Activity
595 followers
-
Zhouhan Lin shared thisWow I am super excited to see this picture posted! The guy thumbing up in the front is actually me! And it was nice to talk with you about our JEPA-related idea on language models!
Zhouhan Lin commented on a post
6mo
Congrats to you and your company! Bringing AI to the core industries that run our economy is such a great goal!
-
Zhouhan Lin liked thisZhouhan Lin liked thisYann LeCun has long argued that predicting pixels is the wrong objective for computer vision. The same logic likely applies to text. Current LLMs spend vast resources predicting functional morphemes rather than abstract ideas. A new architecture, ConceptLM, attempts to align NLP closer to the "World Model" paradigm. PAPER: Next Concept Prediction in Discrete Latent Space Leads to Stronger Language Models This is a move away from pure surface-form autoregression toward hierarchical planning. * This is not the first paper with such an approach. There are many others, e.g. LLM-JEPA, BLT, Thought Gestalt, NextLat, etc. KEY INSIGHTS: 🧠 The "World Model" for Text Similar to JEPA architectures in vision, ConceptLM predicts in a latent abstract space. It separates high-level semantic transitions (Concepts) from low-level syntactic realization (Tokens). This creates a dedicated "reasoning path" that is distinct from the generation path. 📉 Shifting the Scaling Curves The efficiency gains here are structural, not just distinct optimizations. The model achieves comparable performance to standard baselines using **24% fewer training tokens**. If you are optimizing for training compute efficiency, this architecture changes the calculus. 🔄 Retrofitting Open Weights Crucially, the authors successfully applied this method to Llama-3-8B via continual pre-training. This suggests we don't need to train from scratch to get these benefits—we can potentially "upgrade" existing dense models into hierarchical planners. THE TAKEAWAY: As we hit the ceiling of what pure Next Token Prediction can achieve regarding reasoning and planning, architectures that force the model to predict distinct "thoughts" (discrete latents) before "words" will likely become the standard for the next generation of reasoning models. Links in the comments 👇
-
Zhouhan Lin liked thisThank you Alona Fyshe, Scott Lilwall and the Approximately Correct podcast team at AMII for hosting me for this fun conversation! I continue to believe that responsible development and sharing of open-weight models is an important way to ensure that we are make meaningful progress in AI.Zhouhan Lin liked thisThere are a lot of benefits to open-source models in AI. But the more open a model is, the more vulnerable it is to being misused. In the latest episode of Approximately Correct, Mila researcher and Chief AI Officer at Cohere Joelle Pineau talks about the marginal risk approach she uses, focusing on measuring the change in a specific harm before and after a model’s release, to lead to safer, more secure AI models. Have a listen: https://hubs.la/Q03PJx6k0 #AI #OpenSourceAI #MachineLearning #ResponsibleAI
-
Zhouhan Lin liked thisIt is coming! CenoBots L3! CenoBots L3 is compact in size, but is equipped with a 100-TOPS NVIDIA AI chip and a 96-beam 3D LiDAR. We are committed to building the most advanced cleaning robots in the world. Furthermore, it delivers superior cleaning performance by employing two disk brushes with a down pressure of 18kg (39.7 lbs). I am confident that it will be the best compact cleaning robot you have ever seen. Reach out to us if you are interested in discovering more about it.Zhouhan Lin liked thisDuring the forthcoming ISSA Show North America in November, CenoBots will proudly introduce the L3, our latest innovation. Developed with the vision of redefining compact scrubbers, the CenoBots L3 demonstrates our commitment to delivering cleaning robots that are not only compact in size but also exceptionally reliable, easy to operate, highly intelligent, and capable of providing superior cleaning results 1) Compact and Versatile With a pass-through width of just 700mm/27.6in, the CenoBots L3 is highly maneuverable in tight and crowded spaces. Thanks to its compact design, it cleans edges and corners efficiently, making it ideal for facilities such as grocery stores, hospitals, office buildings, and schools. 2)Enhanced Productivity and Cleaning Performance With a 25L/6.6gal solution tank and 400mm/15.7in cleaning width, the L3 delivers greater coverage and efficiency than comparable robots. The 18kg/39.7lbs pressure dual disk brushes ensure superior cleaning results with fewer passes. 3)Powered by NVIDIA AI Chip for 100 TOPS and 96-beam 3-D LiDAR Equipped with NVIDIA’s advanced AI chip delivering 100 TOPS of processing power, combined with 96-beam 3-D LiDAR, ensuring fast, stable, and intelligent environment understanding and navigation. 4)Enhanced Safety with Intelligent Detection System Our advanced AI system detects and avoids not only standard obstacles but also challenging small objects like temporary rugs and cables on the floor, thereby preventing entanglement and protecting both the machine and the premises from damage. 5)Voice-Command Operation for User-Friendly Control Use simple voice commands to operate the robot without touching screens or complex interfaces, reducing the difficulty of operation and making operation more comfortable and accessible. 6)Extended Runtime with Mobile Water Tank Compatible with a mobile water tank and workstation for longer operation in facilities without easy access to water plumbing. Should you require further details regarding the CenoBots L3, please do not hesitate to contact us: info@cenobots.com
-
Zhouhan Lin liked thisi have learned about and implemented Openrouter's provisioning API. happy to provide any user, who requests (email or DM me), with an openrouter api key charged with a few $$$ on MyChartExplorer.Zhouhan Lin liked thisthe past few weekends here and there spread through a few months have been fun for me. i vibe-coded and along the way learned various services that enable one to rapidly deploy a trial version of a web service. an awesome era for software development! the link to the blog post below.
