AI is Loud. The Work Keeps Going.
AI's everywhere today — brain implants, query engines, ad blockers falling. But the quiet stuff matters more: databases that don't break, agents that stop you from taking notes, small businesses finally getting their data right. La gente keeps building.
How Long Until AI Doesn't Need Humans?
AWS just showed how a company can dominate one thing so thoroughly that it accidentally ruins another. Amazon Web Services built Bedrock to be the go-to AI model platform—let you plug in OpenAI, Anthropic, Meta, and others, all through one API. The theory was simple: everyone's going to run AI on AWS, so why wouldn't they buy AI models there too?
The problem is exactly the thing that made AWS strong in the first place. When your customers can run their own models on your infrastructure, they start asking why they should pay extra to access models through your platform at all. AWS destroyed the value proposition for Bedrock by making itself too good at what it does. La migra app of cloud computing, la migra app of AI models. The more AWS scales, the more it undercuts its own play.
This is a real lesson for anyone building AI tools on top of cloud platforms. Your biggest strength can become your biggest liability when the platform you depend on decides to become your competitor. La gente running side businesses, la gente building apps on AWS—they should pay attention.
Why this matters for us: When the platforms we rely on start competing with us instead of serving us, the ones who adapt fastest are the ones who stay sharp and don't get locked in.
ANN search: finding similar things without checking every single one
Elastic just published a no-bullshit explainer on approximate nearest neighbor search. The short version: instead of checking every item in your database to find the closest match, you use shortcuts to get close enough, fast enough. The kind of thing you need when your recommendation engine is searching through millions of products or your AI model is matching embeddings.
What made me stop scrolling through this one is how Elastic frames the trade-offs. Exact search gives you precision. ANN gives you speed. For most real-world use cases, the difference between 99.8% and 99.9% accuracy doesn't matter — what matters is that your API responds in milliseconds instead of seconds. The post walks through the main approaches: tree-based methods, hash-based methods, graph-based methods, and quantization. Each has its own sweet spot depending on your data shape and your latency requirements.
Why this matters for us: if you're building anything that surfaces results — a product catalog, a job board, a service marketplace — understanding ANN means you can ship faster without breaking the user experience, and you'll know what to ask for when you're hiring or shopping for tools.
Close enough, and fast — that's the whole trick.
— elastic.co
#how-your-apps-find-what-s-close-enough-without-checking-everything-59a8f6PostgreSQL Anonymizer 3.1 brings local differential privacy to your database
PostgreSQL Anonymizer 3.1 shipped with local differential privacy built in. The tool sits on top of PostgreSQL and scrambles your data so people can't tell if a specific person's record is in the dataset. This version adds the "local" variant — noise is added at the row level before aggregation, not after — which means you get stronger privacy guarantees even on smaller datasets.
Differential privacy started in academia and is now showing up in government and tech. The idea is simple: add enough random noise to a dataset that a single record can't be identified, while still letting you run meaningful queries.
Inside QuestDB's Query Engine
TLDR Data's Dan wrote up what makes QuestDB's query engine tick. The database is open-source, built for time-series data, and this piece walks through how it processes queries under the hood.
QuestDB is the kind of tool that shows up when you're building dashboards or monitoring systems and need something that doesn't choke on millions of rows. The query engine is where the real work happens — taking your SQL and turning it into fast results without waiting on disk like older databases do.
AI Keeps Breaking. We're Tired of Fixing It
The latest issue from TLDR makes a quiet but sharp point: no matter how much we train these models, they keep breaking in the same stubborn ways.
It's not a new problem. We've been through the cycle before — more data, bigger models, more fine-tuning — and each time the fixes feel temporary. The models get better at following instructions, but when the context shifts even a little, they drift. They hallucinate with confidence. They forget what they knew yesterday.
What's different now is the scale of the gap between what these models can do and what they actually do. We've got systems doing complex reasoning, writing code, translating languages, and yet the same old quirks keep showing up in production. The models are powerful but fragile in specific, predictable ways.
TLDR's point is worth sitting with: maybe the issue isn't that the models are broken. Maybe it's that we keep trying to force them into the same boxes, expecting different results.
Why this matters for us: the AI tools we're told to adopt are still the same glitchy systems — useful enough to try, but not reliable enough to trust blindly.
Untrainable: The AI Model That Refuses to Forget
Most AI models degrade when you update them. Train them on new data — a new language, a new task — and they start forgetting what they already know. That's called "catastrophic forgetting," and it's been a headache for developers since the early days of neural networks. There's a whole field devoted to coaxing models back to their old selves.
Enter Untrainable, a model from 2024 that refuses to forget. Instead of fighting the forgetting, it uses a mathematical trick to keep old knowledge intact even as new data flows in. The result is a model that can learn new things without losing the things it already knows.
TLDR newsletter: what's actually new
Dan at TLDR sent out another roundup of tech news for the week. Not a deep dive on any one story — more like a scan of what moved in the space.
