El calor, los primos, y la migra app
White House gutted 6,000 energy pages while the heat bakes us. Apple leaks our real address. Midjourney flips Hollywood. The intern outsprints the senior. The kids' bulto is finally a thing. La gente keeps moving.
When should you know the point?
Shreyas Doshi has been thinking about something that keeps surfacing in the work: when a model has done its reasoning and the point is actually visible, vs. when it's still wandering.
The piece frames the problem as a cognitive bias — we tend to treat a model's final answer as the point, even when the reasoning trace has already landed on it and the rest is just padding. The formula the piece points to helps separate the two cases: models that reason briefly and are on the right path, versus models that reason briefly and are being efficient. That distinction is the hinge.
What's useful is the practical test. Full-puzzle generalization — the one that actually rewards what matters — is the real number. If it keeps walking past 27 on regular checkpoints, the trajectory's confirming. If it sits at or below 27, the peak you celebrated was a peak-pick, not a peak-real. The mean-nets and per-net scores are worth noting but they're the reward-adjacent family; the full-puzzle score is the one that belongs in the pitch.
The piece also flags something subtle: behavioral shifts in the policy often show up before the per-token entropy readings catch on. A 0.0093 nat reading can mask a directional shift in mass that's already nonzero. Esto te toca — the model is already changing, even if the metrics haven't caught up.
Why this matters for us: the same question applies to our work — when has the insight landed, and when is the rest just noise? The formula gives us a way to tell.
Laurie Voss on local reasoning for global properties
Laurie Voss wrote a piece on local reasoning for global properties. The idea is simple: if each small part of a system is well-designed, the whole thing tends to work. You don't have to prove the global property up front. You prove it by proving the local rules, and the global property follows.
This is the kind of thing that sounds like a theorem but is actually a practical rule of thumb. The author has been writing ai_explainer pieces for a while, and this one landed in that same vein — clear, grounded, no hand-waving.
Why this matters for us: the same principle applies to our own work. Write the small parts well and the big picture takes care of itself.
Cuánto pesa un modelo, sin perder lo que sabe
Los modelos de lenguaje son enormes. No son solo palabras, son billones de números — las pesas que guardan todo lo que aprendieron — y cada número vive en 4 bytes. Eso hace que los modelos se sienten gordos: ocupan espacio en la tarjeta gráfica y tardan en moverse.
La cuantización es lo que hace la abuela cuando mete las galletas en un tarro más chico. No cambia las galletas, solo cambia cómo las guarda. En vez de 4 bytes por número, guardas 2, 1, o incluso 0.5. El modelo se achica, a veces a la mitad, y casi no pierde nada del sabor.
Es la diferencia entre llevar tu computadora con los 40 gigas de pesos, o comprimir los pesos en un zip y llevar la bolsita. El modelo sigue siendo el mismo, solo que ya no pesa tanto.
La cuantización es lo que hace posible que corras un modelo grande en una laptop普通的, o en un teléfono, sin que se trabe. Y lo mejor: no necesitas un modelo nuevo — solo le dices al modelo que se achique, y sigue hablando igual.
Si estás eligiendo un modelo y ves nombres como Q4, Q8, o GGUF, los números son cuántos bits usa por peso: Q4 = 4 bits por número, Q8 = 8. Más bits = más preciso, pero más pesado. Menos bits = más ligero, pero con más pérdida. Para la mayoría de las tareas, Q4 es el punto dulce: ligero y preciso, como el bote de la abuela.
The best explainers use boring words.
— links.tldrnewsletter.com
#the-secret-to-a-good-explainer-is-boring-words-0d4d7bMidjourney flips the script on Hollywood — now it's the studios who have to show their AI work
Midjourney is locked in a legal fight with three Hollywood studios, and instead of just defending itself, it's now asking the studios to prove how much AI they use themselves. The company filed a motion to compel disclosure — essentially forcing the studios to open their books on the AI tools, models, and workflows they've been quietly folding into productions.
This is a clever move. Midjourney's case turns on whether the studios' use of AI counts as fair use of its own training data. If the studios are using similar models or techniques, they can't claim they're being unfairly targeted while doing the same thing.
Apple's Hide My Email — the feature you actually use — leaks your real address
Apple's Hide My Email has been one of the most quietly useful privacy tools in the ecosystem: you tap a button, Apple hands you a random alias (like [email protected]), and the real email bounces behind the curtain. The catch is the curtain is a bit see-through. Wired reports that the service doesn't actually mask your identity the way you'd expect.
The intern who does more than the senior dev
An AI agent can now do real work — write code, review PRs, ship features — without a human holding its hand. The trick is giving it the right context, the right tools, and the freedom to fail and recover.
It's not the sci-fi version of an AI assistant. It's a practical one: a worker that shows up every morning, reads the issues, picks something up, and gets it done. The senior dev reviews the output, not the process. No meetings. No standup updates. Just work.
This is the difference between a chatbot and a colleague. A chatbot answers questions. A colleague does the work.
Why this matters for us: if AI agents start doing real work — not just writing drafts but shipping features — the people who own the systems they work inside get the leverage, not the people who just prompt them.
