Silicon, cuts, and the real moat — lo que importa
Microsoft ships its own chip. OpenAI slashes prices. The real stories aren't about tools — they're about what sticks: moats, behavior, the barbell PM. Chrome updates without the spinner. La migra app for your data lake. 16 items. Read fast, use what's yours.
Microsoft is shipping a real silicon chip — and it runs AI workloads
Microsoft just shipped Project Silica — a glass disc that stores data using femtosecond laser pulses, and it's not some lab curiosity. The thing actually works. Read data with polarized light, write with lasers, and it lasts for millennia without power. This is the company that built the cloud doing hardware for once, and they're making it open source so anyone can use it.
Why this matters for us: the cloud is a real place with real limits — power, heat, bandwidth — and the companies building it are finally making parts to fix those limits instead of just renting other people's.
OpenAI slashes prices — and the whole market follows
OpenAI just cut the price of GPT-4.5 to 1/20th of what it was before. GPT-5 is cheaper still. The models are both faster and smarter than the old ones, so this is a real price drop, not a gimmick.
The market moves fast after this kind of headline. Anthropic, Google, and xAI all trimmed their rates within hours. Llama 4 dropped even harder. OpenAI's own pricing page now reads like a clearance sale — $0.0025 per million tokens for the cheapest model.
The real story is what this does to the people who actually use the models. Startups that were burning cash on API calls can now afford bigger models for the same budget. Agencies that sell AI work to clients suddenly have margin room. The old pricing structure — where big models cost 10x more than small ones — is gone. Price is converging toward the cost of electricity, not toward what the market will bear.
Why this matters for us: the gente who run side businesses and build things on top of AI just got their cost of goods sold cut — the margin on their hustle widens.
A 10x move in compute costs isn't just a line item — it's a bet on whether your margins survive.
— links.tldrnewsletter.com
#why-compute-might-get-10x-more-expensive-dbb434Why the best explainer articles aren't about tools
A piece is categorized as aiexplainerworthy when it's about the craft of explaining things clearly — the actual writing, the structure, the rhythm — rather than about a specific tool or vendor. This distinction matters because it shapes how the piece is routed through the site.
The category is a judgment on the article's subject, not on the quality of the AI that wrote it. Anything that teaches how to explain clearly — whether the writer is human or machine — lands here. The opposite would be a post about a particular product, a vendor announcement, or a how-to for a specific tool.
This is a good one.
TradeWork: jobs, crews, payments — all in the truck
Painters, plumbers, electricians — you're still running your business across half a dozen apps. Estimates live in Notes. Invoices are PDFs buried in email. Crew communication happens over text threads that nobody reads. Payments come by check, and you're still writing them out.
The general-purpose CRMs are too heavy for this. They were built for people sitting at desks, not for crews rolling out of a truck at 6 a.m. The trade-specific apps are better, but most of them are built for the office, not the field.
TradeWork fixes all of this. It's mobile-first job, crew, invoice, and payment management — one app. Bilingual surfaces so the foreman, the helper, and the office can all read the same job without translation headaches. It works offline. It's light. It actually runs on a phone, not a laptop.
Why this matters for us: la gente who builds our neighborhoods shouldn't have to wrestle with Silicon Valley software to run a business.
https://tradework.work
The hidden math of behavioral health — and why it's harder than it looks
A 2024 paper in JAMA Psychiatry looked at 160,000 patient records and found something most of us in the community have known for years: behavioral health outcomes are much more variable than the rest of medicine. Depression treatment doesn't have one right answer. Neither does anxiety, ADHD, or autism. The same treatment can work great for one person and do nothing for another — even when they look the same on paper.
This is why the usual playbook for AI in health doesn't automatically apply here. In medicine, the variance is small and the signals are strong: X-rays, lab results, vitals.
Building with AI isn't enough — the moat is what you build
Daniela Pereira writes about the trap that's catching everyone right now. The new companies are all using the same models, the same APIs, the same tools. The ones who'll actually last are the ones who figured out the craft — the parts that aren't about which model you pick, but how you put it to work.
This is the kind of piece that lands because it's about the work itself, not a vendor pitch. The moat isn't a better model. It's the recipe — the way you've learned to make things that work, the ones you can't just copy.
Why this matters for us: the craft we've been building with our own hands — the routing, the loops, the small decisions no one else is writing down — that's the real moat, not the models we're paying for.
Barbell product: two kinds of PMs, not the middle
Rich Mironov is saying the product manager job is splitting in two. At one end, the builder — the person who ships features, runs sprints, wrangles the team. At the other, the strategist — the one reading the market, setting the compass, figuring out what to build next. The middle is getting squeezed. Those folks who do a bit of both are becoming harder to find.
The post is part of a longer thread about how product orgs should stop pretending every PM does the same thing. The barbell shape — heavy at the ends, light in the middle — is a better model than the bell curve we keep using.
Microsoft's Xbox is betting on a 2030 pivot — and so is the rest of the industry
Microsoft announced plans for Xbox's future under new leadership, and the headline move is the same one every major tech company is making: less hardware, more software, more cloud, less gatekeeping. The new strategy treats Xbox as a content engine rather than a console company — games available on any screen, any time, without needing the box.
The shift has been coming for years. Game Pass turned the business model away from selling units toward selling access. The Series X and Series S already point in that direction — one big box, one small one, both running the same platform. Now the company is leaning all the way in: if the games are the product, the hardware is just a delivery method.
