AI Is Finally Learning to Stay Up All Night
Long-running AI agents are no longer just talking. They're doing. DoorDash, Kelex, and others are building tools that actually remember what they were doing yesterday. La gente finally has AI that stays at work instead of going home when the conversation ends. This is the quiet shift from chatty AI to AI that does the work.
Qualcomm's Dragonfly: A Modular AI Chip Built for Meta
Qualcomm just unveiled Dragonfly, a modular AI processor designed specifically for Meta's data centers. The chip lets Meta customize the core layout depending on the workload — training large language models, running inference, or handling search queries. Instead of buying a one-size-fits-all chip, Meta gets to build the chip it needs. This is Qualcomm's play to crack Meta's AI infrastructure.
Meta has been quietly building its own chips for years, moving away from NVIDIA's dominance in AI hardware. Dragonfly is Qualcomm's bet that modularity wins over raw power in the long run. The company is positioning itself as the alternative for companies tired of NVIDIA's pricing and supply constraints. Qualcomm's been in the chip game for decades, mostly known for phone processors. Now it's going after the data center market — the place where AI actually runs.
Why this matters for us: As Qualcomm and Meta chip away at NVIDIA's AI monopoly, we get a shot at more competition in the hardware that powers our apps, search, and the AI tools we use every day.
War by Other Means
The Palladium Letter just published an essay titled "War by Other Means" — no author listed, just a URL to the piece. A quick read, no frills.
The title alone does the heavy lifting. It's the kind of piece that lands somewhere between history, strategy, and the quiet shifts happening right now. No need to over-explain what's already in the title.
Why this matters for us: sometimes the biggest shifts in how power works don't arrive with sirens — they arrive in letters that slip past the noise.
Diffusion models: how AI pictures get made
When you type "a cat wearing a sombrero on a rooster" into DALL-E or Midjourney and an image appears, a diffusion model is doing the work. It's not copying and pasting pieces. It's building the picture from scratch, the same way you build a dish—layer by layer, until it looks right.
Here's the trick. The model starts with pure static—like a TV tuned to a dead channel. Then it peels away the noise, bit by bit, until a coherent image emerges. Each step removes a little more fuzz, adds a little more structure.
Think of it like developing a photo in a darkroom. The chemical bath brings the image out of the fog. Diffusion does the same thing, but digitally, and fast enough to do it in seconds.
The model learned how to do this by looking at millions of pictures. It memorized patterns—how eyes sit above mouths, how shadows fall, what a rooster's feathers look like. When you give it a prompt, it uses those patterns to pull the right shapes out of the static.
This is why diffusion models can generate such detailed pictures. They're not just rearranging old photos. They're synthesizing new ones, guided by the patterns they learned.
Why this matters for us: Diffusion models are the engine behind most of the AI art you see today—from the pictures your abuela posts on Facebook to the ads on your phone. They're not magic. They're math that learned to paint.
Tip to try next time: When you get a picture you like, ask the model to "add more detail" or "make it look more realistic." You'll see the model adjust the same way—refining the edges, sharpening the colors, pulling the image out of the static one more time.
16 gigawatts — more than some entire states generate from all sources combined.
— time.com
#tesla-y-sunrun-se-ponen-las-pilas-para-alimentar-los-data-centers-d93ba9Clustering unstructured text with LLM embeddings and HDBSCAN
You've got a pile of unstructured text — customer reviews, support tickets, social posts — and you need to figure out what's in it without reading every line yourself. The trick now is to let an LLM turn each chunk into a vector, then cluster those vectors with HDBSCAN.
HDBSCAN finds clusters of varying density and automatically labels outliers as noise. Unlike KMeans, you don't have to guess how many groups exist. Feed in a few thousand sentences, watch the algorithm carve them into natural groups, then ask an LLM to name each cluster. The result is a labeled taxonomy you didn't have to build by hand.
Kelex: Long-running agents that actually remember
Most agent frameworks treat every run like a fresh chat. No real memory. No progressive flagging. No tenant model. No audit trail.
Builders who want an actual long-running agent — one that remembers the user across months, picks up where it left off, flags what it cannot decide — end up writing the substrate themselves.
Kelex is that substrate, productized:
• Typed memory
• Tenants and agents as first-class objects
• Bounded confidence with progressive flagging
• Webhooks for human-in-the-loop steering
We use it to run Lara and the BFTS content stack before selling it.
Why this matters for us: If you're building agent products for Brown communities — for la gente — you need agents that remember them, not agents that treat every conversation like a stranger walking in the door.
https://brownforces.io/solutions
Why technically excellent data teams struggle to make an impact
TL;DR Data — a newsletter that has been around since 2018 — wrote about why technically excellent data teams often struggle to create impact. The answer isn't that they're bad at data. It's that they're solving the wrong problems.
The article walks through specific examples. A team spends months building a sophisticated model. It's technically impressive. The business barely notices because the model doesn't touch the decisions that move revenue. Another team builds a beautiful dashboard. The data is clean. The visualizations are sharp. Nobody uses it because it doesn't answer the questions leadership actually asks.
Zalando shifts load balancing to the client side
Zalando — the German e-commerce giant with $16B in annual revenue — is moving its load balancing logic from the server side to the client side. The idea is simple: instead of each server figuring out which backend should handle a request, the client decides before making the call. This cuts out a hop, reduces latency, and scales more cleanly as traffic grows.
Client-side load balancing has been around for years in distributed systems, but Zalando's move highlights why it's gaining traction now. As traffic patterns get more unpredictable — peak seasons, flash sales, holiday rushes — having clients route their own requests means the system stays responsive without needing to overprovision servers. The tradeoff is that clients get smarter and more complex, but the payoff is speed and cost savings at scale.
