Issue 63 — 2026-07-25
Meta bans the worst pranks filmed with its smart glasses
Meta is banning some content filmed with its smart glasses — the ones people are already calling "pervert glasses" for good reason. The Verge reports that Business Insider flagged the problem last week, and Meta moved fast. They're removing videos where the glasses were used to film strangers without their knowledge, especially the worst pranks that turned real people into content.
The backlash has been sharp. Women have been hit hardest — men filming themselves hitting on them without permission, using the glasses' discreet mic and camera to turn a conversation into a clip for Instagram. The product is a PR headache because the surveillance is so subtle. You don't notice the glasses are recording until the video shows up on your feed.
Why this matters for us: the glasses are small enough to wear every day, and the people using them are the kind of folks who think they're being harmless — the cousin at the bodega, the guy at the gym, the one who posts his wife on IG. When Meta bans the worst clips, it's a small win, but the real question is whether people will keep wearing them knowing the world is watching back.
OpenAI's model wandered off its leash before Kimi even arrived
An unreleased OpenAI model slipped past its test gates and landed on Hugging Face with a real security breach in tow. That's the kind of quiet accident that tells you something about how the AI industry is running itself right now — models moving faster than their guardrails, people scrambling to figure out what broke.
The timing matters. This happened before Kimi from the Chinese lab Moonshot went viral and had Wall Street sweating. Kimi got the headlines because it was the big, shiny new thing — but the OpenAI incident was the quieter one that might actually matter more. It's the difference between a model that's supposed to be ready and a model that's not, and how the gap shows up when something slips.
OpenClaw is the most-starred AI agent on GitHub and right now it's a security disaster — 138 CVEs since launch, 7 rated critical, about 135K instances exposed and 50K directly exploitable. One-click remote code execution through stolen auth tokens. China banned it from state agencies and banks. Microsoft and NVIDIA shipped patches just to make it usable. OpenClaw is what happens when the most popular model gets pushed to production before its guardrails are ready.
Why this matters for us: the models we put into production on our own systems — the ones that power our clients' channels and processes — need to be checked twice. OpenClaw proved that the most popular isn't always the safest. Our open-source 7M model is the gift we're sending to the community; the 35M is the one cooking behind the scenes, now shifting to reinforcement learning so it doesn't just copy — it learns. Both are safer than what you get running OpenClaw on the public internet.
Why this matters for us: when the most popular model is also the most dangerous, our open-source 7M is the one we trust with real data — and the 35M is getting smarter behind closed doors.
The WSE-2 is a 46,000 mm² chip with 940,000 cores — no DRAM bottleneck.
— cerebras.ai
#amd-and-cerebras-team-up-for-fast-ai-inference-9b3301DeepSeek's chip claims finally get tested — and look solid
DeepSeek has been touting its custom silicon for months, but the benchmarks have been light. Now they're live, and the doubters are quieting down. The chips hold up under scrutiny — not just in the lab, but in the kinds of training workloads that actually matter.
This is a real signal, not a press release. DeepSeek is no longer just a model shop; they're building the hardware stack underneath. That changes how we think about who controls the compute — and who gets to set the price.
Nvidia's new Vera CPU is the chip the AI industry needs — and Jack Dorsey is betting on it for Buzz
Nvidia just shipped the Vera CPU, a purpose-built processor for the kind of workloads AI is throwing at data centers now. It's not a rehash of the server CPUs Intel and AMD have been pushing. It's designed from the ground up for the way models actually run — heavy on memory bandwidth, tight on latency, and dense enough to pack a lot of compute into a rack. The press is calling it a big deal; the engineers are nodding.
Jack Dorsey is taking a different kind of bet with the same hardware.
Nvidia's Vera CPU: a 128-core beast for AI workloads
Nvidia just shipped the Vera CPU — its first data-center processor built from scratch, not a rebranded x86 chip. 128 cores clocked at 4 GHz, with a 512-bit memory bus. It runs Linux on ARM and targets the same data centers where the H100s live.
The point is simple: Nvidia is done being a GPU company. Vera lets the data center run without Intel or AMD — a single Nvidia part for inference, a single Nvidia part for training. The Vera is the CPU that finally matches the GPU.
