Issue 66 — 2026-07-28
Brute intelligence — and why it's the real edge right now
Benn writes up a simple idea that's been hiding in plain sight: the models that are winning aren't the ones with the slickest architecture. They're the ones that brute-force the problem hard enough to make it work.
The trick is knowing what to brute-force. Running more compute on the wrong thing is expensive and gets you nothing. Running more compute on the right thing — the thing that actually breaks the ceiling — pays for itself. That's the difference between burning capital and building something that sticks.
What changed is the cost curve. Compute is cheap enough now that brute force is a real strategy, not a fallback for when you don't have a clever idea. The models that can afford to try harder, and the ones that know which direction to push, are pulling ahead. The rest are still arguing about which transformer variant is best.
Why this matters for us: la comunidad is full of gente who work hard and get by — the ones who don't have fancy degrees but show up and get the job done. Brute intelligence is what they do. The models catching up to that are the ones that'll actually be useful to people like us, not just to the folks writing about them.
The browser is the new OS, and Claude is taking over it
Lenny's writing about how Claude Opus is becoming a browser agent — not just a chat box, but something that actually navigates pages, fills forms, clicks buttons, and reads what's on screen. The key detail: it's not a wrapper around ChatGPT. It's built on Claude's own reasoning engine, which means it can plan multi-step tasks instead of just answering questions.
What changed is the interface. You used to open Claude, type a prompt, and get text back. Now you open Claude in the browser, and it is the browser — it sees tabs, reads DOM elements, and does things. For the working person who's juggling a dozen tabs between the bank, the gig apps, and the spouse's calendar, this is the difference between a smart chatbot and a real assistant.
The bigger picture: search is becoming less about finding links and more about doing things. Google's still the king of links. Claude is building the king of doing. The companies that ship agents that actually work — that can check your email, book a flight, read your pay stub — will eat the companies that just write blog posts about AI.
Why this matters for us: la gente who work with our fingers and our phones are the ones who'll benefit most — a browser agent that reads your pay stub, files your paperwork, and spots scams before they hit your account is worth more than any Silicon Valley demo.
Speculative decoding: el draft que acelera tu LLM sin perder calidad
Speculative decoding es un truco que hace que un modelo grande responda más rápido sin cambiar su respuesta.
El modelo grande —digamos un Llama 70B— es como un abuelo que piensa cada palabra antes de decirla. Es preciso, pero lento. El truco es meterle un ayudante pequeño (un modelo de 1–3B) que propone los próximos tokens de corrido. El modelo grande recibe la propuesta en un solo batch y verifica cada token a la vez — si uno no le convence, lo rechaza y sigue.
Es como cuando la mamá cocina para toda la familia: ella pone la carne al fuego y los primos cortan las papas al mismo tiempo. No se pisan; cada quien hace lo suyo. El resultado es el mismo plato, pero llega antes a la mesa.
Lo que hace este truco en la práctica: 7–10 tokens por segundo en una sola tarjeta — casi el doble — y, lo mejor, cero pérdida de calidad. Cada token es idéntico al que habría escrito el modelo grande solo; la diferencia es que viaja más rápido.
¿Cómo se activa? Con un flag en llama.cpp: --spec-default y el modelo draft que comparta el mismo vocabulario. En tu caja, el Llama-3.2-1B trabaja con el mismo tokenizer que R1, así que no hay sorpresas.
Lo que vale la pena recordar: el draft no adivina por arte de magia — propone, y el grande corrige. Si el draft se equivoca en un token, el grande lo descarta y sigue adelante. La respuesta final es siempre la del modelo grande, no del draft.
Tip: si tu servicio corre con flags viejos como --draft-max, va a fallar al iniciar. Usa los nombres nuevos — --spec-draft-n-max — o el sistema no arranca.
The big companies have an incentive to overstate the threat.
— wired.com
#silicon-valley-is-split-over-chinese-ai-and-the-gap-is-real-d2e874Phased design has no finish line — it just stops when it needs to
Phased design is a simple idea: ship the thing, then keep improving it in small slices instead of waiting for perfection. Each phase is small enough to test with real users. Each slice is a real improvement, not a cosmetic tweak.
The trap is thinking phases have an end. They…
Customer interviews when nobody's used your kind of thing yet
Jeff Gothelf has a post about a customer interview technique that sounds a little counterintuitive at first: instead of asking people what they want, you ask them to walk through a scenario they've never actually experienced and watch what they do.
