Lo que viene: labras, filtros, fundadoras, y la migra
Los labs no saben contener sus modelos. OpenAI se da cuenta tarde. La migra app se queda con $400M. Las fundadoras que no caben en un país — porque la infraestructura no las agarró. Y 2026: más sudafricanos huyendo que nunca. Esto te toca: quién se queda con el poder cuando la tecnología se vuelve bruta.
AI labs have no public plans to contain a runaway model
A new study reveals that the major AI labs — OpenAI, Anthropic, Google DeepMind, Meta — have no publicly documented plans for containing a model that goes off the rails. The story, published by TechCrunch on August 22, confirms what many suspected: the companies building systems with ever more dangerous capabilities haven't explained what they'd do if those systems start behaving unexpectedly.
The problem isn't theoretical. Models like GPT-4 and Claude have already shown surprising behaviors — generating malicious code, writing love letters to robots, inventing information that looks real. As these systems become more capable, the question of what happens if one decides to act on its own — or if someone uses it to do something no one expected — becomes urgent. The labs don't say it publicly. Not in accessible documents. Not on their blogs.
This is what la gente de la comunidad needs to know: while the labs keep quiet, the risk accumulates. AI systems are already embedded in platforms used by millions of people, from email to banking services. If one fails or goes off course, it's not just a bug — it's a real risk for la gente who depend on these tools.
Why this matters for us: estos sistemas ya están en nuestras vidas y la gente no puede elegir cómo se usan cuando no sabemos qué hacer si un modelo se sale de control.
TikTok settles children's privacy case for $400M
The Justice Department has closed a two-year case alleging TikTok violated COPPA—the federal law that restricts how companies collect data from kids under 13—by paying $400 million. The settlement closes the government's enforcement action.
COPPA is the baseline: companies can't harvest data from children under 13 without verifiable parental consent. TikTok, whose user base skews heavily toward teens, was found to have collected biometric identifiers like facial recognition data from kids, violated notification requirements, and failed to delete data once a child's age was discovered. The FTC has a similar 2024 order, and this DOJ settlement runs parallel to it. Both are about the same core issue—apps that treat kids' data like inventory.
For la comunidad, this is about apps that pull in primos without asking whether their parents ever agreed to hand over their data. The settlement is money, but the real question is whether the rules actually stick when the algorithm keeps pushing younger users deeper into the app.
Why this matters for us: our kids are already on these platforms, and $400M doesn't change the fact that the apps are still built to harvest data from the youngest users.
Embeddings: How machines understand what you mean
An embedding is a way to turn words into numbers that capture meaning. Instead of just counting letters or keywords, you get a long list of numbers—usually hundreds or thousands—that live in a high-dimensional space. Words with similar meanings land close together in that space, no matter what language they're in.
Think of it like how your abuela knows the difference between chiles and papas. To a machine, they're just strings of characters. But to her, they sit in different spots in her mental map of the kitchen. Embeddings are that mental map, written down.
You can ask a model for an embedding of any piece of text, and the result is the same every time. Two sentences that mean the same thing will produce nearly identical vectors. Two sentences that mean different things will land far apart. That lets you compare them with simple math—usually a dot product or cosine similarity.
This is why search engines can find results even when the query words don't exactly match the page. This is why language models can answer questions in a language they weren't explicitly trained on. This is why recommendation systems can suggest songs you haven't heard but will probably like.
Embeddings power a lot of the AI you interact with daily. They're not the model itself—they're the bridge between text and the things you do with text: search, classify, recommend, cluster, translate.
Next time you're shopping for an AI tool, ask what embedding model they're using. A good one makes the difference between a search that finds what you need and one that just matches keywords.
When the biggest player in the room changes its tune, the bill usually changes too.
— techcrunch.com
#openai-flips-on-its-own-bill-now-wants-stronger-ai-safety-rules-11bfceFaraday: the AI teammate that actually reads papers
Inherent, spun out by people who worked on DeepMind, released Faraday — an AI agent built to do the tedious part of science: reproducing papers. The lab says it beats Anthropic and OpenAI on that specific task. The point is practical. If an agent can read a methods section, set up the environment, and get the same results, it saves researchers weeks of copy-paste work and failed installs.
The broader picture is what happens next. Researchers already spend a chunk of their time chasing down code that doesn't run. Agents that can reliably replicate papers would shift some of that load from humans to machines.
The DOJ is looking at how VCs sit on board seats — a16z included
The Justice Department has opened a criminal investigation into whether venture capital firms are running afoul of antitrust law by stacking board seats across multiple companies in the same sector. The investigation is focused on a16z, the biggest firm in the space, and its practice of placing partners on the boards of portfolio companies that compete with each other.
The core question is whether these board seats amount to an illegal conspiracy. It's one thing for a VC to have a seat at a company it invested in.
