Acabo de presentarle al CTO de una empresa de ~9,000 empleados los resultados de la encuesta interna de AI que conduje.
Llegué esperando contar la historia de siempre: falta adopción, hay resistencia, se necesita capacitación, etc...
Pero esa historia ya no existe: 96% de la organización ya usa AI de una manera u otra, el año pasado, el número era ~44%
Lo que me llamó la atención es que la empresa se partió en 2 grupos que usan la misma herramienta para cosas que no se parecen en nada.
Un grupo la usa como godin sin paga: resume juntas, limpia correos, busca cosas.
El otro la usa de multiplicador de productividad: agentes, integraciones, apps internas integradas las unas con las otras. 63% hace trabajo que hace un año era imposible sin un developer.
Entre developers el uso es de 88%.
Y al fondo de los datos, una frase que los dos grupos repiten palabra por palabra: "No confío en el output"
Los que menos producen la dicen como excusa.
Los que más producen verifican todo, tiran a la basura lo que no pasa su barra de calidad, y diseña mecanismos para que la probabilidad del error disminuya en la siguiente generación.
Desconfiar y producir más no es una contradicción. Es una señal de que los devs están usando la herramienta correctamente.
Confiar a ciegas en el output y descartar por completo la herramiente, son en esencia la misma cosa: los usuarios estan pendejones y no quieren pensar.
El error que veo yo es que muchos esperan que la IA democratize la productividad...
pero no, en la oficina siempre va a haber calienta sillas huevones.
Lo que SI hace la IA, es que amplifica la productividad que ya existía y lo que la encuesta volvió imposible de ignorar es que la empresa ya sabe quién es quién.
For over 20 years, we've dedicated ourselves to removing language barriers so people can learn, speak and connect more deeply than ever before.
Today, we’re taking our next step with the release of Gemini 3.5 Live Translate — our latest audio model for live, speech-to-speech translation across 70+ languages. 🧵
Hoy publico mi primer libro: «Menos software, más impacto». Da un poco de vértigo.
La tesis: tu equipo no va lento por escribir mal código. Va lento porque escribe código de más. 🧵
llama.cpp now has an official website: https://t.co/vztdUpdBWL
Our goal is to make local AI accessible to everyone, and improving the user experience is a big part of that. On the new landing page you’ll find a single-line cross-platform installer. The installation provides a single unified `llama` entrypoint which you can use to run/serve models and interface with 3rd-party agentic applications.
While oriented towards simplified user experience, the new `llama` application also provides all the advanced functionality of the existing llama.cpp tooling with which experienced users are already familiar. Also note that all GGUF models that you might have already downloaded with llama.cpp in the past will be automatically available to use without downloading again (they are stored in the common HF cache on your machine).
We have many improvements in the pipeline both at the UX and at the engine level and we plan to iteratively ship new things over the coming months. One of the main focuses will be seamless integration with local-friendly 3rd-party agents (such as Pi). In the meantime, we’ll continue to listen for feedback from the community and adjust accordingly, so keep letting us know what you think and need.
I've got an agent in a loop optimizing a renderer with the goal to minimize frame times (and tests to measure). It got times down from 88ms to 2ms and allocations down from ~150K to 500. Sounds good, right? Wrong. This is exactly why agent psychosis is a big fucking problem.
As an experiment, I rewrote the Ghostty core render state in Go, with access to identically laid out data structures as Ghostty and the exact same validation tests. I made a purposely naive renderer (simple, correct, but slow). 88ms per frame with 150,000 allocations (horrendous, lol)!
I then kickstarted a Ralph loop to bring the frame times down. I told it it can't modify input data structures or the public API or tests (they're correct), but it can do anything else it wants. It got to work.
It has worked for about 4 hours. I've spent around $350 on this experiment so far. The results?
88ms => 1.5ms
150K allocs => ~500 allocs
Incredible right? Nope.
My hand-written renderer I ported has frame times (same benchmark) of ~20us (0.020ms) and 0 allocations in the update path.
This is the problem with psychosis and lacking systems understanding. If you don't understand the system, you're going to accept that this is an incredible result. If you understand the system, you'll see better solutions immediately and can do roughly 75x better on throughput.
The people who blindly trust agent output are in the former camp. They're sheeple, overdrinking from a fountain of mediocrity.
Standard disclaimer: I use AI all the time. I like AI. The point I'm making is to not blindly accept results. Think. Analyze. Learn.
When an early-stage YC company has real traction, they talk about retention
If they don’t have strong retention, they talk about growth.
If they don’t have strong growth, they talk about revenue / ARR run rate or early profits / unit economics.
If they don’t have revenue, they talk about their Demo Day valuation and how many funds were fighting to get in.
If they don’t have strong Demo Day traction, they talk about how many other YC companies they’re partnering with (“portfolio synergies”).
If they don’t have that, they name-drop famous YC alums who advised them
I haven’t read the paper, but I have noticed that attempting to constrain an agent with rule files is a fools errand. They will break any rule, and overturn any stated constraint.
So I use physical constraints instead. Those constraints are things like acceptance tests, unit tests, mutation tests, crap analysis, dry analysis, property tests, etc.
The agents cannot overturn those constraints. Therefore they become zealous — sometimes too zealous — in conforming to them.