La migra se mueve: chips, IA y la infraestructura real
Microsoft, Ruben Llorach y la nueva capa de routing se pelean por lo mismo: que la IA funcione de verdad, no en slides. Los chips se rediseñan con modelos chicos; la infraestructura migra a datos locales; los chatbots se vuelven viejos. Esto te toca a vos, primas, porque el hardware y el software que compra tu empresa ya lo están poniendo.
La nueva apuesta de Microsoft: $2.5B para meter ingenieros de IA dentro de tu empresa
Microsoft gastó $2.5 mil millones en una empresa frontier y ahora quiere meter ingenieros de IA dentro de las compañías que compran su nube. No es un producto — son personas. La idea: que el equipo de IA de la startup trabaje lado a lado con el equipo de ingeniería del cliente, no solo entregando APIs que se instalan y se olvidan.
Esto es diferente a los agentes autónomos que se prometen por todos lados. Son ingenieros — gente que escribe código, revisa PRs, se sienta en la misma sala o Zoom con tu equipo. Si funciona, es el modelo que realmente escala: la IA no reemplaza al ingeniero, lo multiplica. Si no funciona, es otro contrato caro con un logo bonito.
Porque al final, lo que importa para la comunidad Brown y Black que está construyendo startups en el Valle y en el Eastside — la gente que trabaja con Azure y AWS todos los días — es si este modelo reduce el gap entre promesa y realidad. Porque los ingenieros que se sientan dentro de tu empresa son los que realmente hacen que la tecnología funcione.
Why this matters for us: los ingenieros de la comunidad que construyen con Azure y AWS se van a sentir el efecto directo — es el modelo que decide si la IA es un gasto o un equipo que trabaja contigo, no contra ti.
Chip routing is one of the hardest problems in chip design — and a tiny AI is finally beating the big ones at it
Tom Tunguz writes about chip routing, the 3D puzzle of laying thousands of wires between metal pins without them crossing or doubling up. Chip companies have spent billions on software to solve it. That's the kind of problem most people assume needs a big AI — the ChatGPT-scale ones — but the BFTS experiment shows a tiny model, one-hundred-thousandth the size, is doing it better with the right training recipe.
The proof is in the numbers. A small model started solving only 12 puzzles out of 5,008. After 30 minutes of training, it hit 85 — seven times better. For the first time ever, it solved a puzzle completely, all six wires in place. A bigger version of the same model broke at the starting line, but the diagnosis is clear: a known configuration bug, not a capacity problem. A $15 test is running now to confirm the fix.
The bet here is real. BFTS is arguing that small focused models, trained with the right objective, beat giant general ones at specialized hard problems. The 7M model is going open-source. The 35M is cooking behind the scenes with a new RL objective — it's stopped copying demonstrations and now plays 3D Tetris against itself, finding wire paths it wouldn't have discovered by imitation. This is the difference between memorizing and reasoning.
Why this matters for us: the next wave of AI isn't about bigger models — it's about the right model, trained right, doing real work in the world, and the open-source gate to that is opening now.
No consulting. Not support tickets. Actual engineers, sitting with the teams.
— geekwire.com
#microsoft-puts-ai-engineers-inside-your-company-1a7da2Smart model routing is the quiet infrastructure shift of 2026
Gergely Orosz just published a piece on The Pragmatic Engineer's blog about how teams are routing AI requests across models instead of hardcoding everything to one. The idea is simple: send cheap queries to a cheap model and big jobs to a beefy one, and let the routing layer decide which is which. It's not a flashy product — it's plumbing — and that's exactly why it matters.
The pieces already exist. OpenAI, Anthropic, Google, and the open models all ship APIs that accept the same input and return the same output. The routing layer sits in the middle, watches latency and cost, and picks the best model for the job.
Ruben Llorach's engine can design chips — and it learns fast
Ruben Llorach is building a creative engine with Claude that can take a chip design and lay out the routing — the wires that connect every pin — all in 2 seconds on a single RTX 5090. The model is only 34M parameters but it learned 3D spatial reasoning from examples. It doesn't need hand-written rules; it learned the geometry, obstacle avoidance, and wire-following by watching examples.
What's striking is that the hard part of chip routing — coordinating hundreds of nets across multiple layers without collisions — is the same problem that makes EDA software cost $50k a year.
Chatbots are getting old — and the next wave is already here
Simon Willison is writing about what happens when the novelty of chatbots wears off and the real work begins. The piece tracks the shift from chatbots as a novelty to chatbots as infrastructure — the way we used to talk about APIs, and then they became invisible. The ones that survive are the ones that stop trying to be chatty and start doing actual work.
It is a quiet piece. No hype, no big claims. Just an honest look at the current state of the technology and where it is heading. The kind of thing that matters more than the announcements.
Why this matters for us: this is how the tools we use will change — and the ones that matter for our communities are the ones that stop talking and start doing.
We like Anthropic more than OpenAI
Kristen Berman writes why she's picked Anthropic over OpenAI — and she doesn't do it lightly. She's been around the block. Her point is practical: Anthropic's model is more honest about what it knows and what it doesn't. OpenAI's models are bigger, flashier, and more willing to bluff. For work that matters — the kind where a wrong answer costs you — Berman says you want the one that tells you when it's guessing.
The piece sits in a longer conversation about trust in AI. OpenAI's gpt-4 was the first model that felt useful for most people. Anthropic's claude has been the one that feels honest.
Stop the ai confidence theater on LinkedIn
Diandra Escobar is calling out the LinkedIn performance art where everyone sounds like they've been using AI for years — and the copy is starting to show the seams.
Her point is simple and sharp: the posts that sound confident but say nothing are the ones that signal the writer's field is full of people performing expertise rather than doing it. The confidence is theater. The substance is thin. When the copy starts to sound like the same template with different nouns, that's the tell.
This is a useful filter. If you're writing about AI for a Brown audience and you find yourself saying the same things as every other voice — "the future is here," "transform your workflow" — you're probably doing the theater. The fix is to say one concrete thing and say it plainly.
Why this matters for us: a lot of the AI noise is just people pretending to know what they're talking about — la misma cosa, different font. If we cut through that, the real signal gets louder for la gente who actually need it.
Data residency is an infrastructure problem, not a legal one
The piece at Hacker Noon — the one TLDR Data flagged — makes a clean point: keep your data where you want it, and design your stack to match. Data residency isn't a compliance checkbox. It's a plumbing question.
The writer calls this out in a piece the system marked aiexplainerworthy. The gist is simple enough. A lot of teams treat residency like they're filing paperwork for a lawyer. In practice, it's about where the bytes live and which network they cross. Pick your region. Route your writes. The rest follows.
Chrome finally ships the <usermedia> element for camera and mic access
Chrome has landed the <usermedia> element — a native, declarative way to grab camera and microphone streams without JavaScript. Instead of the old navigator.mediaDevices.getUserMedia() boilerplate, you write a single tag and the browser handles the permission prompt, the stream, and the cleanup. It's the kind of thing that sounds small but saves developers a lot of fiddly code.
The element lets you set mediaType to 'video' or 'audio', point it at a specific device, and even restrict which track to use. Permission is requested once and cached, so the user doesn't get pummeled by prompts on every page load. If the camera is unplugged or permission is denied, the element stays in a sensible degraded state rather than breaking the layout.
Why this matters for us: la gente's phones have cameras — now websites can use them without the old JS mess, and that means cheaper, simpler tools for our communities.