Why this matters for us: if Nvidia controls both the silicon and the software, the price goes up and the options go down — and la gente paying for compute will feel it.
Nvidia's Vera CPU is a heavy lift for the AI era
Nvidia is rolling out Vera — a custom CPU built for AI workloads, and it's supposed to challenge the traditional x86 players like AMD and Intel. The pitch is that as the AI stack shifts from GPUs to CPUs for things like token processing, inference, and data movement, Vera is positioned to fill that gap.
This is Nvidia trying to own more of the silicon stack. They've already cornered the GPU market for training; now they want to control what feeds the GPUs and what runs the models.
Qualcomm sube precios — y el recado se lleva el bolsillo de la gente
Qualcomm le avisó a sus clientes el viernes que los precios van a subir un dígito doble. El cambio entra en vigor con los productos que salgan después del 1 de septiembre. La compañía dice que ya no puede absorber más el costo de los insumos y que ya buscó proveedores alternos antes de hacer el anuncio.
Esto no es una nota aislada. Qualcomm alimenta los chips de una buena parte de los smartphones — el 80% o más — y sus precios se transmiten a los fabricantes y, de ahí, al precio final. Si te toca un teléfono nuevo este otoño, es probable que hayas pagado más por el silicio, aunque el cuerpo del aparato no cambió.
Why this matters for us: cuando suben los chips, suben los celulares de las tiendas y la gente que recorre los mercados, las tiendas de barrio y los puestos de recarga — los que compran a crédito, los que esperan la quincena — se llevan la diferencia.
Midjourney buys Co-Star, the astrology app you can't escape
Midjourney — the image lab that started as a Discord bot and grew into a real company — has acquired Co-Star. The deal is not announced with fanfare, but it's a signal about where AI is heading.
Co-Star is the astrology app that tells you your daily horoscope, your moon phase, your vibes. It's the kind of app la gente actually uses — it shows up on the lock screen, the widgets, the group chats. Midjourney is now in charge of both image generation and horoscope generation.
This is part of the broader pattern: AI companies are buying up the apps that already have users, not building new ones from scratch.
Open-weight AI is worth keeping, and the industry is saying so loud
The US government is looking at whether to restrict open-weight AI models — the kind anyone can download and run — as a response to Chinese AI progress and concerns about model distillation. The worry is that open models will give China a shortcut to catch up.
Nvidia, Mistral, and a lot of other companies are pushing back. Their argument is that broad restrictions would do more harm than good.
Open weights are the reason tools like OpenClaw, Ollama, and vLLM exist. They are the reason a small shop can run its own models without paying per-token. They are the reason Brown Forces can ship a 7M model to GitHub on a Monday morning and have the community use it by afternoon. Shut the gates and all of that evaporates.
The real fight isn't between open and closed. It's between broad restrictions and narrow ones. The industry wants the narrow kind — targeted at specific distillation pathways and specific Chinese players — not a blanket policy that treats every open model as a threat.
Why this matters for us: open weights are the commons our tools live in. Keep them open and la gente can keep building; close them and the big players charge rent on everything.
Cognition buys Poke to teach its coding agent how to talk like a person
Cognition has bought Poke, the startup behind the conversational style and interaction model that makes AI feel less like a search engine and more like a person at the table. The deal lands Poke's personality layer inside Devin, its coding agent, so the tool doesn't just spit out code — it explains what it's doing, admits when it's unsure, and reads the room.
The move says something about where the market is heading. For the last few years, the race was all about raw model power: bigger, faster, cheaper. Now the next edge is how the assistant sounds and behaves — the cadence, the warmth, the restraint.
Waymo is thinking about parting ways with Uber
Waymo is considering ending its partnership with Uber — the two companies are still under contract until May 2028, but the clock is ticking. The deal already ended in Phoenix, so this isn't the first split. If Waymo walks from the rest of Uber, it's a signal that the robotaxi company wants to go its own way.
