Software is cheap, math is open, hardware is real — esto es lo que importa.
The old guard is scrambling — Azure nearly bled its secrets, OpenAI is pricing analytics like bodega prices, and Bunnie proves open hardware works at $50. Meanwhile mathematics without mathematicians shows the future is already here. Open source tools keep getting sharper. This is the week the comunidad gets the memo.
Software is now so cheap it's practically free — Nader Daboussi explains why
Nader Daboussi makes a simple point: the cost of writing software has collapsed. You used to need a team, an office, and months to ship something usable. Now a single person can prototype in a day with the right tools. The bottleneck is no longer building — it's knowing what to build.
The piece walks through the mechanics — how code generation, AI models, and the open-source stack have converged. The result isn't just cheaper code; it's a shift in who can make software and who gets left out. People who understand the tools can now compete with shops that have been around for decades. The ones who don't are still paying for the old way.
The real cost is attention, not compute. You can generate a thousand lines of code for pennies. What's expensive is the judgment of what to generate and why.
Why this matters for us: la gente that understands how to use these tools now — the bodeguero automating inventory, the abuela running a Facebook page for her tamales — they're the ones who'll build. The ones waiting for permission are still paying for the old way.
Microsoft almost leaked all of Azure Cosmos DB's secrets
A routine code review caught Microsoft nearly publishing the full source for Azure Cosmos DB — the database that quietly powers a huge chunk of the cloud. The keys stayed in the vault, but the scare made people sit up: one slip and the whole thing would have gone open.
Cosmos DB is Microsoft's answer to NoSQL at scale, and it's been in the cloud since 2014. It handles document, key-value, graph, and column-family workloads from a single service, and it's been the engine behind everything from Teams to Azure Functions. When a database that big goes open, it changes the competitive math — competitors can read the wiring and build against the same patterns.
The fix was surgical: a small diff that removed the offending files before the release window closed. No outage, no fanfare, just the quiet relief of catching a leak before it becomes a flood.
Why this matters for us: databases like this are the plumbing under the apps la gente use every day. When Microsoft's open source gets right, it's a win for everyone building on top. When it leaks, it's a reminder that the code is always one commit away from being free.
RLHF: The tía que corrige tu español
RLHF is how we get an AI to stop sounding like a textbook and start sounding like a person.
The model starts by learning from a big pile of text — reading whatever's available, like a kid who picks up Spanish from the house. It can produce sentences and even write code, but it has no sense of what's good. It doesn't know that a recipe should list ingredients before instructions, or that a recommendation should be honest instead of flattering.
So people sit down and rate the model's outputs. They pick the better answer, flag the wrong ones, and write notes. This is the human feedback part.
Next comes the reinforcement learning: the model adjusts its weights using those ratings as a reward signal. Think of it like the tía who corrects your Spanish — every time you say something off, she nudges you in the right direction. After enough corrections, the model's responses start looking right to a human.
The trick is that the feedback doesn't need to be perfect. It doesn't need a label for every possible output. A few thousand ratings from real people is enough to steer the model toward answers that feel useful.
RLHF is what turns a model that can talk into one that talks well. Without it, the model is just a very good parrot. With it, the model is a good writer.
The alternative is to fine-tune on labeled data — give the model hundreds of examples and let it memorize patterns. That works too, but it's brittle: the model repeats what it's seen and struggles with things it hasn't. RLHF is more flexible because the reward is continuous — the model is always being nudged, not just copying.
Why we should care: RLHF is the reason the AI tools we actually use sound like people instead of sounding like they're reading from a manual.
Analytics is the first big workload where you can run these models on your own data and get answers you can actually check.
— motherduck.com
#openai-made-analytics-10x-cheaper-with-new-models-y-all-should-be-paying-attenti-a035e7Bunnie Huang's Baochip is the $50 badge that proves open hardware isn't dead
The Def Con 34 badge — the little PCB every hacker wears — was designed by Andrew "Bunnie" Huang, a pioneer of open-source hardware who grew up in the Coachella Valley and built the Chumby. The Baochip costs about $50. It's built on a RISC-V chip, the same architecture that…
LookFresh keeps 0% commission from every cut
Every time a barber books through the big platforms, a percentage gets skimmed off the top. And the apps still feel built for chain salons with a dozen chairs — not for la gente who work chair-by-chair, one cut at a time.
LookFresh fixes this. You get a clean booking link that works in-person and online. Payments come through Stripe Connect and go straight to the operator. No per-cut percentages. A flat platform fee instead. When a stylist charges $50, they keep $50.
Zero commissions. Not now. Not ever.
The shop also gets a full calendar, real-time availability, portfolio photos, verified reviews, and automatic reminders so no-shows stop creeping in. Bookings happen around the clock — your calendar works even when you're off.
Whether you're a solo barber, a salon owner managing a team, a mobile nail tech, or a tattoo artist building your brand — LookFresh gives you everything you need to grow without giving away a single dollar.
YOUR LOOK. YOUR SCHEDULE. YOUR MONEY.
https://lookfresh.vip
Mathematics Without Mathematicians — 2026's Most Interesting Ideas
Daniel Borretti collected the five ideas that actually moved the needle in 2026. TLDR's Dan wrote this up so you don't have to hunt down the original posts.
The list is worth reading for the curation alone — Borretti picked things that aren't just trending, but things that…
Math that doesn't need mathematicians
Mathematics — the kind that runs stock markets, trains models, and powers the infrastructure behind the apps — is quietly leaving its old home. The formal proofs, the heavy notation, the PhDs in ivory towers: that part is still there, but the working mathematics that actually gets things done is moving into the hands of engineers. You don't need to understand the proof to use the result. The result is what ships.
