The internet's noise floor is rising — and la gente is paying for it
Cold email's dying, the internet's dying, and AI agents are quietly eating the world. Meanwhile delivery apps are raising prices on the working class. La teoría del internet muerto is real now. We're not just watching the shift — it's happening in our pockets.
Cold email is getting harder, and it's not the content — it's the noise floor
Zach Holman has been sending cold emails for years. He writes them directly, no AI polishing, no templated fluff. His latest post is a no-bullshit read on what actually works when you're trying to reach someone who doesn't know you — and what's getting worse.
The problem isn't that people are ignoring emails. It's that every inbox is now a flood of AI-generated outreach, newsletters, and automated pings. The signal-to-noise ratio is collapsing. Holman notes that the same email that would have gotten a reply in 2018 now lands in the 200th unread message by the time the recipient gets to it. The content matters less than the timing and the subject line — because the subject line is the only thing standing between your note and the delete button.
His rules of thumb are practical, not philosophical: write to one person, not a company; lead with a specific ask; keep it under 100 words; never use "I hope this finds you well" — it's the cold-email equivalent of a LinkedIn handshake. He also calls out the new anti-pattern: AI-polished emails that sound too smooth. The ones that feel real — a little rough at the edges, specific, personal — are the ones that get opened. The ones that feel like they were generated by a model that read a lot of sales copy will be scrolled past.
This is an aiexplainerworthy piece because it's not about a tool or a vendor. It's about the craft of writing clearly in an age of noise. The lessons apply to anyone — not just founders and PMs, but the cousin who's trying to get her catering business noticed, the auntie on Facebook posting her side hustle, the new grad sending out resumes. The rules are the same: be specific, be short, be real.
Why this matters for us: cold email is how we'll reach schools, clinics, and law firms — the people who decide whether they need Obsidian Comms. If we write like we're talking to one person, not pitching a company, our notes will land.
Zach Holman on cold email — the kind that actually gets replies
Zach Holman wrote a practical post about cold email, the kind that works — not the polished LinkedIn version but the real thing: short, specific, and sent to the right person. The post covers the mechanics — subject lines that don't get trashed, how to write the body so the recipient actually finishes it, when to follow up and when to stop. It's the kind of advice you can use today.
What makes Holman's version useful is that he wrote it from doing the work, not theorizing about it. The tips are concrete enough to apply directly: how long the email should be, what the subject line should say, when to mention a shared connection. No fluff.
Why this matters for us: cold email is how we reach the people who run the businesses and organizations that matter to our comunidad — the abuelas, the shop owners, the community leaders — and Holman's rules work for that world as much as for tech.
Why this matters for us: the best cold emails sound like one person writing to another, not a company broadcasting. That's the same voice this news uses every day.
Why this matters for us: this is about the craft of explaining things clearly, not about any specific tool — the same craft that shapes how we write every item here.
The real cost isn't the call. It's what the call makes the system do next.
— martinfowler.com
#the-llm-tax-on-every-orchestration-call-dc6fa6What the discourse got right about LLMs — and what it got wrong
The Discourse is a forum for people who actually build LLMs — researchers, engineers, founders, the ones who have been through the fire. Their recent thread on what's working and what isn't is one of the clearer summaries of the field I've seen.
The big takeaway is that the old narratives are fraying. People used to say the frontier is models, or compute, or data. Now the consensus is shifting: the real bottleneck is getting models to do what you want them to do, not the models themselves.
Obsidian AI — La IA que no sale de la oficina
IEPs, expedientes médicos, discovery legal, expedientes del condado — la gente que maneja estos papeles necesita IA para redactar, resumir, traducir y buscar. Pero las reglas son claras: los datos no pueden salir del edificio. FERPA para escuelas. HIPAA para clínicas. CJIS para la ley.
Y sin embargo, la mayoría de las soluciones de IA «privada» todavía hacen llamadas a la nube. O bien son proyectos de investigación con una pila local que requiere un ingeniero para mantenerla viva.
Obsidian AI resuelve esto. Es una unidad completa — GPU, modelos, agentes, voz, consola de administración — que se instala en la red de tu organización. No hace llamadas salientes. La superficie es la misma que BFTS Chat, pero los datos y el cerebro no salen de la pieza.
