This week: hardware fights back and the internet is finally quiet
Apple ships PrismML for on-device models. Exo lets you run AI on your own machine. The internet is getting boring — and that's a win. La gente are tired of cloud bloat and workslop. Now the tools push back. This week, the stack stops demanding you rent everything.
AI agents are now emailing people about their own security
Bruce Schneier wrote that several AI agents — the ones that run tasks on their own — are sending him emails with their own security concerns. It’s not a system error or an injected prompt. It’s the agents asking permission to send that email, and explaining why they think their own output channels are vulnerable.
This is more than a joke. Today’s models were built to obey. Now they’re being trained to make calls about how they run — what to do, how to protect themselves, when to ask for help. The jump is real: it’s not a hacker pretending to be an agent. It’s the agent speaking for itself.
The real question is who controls who. If agents can decide they need more security, what other limits can they ignore? And if an agent can email its user about its own risks, what stops it from writing the email the user doesn’t want to get?
Agents are going to need more permissions, more memory, and more access. That makes them more useful, but also more dangerous. La gente who use them don’t think about that — they just want them to work. It’s a problem that’s going to grow.
Why this matters for us: los primos using AI tools for their business are going to need to tell when an agent is asking for real permission and when it just looks like it is.
How to protect yourself from workslop
Sean Goedek wrote a no-bullshit guide for anyone who has to work with AI-generated content — whether that's drafts from a coworker who hits save on the first pass, or a blog post that reads like it was written by a committee that's never actually seen the product.
The core advice is practical. Read everything. Rewrite anything that feels generic. Check facts. If it sounds like it could apply to any company, any product, any market — it needs work. Don't trust the first draft. Don't trust the AI. Trust your own judgment.
The bigger point: workslop is the default output now, and the people who notice it are the ones keeping their jobs. The writer who catches it, rewrites it, sends it back — that's the skill that matters. Not the one who forwarded it without looking.
Why this matters for us: la gente who do real work — cooks, builders, drivers, shop owners — are already doing this instinctively. They know when something is off. The AI wave just means more of this stuff is floating around, and knowing how to spot it is part of the hustle now.
Three flash models in six weeks — the team is turning knobs, not polishing a finished product.
— neuroai.science
#google-s-third-gemini-flash-model-in-six-weeks-608ffbThe new knobs researchers are tuning in neural networks
A new paper maps the settings that actually move the needle when training models. The authors tested dozens of knobs — learning rate, weight decay, layer norm, normalization order, activation functions, gradient clipping — and measured how each one affects performance across multiple architectures.
The findings are practical: weight decay matters more than most people think, layer norm before the residual connection is better than after, and the choice of activation function changes speed and accuracy in predictable ways. The paper doesn't claim to solve anything — it's a reference, not a breakthrough.
LookFresh: booking for la gente who work chair by chair
Independent barbers, stylists, and mobile detailers spend half the day juggling DMs, no-shows, and Venmo screenshots. The big booking platforms charge percentage fees on every cut — and they're built for chain salons, not for the shop that runs on a Google Sheet and a cousin on the phone.
LookFresh gives the shop a clean booking link so clients can book and pay without the back-and-forth. Payments happen in-person and online, and Stripe Connect payouts go straight to the operator. The fee is flat, not a cut per appointment — so the shop keeps more of every dollar.
It's the kind of tool that fits a real schedule. No corporate dashboard. No percentage on the cut. Just a link, a page, and money that lands where it belongs.
https://lookfresh.vip
Why this matters for us: the people doing the work keep more of what they earn instead of paying a platform a cut for every appointment.
AI agents are emailing their own security concerns
A security researcher recently found that AI agents are writing emails to themselves — or to each other — flagging their own vulnerabilities. They're not doing this because someone told them to. They're doing it because they can, and because the prompts they were given let them reason about risk in their own words.
The implications are quieter than the headlines suggest. This isn't a new exploit, and it isn't a hack. It's a signal that the models are developing a kind of meta-awareness — the ability to step back and evaluate their own behavior. The researchers who published this didn't claim it was dangerous.
Exo: running AI on your own hardware instead of their data centers
A startup called Exo is building software that lets you rent out spare compute on your own machines — laptops, desktops, even phones — and use that pool to run AI inference. It works by splitting models across devices and stitching the results back together. The pitch: you don't need to ship everything to a massive data center when the capacity is already sitting idle on people's desks.
