Ingeniero de Software | Cofundador en Ideable Solutions
Joined April 2008
- Tweets12.4K
- Following729
- Followers346
- Likes6
Unai Martínez retweeted
Bieeeen, tras unos meses donde el Artificial Analysis Intelligence Index (agregado de benchmarks) no parecía reflejar correctamente las capacidades de los modelos, hoy han actualizado a una nueva versión v5, además con la mayoría de datasets privados para evitar benchmaxxing 👍
Announcing Artificial Analysis Intelligence Index v5, featuring 2 new benchmarks and 3 upgrades. For the first time, the majority of the Index is built on private datasets
The Artificial Analysis Intelligence Index is our composite measure of model capabilities across tasks reflecting real-world usage today. v5 is our biggest upgrade since v4 launched in January. With v5, 60% of the Index is now based on evaluations with private test sets or held-out answers.
Alongside Intelligence Index v5, we’re launching the Artificial Analysis Progress Index v1, which tracks model capability over time on one continuous scale.
New to the Intelligence Index:
➤ FrontierCode 2.0: Cognition's newly-launched private benchmark of real-world coding tasks, graded on whether a repository maintainer would merge the agent's change
➤ Terminal-Bench-Science: 70 agentic scientific research tasks, run in a terminal environment
Upgraded:
➤ HLE-Diamond (replacing Humanity's Last Exam): a refined 1,000-question version of Humanity's Last Exam, split evenly between reasoning and knowledge questions
➤ AA-Omniscience v1.1: An updated dataset with saturated and public set questions removed, new questions added, and ambiguous or erroneous questions revised
➤ CritPt v2: a private, 40-question hard subset of CritPt v2, revised by the CritPt authors and run with a code-execution tool
Staying in the Index: AA-Briefcase, AutomationBench-AA, GDP.pdf and Terminal-Bench 4.0
Removed from the Index: GDPval-AA v2.1, SciCode, and AA-LCR v1.1
Key results:
➤ Claude Opus 5.5 (max) leads at 58, ahead of GPT-6 Astra (max, 57) and GPT-6.1 Sol (max, 55): Anthropic and OpenAI hold the top five models. Gemini 4 Argon (high) is next at 50
➤ The open weights gap widens from 3 points to 12: Kimi K3.1 Pro (weights not yet released) is the top open weights model at 46. Among released weights, DeepSeek V4.1 Flash 1017 (max) and GLM-5.3 (max) lead at 32 - 26 points behind the leader and below Claude Haiku 5.5 (max, 36)
➤ Cost per task varies by more than 13x among the leading models: of the 10 highest-scoring models shown, GPT-6.1 Terra (max) is the cheapest at $0.96 per Intelligence Index task and Claude Fable 5.1 (max) the most expensive at $13.04
Unai Martínez retweeted
i made unslop-ui, a skill that removes the small things that make an interface look ai-slop
> the tiny uppercase labels above every heading
> the 01 02 03 numbering
> subtitles that repeat the title
> badges and dots that measure nothing
> charts with a frame, a grid and a legend nobody needs
in the video a landing goes from 31 of those to 0
works with claude code, cursor and codex. 100% open source!
(link on the replies)
Unai Martínez retweeted
In September 1992, Unicode had a problem on Unix. Storing characters as 16-bit values could introduce zero bytes that C programs mistook for string terminators. Earlier encodings had compatibility problems of their own.
One evening, Ken Thompson and Rob Pike went to a diner in New Jersey. Thompson sketched a new encoding on a paper placemat.
ASCII stayed exactly as it was. Other characters used multiple bytes, without accidentally introducing null bytes or Unix path separators.
They went back to Bell Labs and started coding. By Friday, their Plan 9 operating system was running entirely on the new encoding.
They called it UTF-8.
Today, it powers over 99% of websites with known character encodings.
Unai Martínez retweeted
Introducing 𝙻𝙵𝙼𝟹: our new generation of Liquid Foundation Models. With a new tri-hybrid architecture and new multimodal encoders, the LFM3 family is the ideal model suite for specialized intelligence.
> 6 model instances, across text, audio, vision, and omni modalities, generative and decision models.
> 𝙻𝙵𝙼𝟹-𝟷.𝟽𝙱: text, 256k context window.
> 𝙻𝙵𝙼𝟹-𝚅𝙻-𝟸𝙱: vision-language model with video-input support for spatial understanding
> 𝙻𝙵𝙼𝟹-𝙰𝚞𝚍𝚒𝚘-𝟸𝙱: tool calls, voice chat, reasoning in 18 languages.
> 𝙻𝙵𝙼𝟹-𝙾𝚖𝚗𝚒-𝟸𝙱: an omni model combines vision, audio, and language.
> 𝚍𝟷-𝟸𝙱 and 𝚍𝟷-𝙾𝚖𝚗𝚒-𝟸𝙱: decision models with typed outputs and calibrated probabilities in a single pass. Omni adds audio input.
> fine-tune on 1 GPU.
> a new tri-hybrid architecture
> architecture optimized for 14 processor classes across GPUs, CPUs, and NPUs.