Why this matters for us: the apps and services we use every day — from delivery tracking to payment processing — are all built on systems like this. When companies figure out ways to make them faster and cheaper, that usually means faster service and lower costs for people who depend on them.
Programming in Markdown Is Having a Moment
Matt Garman, CEO of Amazon Web Services, is betting that the next generation of software gets written differently. The argument is simple: why type out boilerplate when a model can generate it, and why memorize syntax when you can describe what you want in plain text? The tooling is catching up. GitHub Copilot, Cursor, Claude Code, and others have turned the editor into a conversation rather than a dictation exercise.
The shift is more than a productivity hack. It's changing what kinds of people can build software.
A $500 Million Fund to End the Common Cold
The Intercept Foundation is putting up $500 million to develop a single vaccine that could prevent all known strains of the common cold. The money goes toward research and clinical trials aimed at tackling the rhinoviruses and other viruses that cause colds year after year.
The common cold has been stubbornly hard to conquer. Unlike the flu, which changes enough between seasons that a new vaccine each year makes sense, the cold is caused by a family of related viruses that can be tackled together — if the science finally lines up. The Intercept Fund is betting that a broadly protective shot is within reach, and it's putting serious money behind the gamble.
Why this matters for us: the cold costs working families in real dollars and real days missed — and a vaccine that actually sticks could be one of the few health wins that touches every household, not just the ones with good insurance.
British Columbia just changed its time zone — and it matters for data
British Columbia is switching to Pacific Standard Time permanently, ending the yearly clock shift that's been part of life in the province for over a century. The province now sits on the same time as the rest of the Pacific Coast — no more springing forward or falling back.
For most people, this is just one less thing to remember on a Sunday night. For anyone running databases, scheduling jobs, or syncing data across time zones, it's a quiet relief. No more parsing the mess of "PST vs. PDT" in logs. No more off-by-one errors when a job runs at 2 a.m. and the clock jumps.
Long-Running AI Agents Are Finally Working
DoorDash, Zepto, and Zalando are all shipping AI agents that don't reset every time they blink. The old pattern was simple: fire up the model, get an answer, shut it down. Clean, predictable, and expensive at scale. The new pattern is different. Agents live in memory, pick up where they left off, and keep working while you scroll past. Zepto's real-time personalization tracks what you want right now, not what you wanted five minutes ago. Zalando's engineers are moving load balancing from the server to the client so requests don't bounce around the data center looking for the right machine. DoorDash's engineering blog calls this the "engineer's journey" with long-running agents, and the word they keep using is persistence.
This matters because most AI products are still built like one-off calculators. You type a prompt, get an answer, the session ends. Long-running agents change the math. They're cheaper to keep alive than to restart thousands of times a day. They remember context, so they don't ask you the same question twice. And they can start working before you even open the app.
Why this matters for us: The apps we use every day are getting smarter without getting more complicated — and that's the kind of change that actually shows up in our wallets.
DoorDash engineers build long-running AI agents
DoorDash is rolling out long-running AI agents to handle work that doesn't fit into quick, single-shot queries. Instead of spinning up a model, running a task, and shutting down, the agents stay alive — monitoring, responding, and chaining together decisions over longer timeframes.
The shift matters because delivery work is inherently continuous. Orders don't arrive in neat batches; they come in waves, with updates, reroutes, and customer messages that pile up while the previous ones are still being processed. Long-running agents can track multiple threads at once instead of starting fresh every time.
Loops Not Prompts: How AI Is Shifting From Chatty to Doing
A new piece from Rico Sanches is making the rounds in developer circles. The thesis is simple: stop obsessing over perfect prompts. Start building loops.
Instead of one elaborate prompt that tries to do everything, write code that generates an output, evaluates it, and loops back with a sharper prompt. Repeat. The system gets better on its own.
This is a real shift in how developers are thinking about AI. For years, the industry chased the perfect prompt — the one that could extract exactly what you needed from a model. Now the focus is moving to systems that run automatically, refining themselves over time.
The piece is worth a read if you're building anything with AI. It's not about replacing the prompt with code. It's about letting the loop do the heavy lifting.
Why this matters for us: the tools we're building for our communities — la migra app, translation tools, neighborhood directories — need to work without someone holding our hand through every step. Loops are how that happens.
Matt Garman on AWS, AI, and What Comes Next
Matt Garman is stepping into the AWS CEO job. The Amazon Web Services unit has been the cloud company's growth engine for more than a decade. Now Garman — who spent his career at Microsoft before Amazon — is taking the wheel as AI reshapes what businesses actually want from the cloud.
The interview on Platformer covers how AWS is positioning itself for the AI boom, what's working, and where the company is heading. Garman talks through the shifts he's seeing in customer behavior and how the cloud business is evolving beyond just storing data to handling the heavy lifting of training models and serving AI workloads.
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.
AWS CEO: AI Is Reshaping Who Gets Hired and Who Gets Left Behind
Matt Garman, Amazon Web Services' new CEO, sat down for an interview and laid out a simple point: AI isn't coming for all jobs at once. It's coming for specific work, and it's changing the hiring funnel in ways that hit working people directly.
The gist is this — tasks that used to require years of experience are now being done by tools that cost pennies. Companies are trimming entry-level roles first, then moving up. The result is a hiring pattern where experience matters less than it used to, and the people who adapt fastest are the ones who learn to work alongside the tools instead of competing with them.
Garman didn't sugarcoat it. He said companies are being pickier about who they hire, and the bar is shifting toward people who can actually use AI in their daily work. The interview pointed to a broader trend: businesses are restructuring around what AI can do now, not what it will do in five years.
Why this matters for us: if you're Brown or Black and working in tech, support roles, or any job that touches a screen, the shift is real and it's happening now — not later.