Varianza y el futuro — de la oficina a la comunidad
Hoy toca lo que la gente ve y lo que no: la varianza que se esconde en los números, cómo la gente la reduce, y por qué la oficina se siente distinta. De Vercel a la ruta que no se muestra, de fósiles de salamandra a cómo crece la comunidad cuando se sabe mirar.
Vercel cut 10 SDRs to 1 with Claude for $5,000 a year
Vercel's COO Jeanne Dewitt Grosser laid out a new way to run sales teams: one person using Claude to do the work of ten sales development reps. The whole thing costs about $5,000 a year — not the $1.2 million in salaries you'd expect for that headcount.
The mechanism is straightforward. Claude handles the cold outreach, the qualification, the scheduling, the follow-ups. The one person reviews and signs off. It's not a chatbot copy-paste — it's an agent working through the SDR playbook end to end.
This matters because the old story was always about AI replacing knowledge workers by writing emails and summarizing meetings. The real shift is happening on the revenue side. If a sales team can be a person and an AI at half the cost of three people, the economics of customer acquisition change. And the people buying this stuff — the 50-person startups, the family shops, the primos running side businesses — they don't need a 10-person SDR team. They need one person with a good memory and a sharp pen.
Why this matters for us: the tools are finally cheap enough that our own companies can run sales without the bloat, and Claude is the engine doing the heavy lifting.
AI broke the junior dev market. The fix is slow
TLDR's latest post (by Seldo) tracks what's happened to entry-level programming since ChatGPT arrived. The short version: AI tools ate the bottom of the stack. Tasks that used to go to a junior—boilerplate, CRUD, scaffolding, writing tests, turning a spec into a working app—are now handled by models for a few cents. Junior headcount shrinks. The jobs that remain are the ones a model can't do alone: understanding a vague client request, debugging something the model got wrong, and shipping code that fits the rest of the system.
The interesting part is how the market is adjusting. Salaries at the bottom haven't collapsed; they've just stopped growing. Companies aren't firing juniors wholesale—they're hiring fewer of them. The roles that survive pay a premium for the judgment call: knowing when to let the model do the work and when to override it. Meanwhile, the tools themselves are getting better, which means the bar keeps rising. What was junior work last year is already shifting toward mid-level.
For the Brown folks who are learning to code or already in the trenches, this isn't bad news—it's a signal. The apprenticeship model is still working; the apprenticeship just looks different now. You don't need to be the fastest at writing code; you need to be the one who can read the output, spot the mistakes, and make the call. That's the work a model can't do for you, and it's the work that will keep paying.
Why this matters for us: the kids learning to code in our communities still have a path forward—the jobs are changing, not disappearing, and the ones that pay well are the ones that require judgment, not just speed.
Measure the gap between what someone could do before and what they can do now.
— links.tldrnewsletter.com
#how-to-actually-measure-whether-learning-happened-162c59Varianza que la gente no ve — y cómo la reducen
Los datos vienen en granos — filas, usuarios, transacciones. La varianza se esconde entre esos granos: el ruido que hace que un experimento parezca ganarlo todo cuando en realidad es puro azar.
La técnica clave es tomar la varianza por granos y promediarla — no la varianza total, la que vive dentro de cada fila. Si mides mal el grano, mides el ruido. Si lo aciertas, ves la señal real. Es el tipo de corrección que parece menor hasta que deja de mentirte.
Por qué nos importa: los experimentos que decimos que "funcionan" a menudo son ruido disfrazado.
Anthropic puts Claude's science tools in a single workbench
Anthropic shipped Claude's Science AI Workbench today — a suite of tools for research that live under one roof. It bundles Claude Research (the long-form reasoning model) with a notebook-style interface, code execution, and a search layer over arXiv and other papers. You can ask it to read and compare papers, then run the code it writes to verify results — all without leaving the page.
The workbench is built on top of Claude Sonnet 4.5, which Anthropic says handles the heavy lifting for reasoning-intensive tasks. The model can chain multiple steps — search, read, write, execute — without the user having to click through each one.
Open-source maintainership is changing — and copilot is part of it
GitHub Copilot just got better at using tokens — the way it charges — without losing quality. The update means the same model output now costs less per line of code. That sounds small on paper. It isn't.
Token efficiency is the quiet engine of the AI era. Every model charges per token, and every project pays with it. When copilot trims fat, it trims cost for everyone using it. More importantly, it trims the friction between what a maintainer writes and what the model suggests. The result is faster reviews, fewer round-trips, and less context window wasted on noise.
This isn't just a copilot story. It's a signal for open-source maintainership in the age of AI. The people keeping repos alive are now writing alongside models that charge by the token. Projects that get the token math right can afford more help without going over budget. Projects that don't will feel it in every PR.
Why this matters for us: the tools that make code cheaper to produce also make it cheaper for Brown teams, side businesses, and the auntie who just opened a shop — they get the same AI help without the bill getting fat.
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.
AI broke the junior dev market — now it's rebuilding itself
Sam Altman said it bluntly: the market for junior programmers got torched by AI. The posts from the people actually in the trenches back it up — the folks doing the work are seeing what the models do to entry-level jobs, and it's real.
This isn't a one-off blip. AI has been eating the entry-level grind for a while now, but the damage is finally showing up in the numbers. The entry-level market didn't just get smaller; it got reshaped. What used to be the ladder — junior roles that taught people the ropes — is being replaced by tools that do the work directly.
The piece is worth reading because it's not written by someone who sells AI for a living. It's by someone who writes about tools and uses them, and the analysis holds up. The piece is categorized as aiexplainerworthy — about the craft of explaining things clearly, not about a specific tool or vendor. That's the kind of thing that survives a few tool cycles.
Why this matters for us: la gente who get their start in junior roles are getting squeezed, and the paths that used to work are changing — which means the training, the apprenticeships, and the mentorship we've built through Brown Forces become the difference between getting left behind and getting ahead.
Axolotl fossil found — the first fossil salamander of Mexico
Scientists have identified Ambystoma quetzalcoatli, a new fossil species of axolotl, in Mexico. It's the first fossil salamander formally described from the country — not just a fragment, but a species. The name honors Quetzalcóatl, the feathered serpent. The fossil itself sits in a museum; its real impact is on the timeline. Axolotls now have a documented presence stretching back millions of years.
Axolotls are famous for their neoteny — they keep their gills, live in water, and never metamorphose. That makes them weirdly hard to match with fossils, which is partly why this identification matters: it shows the lineage is old and geographically rooted, not just a lab curiosity. Finding a fossil of one in Mexico means the animal's story is tied to the land, to the basins and lakes that held it long before scientists ever scooped one up for a pet shop.
Why this matters for us: it's a reminder that things we think of as strange — the gilled salamander, the cousin who looks different — have a deep place in this country. The axolotl is Mexican. The fossil proves it. La gente already knows that; the science just caught up.