Issue 38 — 2026-06-19
You got faster, your company didn't
A new piece from TERRESTRIAL SOFTWARE is making the rounds, and it's the kind of thing that hits different when you're the one actually doing the work. The basic claim is simple: over the past few years, most of us got faster. Better tools, AI, smarter workflows — the list is long. But our companies didn't pass the gains back. No raise. No shorter week. No less grind.
The piece is worth reading because it names something a lot of us have felt but couldn't quite pin down. Your inbox is fuller, your Slack is louder, but the output is sharper and the work is quicker. You're doing more in less time, and your boss is happy about it. Meanwhile, the leverage that was supposed to go your way is staying right where it started — in the spreadsheet.
This fits a wider pattern across tech and beyond. Productivity gains have been real, but the distribution has been off. The companies that got better at extracting work got richer. The workers got used to moving faster without getting paid to. This is the quiet part of the story about AI and automation that doesn't get enough attention — it's not just about what technology does, it's about who gets to keep what it does.
Why this matters for us: when we're the ones doing the lifting, the gains don't automatically become ours — and the people who control the work decide where the rewards go.
Anthropic Fixes Claude's Token Burn and Shakes Up the Design
Anthropic shipped a major update to Claude this week, and it's actually fixing something most people noticed but couldn't quite name. The AI was burning through tokens — the unit that measures how much Claude consumes — faster than it should. Now it's more efficient, and the whole interface got a redesign.
The update adds design system imports, meaning Claude can now pull in your existing design files and work with them more naturally. Code round trips are smoother too, so if you're writing code and want Claude to refine it, it doesn't have to re-read everything from scratch. The whole thing is a design overhaul, not a patch.
Claude is becoming one of the most widely used AI tools. When a tool like this gets more efficient, it's not just a tech fix — it's your tool working faster and cheaper, so you can get more done in a day. Esto te toca. Why this matters for us: Every time a major AI tool gets more efficient, it becomes cheaper and faster for working people to use it in their daily hustle — whether that's drafting a message, reviewing a document, or building something from scratch.
Less hardware, more roads — that's the play.
— arstechnica.com
#mobileye-is-launching-robotaxis-in-the-us-e65314How Heidi Health is teaching clinical AI to stop guessing
Heidi Health is rolling out a clinical AI model that's been fine-tuned on real patient encounters instead of generic text. The company says the model now handles clinical documentation with more accuracy because it was trained on actual doctor-patient conversations — not just…
Obsidian AI: Private LLM that stays on your network
IEPs, medical records, legal discovery — you know the drill. The rules say this stuff can't leave the building. So when your team wants to draft, summarize, translate, or look something up, they hit a wall.
Most "private AI" still phones home to the cloud anyway. And the fully-local stacks that promise no outbound calls? They're research projects. Fun to read about. Not ready for the office.
Obsidian AI is different. It's a turn-key appliance — GPU, model, agent runtime, voice, and a hardened admin console — that you drop on your own network. Everything runs in-house. No outbound calls. No data leaving your building. Same toolbox surface as BFTS Chat, but the data and the brain never leave the room.
Why this matters for us: School districts, clinics, county legal — we're the ones stuck between the rules and the work, and this puts the tools where we already are.
https://brownforces.io/solutions
The repo that catches data mistakes before they cost you
Bjones has put together a GitHub repo called "Data Pitfalls" — a collection of common mistakes people make when working with data, laid out clearly so you can spot them early. If you've ever stared at a spreadsheet and wondered why the numbers don't add up, this is your spot.…
Alex Ellis is betting big on Loops
Alex Ellis is writing a new post on his blog arguing that Loops—the product he's been building—is the play of the moment. He's been thinking about Loops for a while now, so this isn't a sudden revelation. It's a steady accumulation of conviction.
Loops is what happens when someone who actually builds things decides to build their own solution instead of patching together someone else's tools. That's the kind of thing that catches fire when the alternatives keep getting more expensive and more complicated.
Why this matters for us: when a builder like Alex starts calling the shot, the tools shift, and so does who gets to compete on their own terms.
Anthropic gave Claude a major design overhaul
Anthropic shipped a big update to Claude's design system — the visual and interactive pieces that shape how you actually use it. New imports let you pull in components more cleanly, and Claude now handles code round-trips, meaning the model can write code and then read back…
Mobileye Is Launching Its Own Robotaxi Service in the US
Mobileye, Intel's vision and self-driving subsidiary, is entering the US robotaxi market directly — not just selling sensors to carmakers, but running its own service.
The company's been building the pieces for years: cameras, radar, and its SuperVision driver-assist system that's already in cars from Ford, BMW, and Nio. Now it's flipping the model. Instead of licensing that stack to OEMs and waiting for them to build robotaxis, Mobileye is operating one itself. That's a different kind of bet — you have to run the fleet, handle the maintenance, and deal with regulators on the ground.
This isn't the first shot at robotaxis (Waymo and Cruise have been at it for years), but Mobileye's approach is notably different. They're going standalone instead of riding on someone else's platform. Whether that pays off depends on a few things: can they keep the sensors cheap enough to scale, and can they actually run a fleet that people trust.
Why this matters for us: if robotaxis ever roll out in our neighborhoods, the companies that own the technology on the ground — not just the ones with the fanciest branding — are the ones that set the rules for pricing, access, and whether our communities get left behind.
Google speeds up AI responses with prefix caching
Google just rolled out prefix caching for its GKE Inference Gateway, a feature that stores repeated API responses so the same queries don't have to hit the model again. The result is faster responses and lower bills for whatever's calling the API.
What prefix caching…
Your SQL NOT IN is broken (and nobody told you)
If you've ever run a NOT IN subquery and gotten back nothing — not a single row — you're not crazy. The database isn't broken. It's doing exactly what SQL spec says it should do.
