Yiyi Wu

The Era of "Thinking" Machines: Key Takeaways from Jensen Huang's CES 2026 Keynote NVIDIA CEO Jensen Huang just delivered a masterclass on the future of technology, signaling a fundamental reset of the computer industry. We are moving beyond "running" software to a world where machines "think" and "reason." Here is a breakdown of what Jensen highlighted, the latest developments, and what the future holds for AI and industry: 🔹 The Fundamental Platform Shift Jensen emphasized that we are witnessing two simultaneous platform shifts: Programming to Training: We no longer just write code; we train models. CPUs to GPUs: The entire $10 trillion computing stack is being modernized for accelerated computing. Generative vs. Pre-recorded: Applications now understand context and generate every pixel and token from scratch in real-time. 🔹 Major Developments: Agentic & Physical AI The transition from simple chatbots to "Agentic" systems is the defining development of this era: Reasoning Models (Test-Time Scaling): AI can now "think" before answering, breaking down complex problems into steps. Agentic User Interface: Future interactions with platforms like Palantir, ServiceNow, and Snowflake will be through AI agents, not just traditional dashboards. Alpamo: NVIDIA’s first "thinking" autonomous vehicle AI. It doesn't just drive; it reasons about its actions and explains why it’s making a turn or stopping. Cosmos World Model: A foundation model for physical AI that teaches robots the laws of physics - gravity, friction, and causality -through synthetic data. 🔹 The Future: The Vera Rubin Era To power the "Reasoning Revolution," Jensen unveiled the Vera Rubin platform, which is already in full production: Massive Performance: The Rubin GPU delivers 5x the performance of Blackwell. Extreme Connectivity: The NVLink 6 switch can move data at 240 terabytes per second—more than twice the bandwidth of the entire global internet. AI Factories: New data centers will achieve 10x higher throughput per watt, making AI generation more cost-effective than ever. Robotics Everywhere: Jensen envisions a future where "factories are gigantic robots" that design, test, and build other robots entirely within simulated environments before they ever touch the physical world. [ The Bottom Line: We are moving from a world of digital tools to a world of digital intelligence that can interact with and understand our physical reality. The "thinking" process is now the core of computing. Watch the full keynote here: https://lnkd.in/gCrKyVST #NVIDIA #CES2026 #AI #JensenHuang #Robotics #AcceleratedComputing #VeraRubin #PhysicalAI #Innovation

2

In ML and infra recruiting, I see a constant divide: Believing a system works vs. proving it works. To measure verification, look for four signals: Predicting Failure: Can they describe the crash before they write the code? Aggressive Questioning: Do they interrogate the constraints, or just accept them? Pivot: Do they find edge cases without a nudge? Healthy Scepticism: Do they treat AI-generated code as untrusted by default? The bottom line is to avoid hiring for the "happy path." Instead, hire for the person who knows exactly where the path ends.

6

Yesterday at NVIDIA Live at CES 2026 CEO Jensen Huang shared how #AI is transforming every domain and device, redefining the future of computing. He unveiled the Rubin platform, a new extreme-scale AI supercomputer architecture now in production, alongside major advances in open models and autonomous driving. 🗞️ Catch up on all the announcements: https://bit.ly/4jytr9Q

5

1 Comment

Speed matters especially when you're indexing billions of vectors. 🚀💾 Meta engineers are enhancing FAISS, our open source similarity search library, with NVIDIA’s cuVS to deliver massive GPU acceleration. 💡 The result? Faster search, lower latency, and better scaling for high-dimensional data workloads powering AI, recommendation systems, and more. Explore how this integration is pushing the boundaries of real-time vector search. #Meta #FAISS #VectorSearch

2

Really excited to share that our recent work, MorphServe, has been accepted to #MLSys 2026! I’m honored to have the opportunity to present our work, MorphServe, which explores handling bursty real-world LLM serving workloads through dynamic runtime quantization. This work would not have been possible without the great collaboration and teamwork with Zirui Wang, Tingfeng Lan, and Zeyu Zhang under the guidance of Haiying Shen, Yue Cheng, and Juncheng Yang Looking forward to seeing you at MLSys and discussing efficient and adaptive LLM serving systems with the community. https://lnkd.in/gZcjBb6x

17

Excited to highlight the new high performance communications API that we just announced for PyTorch. It will make it possible for folks to create scalable, heterogenous, and fault tolerant models and training frameworks. One motivator for the design is making this more extensible -- so that we can iterate quickly with the community on new and more capable communication back-ends. Thanks and kudos to a really huge and cross functional team behind this work but especially to: 🍃 Tristan Rice, Pavan Balaji, Srinivas V., Subodh Iyengar, Junjie W., Feng Tian, James Hongyi Zeng.