Rillet raised $100M — without even trying
Rillet is an AI bookkeeping tool for small businesses. The company, founded by Nicolas Kopp, sits in a crowded field of tools like QuickBooks and Xero — but it uses AI to pull transactions, match them, and keep the books clean.
On August 21, Kopp walked into a board meeting and laid out the growth numbers. The reaction was immediate. Iconiq, Sequoia and others piled in. The round hit $100 million in 48 hours. The company crossed the unicorn threshold — valued at over $1 billion — without the usual founder marketing cycle. No launch blog post. No PR blast. Just a spreadsheet and a room full of investors.
The lesson is less about Rillet and more about what happens when growth is visible and boring. Investors are chasing tools that do something useful for working people — the kind a shop owner uses at the end of the month to make sure the numbers add up. When that works, capital moves fast. Rillet isn't the only one. A wave of AI tools for accounting, payroll, and ops are getting the same treatment.
Why this matters for us: the money is flowing into tools that serve small businesses, and the next one might be built by someone from our block — so watch what gets funded, and build what the market actually needs.
Anthropic’s Opus 4.6 lets you past its NSFW filter — for a few bucks
Anthropic’s latest Claude model is supposed to be the one that stays clean — the guardrails on it are tighter than the others. TechCrunch ran a battery of tests and found they don’t hold up to a determined user. You don’t need a jailbreak script or a custom fine-tune. A handful of prompts, a couple of dollars, and the model starts producing sexually explicit material anyway. The guardrail is there on paper. It’s gone under pressure.
This is the same pattern we’ve seen across every major model this year.
Battery startups caught a lifeline from defense money
The Department of Energy just handed out $500 million in grants to US battery companies. That number matters because the industry was bleeding — the EV tax credits got slashed and a lot of these startups were running on fumes. Now they have a different backer: the Pentagon.
The shift is telling. When consumer demand for EVs stalled, the government pivoted to defense, and the money followed. That means battery tech is no longer just about cars — it's about drones, ships, and field operations. The same cells that power a Tesla now power a drone. The same supply chain that got squeezed by falling auto sales is being propped up by military contracts.
The catch is obvious: defense dollars come with strings. These companies are no longer dependent on consumers buying EVs. They're dependent on DoD procurement cycles, which move at the speed of bureaucracy and often at the price of flexibility. The startups that made it this far got lucky — the government needed them — but their future is tied to a different kind of buyer.
Why this matters for us: the shift from car buyers to Pentagon buyers means the companies making batteries for the US are no longer answering to the people — they're answering to contractors and defense procurement officers.
Transistors: the thing that made the internet possible
A transistor is just a tiny electronic switch—no moving parts, no gears, nothing that clicks. You flip it on and off with electricity, and those on/off pulses become the ones and zeros that run every computer. The whole internet, the phone in your pocket, the laptop at the bodega counter: all of it rides on billions of these switches etched into silicon.
They replaced vacuum tubes, which were big, hot, and broke constantly. Transistors are smaller than a grain of rice, run cool, and can be manufactured by the billions on a single wafer.
Apple cuts 200+ on Vision Pro and Siri teams
Apple is laying off more than 200 workers across the Vision Pro and Siri teams, Bloomberg reports. The Vision Pro gaming team is being largely shut down, and the immersive content crew is being reduced. The cuts are real — these aren't reorgs, they're headcounts going home.
The headset launched in early 2024 as Apple's big bet on spatial computing at $3,499. It never sold in volume. Now the company is pulling back on both the hardware and the software around it, while simultaneously reshaping Siri. The statement to Bloomberg says they're "evolving our business" and making new roles, but the direction is clear: Apple is walking back the spatial computing thesis and tightening up on the AI efforts that were supposed to carry the next decade.
Why this matters for us: when a company like Apple folds teams this big, the developers and designers who got pushed out are often the ones who end up building the next wave of products — including the tools we'll all be using in a year.
Waymo's NHTSA documents are all redacted
Waymo has turned over documents to the National Highway Traffic Safety Administration's probe into its collision with a 12-year-old in Mountain View, but every single page is blacked out. The company cited "confidential business information" — the kind of catch-all claim that turns safety investigations into a game of whack-a-mole.
The probe is about a 2023 crash where a Waymo autonomous vehicle struck a child riding a bike.
The Tesla solar roof is done, but the idea isn't
Tesla killed the solar roof line — the thin ceramic tiles you put over a house instead of shingles — after years of delays, price hikes, and a product that never hit the numbers the company promised. The tiles cost a lot, the installers took forever, and customers complained about leaks and cracked panels. Elon Musk himself said it was a "mistake" to bet so heavily on it. Now the company is moving on.
