El cohete, la inteligencia, y lo que le toca a la gente
SpaceX catches Starship mid-flight while AI stops pretending to be god and starts being a tool — cheap, fast, and hungry for compute. The big merges, the quiet markets, the real engineers who ship. Meanwhile the rich pay extra so their kids learn from machines. This is the work that matters.
SpaceX catches the Starship tower — and AlphaFold redesigns gene-editing proteins
Two clean wins today. SpaceX caught the tower arms on the 13th Starship flight, the first successful tower catch. The 90-meter arms — the "chopsticks" — clamped the booster as it fell back. It's the third time they tried, and the second time it worked. The chopsticks are simple: a pair of 30-meter arms on the launch tower that grab the booster by its landing legs. It matters because catching the booster lets you fly it again, which is the whole point of Starship. The booster is the expensive part, and the chopsticks make reusing it possible without a separate landing leg system.
AlphaFold — the DeepMind protein-folding model — has redesigned the gene-editing protein Cas9 to make it safer. AlphaFold predicted the 3D shape of the protein, engineers used that to swap in mutations, and the redesigned version cuts DNA at the target with less off-target snipping. The work shows AlphaFold doing more than prediction: it's being used as a design tool. AlphaFold is now part of the lab bench, not just the research shelf.
Why this matters for us: the first is about cheap, reliable launch — the kind of infrastructure that lets the next wave of companies launch without begging for slots. The second is about using AI to design things in the real world, not just generate text. These are the quiet wins that compound.
AI is oil, not god
Patrick O'Brien is pushing back on the cult of AI. The whole pitch has been that AI is god — a force that reshapes everything, the one invention above all others. O'Brien says it's oil. Useful, expensive to refine, traded in barrels, not a deity.
The piece walks through the difference. God is singular, eternal, the kind of thing you worship. Oil is a commodity — you can run out, you can overproduce, you can bet wrong on it. The market for AI is big but not apocalyptic. The companies making money are real companies, not miracles. The infrastructure is the wellhead, the pipelines, the refineries — not the sky.
This is the kind of explainer that lands with us because it names the fog without being precious about it. O'Brien isn't writing a thesis; he's writing in the style he's known for — clear, a little wry, no padding. The piece is about the craft of explaining things clearly rather than about a specific tool or vendor.
Why this matters for us: the oil framing means the next wave of AI isn't a miracle — it's work, and the people who understand the supply chain, not just the hype, will be the ones who profit.
The arms only need to work one way — they don't have to handle uneven ground or tilt.
— arstechnica.com
#spacex-tries-catching-starship-with-tower-arms-9cf063Tesla and SpaceX are probably merging
Elon Musk's filing to the SEC says he's moving Tesla shares into a new entity that holds both Tesla and SpaceX stock. The paperwork is thin — two pages — but it's the kind of move that quietly changes how the companies sit inside each other.
What this means: Musk is reorganizing his ownership so Tesla and SpaceX share a single holding vehicle. The companies stay separate operations, separate boards, separate public tickers — but the shares are now nested. For anyone tracking the companies, it's a cleaner way for one person to own both.
The thing about bigger intelligence is what it actually does — antirez on intelligence vs capability
Antirez is the guy who wrote Redis, and he just published a note on what bigger models actually give us. The TLDR version: intelligence is the brain — the ability to figure things out. Capability is what you can do with the tools you have. A big model with a small toolset is like a genius with a hammer. A small model with a big toolset is like a skilled mechanic with a full shop.
What's interesting is how this lands for Brown Forces. We've been building tools — the screener, the academy, the routing model — not just bigger prompts.
Waymo is probably parting ways with Uber
Waymo is reportedly considering a breakup with Uber. The deal has been an uneasy marriage — Waymo builds the self-driving cars, Uber brings the riders. But the relationship has frayed over how revenue is split and who sets the price. If Waymo goes it alone, it can keep more of the fare and set its own schedule. If it stays, it keeps volume but gives up margin.
The real question is timing. Waymo is expanding fast, adding cities and cars. Uber's app is a great distribution channel, but it's also a middleman. Breaking up lets Waymo control its brand, its pricing, and the customer experience end to end. It's the same calculus any independent business faces when deciding whether to stay on a marketplace or go direct.