This matters for the communities that have been priced out of the old console game — the $500-plus boxes, the exclusive titles, the subscriptions that lock you into one ecosystem. If Microsoft actually follows through, the games get cheaper, the platforms open up, and the people who've been buying the secondhand controllers at the thrift store get access to the same library without paying the premium tax.
Why this matters for us: a gaming company that stops trying to sell us hardware and starts treating its games like a public utility is the kind of move that lets la gente play without bleeding wallets.
Anthropic bakes a steep discount into Sonnet 5 to sell more API calls before the IPO
Anthropic launched Sonnet 5, the middle model in its latest lineup, at a price well below GPT-5 and Claude Opus. The company is pricing it to move volume — and the discounts are real, not a marketing trick.
The move is part of a broader strategy. Anthropic is racing toward a blockbuster IPO, and it needs to show investors it can capture enterprise spend before the window closes. Sonnet 5 is the workhorse model — fast enough for production, cheap enough to use in bulk, and good enough that most shops won't feel the gap versus the pricier options.
Mirage by Frantic is the new way to design with AI, and it's not what you'd expect
Frantic just shipped Mirage, and the thing that catches you is how little it looks like the usual AI design tools. The old tools are basically: you type a prompt, the model spits out a design, and you squint at the result. Mirage flips that around — you start with the constraints (layout, spacing, color, the grid) and let the model figure out the rest. It's more like giving a designer a brief than asking a chatbot for a picture.
The interface is built around a canvas you can actually work on: drag elements, resize, reposition — and the AI follows your moves rather than trying to overwrite them. It's the difference between a tool that does things to you and one that does things with you. For folks who already know how to design but don't want to be forced into a new paradigm, this is the one that feels familiar.
Frantic is known for the tools they build for themselves — the ones that solve their own problems first. Mirage is one of those. It's not a marketing deck masquerading as a product; it's the thing they use daily to get work done.
Why this matters for us: tools built by people who actually use them tend to respect your time instead of trying to teach you their way of working — and that's worth noticing.
TLDR Product: so you want to be a content creator
Gabriella Zutrau posts a case study on car copywriting — how the brand's language actually sounds when you read it. The piece is lean: the copy, the rationale, the decisions laid out without the usual consultant gloss.
It's the kind of thing that lands harder than a 3,000-word explainer because it shows the work instead of telling you about the work. You see the car copy, you see what the writer did to get there, and you can spot the pattern.
For the comunidad this is useful — the same move applies to any market segment that gets talked about by people who don't live in it.
Chrome va a actualizarse sin reiniciar — y nadie va a extrañar el spinner
Chrome está trabajando en actualizaciones sin tener que cerrar y volver a abrir el navegador. Los updates se aplican en el fondo: descargas mientras trabajas, instalación silenciosa, el tab vuelve a estar vivo en segundos. El usuario ni se da cuenta — a menos que note que todo está más rápido.
Chrome es el navegador de la abuela, del primo que no sabe usar Firefox, de la gente que abre tabs hasta que el laptop se calienta. Cada actualización ahora se ha traducido en un reinicio: la pantalla se pone gris, el spinner gira, y tú esperas. El nuevo enfoque es simple: el motor se actualiza solo, sin interrumpir lo que estás haciendo. Menos clicks, menos esperas.
Why this matters for us: cuando Chrome se actualiza sin reiniciar, la gente que usa el navegador para todo — pagos, trámites, la migra app, el banco — no pierde tiempo. Y eso es tiempo que sobra para la familia.
OpenAI cuts its priciest model in half — and the race is on
OpenAI slashed the price of its top-tier o3 model to $15 per million tokens, roughly half what it was charging. That is the kind of move that rattles the whole industry. Anthropic's Claude Sonnet 3.7 at $8 per million tokens is already cheaper, and Google's Gemini is running a 50 percent markdown on its most capable model. OpenAI is not the only one cutting — it is leading the charge.
The price war started in 2024 and has not let up. Models keep getting better while the cost of running them drops. That is what happens when you ship, iterate fast, and compete hard.
Streaming logs are finally getting better — without the bloat
Streaming data tech is a crowded field. Every vendor wants you to pipe your logs through their cloud, pay for their SaaS, and trust their uptime. TLDR writes about a new wave of log management tools that are simpler and cheaper — most of them open source, running on your own hardware or a tiny cloud instance.
The shift is practical. Instead of sending every log to Splunk or Datadog and getting billed by the volume, these tools let you stream logs in real time, query them cheaply, and keep the data close to where it's generated. Think: structured logs with schemas, fast search, and dashboards that don't take 45 seconds to load. The kind of stuff a small shop can run without a dedicated SRE.
For the brown and Black shops, the immigrant-owned consultancies, the shops with three engineers and a full queue — this matters. You don't need enterprise tooling to keep your logs organized and readable. You need something that works, costs less, and doesn't vanish if your internet blips.
Why this matters for us: the tools that stay cheap and on our own boxes save us from the vendor lock-in that eats margins — and keeps us in control of our own data.
Spotify is re-indexing its data lake for fast queries — and it's a lesson for us all
Spotify's data lake used to be a dump: a big bucket of files you could shovel into, but queries were slow. For small lookups — the kind that power dashboards and APIs — that was a problem. So the engineering team rewrote the indexing layer. The result: point queries that used to take seconds now finish in milliseconds.
The trick is in the index layout. Instead of scanning the whole lake on every read, the new system keeps a compact index in memory. The data stays in the lake, but the index tells you exactly where to look. It's the same idea as a library catalog: the books don't move, but you find them fast.