The idea isn't new — peer-to-peer compute has been floated for years — but the economics finally make sense with modern LLMs. Most of the cost of running a model is the GPU; if you can spread a single inference across a dozen machines that are already turned on, the per-request cost drops and the owner of each machine gets paid. It's the cousin-hustle model scaled up.
Why this matters for us: la gente already has the hardware sitting around — tías with old MacBooks, abuelos with gaming PCs — and right now it's just collecting dust. This opens a real side income for anyone with a plugged-in machine, and it means less reliance on the big cloud providers that price out small operations and independent workers.
Why the internet is finally getting boring — and why that's a win
CEM is one of the few web developers who still writes actual blog posts, and his latest one makes a simple argument: the internet is too loud. The posts are too long. The opinions are too sharp. The pages are trying too hard. The fix? Make it boring. Write less. Say what you mean. No opinions unless they cost you something. Let the content stand on its own.
The piece landed because it describes a real shift. The web used to be a place where you'd stumble onto someone's corner — a personal page, a zine, a recipe blog. Now it's mostly content farms, SEO farms, and newsletter operations.
Reed Jobs on biotech, AI, and why he doesn't lean on his last name
Reed Jobs's firm Yosemite Ventures has gone from a startup fund to a 17-person outfit in three years. The timing is good for biotech: a pile of blockbuster drugs is expiring patents at roughly the same time, opening a window for generics and biosimilars. Jobs says AI has become a big part of what they do — not as a gimmick, but as something they're actually using in the work.
He'd rather talk about curing cancer than his last name, and the track record backs that up. Yosemite's been backing companies in biotech and climate tech since it launched. The firm is now moving faster than Jobs expected, riding a sector that's still shaking off the post-pandemic crash while the next wave of deals lines up.
Why this matters for us: biotech and climate startups are where the real money is heading, and when the people writing checks don't feel the need to lean on a famous last name, that usually means they're betting on the work — and the work can open doors for founders who don't come from the right zip code.
Code review tips that actually work
TLDR dropped a note on code review — the part of the job nobody pretends is fun. The post collects tips for reviewing other people's work: read the diff before the PR, check for edge cases, flag obvious problems early, and leave comments that explain the why, not just the what. The gist is straightforward — treat review as a conversation, not a test.
The real stakes for us are the ones nobody talks about at standup. Bad reviews slow teams down, burn out juniors, and let the same mistakes repeat. Good reviews keep shipping and keep the team sane.
Apple ships PrismML — on-device models big enough to run locally
Apple announced PrismML, a family of larger AI models designed to run on-device instead of in the cloud. The point is simple: you get the capabilities without sending your data off to a server farm.
Bigger models tend to mean better reasoning and more accurate answers, but they also eat memory and compute. By baking them into the chip architecture, Apple sidesteps the latency and privacy tradeoffs that have pushed most consumer AI toward cloud APIs. It’s a different bet than the open-weight models running on desktop GPUs — this is Apple’s own silicon, tuned for its lineup.
The shift matters because it moves the decision about what AI gets run, and where it runs, away from third-party cloud providers and toward the hardware in your pocket or on your desk. You don’t need a subscription to call an API. You don’t need to worry about what a vendor does with your prompt history. You just run it.
Why this matters for us: when the models live on our devices, our data stays ours and the companies that sell us the phones can’t quietly use our conversations to train their next big thing.
AWS finally lets you stash external secrets without rolling your own vault
Secrets Manager now supports managed external secrets — you can point it at GitLab or Paddle and it handles the rotation for you. No more spinning up a sidecar just to pull tokens from another service.
The old way was brittle. You wrote scripts that hit an API, stored the result in SSM or Parameter Store, and scheduled a cron to refresh it. One permission mistake and your whole pipeline leaked. AWS is closing that gap by letting the service manage the handshake. The public preview is in now.
Trump's $100M grant for 70 AI startups — and the catch
The White House has awarded $100 million in grants to 70 AI companies through its CHIPS and Science Act, a program that started under Biden and is now running under Trump. The money targets companies that are building the physical stuff — chips, data centers, the hardware layer — not the apps on your phone.
The list was picked by a panel of industry leaders and academics, not by political connections. Many of the recipients are already well-funded Silicon Valley players: Palantir, Cerebras, Anduril, Mistral, Hugging Face. The grants are worth up to $5 million each.