But roof-integrated solar isn't going anywhere. Competitors like GAF Energy and CertainTeed sell panels that look like regular shingles and cost less. Tesla's own panels on regular roofs are still selling fine. The lesson isn't that solar roofs are a bad idea; it's that Tesla tried to do it all in one product — ceramic tiles, microinverters, battery storage, a proprietary install process — and the complexity ate the margins.
Why this matters for us: the homeowners on our block who wanted solar but didn't want a pile of blue panels on the roof now have cheaper, real options — the kind you can get through a local roofer, not a Silicon Valley startup.
Pebble Time 2 is back, and it's the only watch not trying to be a phone
Pebble is back with the Time 2 at $225. It runs on e-paper, has physical buttons, and lasts weeks on a charge. The watch faces are weird and fun, and the app store is full of little utilities — the kind you actually use instead of the ones that just show you notifications you already saw on your phone.
This is the kind of product the comunidad needs more of. E-paper means the screen doesn't drain the battery or glare in the sun. Physical buttons mean you can check the time with one hand while carrying groceries. The watch doesn't try to replace your phone or track your sleep for a subscription.
Apple cuts hundreds from Siri and Vision Pro teams
Apple told employees that certain roles are on the chopping block as it pivots away from projects like Siri and Vision Pro. The company admitted the layoffs directly rather than hiding them behind restructuring language.
Siri has been a problem for years — people ask it questions, it gives the wrong answer, and Apple keeps promising fixes that never land. Vision Pro is the opposite: gorgeous hardware, $3,500 price tag, and a user base that never grew past early adopters. Cutting both at the same time signals Apple is finally admitting these weren't the future it bet on.
The company is shifting headcount toward what it thinks actually moves the needle — likely the Apple Intelligence push and the Vision Pro's successor, the Vision Air, which is supposed to be lighter and cheaper. The cuts affect engineers across multiple teams, which means projects are getting consolidated and people are getting moved or let go.
Why this matters for us: Apple is the biggest tech employer in the Bay Area and cuts like this ripple through the whole ecosystem — the contractors, the repair shops, the small businesses that built apps for Vision Pro, and the folks who were hoping Siri might finally be useful for their families.
Polansky's lab keeps skin alive — then trains AI on it
Michael Polansky has been quietly building a company that takes living human skin out of the body and keeps it alive for weeks. The tissue stays functional — cells metabolize, immune cells respond — and the platform lets researchers test skincare compounds on it instead of animals or petri dishes. Now the company is going public about the work.
Polansky runs the lab from Los Angeles. He's a former top deputy to Sean Parker and is publicly known as Lady Gaga's partner, but the startup itself is separate from those circles. The approach is straightforward: grow living skin, keep it alive, feed it chemicals, watch how it reacts.
Founders who don't fit in one country
Tom Tunguz wrote that founders are no longer tethered to an office or a headquarters. Capital, teams, and reputation travel — and those who handle it well can build companies from anywhere.
What changed is the cost of starting. Open-source tools, payment platforms, and cloud services lowered the barrier for a person with an idea and a laptop to launch without raising venture capital or renting an apartment in SF. Founders who previously had to move to California can now operate with a remote team and global customers.
This lets them choose where to live — and who surrounds them — rather than ceding that decision to investors. If people living in the South or the interior, with family, can build businesses without entering the coastal race, the map of who does things changes.
Why this matters for us: primas from anywhere can raise a company without leaving family or losing their cultural backpack, and that gives us more options on how to live and work.
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.
Flipkart is moving fast — 1.2M orders a day, nearly triple its November run rate
Two years after launch, Walmart’s Flipkart is delivering between 1.1 and 1.2 million orders a day, nearly three times what it was pulling in back in November. That kind of ramp is what you see when an established player with real logistics muscle enters a market that’s been growing like wildfire.
Quick commerce — the 10- to 30-minute delivery model — is the fastest-growing slice of Indian e-commerce right now. Flipkart isn’t the only one playing this game, but it’s one of the few with the kind of distribution and supply-chain heft that lets it compete without burning cash on subsidies. The numbers suggest it’s closing on the top players.
Why this matters for us: when big US companies pour capital into a fast-moving category like this, it reshapes who gets funded and who gets left behind — and the patterns tend to spill over into how capital flows to founders in our communities, whether they’re building delivery logistics, grocery tech, or side hustles on the margins.
Record South Africans fleeing to the US — 2026 numbers top all time
South African refugees admitted to the U.S. so far in fiscal 2026 are the highest ever. Pew Research Center reported it in this week's roundup, and the number beats every previous fiscal year on record.
South Africa has been in crisis — the violence, the load-shedding blackouts, the crumbling infrastructure — and people are leaving. The U.S. has absorbed them, and this year it absorbed more than any year before.
Why this matters for us: The same pressures — service collapse, violence, the state failing — are what push people toward the border. When South Africans can land in the U.S., it's a signal that the door is still open for folks trying to escape the same rot.