The trick is in the…
How Tanay's robot data pyramid stacks up — and why it's the kind of piece that writes itself
Tanay's post is about the data pyramid for robot learning: the layers of data that go into making a robot actually do something useful. It's an ai_explainer — the kind of thing that helps you understand the craft of explaining things clearly, not a post about one specific tool or vendor.
What makes it worth writing about is the structure. The pyramid maps how raw sensor data gets filtered, labeled, and composed into something a model can actually use. Each layer narrows the signal and sharpens the answer. It's the kind of diagram you want to pin on the wall.
The piece is about the way we describe robot learning to people who aren't deep in the weeds. That's rarer than it sounds. Most robot posts either drown in acronyms or skim the surface. Tanay's sits in the middle — dense enough to be useful, clear enough to share.
Why this matters for us:
it's the kind of writing we should be doing more of — a clean map of something real, not a pitch for anything.
The robot data pyramid — why 3B beats 7B on real robots
Tanay Shah just put out a paper that flips the usual ML intuition on its head. He collected 24 hours of robot arm data — 6M images and 6M joint angles — and trained three models: a 3B transformer, a 7B transformer, and a 7B diffusion model. The 3B model won. By a lot.
The…
Google no castiga el contenido generado por IA
Ahrefs se comió 18 meses de datos de 100 mil sitios y descubrió que Google no baja el ranking del contenido hecho con IA. Lo que sí penaliza es el contenido de baja calidad — y ese se nota tanto si lo hace un humano como una máquina.
El mito de que Google te castiga por usar IA viene de la vieja experiencia de 2023 cuando los sitios llenaron sus blogs con texto genérico. Pero ese no era un problema de la herramienta, era un problema de la calidad. Y ahora que 18 meses de datos confirman que los sitios con contenido de IA no pierden visibilidad, la pregunta es otra: ¿estás produciendo contenido que la gente lee o solo contenido que Google puede indexar?
Lo interesante es lo que esto significa para los que trabajamos con la comunidad — si estás usando IA para explicar cosas en voz propia, con contexto real y sin inflar el conteo de palabras, Google ya no te discrimina por el origen del texto. Te juzga por lo que dice, no por cómo llegó hasta ahí.
Why this matters for us: la gente de la comunidad no nota si el texto lo escribió un humano o una máquina — nota si les sirve.
Zero-copy is the most expensive lie in modern systems
A new post from Iceberg Lake House — written by the team that builds the open-source columnar format — walks through a claim that keeps showing up in system design talks: zero-copy avoids copying data, so it is faster. The post says it is not. The copy is cheap. The cache…
AI is oil, not god — and that's a relief
A writer at Not Boring — the newsletter that actually cuts through the noise — makes the case that AI is oil, not god. A bit of a mouthful, but the idea is straightforward: oil is useful because it powers things. It burns in engines, heats homes, moves trucks. You don't pray to it; you use it. God, on the other hand, is the thing you worship and explain.
The piece argues that the industry has been treating AI like it's divine — a miracle that will transform everything by itself — when it's more like infrastructure. Oil doesn't care how many people talk about it. It just works in the right engine. And if you're building something for la comunidad, you want the thing that works, not the thing that gets the most press.
This is the kind of thinking that lands with people who have actually shipped products and seen the hype cycle eat their lunch. Oil is oil whether you're in Coachella or Palo Alto.
Why this matters for us: Brown Forces is already building with AI as the engine, not the altar — and this confirms the approach we're taking.
Brown forces: the craft of explaining things clearly
This is one of those pieces that explains how to think about writing — not about a specific tool, not about a vendor, but about the underlying craft. The article breaks down the kind of clarity that separates writing people actually read from writing that just fills space.
…
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.
Alibaba bans Claude Code — the linters are at the door now
Alibaba has classified Claude Code as high-risk and told its engineers to stop using it. The company's not balking at the model behind it; they're balking at the way the tool talks. Claude Code sits on top of the editor and sends prompts, diffs, and file contents to the API in real time. For a shop that treats source code as a proprietary asset, that's a leak in the hull.
The ban is a quiet warning shot for the rest of us. Every tool that writes code through an API — Copilot, Cursor, Claude Code, Codex — is quietly shuffling source out to the cloud. The linters are at the door. The question is whether you're using them or locking them out.
Why this matters for us: when the big shops start treating their code like company secrets, the rest of la gente who write code for a living need to pay attention — the tools that talk to the API are the ones moving the work, and the ones writing the prompts are the ones moving it out.