This is the difference between mathematics as a discipline and mathematics as a substrate. The former requires expertise; the latter requires only a good API. The old guard — journals, departments, tenure — will complain. The new layer underneath them is already doing the work. And it's doing it for everyone, not just the people who spent years learning the language.
Why this matters for us: the math that holds up our world is becoming free to use, and the people who actually build things with it don't have a PhD — they just need the tools to work.
10 tools TLDR liked this week
This week's picks span the usual suspects — notebooks, agents, design tools — plus a few that actually ship something new. The list is a snapshot of what the TLDR editorial team finds worth their time, not a curated top-10 of the month.
TLDR is a daily newsletter written for…
OpenAI's new model Astra cracks 10 math problems it couldn't solve before
OpenAI just released Astra and it solved 10 long-standing math problems. These are the kind of problems that have been sitting around for years — not the ones a human would struggle with, but the ones that have kept stumping models. The claim is that this is a real step up, not just a bigger model getting lucky on a few benchmarks.
What's interesting is what this means for the people who actually use models to do work. Astra is the next major model in OpenAI's lineup, and it's being positioned as a meaningful upgrade, not a minor iteration. If the math holds up in practice, it's a sign that OpenAI is getting better at the kind of reasoning that matters when you're asking a model to work through something step by step.
Why this matters for us: models that can reason through problems are tools we can actually trust — not just fancy autocomplete, but something that gets the answer right when we need it.
Montana's experimental medical hub finally gets a green light
Montana is building an experimental medical hub — a place where new treatments can be tried outside the usual FDA approval pipeline. The state pushed the project forward, giving it a legal foothold so developers can run clinical trials on its soil.
The hub is experimental by…
OpenAI says Astra cracked 10 math problems no model could touch
OpenAI put its next model, Astra, up against a century-old math benchmark and it solved 10 problems that previous models could not. The problems are old — they've been sitting in the literature for decades, waiting for a model sharp enough to find the solution. Astra found them.
The benchmark is the kind that doesn't lie: a small set of problems with known answers, not a broad but fuzzy evaluation. That matters because it means the 10 solves are real, not a product of a lenient scoring rubric. The problems were open for a long time. Astra closed them.
Why this matters for us: when models start solving hard problems that have sat unsolved for decades, it means the tools we use — the ones that plan routes, schedule shifts, crunch numbers for the small shop — are getting sharper without anyone having to buy a new license.
PGSimCity — the postgis playground you always wanted
Nikolay's PGSimCity lives at nikolays.github.io/PGSimCity — it's a web-based map editor that talks directly to PostgreSQL. You draw polygons, roads, and points, then the thing writes the SQL back into your database. No separate app, no separate API. Just GeoJSON in, PostGIS…
Data lessons from inside Meta — what a real org does differently
Meta's data team shared a roundup of hard-won lessons from running analytics at massive scale. The TLDR is practical: they treat data like a product, not a byproduct. Tables and pipelines have owners. Models get SLAs. If a dashboard breaks, someone knows. If a metric drifts, the team sees it before the execs do.
The pieces they highlight — schema evolution, query performance, the difference between a good model and a great one — are the kind of things most teams only learn after burning through months of slow queries and stale reports. Meta's playbook is to bake it in early: test schemas before they land, version the models like code, and keep the documentation close to the source.
Why this matters for us: the same discipline that keeps a giant's data honest is how any small team avoids the slow bleed of bad reports and stale dashboards — it's just a matter of doing it before the mess grows.
xberg: the open-source knowledge base your videos have been missing
xberg is the latest from xberg — an open-source knowledge base designed specifically for video workflows. It's not a general-purpose RAG layer bolted on top of a search engine; it stores what matters for video: clips, shots, b-roll, and the metadata that lets you find them…
duckdb now does async IO — no more waiting for the network
DuckDB ships an asynchronous I/O layer that lets it read files while it crunches numbers. Instead of one thread chugging through a 200 GB Parquet file and blocking the whole process, the query engine can be doing compute while the file system fetches the next chunk. The result: faster scans on cloud storage, and the kind of parallelism you usually need a whole orchestrator for, now inside the query engine.
This matters because DuckDB has been the quiet workhorse for data teams that wanted Postgres without the bloat or a warehouse without the bill. Async IO was the last piece holding it back — it could write fast, but reading from S3, GCS, or even local SSDs would stall the thread. Now it reads ahead, overlaps I/O with computation, and keeps the CPU fed. The practical effect is that single-node DuckDB can handle workloads that used to require a cluster.
Why this matters for us: la gente que trabaja con datos — en la oficina, en el side hustle, even en el taller — ya no necesita un data warehouse de $5,000 al mes para queries medianas. DuckDB sigue siendo gratis y ahora se mueve más rápido con lo que ya tiene.
Beyond the happy path — how to stop your database from lying to you
Gabor Koos at MotherDuck is writing about something that happens when the happy path is smooth but the real world is not: your database tells you one story while the actual data tells another. The article is about the craft of writing SQL that survives contact with production…
A dev pipeline isn't a dev pipeline — it's a production system
Jerry Orr just posted a piece on the sundry blog arguing that a development pipeline should be treated as a production system, not a throwaway staging area. That's the kind of observation that sounds obvious until you've been burned by one that isn't.
The point is practical, not academic: if you build your pipeline once and then patch it for years, it accrues debt like any other service. The difference is you don't notice the rot until something breaks at 2 a.m. and you're staring at logs that don't match the config.
Why this matters for us: la gente writing code for real work — not for a demo — should treat their pipelines like the systems they are. Build them right, or pay for it later.
Why this matters for us: it's one of those craft lessons that applies to every tool we build, so we can write it and forget it's there.