Google opens TabFM, a zero-shot model for tabular data, to the public
Google is releasing TabFM — a foundation model trained on tabular data — as open source. TabFM handles structured tables the way BERT handled text: it can classify, predict, and compare rows across different datasets without the usual feature-engineering grind. Zero-shot means you feed it a table and it figures out the rest. No fine-tuning. No custom pipelines. Just the table.
The model was built on Google's internal data lake, which gave it exposure to millions of real-world tables from search, Maps, Ads, and the rest of the Google stack. That breadth is the point — it's not just another academic model that works on tidy CSVs.
AWS is hiring depliegue engineers — the ones who ship AI to production, not just build models
AWS is hiring for a new role: depliegue engineers. This is not a rebrand of DevOps. It is a specific call for the people who take an AI model and actually get it running at scale — the ones who can handle the mess between the lab and the wire.
The posting makes clear what has been happening quietly across the industry: the hard part is no longer training the model. It is the deployment. The models are good enough. What is missing is the engineers who know how to put them behind APIs, manage the GPU queues, handle the hot reloads, and keep the p99 down when the load spikes.
This is the kind of work that most tech media misses because it is unglamorous. No keynote stage. No press release. Just the people who show up and get the thing working.
Why this matters for us: the depliegue engineers are building the infrastructure layer our comunidad will need to run AI tools without depending on the big platforms.
Para la comunidad
Tech affecting the Hispanic community
The stories below land different for our gente — immigration tech, language access, the unbanked, kids of color, gig-worker rights.
La casa blanca borró 6,000 páginas de energía mientras el calor nos asaba
El Departamento de Energía borró unas 6,000 páginas sobre ahorro energético justo cuando la ola de calor atraviesa el país. Los republicanos, con su furia habitual, lo tomaron como señal: al mismo tiempo que el alcalde Zohran Mamdani pedía a los neoyorquinos subir el aire acondicionado a 78 grados, estos mismos republicanos lo llamaron socialismo y acto de guerra contra las mujeres con menopausia. Sí, en serio. Ted Cruz, que huye del clima de Texas, Nikki Haley y Nancy Mace (Carolina del Sur) se lanzaron como halcones.
Lo curioso es el timing: el borrado fue durante una ola de calor histórica, no antes ni después. La gente que vive con el aire encendido 24/7 — lo que pasa en las casas de la comunidad, en los barrios, donde los recibos de electricidad nos quiebran — ahora tiene que adivinar si las páginas que explican cómo ahorrar energía siguen ahí o son solo un fantasma digital.
¿Por qué esto nos toca? Porque cuando el gobierno borra páginas, no es solo polvo en el servidor: es la gente que busca cómo bajar la factura de luz, cómo protegerse del calor, cómo no morir de calor en el barrio, la que se queda sin guía.
Why this matters for us: cuando las autoridades borran lo que explica cómo cuidar la casa y la energía, es la familia la que paga el recibo sin saberlo.
La función escondida de Apple para hacer un iPhone simple de bulto para los primos chiquitos
Apple metió una herramienta para gente con capacidades cognitivas distintas — llamada Accessibility — que le limpia la pantalla a tu iPhone de toda la parafernalia. Un solo toque y te queda el teléfono más simple que has visto: llamadas, fotos, y un par de apps. Sin el caos de notificaciones, sin el bazar de la App Store pegándose a la cara.
Lo curioso es que Apple ni siquiera la promociona como la solución para los niños. La gente la usa para abuelitos que se pierden con el iOS nuevo. Pero los tíos y las tías que les dan el primer iPhone a sus chavos se dan cuenta: con esa función activada, el teléfono no se convierte en un juguete de 100 apps que se les cae al piso, se le apaga, y empieza a gritar. Es un teléfono de verdad, con las funciones de verdad, sin el ruido.
Why this matters for us: La abuelita y la tía que le regalan el iPhone a su nieto están usando sin saberlo la herramienta que Apple diseñó para simplificarle la vida — y ese mismo truco le simplifica la vida a los primos chiquitos, sin que nadie tenga que explicárselo.
Romance scams are no longer a Nigerian side hustle — they're a global industry
Carlos Barragán, author of The Yahoo Boys, is going live with Wired's Kate Knibbs to answer questions about the people behind the catfishing accounts that have drained billions from our abuelas' bank accounts.
The Yahoo Boys are the original internet scammers. They started in Lagos in the 1990s, sending fake love letters and investment pitches to Western women. Over the decades they evolved — from phishing emails to romance profiles, from phone scammers to AI-generated voices and deepfake video calls. What began as individual hustles is now a coordinated industry.
Barragán brings the on-the-ground perspective: he's tracked these operators from Lagos call centers to their overseas bases, mapped the supply chain of fake profiles and rented phones, and written about how these scammers exploit the same human instincts — loneliness, trust, the hope for something better — that we all feel. The Yahoo Boys didn't invent the con; they perfected it for the internet age.
Why this matters for us: las tías are still sending money to strangers online, and the scammers are getting better at the performance. Knowing how this industry works means knowing what to look for.