El hardware se vuelve la barrera — y la gente sigue adelante

A16z escribe que la ventaja competitiva de los próximos años no va a venir de modelos más grandes, sino de chips diseñados para correrlos más barato. La tesis es simple: las GPUs actuales —H100, A100— son genéricas y caras. Si alguien diseña un chip específico para la carga de trabajo de inferencia, el costo por token se desploma.

No es nuevo en teoría. Los ASICs de Google para BERT y los TPU de Google para Transformer ya lo mostraron. Lo que cambia hoy es la escala. La inferencia de LLMs consume más potencia que los centros de datos tradicionales. Y los modelos crecen —70B, 170B, 500B— mientras el costo de la GPU se queda atrás. Si tu modelo es el mejor pero tu chip es lento, pierdes.

La pieza de A16z conecta con lo que ya se está moviendo en el mercado: chips como el Groq, los designs de Cerebras, los chips de Tenstorrent y AMD Strix Halo para inferencia local. La carrera no es solo por el modelo — es por la máquina que lo corre.

Why this matters for us: los chips diseñados para lo que hacemos —inferencia rápida, bajo costo— van a bajar el precio de las herramientas que la comunidad usa a diario, y los que no se adapten van a quedarse pagando de más por lo que otros ya tienen más barato.

ISBNdb quietly bought a book to prove a point about training AI models. Then 404 Media caught the trail and told the story. The company scrambled to contain the damage: it yanked the AI pages off the site, denied it ever scanned or sold a book for training data, and called the whole thing a test of market interest.

The backpedaling is telling. If the book was only a test — a one-off to see if the market would bite — then the real play is the database itself. ISBNdb holds the ISBNs that keep the publishing world from collapsing into chaos. Feed that into a model, and you've got a cheaper, faster way to train AI without paying for a mountain of books. The 404 report exposed the move before the company was ready to admit it.

Why this matters for us: when the big tech companies quietly stockpile our books to train their models, we're the ones footing the bill without knowing it.

Apple's second-gen AirTag just hit $89 for a four-pack at Amazon and Target — $10 off the usual $99 and the all-time low. Costco members are getting five for the price of four.

The Gen 2 upgrade is small but useful. Apple swapped in a new ultra-wideband chip for tighter Precision Finding — the arrows in the Find My app track your item faster, and the Bluetooth range is wider. That matters in multistory homes where the old tag would drift between floors. The speaker is louder too, which means you can actually hear the tag in the couch cushions. And the CR2032 battery is replaceable; you won't trash the whole thing when it dies.

The travel features are the ones nobody mentions until they need them. AirTag sharing lets five people track the same item — the family can all see the luggage. Share Item Location passes your tag's coordinates to United, Delta, and American so the airline can flag it if it gets separated. You do still need to buy a separate loop for hanging it on a keychain — Apple never added a lanyard hole. But at $89, the four-pack is a steal.

Why this matters for us: $89 for four is cheap enough to strap one on every set of car keys, the abuela's purse, and the pickup truck — the kind of hardware that quietly solves problems for la familia without needing an app or a subscription.

Dan Nielsen is writing about what has happened to the look of AI-generated images. For a while, AI pictures had a particular sheen — a certain gloss, a certain softness, the way they looked like they were painted by someone who had seen too many art books. That look became so recognizable you could spot an AI image on sight.

Now, as the models get better, that look is fading. The new batch doesn't have the tell. It's harder to tell what's real and what's machine-made. And in that sense, the AI aesthetic is becoming its own genre — the way the Dutch Golden Age is its own thing, the way anime is its own thing. It has conventions. It has a history. It will have a place in art history.

Why this matters for us: as the line between human-made and machine-made blurs, the community we're building for — la gente who work with their hands, who value craft — needs to understand what they're buying when they buy the new tools.

Seven water treatment systems across the country have been hit by cyberattacks that investigators believe are linked to Iran. The attacks hit critical infrastructure — the kind of pipes and pumps that keep the water turning on for communities. If Iran is behind this, it means state-sponsored actors are actively targeting the systems we actually rely on, not just the tech companies everyone talks about.

The FBI is looking at AI-powered tools to catch these kinds of attacks before they spread. The idea is that pattern-matching algorithms can spot the telltale signatures of Iranian operations — the way they move, the timing, the ports they open — and flag them faster than humans can. Russia has also been busy, charging Telegram's founder, while xAI is fighting a state's nudification ban. The Democrats, for their part, are learning the hard way that getting scammed is part of the game now.

Why this matters for us: when water treatment systems go down, la gente wakes up without water. These attacks aren't just headlines for Silicon Valley — they're the pipes running under our neighborhoods.