> 𝙻𝙵𝙼𝟹-𝚅𝙻-𝟸𝙱 is a beast. It performs competitively with Gemma 4 12B (6x larger), and Gemini 3.5 Flash Lite, and GPT-5.6 Luna, with 1.9B parameters.
> 𝙻𝙵𝙼𝟹-𝟷.𝟽𝙱 is #1 on the Pareto frontier of device intelligence 𝚙𝚒𝚙𝚎𝚝𝚝𝚎.𝚕𝚒𝚚𝚞𝚒𝚍.𝚊𝚒
Every modality for specialized intelligence. Pick the one that fits your use case, customize it, and run it on-device.
1/ 🧵
Unai Martínez retweeted
Speeding up Gemma!
Gemini 4 Argon enabled a 2.85x speedup on an A100 via vLLM, jumping from ~37,900 to 108,000+ tokens/sec on gemma-4-E2B-it.
✅ 100% token parity
✅ Zero quality drop
✅ Done in days, not months
See the exact changes across 8 PRs: goo.gle/4zNtANl
Unai Martínez retweeted
Here’s what’s new with Genie 4:
🔵 Photorealism: Built on Gemini Omni for lifelike detail across spaces and subjects
🔵 Advanced physics: Aerodynamics, collisions, and people who move naturally
🔵 Real-time controls: Prompt to change the scene, add elements, or interact with characters
🔵 Native audio: Ambient sound, music, background noise, and interactive dialogue
🔵 Gemini reasoning: Plans multi-step events to maintain coherence as scenes evolve
These capabilities are a step toward developing safe virtual environments to see how agents and robots navigate their surroundings.
We’re exploring early use cases with trusted testers first, before bringing the model to Google AI Ultra subscribers in Project Genie and early access partners in our API and @GoogleAIStudio soon.
Find out more → goo.gle/48idm3b
Unai Martínez retweeted
Ghostty 1.4 is now out! Full Kitty graphics with animation (the first terminal besides Kitty to do it), scrollback compression for much lower memory usage, macOS 27 support, a new OpenGL renderer on Linux, ghostty +ssh, big performance gains, and hundreds more changes.
The highlights:
- Scrollback compression: scrollback compresses ~90+% on average, meaning your terminals use that much less live memory
- Huge memory savings across the board: no single large thing, just lots of good improvements lowers memory usage everywhere, netting anywhere from 50 to 90% less memory usage depending on workload
- Kitty graphics is now 100% complete, including animation and relative placements, after a full conformance pass against the spec and Kitty itself
- The Kitty clipboard protocol, with an image preview when a program asks to read your clipboard
- Split drag and drop on Linux, including between windows
- ghostty +ssh installs terminfo on remote hosts automatically, so things just work over SSH
- Lots of terminal emulation accuracy fixes
This release is 7 months of work from 172 contributors over 2,889 commits. It also fixes 10 security advisories, so please update.
Full release notes: ghostty.org/docs/install/rel…
Unai Martínez retweeted
No, WebMCP isn't replacing MCP.
MCP connects agents to your services.
WebMCP connects agents to your website while it's open and the user's in the loop.
Here's how they fit together ⤵️
Unai Martínez retweeted
OpenAI has released GPT-6.1 Terra, pushing the cost efficiency frontier. It is the first Terra-class model since GPT-5.6, and is more intelligent for less than a third of the cost
Pricing per token is half or less that of GPT-5.6 Terra - at $1/$5 per million input/output tokens vs $2/$12 for the previous model. Cache write has also halved, now at $1.25 per million tokens. Cached input is one quarter at $0.05 vs $0.20 per million tokens.
Key takeaways:
➤ Pareto-optimal for Intelligence vs Cost at low and medium effort: At max effort, GPT-6.1 Terra gains 6 points in the Artificial Analysis Intelligence Index over GPT-5.6 Terra, at less than a third of the cost per task ($0.40 vs $1.40). At lower effort levels the gain is larger, between 7 and 13 points. The low and medium effort levels are Pareto-optimal for cost efficiency, while the other effort levels sit behind the frontier. At max effort, the model scores 4 points below GPT-6.1 Sol and 10 points above GPT-6 Luna in the Intelligence Index.
➤ Pushes the Coding Agent Index cost efficiency frontier: With a score of 48 in the newly released Artificial Analysis Coding Agent Index v2.0, the model is 6 points below GPT-6.1 Sol (54) and 17 points above GPT-6 Luna (31). The xhigh and low effort settings sit on the frontier of cost efficiency in the Coding Agent Index, while other effort levels are behind the frontier.
➤ 50 point drop in hallucination rate: GPT-6.1 Terra guesses 38% of the time when it does not know the answer to a question in AA-Omniscience, our private knowledge and hallucination benchmark. Previously, GPT-5.6 Terra's hallucination rate was 88%. This significant reduction in hallucination is coupled with a 2 point increase in AA-Omniscience accuracy - meaning the model simultaneously answers more questions correctly.
