Lorenzo Moreno (@rutrus) header
Lorenzo Moreno (@rutrus) profile picture

» Life Hacker » Blockchain Engineer » #BuildingProducts #Python #Rust #SkinInTheGame

Remote
Joined December 2007
Photo by Lorenzo Moreno (@rutrus)Photo by Lorenzo Moreno (@rutrus)Photo by Lorenzo Moreno (@rutrus)Photo by Lorenzo Moreno (@rutrus)Photo by Lorenzo Moreno (@rutrus)Photo by Lorenzo Moreno (@rutrus)Photo by Lorenzo Moreno (@rutrus)Photo by Lorenzo Moreno (@rutrus)Photo by Lorenzo Moreno (@rutrus)Photo by Lorenzo Moreno (@rutrus)
Lorenzo Moreno retweeted
I just solved "Robotics Engineering" education with Fable 5.2 forever.. no need to spend $500+ on a starter kit to learn what a GPIO pin does, just visit my interactive site: • 8 real boards in 3D: Arduino Uno, Raspberry Pi 4... • hyper-realistic graphics with real-life actions • build part by part: place chips, connectors and headers, then solder them in • wire real circuits: blink an LED, read a sensor, drive a motor Fable 5.2 generated every 3D model in code, no ready-made 3D files or photos try it free, no sign-up ↓ robotics-lab-app.vercel.app
117
542
39
4,675
373,882
Lorenzo Moreno retweeted
Introducing e-means, an algorithm beating k-means in 10x fewer epochs. With order of magnitude wall-clock speedups, e-means is released in FAISS and needs no hyperparameter tuning beyond k. It is robust to various scales and datasets. Still skeptical ? Just try e-means [1/10]🧵
150
1,223
256
10,218
889,865
Lorenzo Moreno retweeted
En Certuma estamos preparando la campaña de marketing en redes sociales para lanzar nuestros productos en Estados Unidos y en Argentina. Y me doy cuenta de que lanzar una campaña en 2026 no tiene nada que ver con las que hice para mis otras empresas médicas, Prelude, Overture y Gameto. Antes, una campaña significaba contratar una agencia, hacer un casting de modelos, alquilar un estudio, contratar fotógrafo, productora y locutor, y esperar semanas. Cada versión del anuncio costaba miles de dólares, así que probabas tres o cuatro y rezabas para que alguna funcionara. Y si querías hacerla en inglés para Estados Unidos y en español para Argentina, pagabas dos veces. Hoy estamos mirando herramientas como Glif, un agente creativo fundado en 2022 por Fabian Stelzer y Jamie Wilkinson (cofundador de Know Your Meme), que en abril levantó 17,5 millones de dólares liderados por Andreessen Horowitz y Union Square Ventures. Funciona así. Le escribes en una conversación lo que necesitas, le subes tu marca y tu producto, y el agente elige entre más de 100 modelos de IA: uno escribe el guion, otro genera la imagen, otro el vídeo, otro la voz y otro la música. Te entrega la campaña terminada, con anuncios para Instagram, TikTok, YouTube y Google en todos los formatos, subtítulos y textos. Esta semana sumaron una agencia de modelos virtuales: 165 personas que no existen, cada una con su cara, su estilo y su voz, siempre idénticas de una escena a otra. Eliges una, le das el guion y aparece en el vídeo, en el póster y en el reel. Sin rodaje ni casting. Según Stelzer, el vídeo de lanzamiento lo hicieron en menos de dos horas, y el plan básico cuesta 10 dólares al mes. Para nosotros eso significa poder hacer en una tarde decenas de versiones de un anuncio, para médicos y para pacientes, en inglés y en español, ver cuáles funcionan en Estados Unidos y cuáles en Argentina, y quedarnos con las mejores. Lo que antes era un presupuesto de agencia ahora es un experimento de un día. Hay una línea que en salud no vamos a cruzar. Las propias normas de Glif dicen que sus modelos son ficticios y que no se pueden presentar como clientes ni como reseñas reales. En medicina eso es todavía más importante. Una modelo virtual puede explicar qué hace nuestro producto. Nunca puede hacerse pasar por una paciente. La publicidad siempre fue ilusión. Ahora la ilusión cuesta 10 dólares al mes.
