In Today’s Issue:
🔓 Elon Musk releases the Grok-based transformer code behind the "For You" feed
🚀 Step3-VL-10B beats Gemini 2.5 Pro on STEM benchmarks
🧠 GLM-4.7-Flash debuts with MLA architecture
⚠️ BlackRock’s Larry Fink warns that AI-driven inequality could delegitimize capitalism
✨ And more AI goodness…
Dear Readers,
Today might just be the day open-source AI proves it can punch way above its weight class - Z.ai's GLM-4.7-Flash activates only 3 billion of its 30 billion parameters per token yet outperforms dense models twice its size on agentic benchmarks, and it's completely free to use.
But that's just the headline act: we've also got STEP3-VL-10B matching Gemini 2.5 Pro with a fraction of the compute, Larry Fink warning Davos elites that AI could widen inequality unless capitalism fundamentally changes course, and NVIDIA declaring that the "ChatGPT moment for robotics" has officially arrived with their new Cosmos and GR00T models now training robots to reason about the physical world. Whether you're building, investing, or just trying to keep up, this issue is packed with signals that the efficiency era of AI has begun - let's dive in.
All the best,
🚀 X Open-Sources Feed Algorithm
xAI has publicly released the core recommendation system behind its “For You” feed on X. The system relies almost entirely on a Grok-based transformer (from xAI) to predict user actions - likes, replies, reposts, blocks - and rank posts with almost zero hand-engineered features. The big takeaway: modern social feeds are now driven by multi-action prediction models and composable pipelines, not rules and heuristics.
🚀 Tiny Model, Massive Performance
STEP3-VL-10B is turning heads by matching - or beating (!) - models up to 20× larger, including Gemini 2.5 Pro, across top-tier multimodal benchmarks like MMMU, MathVision, and MMBench. Trained on 1.2T tokens with 1,400+ RL rounds, it delivers elite STEM reasoning (94.43% on AIME 2025), visual perception, and OCR, all while supporting 128K context via PaCoRe. Open multimodal AI is getting dramatically more efficient, not just bigger.
⚠️ Davos Capitalism Faces Reckoning
At the World Economic Forum in Davos, Larry Fink warned that capitalism is losing legitimacy as growth benefits too few, even as wealth creation since the Cold War has exploded. The BlackRock chief says AI could widen inequality like globalization did unless leaders move beyond GDP and listen to communities far outside the Alpine bubble - amid rising populism and distrust of elites.
State of the AI Industry — the OpenAI Podcast Ep. 12
GLM-4.7-Flash Dominates 30B Class
The Takeaway
👉 GLM-4.7-Flash activates only 3B of 30B parameters per token, delivering flagship performance at a fraction of the compute cost
👉 The model leads its class on agentic benchmarks (SWE-bench, τ²-Bench, BrowseComp), making it ideal for coding assistants and tool-heavy workflows
👉 Free-tier API access and open weights on Hugging Face remove cost barriers for developers and researchers
👉 Sparse MoE architectures are now proven competitive with dense models—expect more teams to follow this direction
The open-source AI race just got a lot more interesting. Z.ai dropped GLM-4.7-Flash, and it's shaking up what we thought was possible with lightweight models.
Here's the deal: This 30B parameter model only activates around 3 billion parameters per token thanks to its Mixture of Experts (MoE) architecture. Think of it like a” team of specialists where only the right experts jump in when needed”. The results are staggering: Flagship-level performance without melting your GPU.
What makes this exciting for our community is the benchmark dominance. GLM-4.7-Flash crushes tasks requiring multi-step reasoning and tool use, including SWE-bench Verified, τ²-Bench, and BrowseComp. It's not just competitive with dense 20B models—it often beats them.
The cherry on top? It's free at one concurrency, with weights available on Hugging Face.
Could sparse architectures finally make powerful local AI assistants a reality for everyone? The evidence is piling up.
Why it matters: GLM-4.7-Flash proves that efficiency and performance don't have to be trade-offs. This shifts the game for developers who want powerful AI without expensive cloud bills.
Sources:
🔗 https://huggingface.co/zai-org/GLM-4.7-Flash
🔗 https://x.com/Zai_org/status/2013261304060866758?s=20
What stands out about Emergent isn’t just how fast you can build, it’s how well what you build actually holds up.
I tested Emergent and it’s fantastic! In a space crowded with AI app builders that look impressive in demos but struggle under real usage, Emergent has managed to do something rare: deliver software that’s production-ready from day one.
From backend logic and integrations to deployment and scaling, the platform feels intentionally designed for real-world building, not experimentation alone. Apps don’t fall apart after a few iterations, and they don’t require constant rebuilding once users arrive.
That reliability is why Emergent has quickly become one of the most talked-about platforms among founders and creators, trusted by millions of builders and backed by some of the world’s leading investors.
If you’re looking to turn an idea into real software, whether that’s a SaaS product, internal tool, marketplace, or mobile app, Emergent stands out as a platform built not just to create, but to last.
“The ChatGPT Moment for Robotics":
NVIDIA Unleashes physical AI and rewrites the rules of automation - a little review.
At CES 2026 (I know it's been a few days already ;)), Jensen Huang made a bold proclamation that's been on everyone's lips ever since: "The ChatGPT moment for robotics is here.” And NVIDIA brought the receipts. The company dropped a sweeping suite of open-source AI models, new frameworks, and edge computing hardware designed to accelerate what it calls "Physical AI" - systems that don't just process data but actually understand the physical world, reason through problems, and take action.
The star of the show? The new Cosmos and GR00T models, now available on Hugging Face, which enable robots to learn faster through synthetic data, evaluate their own performance in simulation, and even reason about their environment like humans do.
Meanwhile, the Jetson T4000 module brings Blackwell-architecture AI to robots at scale, delivering 4x the compute efficiency of the previous generation for just under $2,000. What makes this genuinely disruptive isn't one product - it's the ecosystem play. Boston Dynamics, Caterpillar, LG Electronics, NEURA Robotics, and a dozen others are already deploying NVIDIA's stack in next-gen humanoids and industrial machines.
Salesforce is using Cosmos Reason to let enterprise robots analyze video footage and resolve incidents twice as fast. LEM Surgical is building autonomous surgical robots with Isaac for Healthcare. NVIDIA and Hugging Face are integrating Isaac and GR00T directly into LeRobot, merging NVIDIA's 2 million robotics developers with Hugging Face's 13 million AI builders. If 2025 was the year AI learned to talk, 2026 is the year it learns to walk- and work.
The Headlines Traders Need Before the Bell
Tired of missing the trades that actually move?
In under five minutes, Elite Trade Club delivers the top stories, market-moving headlines, and stocks to watch — before the open.
Join 200K+ traders who start with a plan, not a scroll.