Idea de #mcp O #Agente. Revisa tu conversaciones al final del día y ajusta tus agentes arreglando los: NO! ASI NO! NO USA ESTE PATRÓN! Una IA que aprende sola y se mejora a si misma. Antes de irse para cama: "Usa el mcp Inteligencia evolutiva para mejorar los agentes para mañana"
@aena Una vergüenza, T1 de Barcelona temporada alta y solo un control de seguridad abierto. Tercermundista y gestión nefasta. Como es eso posible? No hay planificación alguna?
I'd like to propose @OpenAIDevs / @thsottiaux a modification to the Usage Limit functionality for "ChatGPT Codex".
Current Problem: When the 5-hour usage limit is reached, users are blocked even when they still have plenty of weekly usage remaining. The current workaround is to use a full reset, but that feels wasteful because it resets both limits when only the 5-hour allowance is exhausted.
I’d like to propose a proportional option for banked resets, so users don’t have to spend a full reset just to recover from the shorter 5-hour limit.
For example, each banked reset could be used as either:
- 3–4 five-hour resets, or
- 1 full reset covering both the weekly and current 5-hour limits.
Once that’s in place, an optional opt-in Auto-Use setting could automatically apply a five-hour reset when the 5-hour limit is reached.
Growing up hearing “Wikipedia isn't a valid source” and then entering a workplace where people say “just ask ChatGPT” is a surprisingly strange timeline
we're helping a customer spending $60k/mo move from OpenAI & Anthropic to open source models
they use almost every model offered by the labs, so we needed to find replacements for all of them
after generating evals, this is what we landed on
new cost: $12k/mo, 80% savings
For agentic coding, one can say:
- Unless you need Terra Ultra perf, it's always better to use a Luna model with higher effort setting (same or better performance but cheaper).
- Forget everything below Sol High, use Luna with higher effort settings here
- Forget Sol Extra High, use Terra Ultra here
- The extra cost of Sol Ultra is probably not worth it over Max
Me: How long will this project take.
Claude: About 12-18 months.
Me: But you're an AI
Claude: Ah yes if you leverage AI Coding then the estimate would be 6-9 months.
Me: But you're an AI
Claude: Even with AI, there are many aspect of coding that take time. AI can only do much to accelerate coding but it is not a magic tool. Now that we have a realistic estimate, would you like to get started?
Me: Yes
*3 hours later*
Claude: Ok I'm done.
Ha ido todo genial, ahora que se recupere Patri y en unos días pa casa! Muchas gracias a todos por los comentarios.
Es muy loco que spawnee una persona así de repente, flipas pepinillos.. con lo fácil que seria poner huevos!
A las mujeres os ponía un monumento del pateo que os metéis. Lof iu!
Update on giving Grok 4.5 control of my llama.cpp repo to optimize Gemma4-12B-QAT on my 3090.
After 55m we have results.
It spent the session repeatedly rebuilding llama.cpp, running controlled A/B bench, rejecting regressions, and narrowing down a better configuration.
There were constraints set by me:
- Same model
- No pl increasing (capped at 250w)
- f16 KV
- No MTP or speculative decoding
- No reducing context length
- No quant swaps
The benchmark harnesss was also fixed across builds
ppp4096 + tg256 at depth 0, 8k, 32k and 64k, with 5 repetitions each.
It tested three new configurations against the stock build:
A: Stock baseline
B: sm_86 explicitly targeting Ampere, with CUDA compresssion mode set to speed
C: FA_ALL_QUANTS, was skipped since there was little expected benefit for pure f16 KV and a significant compile-time cost.
D: FORCE_MMQ
E: FORCE_CUBLAS
The interesting part is that forcing the supposedly faster kernel paths did not win.
FORCE_MMQ slightly improved short-context performance, but regressed long-context generation.
tg256 @ 32k:
71.71 to 66.52 t/s
tg256 @ 64k:
65.96 to 63.22 t/s
FORCE_CUBLAS was even worse for prefill.
pp4096:
2624.8 to 2140.9 t/s
The best overall build was much simpler:
- sm_86
- CUDA FA ON
- CUDA graphs ON
- compression mode = speed
- default MMQ/cuBLAS auto-selection
The biggest same session gain was at 8k depth:
tg256 @ 8k:
72.03 to 76.76 t/s => +6.6%
Not a massive breakthrough, but a real measured gain from build configuration alone. Same model, same hardware, same 250W limit, same f16 KV, same benchmark harness.
Then it moved on to a batch/ubatch sweep.
ESPAÑA: Transparent Ranking of Universities (2026) by the number of Papers in the Top 1% and Top 10% most cited papers (2021-2025) in the OpenAlex database
https://t.co/SIEdGOgaJE
🚨 Windows tiene más de 200 servicios corriendo en segundo plano ahora mismo.
La gran mayoría existen solo para recopilar datos sobre ti.
Alguien se hartó y creó una herramienta 100% gratis y de código abierto que elimina casi todas las mierdas anti-usuario que Microsoft ha metido en los últimos años:
•Cortana
•Copilot
•Anuncios en el menú Inicio
•Telemetría agresiva
•Seguimiento de ubicación
•Aplicaciones sugeridas que nunca pediste
Todo desaparece con un clic. Un solo archivo de 8 MB.
Se llama optimizerDuck.
Es lo más parecido que he visto a tener Windows limpio, como debería haber venido de fábrica.
Lo que trae:
→ Más de 35 optimizaciones reales (rendimiento, privacidad, GPU, batería, servicios y UX) → Control total de +200 servicios con etiquetas de riesgo (bajo / medio / alto) → Ajustes específicos de GPU para AMD, NVIDIA e Intel → Removedor de bloatware con vista previa de lo que vas a eliminar → Cada cambio genera archivo de reversión automático → Te obliga a crear un punto de restauración antes de tocar nada
Sin instalador. Sin telemetría. Sin anuncios. Sin versión premium de mierda.
GPL v3 • Se ejecuta sin internet • Windows 10 y 11 (x64)
¿Cansado de que Windows te espíe aunque no quieras?
Descárgalo, pruébalo y dime en comentarios qué tal te fue.
Si te sirvió, comparte. Que esto llegue a más gente.
📊 Aston Martin necesita recortar 3,17 segundos respecto al presente para intentar puntuar
👉 De media, están a 1,3s del corte para pasar a Q2 (y subiendo), y a un mínimo de 3s para encabezar la Q2, para desde ahí aspirar a asomarse hacia los puntos".
😅 "Como se diría vulgarmente, 'la leche'. Eso sí, si Aston Martin pegara ese salto cualitativo que insinúan, menuda historia".
[@JavieRubioF1 - @elconfidencial]
#F1