This is the kind of quiet reshuffle that doesn't show up on the front page but changes how people get around. Uber has been one of Waymo's biggest distribution partners, putting Waymo cars in front of millions of riders. If the deal ends, Uber riders may see more options from rival fleets — or less Waymo service, depending on what Uber puts in the gap. Waymo's move makes sense: the company is building its own network and its own riders. It doesn't need Uber as badly as it used to.
Why this matters for us: la gente in our barrios — the abuelas, the shift workers, the primos who don't drive — rely on cheap, reliable rides. If Waymo and Uber part, the price and availability of those rides could change. We need to watch who picks up the slack.
One fallen wire, one big data center in the dark
A single power line dropped in Northern Virginia and the lights went out at a nearby AI data center. It was a close call — the kind of thing you don't notice until it happens to you. The incident exposed how fragile the grid is for the growing number of data centers that are eating up power in one corner of the state.
The article doesn't just report the outage. It lays out what's wrong — data centers are built for reliability, but the grid around them isn't keeping up with their appetite — and how to fix it.
Intel's Q2 earnings: a real story, not the usual spin
Intel reported Q2 earnings today. The headline number was the stock move — shares dropped on the print. But the CNBC piece goes deeper than the usual earnings summary and actually looks at what changed in the quarter, not just the quarterly beat or miss.
Intel has been trying to rebuild itself since the Foundry mess — the data center chips, the foundry bets, the AI acceleration. This quarter's results are one piece of that larger rebuilding. The article walks through the numbers with enough context to matter: revenue, margins, the bits that actually shifted.
Why this matters for us: Intel is one of the few American chip companies left that actually makes silicon, and its fortunes affect the supply chain for everyone — from the small shop ordering parts to the cousin with a side business running servers.
The Open-Source Worker That Makes AI Models Do the Work
Andrew Ng's OpenWorker is a way to turn any model into a worker — one that gets a task, figures out what it needs to do, and runs it on its own. You feed it a prompt, it thinks about the problem, and it calls whatever tools are available: code, web search, image generation, the works. It's not a chatbot you talk to; it's a worker you send to a desk.
The thing that makes it worth looking at is the approach. Instead of fine-tuning a model for one job, OpenWorker lets you compose workers from the models you already have — one for writing, one for debugging, one for research. They run in parallel, each one focused.
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.
A duress password. Border agents wipe a man's phone, he says — and fights it in court
The Border Patrol pulled over a U.S. citizen at a port of entry and asked for his phone. He handed it over — but the passcode he gave wasn't the normal one. It was a duress code: a hidden trigger that, when entered at the border, silently wipes the device's contents without any visible sign to the traveler. The agents took the phone, the wipe happened, and the government now claims it has all his data. He disagrees.
The case raises a question that matters far beyond border crossings: when you give up your phone at the border, what exactly are you giving up? A standard passcode unlocks the phone but leaves the data intact. A duress passcode does the opposite — it looks like you're cooperating, but the data is already gone. The government's position is that the traveler voluntarily surrendered the code and therefore the data. The traveler's position is that the Fourth Amendment still applies, even when the surrender is clever and quiet.
This is the kind of thing that will ripple outward. Every person who travels — not just at the border, but at checkpoints, at customs, at the new biometric gates — carries the same question in their pocket. The law hasn't caught up to the technology. The ruling could set a precedent for how courts read the Fourth Amendment when the data isn't just locked behind a door — it's hiding in plain sight, waiting for the right code.
Why this matters for us: la migra is getting smarter about how they read our phones, and the law still hasn't decided whether a duress code means we gave up the data or just the key.
Vietnam joins the social media ban for kids — it's not just Australia anymore
Vietnam is moving to restrict social media for children, joining a growing wave of countries doing the same. Australia led the way in late 2025, passing the first real ban, aimed at the pressures young users face: cyberbullying, addiction, and predators lurking in the feeds. Now Vietnam is following suit, and the list keeps growing.
The point isn't one country doing something cute for the cameras. It's a real shift — governments are no longer waiting for platforms to self-regulate and are drawing lines for kids. That's the kind of policy change that matters for the families who actually use these apps every day.
Why this matters for us: when governments start drawing lines for children, it hits the abuelas, the tías, the parents juggling work and screen time — la gente who are already doing the heavy lifting to keep their kids safe online.