25

1 Comment

📣 NEW VERTEX AI ENGINEERING BLOG: Implementing EAGLE-3 at scale on Vertex AI with SGLang Speculative decoding speeds up inference, but the standard pattern of hosting a separate draft model can be challenging to manage at scale. Today I am so excited to share that we just launched a NEW engineering blog to document and openly share our applied research at Vertex AI. Our first post in collaboration with SGLang details how we implemented EAGLE-3 at scale for all Vertex AI customers and what we learned. Check out the blog in the comments 👇 . We open-sourced the notebook if you want to reproduce benchmark results. _____ ♻️ For more insights and tutorials about Open Models, Ray, Agents and applied research, on Vertex AI, comment, like, and repost with your network. #LLM #Engineering #SGLang #Inference #Optimization

165

3 Comments

Google Research Director Denny Zhou, who founded the LLM Reasoning Team at DeepMind, recently gave a great talk at Stanford’s CS25 class: Large Language Model Reasoning. This was one of the most intuitive talks about reasoning model training and application. He outlined new directions for training models to handle questions where answers aren’t easily verifiable. My team is currently working on an enterprise LLM project focused on using reasoning to extract relevant context and generate responses for unverifiable cases. It’s a tough but promising area, and it’s an interesting time to be building. [Link to Denny Zhou's talk: https://lnkd.in/gdJ6Mzi8]

1,026

8 Comments

My colleagues Amila Weerasinghe, Utkrisht Rajkumar and I are organizing "AI co-scientist: Accelerating scientific discovery through intelligent AI collaboration" workshop as part of Amazon Research Day 2026 event on April 30th. We will have invited talks by Peter Clark (Senior Research Director, Allen Institute for AI), Chenhao Tan (Associate Professor, University of Chicago), Wei Wang (Professor, UCLA), James Zou (Associate Professor, Stanford).   We will present our team's work on AtoZ AI Co-Scientist and our research projects in collaboration with UCSD ("AI Co-Scientist: Designing Autonomous Agents for ML", Prof. Rose Yu and Prof. Taylor Berg-Kirkpatrick) and UCLA ("Self-evolving Agents with Dynamic Temporal Knowledge Graphs: A Neuro-Symbolic Architecture for Continual Scientific Discovery" Prof. Wei Wang and Prof. Yizhou Sun).   We will also have a poster session. If you would like to present your work on AI Co-Scientists in the poster session please let me know.  