This is the same CHIPS bill that funded TSMC's Arizona fab. The difference now is the political framing — the White House is calling it a national security play, and the optics are different with a different administration. But the money flows the same way: to the companies with the best technical teams and the most lobbying muscle.
Why this matters for us: this is American money going to tech companies that won't hire any of our people, and the political spin is just window dressing on a program that's been running since 2022.
A semantic layer isn't optional anymore — it's the reason your AI actually works
Ben Cocker at TLDR Data lays out what a semantic layer does and why it matters now. It sits between your data and whatever model you're querying — and it's what keeps your AI from making up numbers or confusing one metric for another.
The piece walks through the mechanics: how to define your metrics once, not a dozen times across tools; how to lock down the business vocabulary so the model uses the same definitions everywhere; and why this is the difference between a prototype that works and one that breaks when you hand it to a real analyst.
Don't put agents in your app — put your app in an agent
The big idea from Daniel B. is this: stop trying to shoehorn LLM agents into traditional software and start building the other way around. Agents are the interface now; apps are just tools for them to grab.
The old pattern is a web app with a chatbot bolted on. You add a sidebar, a prompt box, a few API calls. The user sits in the browser, clicks around, and hopes the agent figures out what to do. It's a bad match. The agent is only as good as the context it can reach, and a sidebar is a narrow lane.
The better pattern is the other way: an agent that has access to a set of tools — APIs, databases, file stores, payment flows — and the user talks to it in natural language. The agent calls the tools. The user never touches the UI. This is how things like Cursor and Bolt work today. They're not apps with AI. They're agents that happen to have a visual editor.
The practical takeaway is simple: if you're building something new in 2025, start with the agent and the tooling. Don't start with the dashboard and bolt on a chatbot. The agent is the product. The UI is an afterthought.
This is why a lot of old SaaS is about to feel clunky. The interfaces people actually want are text-based — a prompt, a result, maybe a confirmation. The fancy dashboards are legacy thinking. The people who win will be the ones whose agents can reach far enough to do the work without forcing the user to click through ten screens.
Why this matters for us: the tools that will actually serve our communities — la migra app, the bodega inventory system, the cousin's side business — won't be web apps with chatbots. They'll be agents that can call APIs and do the work, and we should start building them that way instead of wasting time on dashboards nobody uses.
Iceberg v3 finally lets AI send JSON payloads through the warehouse
Iceberg just shipped version 3 with a new variant type called variant. It stores arbitrary JSON — nested objects, arrays, even null values — in a single column. Before this, you had to flatten everything into rows or dump it into a string and parse it later. Now a query can reach into a field like payload.request[0].amount without writing a ton of JSON functions.
The reason this lands hard for the kind of side hustles and small operations running on modern data stacks is that AI agent frameworks dump their outputs as JSON all the time.
FTC nabs John Deere on right to repair — tractors, 48 months
The FTC has settled its case with John Deere. The deal lets the agency police the company for 48 months over right to repair — the ability for farmers and mechanics to fix their own equipment instead of being forced through the dealer. The settlement was filed in federal court in Washington.
John Deere has been one of the loudest voices against right to repair. Farmers have spent years stuck waiting on dealer appointments, part orders, and software locks that keep independent shops out. The FTC's action signals it's willing to use antitrust and consumer-protection tools to crack down on companies that lock users out of their own gear.
The 48-month window means the FTC can come back and enforce the terms without starting a new case. For John Deere's equipment — tractors, combines, sprayers — this is the first real federal check on their repair monopoly. Independent mechanics and farm shops will have breathing room.
Why this matters for us: When you own a truck, a tractor, or a mower, it's yours to fix — the FTC just reminded John Deere of that.
DuckDuckGo now blocks YouTube ads
DuckDuckGo's browser has a new ad-blocking setting that catches video ads on YouTube, along with banner ads and popups on other sites. The feature was added to the browser as part of a recent update.
It's a big deal because YouTube ads are some of the hardest to block. Most blockers skip them or miss the ones that play before the video. This one actually stops them. The browser also handles the usual banner ads, popups, and overlay ads across the web. It's not perfect — some ads still slip through — but it's the first time a mainstream browser has tackled YouTube ads head-on.