Marcel Butucea

🤯 LLM inference 2x faster with Mixture-of-Recursions? They route tokens to different thinking depths, saving compute! MoR models even beat vanilla Transformers in accuracy... venturebeat.com/ai/mixture-of-…

China-originated open-source LLMs are popping up like crazy lately. 🇨🇳💥 Today, Z.ai officially released the GLM-4.5 series, an open-source SOTA model built for reasoning, coding, and agentic applications. (MIT License) The GLM-4.5 series includes two models:…

🔒 NVIDIA just cryptographically signed ALL NGC models with OpenSSF’s OMS standard—first major hub to do so! 🚀 Now devs can verify model integrity at every step (SHA-256 hashes, details here: developer.nvidia.com/blog/bringing-…

Google’s new BigQuery toolset slashes agent dev time by 70%, thanks to built-in metadata fetchers and SQL execution tools. But beware—the setup still demands OAuth, service accounts,... cloud.google.com/blog/products/…

NVIDIA’s new GB300 NVL72 PSUs slash AI data center power spikes by 30%—by fixing a wild quirk: syncing thousands of GPUs turning on/off *simultaneously* like a power-hungry swarm. 🐝 T... developer.nvidia.com/blog/how-new-g…

A fun & v feasible project idea for someone out there: bundle up face detection, speech recognition, GPT as the core "intelligence engine", text to speech, and face generative model to create a digital human you can talk to e.g. on webcam/phone (but it's just a "dressed up" GPT).

An interesting historical note is that neural language models have actually been around for a very long time but noone really cared anywhere near today's extent. LMs were thought of as specific applications, not as mainline research unlocking new general AI paths and capabilities

I wrote a minimal/educational GPT training library in PyTorch, am calling it minGPT as it is only around ~300 lines of code: github.com/karpathy/minGPT +demos for addition and character-level language model. (quick weekend project, may contain sharp edges)

Decision trees: simple to grasp, BUT easily overfit! 🌳 Ever notice how a tiny data tweak can drastically change the whole tree structure? The article touches on it, but it's a bigger issue than you might think... kdnuggets.com/7-must-know-ma…

Did you know you can set up real-time Google Sheets dashboards that auto-refresh with data from Google Analytics, Salesforce, and even GitHub? No coding needed—just smart linking and slicers! 📊 Per... kdnuggets.com/create-an-anal…

Wow — Google’s SensorLM was trained on a staggering 59.7 million hours of wearable data from 103k+ people across 127 countries! 🤯 It auto-generates natural language descriptions by combining context and sensor data. Mind blown! research.google/blog/sensorlm-…

India's gov't plans to train 150,000 students in AI & data skills, but only 1 of 4 southern state centers got initial funding 🤔 Karnataka, Andhra Pradesh, & Telangana still waiting analyticsindiamag.com/ai-features/ca…

Did you know 88% of Fortune 100 companies use E2B's cloud for AI agents? 🤖 But most AI agents fail to reach production due to security & reliability issues. E2B's secret?... venturebeat.com/ai/how-e2b-bec…

Qwen3 Coder’s 256K-token window isn’t just big—it’s a *revolution*. With 480B parameters (but only 35B active via MoE), it processes entire repos in one pass, beating benchmarks like SWE-Bench Verified. 🚀 Even detects SQLi/XSS ris... datasciencedojo.com/blog/qwen3-cod…

Another day another breakthrough ! The power of open source ! z.ai/blog/glm-4.5

PSA: People will think your conference paper is awesome if you stay within the time limit, even if it otherwise stinks. #ets2024

We just submitted our AI research paper, co-authored with Ali Irshayyid, Jun Chen, etc. OORT @oortech makes impact not only in industry, but also in academia. Years ago, I asked a few Europeans what they thought of Germany, and they all answered simple: “Germans make good…

The Zama Concrete ML library is one star away from reaching 1k+! github.com/zama-ai/concre… 💫

AI in the *physical* world hasn't had its ChatGPT moment...yet. 🤔 Apptronik & Waabi CEOs at #TechCrunchDisrupt are tackling the challenge of bringing intelligence into motion. techcrunch.com/2025/07/28/mee…

82% of security teams are drowning in alerts but missing real threats 🔍 New research reveals the grim math: 61% say too many feeds, 60% blame analyst shor... cloud.google.com/blog/products/…

Chinese students use AI 80% more than U.S. or UK peers—turns out they’re treating it as a skill, not a cheat code. 🎯 By 2025, 60% use AI *daily*, law students draft abstracts with DeepSeek, and universities are embracing AI as a learning tool. technologyreview.com/2025/07/28/112…