TLDR has built a reputation for keeping things tight. No fluff, no padding. Just the headlines that matter to people who work in tech and want to stay sharp without reading through a dozen substacks.
Why this matters for us: when the newsletter crowd talks about tech, it's usually Silicon Valley talking to itself — so we keep our eyes on what actually lands in our pockets.
China's got a brain implant that's already working — before Neuralink even launched
A patient in China just received the world's first brain implant designed to treat Alzheimer's, and the device is already being marketed commercially. This is a big deal because the device works by delivering targeted electrical stimulation to parts of the brain where memory-forming cells tend to die off. The patient's cognitive function has been improving.
Here's what makes this interesting for the bigger picture: while Neuralink has been the Silicon Valley poster child for brain-computer interfaces, China is actually racing ahead on the commercialization side.
Schema.org's new dataset shows how small businesses are structuring their data
Schema.org has released a new dataset tracking how widely its structured data markup is being used. This is the same markup that powers Google's rich results — the star ratings, product cards, events, and recipe blocks that show up when you search. The dataset breaks down adoption by industry, technology, and geography, giving a clearer picture of who's doing it and who's not.
For small businesses, structured data is how Google understands what they're selling. Without it, your business shows up as a plain link. With it, customers see your hours, prices, reviews, and availability right in the search results. The new data suggests a lot of small businesses are still figuring it out — and that gap is widening as search gets smarter.
Why this matters for us: la gente searching on their phones want to see your hours, your prices, your ratings — and if Schema.org markup is the way to get there, it's worth knowing.
Metadata needs to handle change without breaking
The Hudi team just published a piece on why metadata — the info about your data — has to be mutation-friendly. If you've ever watched a database table get updated and the metadata trail behind, leaving you with stale schemas or confusing state, you know the problem. Metadata that can't handle changes smoothly becomes a bottleneck for everything downstream.
The argument is straightforward: metadata isn't just a static record. It's a living thing.
QuestDB's Query Engine Gets a Serious Upgrade
QuestDB has rebuilt its query engine from the ground up. The new version processes time-series and relational data faster than before—less time waiting on queries, less money spent on compute. The company has been building a real-time database that sits somewhere between a traditional database and a specialized analytics tool, and this upgrade tightens that positioning.
The query engine is what makes or breaks a database for developers who actually use it. Good queries mean dashboards load faster, reports don't time out, and your team isn't paying for compute you don't need. QuestDB's main use case is real-time data—stuff that arrives continuously and needs to be searched quickly. That's a growing space as companies collect more sensor data, logs, events, and user signals.
Why this matters for us:
Chrome kills uBlock Origin — and your ad blocker's days are numbered
Google Chrome is phasing out the bypass mechanism that lets uBlock Origin hide ads, effectively shutting down one of the most popular ad blockers in the world. Microsoft Edge and Opera are next.
uBlock Origin is free, open source, and one of the most widely used ad blockers on the internet. It's the kind of tool that lets people who don't want to pay for subscriptions still access content funded by ads. Now Chrome is closing the loophole.
This isn't just a browser update — it's another move in Google's long campaign to control what you see on the web.
Stack Overflow is opening its doors to AI agents
Stack Overflow just announced a new product designed specifically for AI agents — not for people, but for the bots. The idea is that AI tools like chatbots and coding assistants can now query Stack Overflow's vast library of developer Q&A to find answers on their own, without needing a human to read through the results.
This is a meaningful shift. Stack Overflow has always been the go-to reference for developers, from the junior dev debugging their first app to the senior engineer writing code at 2 AM. Now the site is building infrastructure so AI agents can tap into that knowledge directly. For the comunidad of developers, it means the AI tools we're already using are getting better at understanding how we actually solve problems — not just the textbook answers, but the real-world workarounds.
Why this matters for us: As AI agents become part of how code gets written, the questions we ask Stack Overflow today will shape what future AI tools learn, which means our community's knowledge is becoming part of the foundation of the tools themselves.
China's First Underwater Wind-Powered Datacenter Is Live
China has opened the world's first wind-powered underwater datacenter. The facility is sitting under the sea, drawing power from wind turbines, and it's now running operations. No one has done this at scale before.
Datacenters are hungry things — they suck up electricity like nobody's business. Cooling them in water is already a trick, but powering them with wind changes the math on what's possible. The question is whether this model scales or if it's a one-off proof of concept.
The people building it are the same kind who keep the generators running when the grid goes down.
PagerDuty's Scribe Agent Cuts Out Manual Note-Taking
PagerDuty updated Scribe, their AI agent, so it no longer relies on manual note-taking to stay on top of what's happening. The agent now captures and preserves context as incidents unfold instead of losing it between handoffs.
For engineering teams that juggle alerts across Slack, Jira, and dozens of other tools, the update means less time spent hunting through chat threads and more time actually fixing things. It's the kind of quiet improvement that makes a difference when you're already running on three hours of sleep.
Why this matters for us: When tools stop making us do the remembering, we get back time to actually work — and that's a win for every Brown dev who's spent too long flipping through Slack while the page stays down.