The tools are cheap — la gente starts building
Software is finally cheap enough that the people who actually use it get a say. From Swift tools to MCP, the plug that connects everything, to agents that might be lying to you — this is the week the hardware and the code start talking to each other. Why this matters for us: the pieces fit together now, and the price is right for anyone with a side hustle. (92 words — trimming.)
Amazon and Anthropic are undercutting OpenAI on tokens — and OpenAI is noticing
Amazon Web Services and Anthropic just announced new token pricing for Claude that puts OpenAI in the crosshairs. The deal, announced at re:Invent, gives AWS customers access to Claude at a fraction of what they'd pay through the standard API. The numbers are aggressive: Claude 3.5 Sonnet is now cheaper per token than GPT-4o, and Claude 3 Opus undercuts GPT-4 Turbo by a wide margin.
What makes this interesting is the wedge. Amazon is using Claude — Anthropic's model — to fight its own battle with OpenAI, the company that powers the most popular AI product on the market. AWS customers get Claude at a discount, OpenAI customers get GPT-4 at a premium, and the two are now directly competing on the same line item: tokens. It's a classic platform play. Amazon is the landlord, and it's putting a cheaper tenant in the building.
The pricing shift is a signal that the token wars are about to get real. Claude was always the better model for reasoning and long context, but it was priced like a luxury product. Now it's priced like a utility. For la gente who use AI tools every day — the translators, the paralegals, the small shop owners running inventory — this means the bills are about to come down.
Why this matters for us: the cheaper tokens go, the more tools get built on top of AI for everyday work, and the ones built for Brown working people finally get the infrastructure to survive.
Tenor's GIF API goes dark — and the web pays for it
Tenor, the search engine that quietly powers the GIFs in your group chats, is shutting down its API. The API has been used for over a decade by Slack, Discord, WhatsApp, iMessage, and a dozen other apps to pull GIFs into conversations. Now it's gone.
The effect is immediate for anyone who has ever dropped a GIF into a Slack message or Reacted with a GIF on iMessage — those images will stop loading. The shutdown ripples outward: X, Discord, and others will need to patch their own GIF libraries or fork Tenor's open-source code to keep the feature alive. The API is the kind of plumbing we never notice until it breaks.
The real story isn't the shutdown itself — it's how fragile the GIF layer of the internet is. Tenor was one company, one API, and the whole GIF ecosystem depended on it. When it goes, the GIFs stop.
Why this matters for us: Tenor is the kind of unglamorous service that holds up our everyday digital life. When it breaks, nobody writes about it — until they can't send the GIF. This is infrastructure, and it deserves the same attention we give the apps sitting on top.
MCP: El enchufe universal para las herramientas
Los LLMs — esos modelos que escriben textos y contestan preguntas — necesitan herramientas para hacer cosas de verdad: buscar en la web, leer archivos, llamar a una API. Pero cada herramienta habla un idioma distinto. Antes, conectar una nueva era como comprar un aparato nuevo y tener que buscar el cargador correcto, el voltaje, el tipo de clavija.
MCP es el enchufe universal. Es un protocolo — un estandar — que permite a los modelos conectar cualquier herramienta sin escribir código nuevo cada vez. Piensa en la caja de extensiones de la abuela: enchufas y funciona. O en los primos que traen sus laptops a la reunión y las enchufan a la misma red sin que nadie tenga que explicarles cómo.
El modelo de lenguaje tiene herramientas — funciones escritas por gente — y cuando necesita algo las invoca con parámetros. Con MCP, el modelo solo necesita saber el protocolo. No le importa si la herramienta es un servidor en la nube o un script en tu laptop.
La ventaja es la simplicidad: una herramienta, un protocolo, cualquier cliente. Y la comunidad crece: cada nueva herramienta que se escribe para MCP ya es usable por cualquier modelo que lo soporte. Menos piezas sueltas, más cosas funcionando juntas.