Es lo que necesitas cuando el sello de confidencialidad no es suficiente: cuando los datos deben estar físicamente en tus servidores, no solo «bajo contrato».
https://brownforces.io/solutions
SpaceX quiere meterse con los carriers — y ya tiene la red
SpaceX está armando su propia red satelital para competir con los grandes carriers de telecomunicaciones, y no lo hace con promesas: ya tiene la infraestructura. Los Starlink satellites — esa constelación que se ve en el cielo cuando oscurece — son el primer paso. Pero la jugada es más grande: una red propia que no depende de las compañías de telefonía tradicionales ni de sus torres ni de sus contratos.
La diferencia es clara. Los carriers actuales mueven datos por fibra y torres terrestres que hay que construir, mantener y pagar. SpaceX lleva su red al espacio.
ChatGPT's agent loop got an upgrade, and it shows
ChatGPT has been quietly patching its agent loop — the mechanism that lets the model chain multiple tool calls together instead of stopping after one. The key change is an eval gate: before it commits to a next step, it asks itself whether it's confident enough to proceed. If not, it loops back for clarification rather than guessing wrong.
The idea is simple but changes the feel of the experience. Instead of plowing through a task with blind confidence and potentially wasting turns, it pauses when it's unsure — a small cost that saves bigger mistakes. The pattern echoes the bounded-confidence approach we're using elsewhere: move fast when sure, flag when not.
Why this matters for us: it's a pattern worth copying — confidence gates in tool loops save time and errors, and they don't require new models, just better wiring.
Why this matters for us: the same pattern — move when sure, flag when not — is exactly what we're building into Kelex so the agent doesn't stall or guess wrong.
OpenAI's revenue is finally catching up to the hype
OpenAI's CFO Sarah Friar told employees that July's ARR topped all of Q2. That's a big jump — they're now doing more revenue in a single month than they did in the entire previous quarter.
The timing is notable. OpenAI has been spending heavily on the new data centers, and the market has been wondering if the revenue would keep pace. This number suggests they're getting there.
Friar is an ex-PayPal exec with a strong background in fintech. Her presence as CFO is part of OpenAI's push to make the business side as serious as the tech side.
La teoría del internet muerto ya es real
Fast Company reporta que la "dead internet theory" — la idea de que la mayor parte de lo que navegamos ya no es humano, sino robots hablando con robots — está ocurriendo de verdad. Los web agents (agentes que navegan la web) crecen más rápido que el tráfico generado por humanos. Lo que antes era un nicho de automación técnica ahora es el tráfico dominante.
Esto no es solo bots generando contenido. Son agentes que leen páginas, evalúan productos, toman decisiones — y lo hacen en el idioma que las máquinas hablan. Si tu página solo está pensada para humanos que leen con ojos, los agentes pueden pasar de largo.
El ajuste ya empezó: páginas optimizadas para ser leídas por máquinas, no solo por personas. AEO (Answer Engine Optimization) compite con el SEO tradicional porque los agentes no buscan como nosotros — leen diferente, comparan diferente, recomiendan diferente.
Why this matters for us: la gente que no optimiza para agentes va a perder terreno sin darse cuenta — sus páginas siguen ahí, pero ya no las está viendo quien decide.
AI is not replacing marketing — it is a new kind of worker
A reader named Almundo wrote on Substack that the usual panic — AI is coming for your job — is backwards. The real story is that AI is becoming a coworker, not a replacement. Think of the way an intern shows up on day one: they can draft copy, sort data, write the first pass at a deck, and do it without complaining. That is what AI does now. It is the intern.
The piece says the old model was a person doing the work. The new model is a person directing a small team of AI interns. You tell each one what to do. You review. You correct. You put the pieces together.
Warner Bros. sues Amazon over its lawless employee shopping spree
Warner Bros. Discovery has filed suit against Amazon, accusing the tech giant of illegally poaching talent. The complaint names Pia Barlow, former senior VP for originals marketing at Warner, who is set to start as Amazon's head of original series marketing on August 3rd. She was still under contract — until October 31st — when she left.