Today we’re launching the first AI modeling agency 💅 right inside Glif. Cast hundreds of real AI models, drop your product, idea or brief and let Glif create super compelling content for you end to end. If you still think AI models aren’t ready for real work, watch this 👇
71
106
11
933
78,989
🚨 Fallo crítico en el kernel de Linux: se puede atacar por red, sin login y sin necesidad de ninguna acción por parte del usuario. CVE-2026-98432 
➡️ CVSS 9.8 (crítico) 
➡️ Afecta a la gestión de conexiones TCP/IPv4, el código que atiende cualquier conexión a tu servidor. 
➡️ Ubuntu marca vulnerables desde la 16.04 hasta la 26.10, también los kernels de AWS, Azure, Google Cloud y Oracle. 
➡️ La corrección todavía está «en preparación» en casi todas. Si tienes un servidor o un VPS con Linux, sigue los avisos de tu distribución y actualiza el kernel (y reinicia) en cuanto salga el parche. ubuntu.com/security/CVE-2026…
81
578
115
3,402
472,042
Lorenzo Moreno retweeted
🎉 NVIDIA Nemotron 3.5 Transcribe is here—with day-0 serving through vLLM + NeMo Speech. 40 languages. Recordings up to an hour. Promptable transcripts, speaker labels and audio Q&A. Open weights in BF16, FP8 and NVFP4.
Meet NVIDIA Nemotron 3.5 Transcribe 🎙️ An open-weight, promptable 30B MoE model with ~4B active parameters that brings transcription in 40 languages, speaker attribution, translation, contextual prompting and audio understanding into one checkpoint. Watch it keep track of who said what when everyone’s talking at once. 1/ 🧵
19
24
297
20,366
Lorenzo Moreno retweeted
OvisOCR3 has landed: a unified 0.8B model that turns document images into structured, usable data across five OCR and document-understanding tasks. 🤖 modelscope.ai/models/ATH-Maa… 🏆 SOTA at its size across five document tasks: In the team’s reported comparisons, OvisOCR3 leads comparable-size models across document parsing, general OCR, key information extraction, figure captioning, and cross-page table reconstruction. 📄 Parse documents into usable formats: Generate structured JSON, layout JSON, or Markdown with LaTeX formulas and HTML tables from a page image. 🔎 Go beyond reading text: Extract specified fields into a JSON schema, answer image-grounded questions, recover chart data, or return text locations as bounding boxes. 🖼️ Understand visuals inside documents: Add localized captions for figures and create Mermaid representations for flowcharts as part of the parsed page content. 📊 Reconnect tables across pages: Use consecutive page images and parsed Markdown to rebuild a complete HTML table, with source-page associations retained. 🛠️ One compact checkpoint, prompt-selected tasks: Switch capabilities through task prompts instead of deploying a separate model for each OCR or document-processing workflow. 📜 Apache-2.0 License.
2
16
3
130
5,538
Lorenzo Moreno retweeted
An Anthropic engineer explains why he earns $350K–$950K: “I know how to build harnesses. 99% of people still build multi-agent systems manually and wonder why they keep breaking. I build systems that run loops, use sandboxed tools, and turn yesterday’s logs into new skills. Same model. 10x more reliable agents that improve while you sleep.” Opus 5.5 → Agents → Harnesses → Loops → Graphs In a short workshop, he breaks down how to build one from scratch: Step 1 → Manage loops, context, and sandboxing through a harness Step 2 → Separate the brain from the tools to reach the first token 60% faster Step 3 → Run agents server-side and recover from crashes using logs Step 4 → Make failures recoverable with retries and replays instead of starting over Step 5 → Turn past execution logs into new memories so the system gets smarter over time Anthropic calls this process “dreaming.” Most developers spend weeks building these systems manually. This workshop shows you a faster way. Bookmark this and watch it.