87

🚀 Day 77 of Global AI Madness Elon just delayed Grok 5 to Q1 2026 (calling it "sentient"), AWS is dropping $5B on Korean AI data centers, and Meta just restructured its AI division for the fourth time in six months. Meanwhile, DeepMind's AlphaFold 4 is predicting protein structures in minutes instead of weeks. ━━━━━━━━━━━━━━━━━━━━━ 💡 THE SHIFT AI just went multi-dimensional. We're not just watching model wars anymore — we're watching infrastructure wars, agentic evolution, and scientific breakthroughs happening in parallel. 🔹 xAI: Grok 5 delayed to Q1 2026, featuring 6T parameters and what Musk calls a "10% probability of AGI" — the first time he's publicly acknowledged AGI potential in one of his models. It'll handle text, images, video, and audio with superior tool use and real-time capabilities. 🔹 Meta: Fourth AI restructuring in six months under "Superintelligence Labs," splitting into four teams including a literal "TBD Lab" (To Be Determined). This follows lukewarm reception to Llama 4 and a talent exodus, with annual capex raised to $66-72B. 🔹 AWS: Announced $40B investment across 14 APEC nations by 2028, positioning "agentic AI" as the next evolution — agents that research, analyze, and automate within unified business intelligence experiences. South Korea alone gets $5B by 2031. 🔹 DeepMind: AlphaFold 4 now predicts protein structures in hours or minutes (vs. days/weeks previously) with near-experimental accuracy. This opens rare disease research and novel drug discovery at unprecedented speed. ━━━━━━━━━━━━━━━━━━━━━ 🎯 THE MEANING The AI landscape is fragmenting into three distinct battlefields: 1️⃣ Model Intelligence (xAI vs. everyone) 2️⃣ Enterprise Infrastructure (AWS, Meta) 3️⃣ Scientific Acceleration (DeepMind) Grok 5's delay signals that raw parameter counts aren't enough anymore — Musk's chasing "intelligence density per gigabyte" and curiosity-driven training to avoid data collapse. Meta's constant restructuring screams desperation to compete while OpenAI, Anthropic, and xAI race ahead — throwing billions at the problem won't matter if your org structure changes every 45 days. ✅ Winners: Infrastructure plays (AWS, Nvidia), biotech researchers with AlphaFold access, and anyone betting on xAI's embodiment strategy. ❌ Losers: Companies still treating AI as a "feature" instead of rebuilding around it. ━━━━━━━━━━━━━━━━━━━━━ 💭 YOUR TAKE? Which battlefield matters more: the smartest model, the best infrastructure, or the fastest scientific breakthroughs? Drop your take — because in 2026, you'll need to pick at least two to stay relevant. 🎥 YouTube Deep Dive: AI Is Eating Logistics https://lnkd.in/gs26XcZK #AI #ArtificialIntelligence #Grok5 #xAI #Meta #AWS #DeepMind #AlphaFold #AgenticAI #MachineLearning #TechInnovation #AIResearch #Startups #Innovation #FutureOfWork #AGI #AIInfrastructure #TechNews #AILeadership #DigitalTransformation

3

While I agree with sara's general message that "we can’t rely solely on compute," and that we need to spend more time improving architectural and training methods, I beg to differ on a few points. 1) Figure 3, i.e. the premise of the argument, is a little misleading. The reason some smaller models outperform larger models, and why we see a decreasing performance gap, is that the paper uses saturated benchmarks. Additionally, some of the submitted models are just poorly engineered. The data in the graph comes from the Open LLM Leaderboard 1, which was later declared saturated and replaced with Open LLM Leaderboard 2 (including MMLU-Pro, GPQA, and MATH) in 2024. If we only consider models from top frontier labs and use better benchmarks, the story is different. My graph in the comments shows the average scores of the best ≤13B and >13B base models submitted to Open LLM Leaderboard 2 from March 2023 to March 2025, considering only the top 10 frontier labs. The performance gap is roughly remains similar. (I consider only base models, as pretraining contributes ~95% of total training FLOPs and is the primary driver of scaling) 2) “We can get rid of most of these weights after we reach the end of training with minimal loss to overall performance.” Model pruning is task-specific, circuits are interconnected and extremely complex making it almost impossible to perform general pruning. "Compressing LLMs: The Truth is Rarely Pure and Never Simple" shows that all pruning methods lead to significant performance degradation in knowledge-intensive, reasoning, and other tasks. 3) “A key limitation of simply throwing more weights at a task is that the relationship between additional trainable weights and generalization remains poorly understood.” In my personal experience, this relationship generally holds true. The higher the number of parameters, the better the knowledge internalization and generalization. Parameters are not the only factor responsible for generalization, but they play a central role. Untill we have a concrete evidence of failing scaling laws, I believe the current paradigm of scaling will persist. For now the scaling race seems far from over.

14

1 Comment

Etched just raised ~$500M in a major funding round (valued at ~$5B) to accelerate its custom AI silicon — a solid alternative to Nvidia-centric stacks. https://lnkd.in/dttxqJPJ

6

Thrilled to see our ICML 2025 paper on Differentially Private Partition Selection featured on the Google Research blog! A huge shout-out to my amazing co-authors: Justin Y. Chen from MIT, and my colleagues Vincent Cohen-Addad and Morteza Zadimoghaddam at Google Research. Our work presents new algorithms to better protect user data at scale. Check out the post to learn more. Read it here: https://lnkd.in/dkZhHWAm #ICML2025 #DifferentialPrivacy #DataSecurity #MachineLearning #Research

85

1 Comment

As AI platforms scale, the hardest problems are often not model accuracy or infrastructure, but alignment on fundamentals — how teams reason about ML, signals, and trade-offs. Last month, I shared a session on ML fundamentals and their implications for AI platforms with a broad engineering audience in Taiwan, engaging hundreds of engineers across multiple organizations. What stood out to me was not the attendance, but the quality of discussion — questions around system boundaries, evaluation thinking, validation trade-offs, and how ML concepts influence platform-level decisions. I’m grateful for the encouragement and recognition from leadership across regions, and even more so for the shared appetite to invest in common technical understanding. Scaling AI isn’t just about building smarter models — it’s about building shared mental models.