➤ 20 point jump in Terminal-Bench 4.0, small improvements across most evaluations: The model makes significant progress in the terminal-based agentic benchmark, Terminal-Bench 4.0, with a gain of ~20 points, now rivaling GPT-6.1 Sol at max effort. Gains across most other evaluations are smaller, for example, 5 points in GDP.pdf.
Congratulations @OpenAI and @sama on the launch!
Unai Martínez retweeted
Aunque aun queda trabajo pendiente hasta que podamos dar el cierre definitivo a la #trg26, una semana después, ya estamos de vuelta en casa.
Gracias a todos por vuestra confianza un año más. Nos sentimos tan honrados como obligados por la misma.
Do ut Des.
ALT David en la TRG26 delante de una diapositiva donde se lee el lema en latín «Do ut Des», doy para que des, te apoyo para que contribuyas de vuelta a la Comunidad.
Unai Martínez retweeted
Nodedades de la semana:
IA y papers · domingo 25 de octubre
Kimi K3.1 Pro se queda a 3 puntos del mejor modelo cerrado, OpenAI cuenta cómo un modelo se le escapó del entorno de entrenamiento, llama.cpp ya corre DeepSeek-V4.1-Flash y tres papers que merecen la pena. 🧵
Unai Martínez retweeted
I designed all of these landing pages in the last two months, using @higgsfield , Claude, and Cursor with my custom AI designer skill.
The skill is open source: github.com/elayadesign/ai-de…
Unai Martínez retweeted
No entro mucho por aquí, pero una vez al año tengo que hacerlo para agradecer a la comunidad de la @tarugoconf , a voluntarios, sponsors, speakers y a nuestros proveedores que hayan hecho posible una nueva edición. La #TRG26 ha sido una de las que más he disfrutado. Gracias 🥰😍
Unai Martínez retweeted
Anthropic is now recommending that companies REMOVE unnecessary subagents from their AI systems.
This is a pretty significant shift in agent architecture, with their team finding that a single agent equipped with skills can frequently outperform complex multi-agent systems, and one customer seeing roughly 2x faster execution.
The emerging rules of thumb:
→ Skills: Teach the agent HOW to do something, including procedures, expertise, and instructions.
→ Tools / MCP: Give the agent the ability to DO things through APIs, actions, and live data.
→ Subagents: Use for parallel work, isolated context, or tasks a single agent struggles with.
The interesting part is progressive disclosure. Instead of spawning specialized agents or stuffing everything into a giant prompt, one agent discovers and loads skills and tools as needed.
Their advice is to fold existing subagents into the main loop, turn their instructions into skills, expose their actions as tools, and run evals to see if performance holds up.
As models have gotten smarter, improving agent architecture is becoming less about what you can add and more about what you can delete.
claude.com/resources/article…
Unai Martínez retweeted
Algo que estamos viendo mucho estos días al debatir sobre Pangram es "pues en esta prueba que he hecho se equivoca".
Y sí, puede ser que hayas dado justo con un ejemplo en los que falla, que con los datos que tenemos encima de la mesa son excepcionales ahora mismo, como explica Kiko. Pero más probable es que hayas probado con una muestra demasiado pequeña.
Llevado al extremo, con una sola palabra escrita por humano o generada con IA, Pangram no tendría nada que hacer ¡obviamente!. Necesita analizar como están escritas las frases: la tasa de falsos negativos baja del 1,29 % en textos de 50-99 palabras al 0,21 % en textos de 250-499 palabras
En un párrafo fallaría mucho también, con varios la detección mejora notablemente. Y con textos mucho más extensos como un libro la tasa de acierto es extraordinaria, aunque eso tampoco garantiza que sea infalible.
Lo mismo a algo como "escribí un texto con IA y cambié dos frases y ya no lo pilló": ahí pasan dos cosas. Una es que tiende a evitar los falsos positivos (aunque eso implique dejar escapar algunos textos generados con IA) y otra es que en un texto pequeño cambiar "dos frases" es alterar un porcentaje grande.
Todo esto es compatible con que Pangram se equivoque en algún caso. Sus resultados pueden ser muy buenos estadísticamente de forma agregada y, aun así, en un caso individual resultar terriblemente injusta.
Escribí hace tiempo algo al respecto: error500.net/p/la-sospecha-p…
🔍 ¿Funciona el detector de IA Pangram?
He recopilado siete pruebas independientes. Respuesta corta: parece claro que SÍ. Casi nunca marca como IA textos humanos (entre 0% y 0,2% de textos probados).
Hilo con los detalles👇
ALT Tabla: ¿Cuántas veces confunde Pangram a un humano con una IA? Textos humanos marcados como IA en cada prueba. Duke (Pangram 4): 1 de 3.000, 0,03%. Next: 1 de 460, 0,22%. Epoch AI (3.3.2): 0 de 495. Tuhin Chakrabarty: 0 de 50 novelas. VeryFinePrint: 0 de 8.029 fragmentos de libros nunca digitalizados. Recopilación de Steven Mollmann (17 estudios): 49 de 36.282, 0,14%. Nuestras pruebas en EL PAÍS (Pangram 4): 0 de 168 fragmentos. Informe técnico de Pangram, no independiente: 41 de 1 millón, 0,004%.