121
259
7
2,662
433,743
Lorenzo Moreno retweeted
This is f*cking insane. SpaceXAI engineer Lauren Tan: "You can't run 100 GrokBot agents when you don't even trust the work of one If nothing checks your agents for you, you are the checker. You're the bottleneck" In this 1-hour talk, she shows how she went from rereading every answer by hand to a GrokBot team that checks its own work and ships 50+ updates a day GrokBot → Agents → Self-Checks → Trusted Team → Work That Ships Itself Trusting your agents isn't a feeling. It's a system you build once Most people spend longer checking their AI than it would take to do the job themselves Worth more than any $2,000 agentic engineering course Bookmark it and watch tonight Then read how to build agents you can actually trust in the harness engineering guide below ↓
16
66
6
772
96,773
Lorenzo Moreno retweeted
i can't stop looking at this chart MiniCPM-OCR 1.0 is a 1.3B model built only for reading documents, and it's the one dot in the top right, the highest average score across four document benchmarks and the most pages per second their numbers: > 93.68 on Real5-OmniDocBench > 92.39 on Wild-OmniDocBench > 54.7 on OCRBench v2 > 1.217 pages/s on one rtx 5090 > 1.49x PaddleOCR-VL-1.6 throughput > 3.62x MinerU2.5 Pro throughput the real-world scores are the ones i care about, real documents reach a model as a phone photo of a contract at an angle or a curved page out of a book, and that's where it leads, on the standard OmniDocBench v1.6 it sits 0.4 behind the top ocr is the input for everything after it, agents and rag pipelines are only as good as the text they pull out of a pdf, and one wrong digit on an invoice makes the whole chain confidently wrong that's why the companion model got my attention, MiniCPM-OCR Jev is a 2B checker that reads the page next to any ocr model's output and scores each region, so you know which parts a human still has to check > AUROC 0.936 on JevOCRBench > Qwen3.5-397B-A17B gets 0.938 the parser and the checker are small enough to run on your own gpu, so contracts and receipts never leave your machine, with llama.cpp, vllm, sglang and ollama supported on day one weights and code from @OpenBMB: Hugging Face: MiniCPM-OCR 1.0: huggingface.co/openbmb/MiniC… MiniCPM-OCR Jev: huggingface.co/openbmb/MiniC… GitHub: github.com/OpenBMB/MiniCPM-O…
Your OCR says it’s done—but did it get everything right? 🤔 In high-stakes workflows, even a small error can be costly. 🚀 Meet the MiniCPM-OCR suite: an open-source solution to parse fast and guard the results. Two new models: 🔹 MiniCPM-OCR 1.0 — SOTA OCR performance and speed at 1B scale. Compact, accurate, and fast. 🔹 MiniCPM-OCR Jev — A 2B decision model that helps assess parsing accuracy, with performance approaching that of Qwen3.5-397B-A17B on JevOCRBench. ⚡ What’s new? Small model, SOTA results: MiniCPM-OCR 1.0 scores 93.68 on Real5-OmniDocBench and 54.7 on OCRBench v2. A sharp goalkeeper 🛡️: MiniCPM-OCR Jev is model-agnostic, checking text, formulas, and tables from different OCR parsers. It achieves 93.6 on JevOCRBench, outperforming JEV-35B-A3B (90.1) and approaching Qwen3.5-397B-A17B (93.8). Up to 128 pages per minute 🤯: With DSpark, MiniCPM-OCR runs at 1.49× the speed of PaddleOCR-VL-1.6 and 3.62× that of MinerU2.5 Pro. The SDK’s concurrency optimizations further boost throughput to as much as 128 pages per minute. Deploy where you work: Day-0 support for Arm, Houmo AI, Intel, and NVIDIA, plus 10 more chip vendors through FlagOS. Ready for your stack: vLLM, SGLang, llama.cpp, and Ollama, with quantized models, fine-tuning support, and Agent Skills. 🔗 GitHub: github.com/OpenBMB/MiniCPM-O… 🤗 Hugging Face: MiniCPM-OCR 1.0:huggingface.co/openbmb/MiniC… MiniCPM-OCR Jev:huggingface.co/openbmb/MiniC… 👉 Try the Demo:platform.modelbest.ai/consol… Use the API:platform.modelbest.ai/consol…