Here's the trick: NOT IN isn't really a special operator. It's NOT (column IN (...)). So when your subquery returns a NULL, the whole thing collapses. 5 NOT IN (1, 2, NULL) is NOT (5 = 1 AND 5 = 2 AND 5 = NULL), and 5 = NULL is NULL, not TRUE. So you get NULL, which treats like FALSE. Everything disappears.
The fix? Use NOT EXISTS or LEFT JOIN ... IS NULL if your data has NULLs. Or wrap the subquery in COALESCE. But honestly, most of us just work around it and move on.
Why this matters for us: a lot of us who code, build side hustles, or manage databases at work have been writing queries that silently return wrong results — and we've been blaming ourselves for months.
PostgreSQL drops pg_kpart v10 for faster partitioning
PostgreSQL shipped version 10 of pg_kpart, its partitioning extension. The release is live on the PostgreSQL news page.
Partitioning is how you slice big tables into smaller pieces so queries don't have to scan the whole thing. pg_kpart makes that easier to set up and…
Databricks launches LakeHouseRT for real-time data
Databricks is rolling out LakeHouseRT, a new version of its lakehouse platform built for real-time workloads. The update lets companies run streaming analytics and queries on data as it arrives, rather than waiting for batch processes to finish.
The lakehouse has been Databricks' play to compete with the traditional data warehouse model — combining the flexibility of data lakes with the performance of warehouses. But real-time has been the tricky part. Most lakehouse setups are built for batch processing, not the millisecond-to-second updates that modern apps need. LakeHouseRT tightens that gap.
Why this matters for us: Databricks is trying to make their data tools work faster for the kinds of businesses that move fast — retail, logistics, delivery, and the rest of the companies that can't wait for overnight batch jobs.
La migra del costo de infraestructura ya no corre con CPU
Data processing is becoming a GPU workload. Anyscale's new research tracks how the task is shifting — and what it means for the people running services on top of that compute.
The shift is real. Data processing workloads, once the domain of CPUs, are increasingly running on…
TLDR Data drops its latest newsletter — what's moving
TLDR's Dan Newman published another round of data and tech updates. The newsletter is known for cutting through the noise with things like AI model comparisons, crypto market moves, and new tools landing in the ecosystem.
This edition covers the usual suspects — new datasets, AI model benchmarks, and startup funding — but the real value is in the curation. TLDR has built a reputation for surfacing stories that Brown working people actually need to know about, not the same Silicon Valley press releases everyone else is reposting.
The newsletter has become a go-to for folks trying to stay sharp without drowning in newsletters. If you're tired of reading the same stories ten times across different outlets, TLDR's approach of picking the best and moving on is exactly what the comunidad needs. No filler. No fake quotes. Just what's happening.
Why this matters for us: when we get accurate, unfiltered updates about tech and markets without the usual PR spin, we can make real decisions about our money, our side hustles, and where the next opportunities are landing.
TLDR Data's latest roundup for the data crew
TLDR Data dropped a new issue for the data crowd. The newsletter — run by Dan, who's been curating data news for a while now — covers the tools, papers, and stories that actually matter to people working with data day to day.
What TLDR Data does differently is the curation.…
Tu IA local ya es buena, y no necesita la nube
El hype de los modelos de IA en la nube (GPT-4, Claude, Gemini) tiene un problema: para la mayoría de las personas, es demasiado. Alex Ellis lo dice claro en su post — los modelos de código abierto y los que corren en tu propia máquina ya están compitiendo y, en muchos casos, superando a los grandes modelos en velocidad y costo. La tendencia es clara: cada vez más gente está corriendo sus propios modelos en casa o en servidores pequeños, sin depender de la nube.
Pero aquí está el punto que no se cuenta tanto: si tu IA corre local, necesitas que todo lo demás también esté listo para ella. Los sistemas de diseño (design systems) que los equipos de ingeniería usan — las reglas, componentes, y patrones que guían cómo se construye software — están empezando a ser "especificados" para que los agentes de IA los entiendan. La gente de Evil Martians muestra cómo diseñar sistemas que funcionan para humanos Y para máquinas.
Esto tiene que ver con algo más grande: la descentralización del poder de la IA. Cuando los modelos corren local, no pagas por cada token. No dependes de una API que sube precios. No estás atado a la infraestructura de los grandes. Y si tu sistema de diseño está pensado para agentes, puedes automatizar más del trabajo repetitivo — el tipo de trabajo que te hace perder tiempo a ti y a tu equipo.
Why this matters for us: cuando la IA se queda en tu máquina y no en la nube de los grandes, el control vuelve a nosotros — y eso cambia quién gana dinero, quién pierde tiempo, y quién decide cómo se construye el futuro de la comunidad.
Zuckerberg threw a party for his workers — then cut most of them
Mark Zuckerberg wrapped up Meta's annual hackathon with a message that landed like a joke no one laughed at: the company had laid off about 1,000 employees since the event started.
The hackathon itself was the kind of Silicon Valley spectacle — employees pitching ideas,…
TLDR's daily digest keeps up with the tech that changes how we work
TLDR is a daily newsletter that's been quietly growing into one of the most read tech digests out there. Instead of dumping you into a 50-article feed, they curate what's actually worth knowing — AI tools shifting how people work, policy changes that ripple into communities, and startups building things that touch the people we know.
The newsletter's strength is in its pacing. It doesn't try to cover everything, and it doesn't pretend every item is a big deal. You get the signal without the noise, which is what we need when we're already juggling family, work, and the side hustle.
Why this matters for us: staying sharp on tech doesn't have to mean subscribing to five different feeds and drowning in noise — TLDR keeps it lean so we can move through the day without missing what actually shifts the ground.