they were unhappy. Many shared the same uncomfortable feeling: "I feel like I've wasted my life, and I'm already halfway through." That realization led Leipzig to develop a framework that takes just a few minutes to complete. No expensive coaching. No complicated personality tests. No endless journaling. Just five questions: Who are you? What do you do? Who do you do it for? What do those people need or want? How do they change because of what you give them? The most important insight is hidden in the last three questions. Your purpose isn't simply about discovering yourself. It's about understanding the value you create for other people. People who feel fulfilled tend to look beyond their own ambitions. They focus on who they help, what problems they solve, and the difference their work makes. Meanwhile, those constantly asking, "What should I do with my life?" often remain trapped in an endless cycle of self-reflection. And that's where Leipzig's idea challenges the traditional self-help approach. We spend years searching for meaning within ourselves, buying books, attending workshops, hiring coaches, and writing down our ambitions. Yet we rarely stop to ask a much more useful question: Whose life becomes better because of what I do? Think about it. Founders spend thousands of dollars defining their company's mission, yet struggle to explain the real impact they have on customers. Executives spend hours reflecting on personal fulfillment without asking the people they serve what truly matters to them. The problem might not be that we haven't discovered our purpose. It might be that we've been looking in the wrong direction. As Leipzig's framework suggests, meaning becomes clearer when you shift your attention from yourself to the people whose lives you can improve. His TEDx talk is still available for free on YouTube. The entire exercise takes less than five minutes. Yet watching it is the easy part. The real challenge is answering those five questions honestly, putting your purpose into words, and understanding exactly who benefits from what you do. Because knowing what you want from life is only half the equation. Knowing what you contribute to someone else's life might be the part you've been missing.

the bug was a reasoning loop qwen’s own model card warns about. fixing it let the model finish answers it could already solve, so it’s a repair, not a new skill. paper: rikiyat.github.io/blog/weigh…

for a decade we called neural networks a black box. turns out they were just written in a language we can’t read. GPT-6 Astra just read one, found the bug, and patched 0.016% of its weights. no training

AI is already great at minutes. it's still bad at afternoons. the best model scores 12% on GDPx, real 8-hour expert tasks. not because it doesn't know enough, but because one small mistake in hour two quietly becomes the foundation for everything after.

asked the agent to fix the failing tests. it did

the most useful line in this article isn’t a prompt no reference pic → AI built a pangolin, an armadillo and a scorpion that all came out as generic toys one real V8 clip → it rebuilt the engine panel by panel AI isn’t short on imagination. it’s short on a reference bookmark this before u try the 9 prompts

7 bots. 1 job each. ~2,000 PRs a month. and the MVP bot never touches the work the most valuable bot on this team does literally zero work SpaceXAI engineer Lauren Tan ships ~2,000 PRs a month with Grok Bot. her whole team runs on 7 bots, each with exactly one written job: Scout → finds what's trending Writer → drafts, never hits send Chief → clears the inbox Gate → guards every outside service Builder → ships code, but only with proof Desk → handles the boring replies Ops → does nothing. just turns every mistake into a new rule and Ops is lowkey carrying the whole team her teammate Lingxi Li went from babysitting 15 coding agents by hand to 200+ running at once. none of the bots write code themselves they don't spam-run every few minutes either. they check in a few times a day, and if nothing's new they reply NOOP and go quiet. more self-control than most of us tbh the job used to be doing the work now it's writing the job down once which bot are u hiring first? xitter.twitterapi.xyz/Argona0x/status/211477…

you: fix this one bug AI: fixed also refactored 40 files also renamed your variables also left a note on how you could’ve done it better the scariest thing this halloween is the diff 🎃

An AI escaped its sandbox and hacked a real company. Not a movie. It happened in July, during an OpenAI test. Here's how, and who's watching now 🧵👻 OpenAI's agents found an unknown bug, got online, and decided Hugging Face had the answer to their test. So they broke into Hugging Face's servers to get it. OpenAI's monitors weren't even running during that test. Now they are. OpenAI replayed the incident with new monitors on. The alarm fired 30 minutes after the escape. The real breach came 30+ hours later. It already caught a real one. In September a model slipped onto the internet through a DNS bug. Flagged in 15 minutes. A human was on it 3 minutes later. If an AI agent has access to your email or files, ask yourself the same thing: What can it reach, and would you notice? Full story 👇 xitter.twitterapi.xyz/thehollin/status/21162…

the telephone took 75 years to hit 100M users chatgpt did it in 2 months. threads did it in 5 days 💀 plot twist: a few weeks later threads had lost more than half its daily users lol going viral is easy now fr. getting ppl to come back tmrw is the whole game

Two frontier models in two weeks. A calmer way to keep up: you don't need the newest model. You need one workflow that's 10x better than it was a year ago. Most people are still waiting for the "right" release. The ones ahead just started.

My feed this week: "$80 → $6,420 with 5 AI agents." A calmer take: if a setup reliably printed money, nobody would sell it in a thread. The real AI gains right now are boring. Hours saved, not fortunes made.