La célula que nace sola, y los modelos que se cansan
Las noticias de hoy van por la vida: una célula hecha a mano crece y se divide sola, el plasma genera electricidad para la red, y los chatbots muestran que el cansancio no es solo humano. En el fondo, lo que importa es que la gente construye sin intermediarios — una caja, dos productos, privacidad intacta.
A cell built from scratch grows and divides
Scientists at the J. Craig Venter Institute have built a cell from scratch — not by editing a living one, but by writing the genome like code and assembling the pieces. The cell was placed in a growth chamber and started dividing. It lived. It reproduced. This is the first time a fully synthetic genome has been put into a living cell and seen it do the one thing that matters: divide.
The previous synthetic-genome work in 2010 was a milestone, but the new cell is different. It is not a copy of an existing cell with a swapped-out chromosome. It is a cell whose genome was designed, synthesized, and assembled — a 1.1-megabase genome — and then dropped into an empty shell. It breathes, it eats, it splits into two.
The Quanta piece frames this as a proof of concept for synthetic biology writ large. If you can write DNA the way you write software, you can start designing cells for specific jobs — ones that break down plastic, that produce insulin, that repair themselves. The bottleneck is no longer whether we can build a synthetic cell, but whether we can design one well enough to be useful.
This matters for the comms and AI hardware box. The same kind of logic — writing a genome, assembling it, watching it work — is what the chip-routing model does for silicon. Both are about encoding rules into a substrate and letting the substrate do the work. The cell is the biological version of a routed die.
Why this matters for us: it shows synthetic biology is moving from lab curiosity to engineering discipline, and the same pattern — design, assemble, observe — applies to chips, cells, and the models that guide them.
Fusión real: la gente genera electricidad del plasma y la lleva a la red
La instalación de Realta en el sur de Inglaterra ha logrado algo que parecía imposible: convertir la energía de una reacción de fusión nuclear directamente en electricidad, sin pasar por las turbinas de vapor que se usan en todo lo demás. Es la primera vez que pasa — la fusión lleva décadas prometiéndonos esto.
El plasma está a millones de grados. En lugar de calentar agua y mover turbinas, los campos magnéticos de la máquina capturan las partículas cargadas y las convierten directamente en corriente eléctrica. Menos piezas móviles. Menos pérdidas. El mismo principio que usan los generadores de naves espaciales, pero aplicado a escala industrial por primera vez.
Esto no es un experimento de laboratorio: la máquina está operando y la electricidad ya está en la red. Si el diseño escala, la fusión se vuelve más barata de construir y más fácil de mantener — lo que lleva décadas siendo el problema.
Why this matters for us: la gente que paga la luz y las empresas que buscan energía limpia por fin tienen una opción que no depende de baterías, de agua, o de que el viento sople.
MoE: No es un modelo gordo, son muchos especialistas
MoE es un modelo grande que no trabaja todo a la vez. Piensa en una cocina mexicana: cuando llega la gente, no hace todo el chile en un solo caldero. Tiene uno para el mole, otro para el adobo, otro para la salsa. Cada uno es experto en lo suyo, y el jefe de cocina decide cuál llamar para cada platillo.
Un modelo MoE hace lo mismo con sus neuronas: tiene varios grupos pequeños llamados "expertos". Cuando recibe un prompt, activa solo los expertos que le importan — no los 70 mil millones a la vez, sino quizás 20 millones. El resto se queda callado.
Esto no es solo eficiencia. Es velocidad. Es que el modelo gasta su fuerza donde la necesita, como cuando tu tía no hace tres días de tamales y uno, sino tres tamales distintos, y cada uno sabe a lo que debe saber. No pesa más. No cuesta más. Solo trabaja mejor.
La próxima vez que uses un modelo MoE, fíjate si responde más rápido con la misma calidad. Si lo hace, es por esto.
}
The product is the model, not the wrapper.
— links.tldrnewsletter.com
#the-chatbots-are-tired-and-they-re-showing-it-90f35bThe new bet: AI routing, not models, wins the next phase
Tom Tunguz is laying out a thesis most founders are already living but haven't named it yet. The argument is simple: the models are getting commoditized, and the real moat is routing — the work of deciding which model does what, when, and why. Not the model itself. The…
Una célula construida a mano por primera vez crece y se divide sola
En el Quantum Magazine sale la noticia de que por primera vez alguien armó una célula desde cero — cada componente puesto a mano, no sacada de otra célula — y la puso a vivir. Respira, crece, se divide. No es un experimento de laboratorio con un frasco; es un cuerpo entero…
Local reasoning for global properties — how to prove the whole from the parts
Laurie Tratt has a new post on proving global properties — things that depend on the entire program — using only local reasoning about individual pieces. The trick is writing assertions that are compositional: each piece carries its own guarantee, and when you compose pieces the guarantees stack up automatically. No global state, no hand-waving.
The practical payoff is that you can reason about a module in isolation and be confident it fits into anything. That's what copybara does when it checks that a migration preserves invariants across the whole repo. You write the checks once, and they hold for every pipeline run.
Why this matters for us: it's a clean way to build tools for la comunidad — prove things once, reuse them everywhere, so the code doesn't rot under the weight of edge cases.
Build without the handoffs — one box, two products, privacy intact
Obsidian Comms is a private communications server that sits on a small box in your building. It handles video calls, voice calls, and encrypted messaging — end-to-end encrypted so the server itself can't read your messages or hear your calls. No cloud, no third-party servers,…
AI companies are maxing out tokens just to squeeze more out of models
An Elastic research paper shows how much extra tokens models are gulping down as the industry chases better answers. The pattern is real and the cost is real: companies are sending 3–4× the tokens a model needs for the actual work. That bloat isn't free. At scale it can be the difference between a model fitting in memory and needing a second GPU.
The trick is tokenmaxxing — the kind where you pad the prompt with extra context, chain multiple reasoning steps, or send the same question twice to different models and compare. It often improves accuracy, but it also means you're paying for words that don't need to be there. The paper tracks this across several models and finds the overhead is nowhere near uniform. Some models eat tokens like they're free. Others choke on it.
Why this matters for us: if you're running models on anything but a data center — on a VPS, in a small team, or for a local business — token waste is a real line item. The fix is usually trimming prompts, not buying more GPU.