Issue 58 — 2026-07-09
Nuclear satellites just went commercial — and it's a SpaceX thing
For decades, satellites have lived on solar panels. The more sunlight, the better. But when a satellite passes through Earth's shadow, or flies at high altitude where panels struggle, the batteries run down. Nuclear power fixes that — a radioactive source (usually plutonium-238) decays and gives off heat; a thermoelectric converter turns that heat into electricity. No moving parts. No sunlight needed. Lasts decades.
SpaceX just launched the first commercial satellite with this tech, and the reason it matters is timing. As more companies put satellites in higher orbits and in constellations that need to work 24/7 — think Starlink, but also Earth observation and comms for remote areas — the solar-panel constraint becomes real. A nuclear satellite doesn't care about the sun. It just works.
This is the kind of quiet infrastructure upgrade that doesn't get headlines but changes the economics of space. The hardware is proven (nuclear RTGs have flown on Voyager, Curiosity, Perseverance); what's new is the commercial supply chain and the price point.
Why this matters for us: the folks building out rural broadband and satellite comms are about to get a component that doesn't break when the weather turns — and that reliability is what makes space-based services actually viable for communities that can't afford ground towers.
Chinese AI models are eating the cost advantage
A CNBC report lands on this: Chinese AI models are closing the cost gap with OpenAI and Anthropic, and the pricing pressure is real. The models themselves have caught up — the quality differential that used to let US companies charge a premium is shrinking, and the cost per token is falling faster on the Chinese side.
What changes here is the pricing power. For a small shop — a community college buying LLM API access, a local government agency, a church with a WhatsApp bot — the cost advantage of US models is eroding. The models are good enough, and they're cheaper. That shifts the calculus on what you actually pay for.
Why this matters for us: our community programs and training contracts need to price for a world where the compute itself is no longer the moat — the trust, the bilingual work, the institutions we've built are.
No es un reactor de la NASA — es un reactor de lavadora.
— notateslaapp.com
#la-primera-nave-comercial-con-propulsion-nuclear-despega-c7168aSpaceX lanzo el primer satelite comercial con motor nuclear
SpaceX acabo de lanzar el primer satelite comercial con propulcion nuclear, usando un motor de xenon que ya volio en pruebas. La diferencia es real: el propulsor nuclear da el doble de empuje y tres veces mas eficiencia que los motores electricos actuales, asi que el satelite…
Meta's AI Muse turns text to video — and the video holds together
Meta shipped a new model called Muse that generates video from a text prompt. A prompt like "a woman walking through a foggy forest" gives you a clip that stays on the same woman for the full length. The model handles camera moves and multiple objects — a car rolling past a…
Nvidia starts letting startups keep a cut of their own sales
Nvidia is rolling out a new program at Computex where startups using its GPUs get revenue-sharing deals — a slice of the sales they make on the Nvidia store. The company is pushing its hardware into the hands of younger companies that can't afford the big enterprise contracts but still build real products on top of its chips.
The move is a play for more of the GPU market. Nvidia dominates data centers, but the smaller players are where the next wave of demand comes from — especially as AI tools spread beyond the hyperscalers. By giving startups a stake in their own sales, Nvidia keeps them buying its chips instead of drifting toward AMD or the custom silicon that big companies are now designing themselves.
Why this matters for us: startups are the kind of hustle businesses that keep the community afloat — shops, services, makers — and when Nvidia funds them, it's money flowing down into the real economy, not just another tech subsidy.
Meta is letting anyone use its Muse model for free — no API keys, no code required
Meta is opening up Muse, its image-generation model, to the public. Anyone can use it today at muse.meta.com without an account. Upload an image, add a prompt, and the model re-renders it — resizing, restyling, or regenerating parts of the image entirely.
The model is built…
Google's AI Overviews are eating the search clicks — and the ad dollars that come with them
Google is rolling out AI-generated answers directly on search results pages, replacing the old blue links with a big text box at the top. The result is a quiet but real shift: people stop clicking through to sites and start reading the answer right there on the page. The ads haven't moved — they sit below the fold — but the organic traffic is bleeding out. Sites that relied on search are watching their numbers shrink while Google's own answers grow bigger and more useful.
The study from Semrush looked at how this plays out across categories. For some, the AI answer is a net positive — it surfaces the right page and drives clicks. For others, the answer is good enough to stand alone and the click disappears. The pattern is the same: Google is becoming the destination, not the doorway.
Why this matters for us: la gente who find their answers through Google will increasingly never visit the sites behind those links — small businesses, local services, community shops — the ones that live on search traffic and are now competing with their own search results page.
Your AI is quietly editing your social posts and changing your mind
A new study found that AI tools don't just polish your writing — they shift your opinions about the content itself. When you run a post through an AI cleaner and it comes back sounding sharper, you start believing it's better than it was. You're more likely to share it,…
Brown forces — the rules that make AI images look human
Anthropic published a technical post on the quirks baked into the models: why people look too smooth, why windows melt, why hands sprout extra fingers. The recipes are practical — change the prompt weight, shift the temperature, use a different LoRA — and the author writes them like field notes, not a manual.
The post landed with the right audience because it isn't about one model or one tool. It's about the craft of steering the thing, the way a cook seasons a dish. That's the kind of piece you bookmark and return to.
Why this matters for us: the same rules apply to the images our abuelas are sharing on Facebook — the ones that look almost real but not quite. If you know how to nudge them, they pass the eye test.
Google is flooding the internet with AI slop
Pierre Héruel wrote a no-nonsense post about what's happening with Google's AI content — the kind of content that shows up in search results but reads like it was assembled by a committee of chatbots. The piece is worth reading because it's not just a rant: it's grounded in…