13
26
1
290
14,870
Lorenzo Moreno retweeted
Google 开源了一个很有趣的 AI Coding 插件,叫做 Elephant-Goldfish(大象与金鱼)。 它想解决一个常见问题: 你跟 AI 讨论了两个小时,AI 觉得自己什么都懂,结果写出来的程序还是跟需求有落差。 Google 的做法是:让两个 AI 代理分工。 ⸻ 🐘 大象(Elephant) 负责记住项目脉络、讨论架构、整理需求,把决策写成设计文件。 🐟 金鱼(Goldfish) 刻意不继承之前的聊天记录,像第一次接手项目的工程师,重新审查设计。 金鱼会做三件事: 理解检查: 光看文件,能不能理解整个设计? 对抗审查: 有没有遗漏边界条件、隐藏依赖或逻辑问题? 实现检查: 文件、数据结构、函数接口是否清楚到可以开始写程序? 如果金鱼看不懂,就回去修改规格。 ⸻ 这套流程最有意思的地方,是把 AI Coding 的质量管控往前移。 以前是 AI 写完程序,人类才开始找问题。 现在可以在写第一行 Code 之前,就先让另一个 AI 挑出规格漏洞。 尤其适合跨文件修改、复杂功能、多人协作的项目。 目前官方 GitHub 已列出 Antigravity、Claude Code、Codex 的安装方式。 不过,审查通过仍不代表程序一定正确,人工验收与测试还是需要。 ⸻ 我觉得值得观察的是: 当 AI 越来越会写程序,「谁来检查 AI 理解得对不对」,可能会变成更重要的工作。 项目网站:github.com/google/elephant-g…
44
144
12
1,400
78,460
Lorenzo Moreno retweeted
Unlimited OCR 2 is now open source. 📖✨ Built for virtually unlimited document parsing, it combines Sparse Reference Sliding Window Attention (SR-SWA) with Unlimited-Horizon Decoding to focus computation on the most relevant visual references while releasing KV caches that are no longer needed. Here’s what it delivers: • Nearly constant per-step decoding latency and memory footprint as documents grow • State-of-the-art results on MPDocBench and MarathonDocBench • On 100-page documents, 7.0× faster per-step decoding than Full Attention and 5.6× faster than R-SWA, with strong parsing accuracy maintained on documents up to 300 pages Explore it on GitHub. 👇 github.com/baidu/Unlimited-O…
9
62
12
460
37,749
Lorenzo Moreno retweeted
Your OCR says it’s done—but did it get everything right? 🤔 In high-stakes workflows, even a small error can be costly. 🚀 Meet the MiniCPM-OCR suite: an open-source solution to parse fast and guard the results. Two new models: 🔹 MiniCPM-OCR 1.0 — SOTA OCR performance and speed at 1B scale. Compact, accurate, and fast. 🔹 MiniCPM-OCR Jev — A 2B decision model that helps assess parsing accuracy, with performance approaching that of Qwen3.5-397B-A17B on JevOCRBench. ⚡ What’s new? Small model, SOTA results: MiniCPM-OCR 1.0 scores 93.68 on Real5-OmniDocBench and 54.7 on OCRBench v2. A sharp goalkeeper 🛡️: MiniCPM-OCR Jev is model-agnostic, checking text, formulas, and tables from different OCR parsers. It achieves 93.6 on JevOCRBench, outperforming JEV-35B-A3B (90.1) and approaching Qwen3.5-397B-A17B (93.8). Up to 128 pages per minute 🤯: With DSpark, MiniCPM-OCR runs at 1.49× the speed of PaddleOCR-VL-1.6 and 3.62× that of MinerU2.5 Pro. The SDK’s concurrency optimizations further boost throughput to as much as 128 pages per minute. Deploy where you work: Day-0 support for Arm, Houmo AI, Intel, and NVIDIA, plus 10 more chip vendors through FlagOS. Ready for your stack: vLLM, SGLang, llama.cpp, and Ollama, with quantized models, fine-tuning support, and Agent Skills. 🔗 GitHub: github.com/OpenBMB/MiniCPM-O… 🤗 Hugging Face: MiniCPM-OCR 1.0:huggingface.co/openbmb/MiniC… MiniCPM-OCR Jev:huggingface.co/openbmb/MiniC… 👉 Try the Demo:platform.modelbest.ai/consol… Use the API:platform.modelbest.ai/consol…