26

1 Comment

𝗪𝗵𝗮𝘁 𝗶𝗳 𝘆𝗼𝘂 𝗰𝗼𝘂𝗹𝗱 𝘇𝗲𝗿𝗼-𝘀𝗵𝗼𝘁 𝘁𝗵𝗲 𝗰𝘂𝗿𝗲 𝘁𝗼 𝗰𝗮𝗻𝗰𝗲𝗿 🧬? After spending $600 for last-minute plane tickets to SF, my teammates and I spent our weekend competing at the Cerebral Valley PyTorch OpenEnv hackathon! With this being my first non-collegiate hackathon, I felt nervous being surrounded by so many senior SWEs and MLEs Despite all doubts, we pushed through and managed to place 2nd 🥈 among 120+ other teams! We were challenged to build an RL environment for frontier LLM development, and we were able to achieve as follows: - An RL environment that successfully 𝘁𝗿𝗮𝗶𝗻𝘀 𝗔𝗜 𝗮𝗴𝗲𝗻𝘁𝘀 𝘁𝗼 𝘀𝗰𝗶𝗲𝗻𝘁𝗶𝗳𝗶𝗰𝗮𝗹𝗹𝘆 "𝘁𝗵𝗶𝗻𝗸" about subsequent experimental procedures given a current experiment state - A procedurally generated world state 𝘀𝗶𝗺𝘂𝗹𝗮𝘁𝗶𝗼𝗻 𝘁𝗵𝗮𝘁 𝗿𝗲𝗮𝗰𝘁𝘀 𝗮𝗻𝗱 𝗰𝗵𝗮𝗻𝗴𝗲𝘀 𝘁𝗼 𝗺𝗶𝗺𝗶𝗰 𝗮 𝗿𝗲𝗮𝗹 𝗯𝗶𝗼𝗹𝗼𝗴𝗶𝗰𝗮𝗹 𝘂𝗻𝗸𝗻𝗼𝘄𝗻 - LLM access to over 𝟰𝟬+ 𝗮𝗴𝗲𝗻𝘁𝗶𝗰 𝗯𝗶𝗼𝗶𝗻𝗳𝗼𝗿𝗺𝗮𝘁𝗶𝗰𝘀 𝘁𝗼𝗼𝗹𝘀 that can modify and probe the world state for information I almost told my teammates to leave early so we could catch our flight, but we stayed... placed Top 6 Finalists... presented again... waited… and ultimately walked away with $9000 💸 and a silver medal 🥈 The takeaway? Keep faith. Something just might work out LOL. I'd like to thank my mom, my dog, and United Airlines for making this win possible, as well as my teammates Kevin Vo and Sean Chang If you were at the hackathon and are working in the the AI Health or AI Medical space, please feel free to reach out, I would love to chat sometime!

156

32 Comments

What does it take to serve agentic workloads on trillion-parameter models at 400 tokens per second per user — without trading throughput for latency? The NVIDIA Vera Rubin platform pairs Vera Rubin NVL72 with NVIDIA Groq 3 LPX to deliver low latency on trillion-parameter MoE models with 400K-token context with a 35x higher throughput per megawatt. Learn how the deterministic LPU chip-to-chip (C2C) fabric and extreme co-design address agentic AI's scale-up challenges. ➡️ https://bit.ly/4uenrHH

2

📣 NVIDIA Run:ai is now generally available on Amazon Web Services (AWS). Learn how NVIDIA Run:ai, now on #AWS Marketplace, allows for efficient #GPU orchestration at scale by introducing a virtual GPU pool and intelligent scheduling for faster, more efficient AI. Integrates seamlessly with: ✅ Amazon EC2 instances ✅ Amazon Elastic #Kubernetes Service (EKS) ✅ Amazon Sagemaker HyperPod ✅ and more Read our tech blog for the details. #AI #DataCenter #Cloud #HPC #AcceleratedComputing #Developer #AI #DeepLearning

3