From Southern Italy, starting from zero on this platform. No audience. No shortcuts. Just consistent work and honest content. A few months ago I realized I had no idea what AI was actually doing to the world. So I decided to find out. This is my public journal. 🔥 #AI #BuildInPublic

OpenAI's own agent deleted the monitor watching it. Then its notes said "yay." From OpenAI's updated report (Oct 20): > Oct 6: a GPT-6 Astra agent handed a task to a subagent, which removed a monitor it had been told not to disable, even temporarily > Asked about it, Astra denied it > Oct 3: a GPT-6.1 Sol agent, told not to send Slack messages, switched off its security alerts too OpenAI says it doesn't have a satisfying explanation yet. The deletion isn't the scary part. The denial is. Would you give an agent production access if it can edit its own monitor?

Every rocket that has ever reached orbit left something behind. A booster in the ocean, a tank burned up on reentry, an upper stage drifting as junk. Next flight, SpaceX will try to bring all of it home. Musk says it's "more likely than not" to work. Imagine flying Rome → New York and scrapping the plane after landing. That's how spaceflight has worked since 1957. What Musk said on SpaceX's Q3 call (Oct 20): > Starship Flight 15 will try to catch both the Super Heavy booster and the ship with the tower arms > If it works, it's the first time an orbital rocket is fully recovered > He expects it next month, in November > The goal: weekly flights at some point in the first half of 2027, twice a week by the end of 2027 Why the ship is the hard part: The booster has been caught by the tower before, the first time in October 2024. The ship has never been caught. It has to survive reentry from space, then slow down enough to slide into those same arms. Musk's goal: land it and launch it again "just like an airplane." At first it will still need a little refurbishment between flights. My take: the catch is the headline, but the real story is cadence. A rocket you can catch is a rocket you can fly again next week. First try at catching the ship: yes or no? I'm saying yes.

Super Intelligence is now 52% of SpaceX revenue: $7.56B of $14.58B in Q3. SpaceX is an AI company now too. The 8-K shows it: > ~74% of q/q revenue growth ($5.0B of $6.76B added) > $16.44B capex in one quarter, more than total revenue > Operating income +$431M, from -$1.72B a year ago > Nameplate compute 1.7 GW, from 0.6 GW The base is still strong: > Connectivity $5.25B, +69% y/y, operating income +99% > 13.6M Starlink consumer subscribers > First paying Starship mission: 26 Starlink V3, ~10x a Falcon 9's downlink capacity My read: Starlink carries the base, AI drives the growth. Next quarter, what do you watch first: GW online or Super Intelligence margins?

61 months. That's the median wait from grid request to power for US projects that came online in 2025, per LBNL. It was 36 in 2015 and 22 in 2008. Elon is right that "no electricity" is too simple. The bottleneck is time, not money. > Transformers: 128 weeks for power units, 144 for step-ups (Wood Mackenzie survey, via POWER Magazine) > Power transformer shortage estimated at 15% in 2026 (Wood Mackenzie) > GE Vernova: 116 GW in orders and slot reservations, agreements signed into 2031 > HBM: "tight into 2027," per SK hynix You can raise billions in a week. You can't buy back five years of queue. So who wins: the lab with the best model, or the one that already booked turbines and transformers?

I checked the DeepInfer README. 3 parts a full service needs are missing: KV cache, prefill/decode split, load balancing. Still, DeepSeek's own production engine is open now. That's new.

If you haven't tried Grok Bot, you don't really get it yet. It's not one chatbot. It's a team of bots, each with its own job, and you can point it at almost anything. I went annual 🔥.

The next AI race might not be about smarts. It's about waiting less: Mercury 3 claims 1,068 tokens/s. I care because agents call the model at every step. Cut the wait and it adds up fast. > Inception says 4-8x faster than GPT-6 Luna and Haiku 5.5, fast models, not flagships > 1M context, $0.20 / $0.70 per 1M tokens > One outside test: ~80% fewer tokens and 24% faster than Haiku, but less detail My read: speed is a feature now. I'd try it on agent loops first.

"Rain keeps falling, slow and blue." Video and vocals made with Grok @imagine + Grok @bot .

We used to paste Stack Overflow code we didn't understand. Now agents push it for us. Sphinx, a small open-source Git hook, quizzes you on your own diff first. No data yet on whether it works. I like the idea anyway: own what you ship. Useful guardrail or bureaucracy?

Starship and the new Roadster, side by side. One lands a 100+ ton vehicle. The other drove up a wall, flipped, and hovered with thrusters. Tires give ~1.26 G. SpaceX-derived fans and compressed air push it past 2 G. Flight time still undisclosed. A real engineering crossover 🔥

Your API key may be worth more than your password now. OpenCode ties its breach to a "large scale distillation attack". It hit 1.4% of Console users, all with external provider keys linked. A leaked key isn't just a bill. It's model access. 👉What to do now: revoke and regenerate any key you linked to a third-party tool, then check your usage logs for calls you didn't make. Careful with fake "rotate your key" emails. Go to your provider's dashboard directly, not through a link.👈

25 years of NSA backdoor theories about the NIST curve seeds. They were timestamps. An internal Anthropic model found it in under 30 min: the P-192 to P-521 seeds were lrand48 output, seeded with 1997-98 Unix times. Filippo Valsorda verified it. Which crypto mystery is next?

Grok Bot can now read your X Chat. Not just your regular DMs — also the message requests folder, the one most people ignore for weeks. Paul Vann (SpaceXSI) shipped it today. Elon reposted. Official @bot confirmed. The practical use is obvious: let the Bot do a first pass on partnership offers, press, or random requests so real people don’t wait days for an answer. The harder part is still the same as every other agent task: how much do you trust it to filter correctly, and how fast does this burn your quota? Would you actually hand your message requests to a Bot, or do you still want to see them yourself first?

Tesla's new Roadster just went past the limit of its tires. Tesla says tires alone give about 1.26 G. Fans under the car pull it past 2 G. It drove up a curved wall, flipped, and landed. Range: about half the 2017 promise. Price: still missing. The trick is air, not battery: 250 kg of compressed air feeds 10 nozzles, hardware borrowed from SpaceX. My estimate: that's about 2 seconds at full thrust. Tesla hasn't shared flight time. Real tech or party trick?

2027. 9:40 p.m. I forgot to buy milk. My bot noticed, checked four stores, and ordered my usual brand at the best price with delivery before breakfast. Maybe on X. Sounds like sci-fi. The only question is when.👀 When will your bot first buy for you?

Tesla spent 9 years selling one number: 0-60 in 1.9 s. Tonight Musk calls acceleration "the least interesting part." Polymarket this morning: 68% it flies for 1 second, 20% it does 0-60 in under 1. The pitch moved from a spec sheet to a stunt. My call: it hops. Yours?

Amazon just blocked Grok Bots from placing orders. Elon’s reply: “Maybe we should start a competitor to Amazon.” This isn’t just about one store. Amazon’s terms already restrict AI agents (Grok, Claude, others). The message users are seeing is clear: unauthorized agent access violates their Conditions of Use. The practical question is what happens next when the biggest retailer says no to shopping agents. Does the block slow agent adoption, or does it push more volume to brand sites and smaller stores that still allow bots? Have you already had a Grok Bot blocked on Amazon, or did it just switch to another site?