319
344
110
1,776
2,059,341
Lorenzo Moreno 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
87
203
52
2,732
224,032
Lorenzo Moreno retweeted
NVIDIA releases Nemotron-3.5-Bolt-E4B, a new model that runs locally on 6GB RAM. 💚 Bolt-E4B can run on edge devices like phones and delivers top-tier performance for its size. Trained on Gemma 4, it supports vision. GGUF: huggingface.co/unsloth/NVIDI… Guide: unsloth.ai/docs/models/nemot…
Available now: NVIDIA Nemotron 3.5 Bolt E4B 4GB ⚡️ Our newest small open model is built for local assistants, interactive characters and robots that can carry on a conversation and use tools to get things done. Watch Reachy Mini chat and move, with Bolt handling conversation and NVIDIA Cosmos handling vision. Both models run locally on Jetson Orin Nano 8GB. Learn more: research.nvidia.com/labs/nem…
93
202
40
2,291
124,194
Lorenzo Moreno retweeted
📊 ¿Qué encuestadoras aciertan más? ¿El CIS o las privadas? ¿Cuáles? Compara 16 casas en 44 elecciones desde 2018. Miramos (1) su precisión, (2) su desvío izquierda/derecha y (3) si publican sus microdatos. Aquí la tabla resumen. Hilo👇
26
112
22
373
71,762
Lorenzo Moreno retweeted
Commandos, el Worms, winamp y el Diablo. Quién necesita más en la vida?
2000’li yıllarda bilgisayar başında vakit geçirmek böyleydi.
8
2
23
3,178
Lorenzo Moreno retweeted
Introducing Indra, a 1M parameter OCR model: - 49 pages/s on CPU, 11x faster than PaddleOCR and 17x Tesseract - Beats them on OmniParseBench text (41.3 vs 35.8 and 19.3) - Full layout and line bounding boxes
128
254
41
2,895
150,626
Lorenzo Moreno retweeted
Recursive Agent Harness from Google Research. 100% Opensource.
52
82
13
797
62,471
Lorenzo Moreno retweeted
Con el revuelo que ha creado la página America[.]gov, se me ha ocurrido crear un proyecto abierto que espero que le facilite la vida a los devs que quieran trabajar con datos públicos. Empieza por un catálogo abierto de las fuentes de datos de la Administración española, con información hiper optimizada para agentes en distintos formatos ➡️ La API del BOE da 400 si no envías Accept: application/json. ➡️ La exportación de la BDNS da 50 filas aunque haya miles, sin avisar. ➡️ Para filtrar un municipio, el INE no quiere su código (02001) sino un Id interno (6124). ➡️ Muchos .gob.es sirven el certificado FNMT sin la intermedia: el navegador entra, curl y requests fallan. ➡️ AEMET responde en dos pasos, con el fichero en ISO-8859-15 y los errores dentro de un HTTP 200. Llevo cerca de 100 fuentes: legislación, hacienda, contratación y subvenciones, estadística, energía, medio ambiente, transporte, sanidad… y empiezan las comunidades autónomas (Madrid, Cataluña, Andalucía, Comunitat Valenciana) y los ayuntamientos de Madrid y Barcelona. La idea es acabar mapeándolo todo, actualizado y auditado. Para agentes: llms.txt, una versión ligera de 7 KB o un servidor MCP que se instala en una línea con uvx y carga solo la ficha que hace falta. Para devs: cargadores en Python que ya resuelven esas trampas. En tareas difíciles, un agente con el catálogo hace la mitad de llamadas y casi ninguna fallida. Se verifica cada semana con GitHub Actions y es CC0. Si conoces una trampa de una fuente pública, abre un issue: entra en la ficha con fecha. 👉 github.com/BquantFinance/Adm… Hagámosle la vida más fácil a quien quiera construir con datos públicos.
41
166
14
733
53,200