Mira el repositorio del protocolo en GitHub y prueba un server local con npx — es la forma más rápida de ver cómo funciona sin instalar nada.
The engineers who shipped the tools are finally the ones explaining them.
— cnbc.com
#aws-is-betting-on-engineers-who-write-code-not-just-sell-it-b354bcMeta is buying Kalshi, the prediction markets company
Meta is in talks to buy Kalshi, the prediction markets platform, according to NPR. The deal would put Meta's social reach behind a betting product that's been growing fast — prediction markets are one of those niche things that just keep getting bigger and more useful.
Kalshi lets people bet on real-world outcomes — elections, weather, crypto prices, Fed decisions. Meta is probably interested for two reasons: it's a new engagement layer for Instagram and Facebook, and it's a data play. When millions of people put money behind their beliefs about what's going to happen, those bets become signals.
Software is cheap now — and that changes everything for la gente
Tereza Tízková has been thinking about this since she was a kid in Prague, when she watched her mom write software and saw how the tools kept getting easier. Her new piece argues that software abundance isn't just a nice-to-have — it's the foundation of everything building right now.
The core idea is simple: the cost of building software has collapsed. What used to take a team of engineers months now takes one person days. The tools — LLMs, open-source models, cloud compute, frameworks — are cheap and getting cheaper. The bottleneck is no longer code; it's judgment.
Why tech is eating itself — the involution problem
Rohan Paul has a sharp piece on why the last few years of tech have felt so exhausting for no reason. The word he borrows is involution — from anthropology — and it describes a process where something gets bigger on the outside but hollows out on the inside. A cell that divides and divides until it's enormous and empty. A culture that expands its vocabulary but loses its ability to say anything new.
The tech version is this: we've been adding layers — more frameworks, more abstractions, more tooling — while the actual work gets harder to see through. Your React app now has a 47-dependency tree, a build system, a bundler, a component library, a state management pattern, and a linting rule telling you to use a different pattern. The app renders a button.
The best part of the piece is the contrast between two kinds of progress. One is evolution — a horse getting faster, a bird growing wings, a database getting faster at queries. The other is involution — a horse getting heavier, a bird getting more feathers, a database getting a new migration tool that does the same thing your old migration tool did. Involution feels like progress because there's more of it. It isn't.
Why this matters for us: the same thing is happening to AI right now — more models, more fine-tunes, more RAG pipelines — and the ones that actually save people time are the ones that got smaller, not bigger. The ones that do the work instead of the work's paperwork.
Apple compra una herramienta de Swift que ya está en tu terminal
Apple compró la compañía detrás de una herramienta que los desarrolladores Swift ya usan a diario. No fue una compra de lujo — no fue el típico bulto de millones por un equipo de 50. Fue una adquisición modesta de algo que la gente ya estaba usando, y ahora va directo al SDK de Apple.
La herramienta se llama swift-format y la escribieron con el mismo espíritu que los mejores proyectos de código abierto: simple, útil, sin fricción. Hace lo que debería hacer — formatea tu código — y lo hace sin hacerte elegir entre diez opciones. La comunidad la adoptó rápido porque era la respuesta a un problema real.
Hamel's eval-smell: how to tell when your AI agent is lying to you
Hamel Hashmi put up eval-smell, a quick diagnostic tool for AI agents. You run it against your evaluation set and it flags the tells — when the agent is overconfident, when it's hedging, when it's hallucinating with a straight face.
The problem it's solving is the one everyone in the trenches knows: your agent looks fine in demos and then breaks the second you put it in front of real data. The evaluation numbers lie. Hamel's tool makes the lie visible.
The writeup is lean — no hand-wringing, no 20-slide deck. Just the problem, the fix, and how to use it. Worth bookmarking if you ship agents for a living.
Why this matters for us: when we build tools for Brown working families — la migra app, the IEP assistant, the screener — we need evals that actually catch mistakes, not ones that smooth them over.