-
Zhouhan Lin liked thisZhouhan Lin liked this💡 This work feels close to CycleGAN moment for the foundation-model era. An idea that aligns with my ongoing research! MIT CSAIL’s new paper: ““𝗕𝗲𝘁𝘁𝗲𝗿 𝗧𝗼𝗴𝗲𝘁𝗵𝗲𝗿: 𝗟𝗲𝘃𝗲𝗿𝗮𝗴𝗶𝗻𝗴 𝗨𝗻𝗽𝗮𝗶𝗿𝗲𝗱 𝗠𝘂𝗹𝘁𝗶𝗺𝗼𝗱𝗮𝗹 𝗗𝗮𝘁𝗮 𝗳𝗼𝗿 𝗦𝘁𝗿𝗼𝗻𝗴𝗲𝗿 𝗨𝗻𝗶𝗺𝗼𝗱𝗮𝗹 𝗠𝗼𝗱𝗲𝗹𝘀”” reopens a key question from classic unsupervised learning: 👉 Can models learn shared structure across modalities, even without paired data? The proposed Unpaired Multimodal Learner (UML) trains on unpaired text, images, and audio, sharing parameters across modalities. 𝙉𝙤 𝙚𝙭𝙥𝙡𝙞𝙘𝙞𝙩 𝙖𝙡𝙞𝙜𝙣𝙢𝙚𝙣𝙩. 𝙉𝙤 𝙥𝙖𝙞𝙧𝙞𝙣𝙜. Yet it discovers cross-modal structure that improves unimodal models themselves e.g., unpaired text boosts image understanding and robustness. Why this matters: 💡 Shows that horizontal scaling (across modalities) can rival traditional scaling by size. 🌍 Offers a new perspective on implicit world modeling: learning the shared latent reality that connects vision, language, and sound. 🧩 Bridges supervised and self-supervised learning, unlocking unpaired enterprise and scientific datasets. 🔭 Points toward modality-agnostic foundation models that learn from the world directly, not from curated pairs. 𝗦𝘂𝗺𝗺𝗮𝗿𝘆: ✨ Sometimes, shared inductive structure plus scale is all the alignment we need. #MultimodalAI #WorldModels
-
Zhouhan Lin liked thisZhouhan Lin liked thisWhat’s happening at Microsoft in the world of Spoken Language Models? Join us this Thursday at 12:00 PM EDT for our #ConversationalAI Reading Group, where we will host Jinyu Li, Partner Applied Science Manager at Microsoft. Talk title: “The Development of Spoken Language Models” More Info: https://lnkd.in/ekeUXdXX This is a great opportunity to: - Learn about the key challenges in building spoken #LLMs - Engage directly with one of a leading experts in our field - Get insights on how Microsoft is advancing conversational AI Will you join us? As usual, everybody is welcome! #LargeLanguageModels #DeepLearning #SpeechAI Pooneh Mousavi Gina Cody School of Engineering and Computer Science Mila - Quebec Artificial Intelligence Institute
-
Zhouhan Lin liked thisZhouhan Lin liked thisAnnouncing another lead for our Mechanistic Interpretability tutorial! 🛠️ Riccardo Ali is an exceptionally talented PhD student at Cambridge, with a strong passion for rigorously understanding the internals of neural networks. And when I say rigorously, I mean category-theory-level rigorously 🐈 In recent months, Riccardo also leveraged his insight to analysing the learning trajectory of LLMs from the lens of entropy, positioning him greatly for this tutorial session! I had the pleasure to work with him on several occasions -- I'm delighted that our attendees will get to experience this also! Join us in Montenegro 🇲🇪 next month for the MMLW, part of EEML workshops, organised together with the Montenegrin AI Association 🚀 (Free) registration link in the comments. There's less than 48h left to register!!! ⏰
-
Zhouhan Lin liked thisZhouhan Lin liked thisToday marks an important milestone. I’m launching Percepta together with Hemant Taneja, Hirsh Jain, Thomas Mathew, Radha Jain, Michael Rochlin, Constantinos Daskalakis and an incredible team, with the goal of bringing AI to the core industries that run our economy. For AI to deliver transformational impact, we need a strategy that orchestrates multiple disciplines. The most meaningful work of my career has been with strong interdisciplinary teams who take on ambitious challenges. That’s the kind of team we’re building at Percepta. The two key pillars of our work: First, we deeply partner. Our researchers, product managers, and engineers work inside our partners’ operations - shadowing workflows, mapping decision points, building AI systems, and ultimately owning outcomes. Second, we believe organizations revolve around decisions, and we’re reimagining decision-making to orchestrate AI agents and humans. This is our core research focus, blending foundation models, reinforcement learning, and optimization. The unique challenges of these domains won’t be solved by mimicking human data or finding a better prompt. Decision-making has many distinct elements, and our AI team brings complementary skills to the challenge: collectively, we’ve built superhuman bots, deployed RL agents at scale, trained large language models to reason, proved fundamental results in learning theory, and released algorithms and tools used by the field every day. We’ve already seen outcomes that change behavior: step-function productivity improvements in financial services, material enterprise value in care delivery organizations, and orders-of-magnitude cycle-time reductions in public-sector processes. A bit over a year ago today, I joined General Catalyst - compelled by Hemant Taneja vision for AI transformation and to partner with him and help establish the early workings of Percepta. Since then, we’ve been quietly building a team of 25+. I’m grateful to the broader GC famiglia and especially Quentin Clark, Jeannette zu Fürstenberg, Marc Bhargava, David Fialkow, and Teresa Carlson for being early partners in our journey. We’re working against an incredibly ambitious mission. We’re based in NYC and Boston and growing quickly. If this excites you, let’s chat!
Experience & Education
-
Shanghai Jiao Tong University
View Zhouhan’s full experience
See their title, tenure and more.
or
Explore more posts
Explore collaborative articles
We’re unlocking community knowledge in a new way. Experts add insights directly into each article, started with the help of AI.
Explore More