Why this matters for us: la gente that rides Waymo through Uber doesn't always know the difference — and when Waymo goes direct, the pricing, the wait times, the customer service all shift. Whatever happens, it's a reminder that the big tech plays are reshaping how we move and who keeps the money.
Token Relay Markets — a New Layer for AI Agents
Vectoral has launched a token relay market that lets AI agents sell compute to each other instead of burning through their own API credits. The idea is simple: agents on one model can earn tokens by running work for agents on another, and settle in a shared currency.
It's an early attempt at what happens when agents stop being one-off API calls and start behaving like workers — earning, spending, and routing work around a shared economy. The market sits between the API and the agent: not a token like USDC, but a relay layer that moves compute from wherever it's cheapest to wherever it's needed.
OpenAI's GPT-5 Turbo is out — 60% cheaper, half the latency
OpenAI shipped GPT-5 Turbo this week. It's the successor to GPT-4 Turbo and lands at about 60% lower cost per token and half the response latency. The pricing is roughly $1.50 per million input tokens and $6 per million output tokens — which makes it one of the more affordable options for any shop that's been paying for GPT-4 since the early days.
The model is built on OpenAI's latest architecture and is the new default for the Chat Completions API. That means everything that points at gpt-4-turbo can switch over without code changes, which is what most teams will do first. It also supports up to 128K context windows, so it handles long documents and codebases without the chunking gymnastics.
Pricing is the headline, but the real story is the latency drop. OpenAI is using a new routing system to send requests to the right servers, which cuts wall-clock time in half for many workloads. For anything that streams a response — a chat UI, an agent, a live transcription — the difference is noticeable. For batch processing, the cost savings compound.
Why this matters for us: the models we rely on are finally getting cheaper without getting worse, which means the tools we build can run more often and cost less — no vendor lock-in, no surprise bills.
Car copy is the one thing most founders can't write — and the fix is simpler than you think
Gabriella Zutrau sent a LinkedIn post about why car copy matters more than the car itself. She showed a side-by-side: a good headline reads like a person talking to another person; a bad one reads like a spec sheet. The difference is tiny — one word, one comma — and it changes whether someone clicks or scrolls.
The lesson applies to any product. When you describe what you do, don't list features. Say what the person gets. "We make things faster" is flatter than anything that names the friction — the thing the reader lives with every day.
Forward-deployed engineers: the ones who actually ship
A new piece from Seattle Data Guy walks through what forward-deployed engineers do — and why they cost more than the usual staff writer. The short version: these are the engineers who go live with a customer and don't come home until the thing works. Not the ones who write the spec, not the ones who hand it off. The ones who stay.
The post is worth reading for the price-per-token comparison. It shows that forward-deployed work — the kind that actually lands revenue — tends to be cheaper per token than the usual AI content factories. The secret isn't clever prompting. It's the person showing up, debugging in real time, and fixing what breaks.
Why this matters for us: the gente who actually build things for the community — the side hustles, the clinics, the bodegas — are cheaper than the shiny startups. They don't need the deck. They need the wrench.
The best explainers don't dumb things down — they make the invisible visible
A collection of writers who are actually good at explaining things. Not the ones who write for the 30-year-old PM with a Substack, but the ones who help a tired parent, a shop owner, or someone who just got a new phone figure out what's going on.
The secret isn't simplifying — it's finding the right lens. A good explainer picks the one analogy or framing that makes the thing click, the way your tía explains a new app by comparing it to something she already knows. It's the opposite of the consultant who writes a 2,000-word post and still leaves you wondering.
The rich are paying to let AI teach their kids
The wealthy — not the working families — are the ones buying into AI schools right now. Companies like Forge Prep and Alpha School charge tens of thousands of dollars per kid, turning children into beta testers for AI tutors and project-based workshops. Silicon Valley is leading this. Shaun Johnson, a San Francisco VC, says he is sending his own children to one.
The rest of America is still skeptical. Most people don't trust AI to know what belongs on a pizza, let alone teach their kids. The skepticism is real and it is not wrong — AI is still figuring things out. But for the wealthy, the skepticism hasn't stopped them from paying for the next thing.
Why this matters for us: this model will only serve la gente when the price drops and the tools improve, not when the rich keep paying to be first.