Francesca Orsi, HBO's head of drama series and films, is believed to be another target. The suit frames it as a broader pattern: Amazon has been raiding well-established Hollywood houses, bringing over executives on the cheap before their non-compete windows close. Warner calls it a "lawless employee shopping spree" and says Amazon has chosen to ride on the coattails of other studio veterans rather than developing its own talent.
The case is about leverage — and timing. If the court agrees, it could make it harder for Amazon to keep poaching senior creative staff without paying for the transition period. For the streaming wars, it's a signal that the old guard is fighting back.
Why this matters for us: the people who build what we watch are worth more than the contracts that bind them — and the companies that ignore that pay the price.
Llama models write blog posts worse than Claude — and here's why
The post on wakamoleguy.com puts LLMs through a quiet stress test: ask them to write a blog post, then look at the result. The headline claim is that Llama models are surprisingly bad at this, worse than Claude for the task — not because they're dumb, but because the way they're trained makes them drift into a certain kind of generic writing.
The piece walks through what actually goes wrong. Llama writes sentences that are grammatically correct but structurally flat. It defaults to the same cadence across topics — the same opening gambits, the same transitions. It avoids mistakes by being safe, and safety looks like sameness.
Wide-column databases are quietly eating the world
TLDR Data's latest post walks through what a wide-column store actually is. Bigtable and Cassandra use the model — wide rows keyed by a primary column, with any number of sparse columns spread across them. It's the difference between a table and a map.
The advantage is real. You can add columns without touching the schema, and queries touch only the columns they need. That's why write-heavy workloads — analytics, time-series data, logs — gravitate toward these stores. They don't need the full relational machinery to be useful.
If your data is naturally wide — one row per user, per device, per session — a wide-column table maps to it directly. If it's narrow and relational, you're paying for columns you don't use.
Why this matters for us: the shift from rigid tables to flexible stores is one of the quieter wins for people running lean — less schema churn, less lock-in, less ops overhead.
Monte Carlo ships a new observability tool for agents, because LLMs are hard to read
Monte Carlo released a tool that watches agents — the kinds that use tools, call APIs, and chain together several model calls — and records what they do so engineers can actually debug them. For a while, observability was built for APIs and microservices: latency, errors, traces. Agents moved the goalposts. A single request can spawn a dozen calls to different models, each with different prompts and different tool calls. The old tools don't capture that.
The new tool sits between your agent and the models it calls.
LangChain is building the data plumbing for AI agents — not the agents themselves
LangChain just put up a blog post describing a new layer of infrastructure for the agent stack. The move is deliberate: they are no longer selling just the agent framework. They are building the pipes that let agents fetch data, store results, and talk to each other.
The new pieces include a vector store for embeddings, a graph store for relationships, and a table store for structured data — all with an agent-friendly API. Agents can read and write to these stores without the developer wiring up separate databases. The idea is that a single agent can query a vector store for a document, drop its findings into a graph, and write a summary to a table, all through one consistent interface.
This is not vaporware. The components are already live in the LangChain release and the code is on GitHub. LangChain is positioning itself as the plumbing company for the next wave of agent tooling, not the agent builder.
Why this matters for us: the people building tools for our schools, clinics, and small businesses are going to use LangChain — so we should know what it's building, before the vendors do.
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.
The delivery apps are quietly raising their prices — and it's about to hit the working class
DoorDash, Instacart, and Uber Eats are raising fees on the very people who rely on them most. The increases are small on paper — a few cents per order, a dollar here, a dollar there — but they add up fast for families who order groceries and meals weekly.
The apps have been subsidizing fees for years to lock in habit. Now that they've won, they're taking back the margin. The real cost isn't just what you pay; it's the quiet erosion of convenience, the point where $15 feels like $18 without anyone saying a word.
This matters for working families who depend on these services for groceries, meals for kids, and the auntie who can't drive to the store. The price hikes hit the people who need the service most — and have the least flexibility to shop around.
Why this matters for us: la gente who relies on DoorDash and Instacart for weekly groceries and meals is about to pay more for the same convenience — quietly, without anyone asking.