Woojong Koh

Inspired by the Anthropic blog post on building a C compiler using long-running AI agents (ralph-loops), Thomas Dybdahl Ahle and Erik Meijer proposed to build a comprehensive SystemVerilog simulator. For the simulator to be practically relevant, certain features are needed -- like a full event-driven simulator, VPI/cocotb integration, UVM runtime support, bounded model checking, logic equivalence checking, and mutation testing. The functional results are genuinely impressive, powered by the 580k LOC generated over 43 days. https://lnkd.in/ex2fPqy4 In my opinion, the implications for legacy EDA software and its "moat" are substantive, especially if we find a path to more open industry collaboration standards like with Silicon Integration Initiative, Inc.. These implications naturally carry over to what is tractable for startups, and what we may see unfold over the next 12-24 months for hardware engineering. Astonishingly enough, this project was completed by Thomas while he was on paternity leave and welcomed a new human to his and Normal's world.

94

10 Comments

What if your entire home acted as your personal assistant? We’re thrilled to announce our investment in Sorcerics(Hyeonjong Ryu), a team building a fully autonomous AI home grounded in the philosophy of “Ambient AI”, where technology truly understands people. Unlike traditional smart homes that wait for commands, Sorcerics’ AI home proactively interprets users’ gestures, behaviors, and spatial context. For example, rather than simply detecting when someone lies down, the system assesses overall sleep readiness and automatically adjusts lighting and the surrounding environment-creating a home that genuinely understands its users. 📍Sorcerics -Uses a single camera and proprietary LLM to analyze gestures, behavior, and spatial context -Delivers highly aesthetic hardware at one-quarter the cost of competitors, ensuring strong product competitiveness -Protects user privacy through on-device AI -Plans a global product launch in Q1 2026, targeting the U.S. market

15

Back in 2023, TGI has initiated a movement where optimized inference engines rely on the model architectures from transformers as the source of modeling truth. This approach is now adopted by downstream inference engines: vLLM, SGLang, as well as local engines with inter-compatibility such as llama.cpp or MLX. Given we at Hugging Face contribute to those community runtimes, we decided to switch text-generation-inference itself to maintenance mode. Thanks for all the tokens 🔥 Hat/tip Lysandre Debut

481

32 Comments

🚀 We’re excited to announce our investment in LinqAlpha In public markets, the edge is no longer just about accessing information faster. It is about finding the right signals before they are priced in. LinqAlpha is building the Alpha Intelligence Layer for global public markets, helping investment teams use specialized AI agents that learn their own investment frameworks, reason across internal research, and surface market-moving insights. What stood out to us is how purpose-built LinqAlpha is for institutional finance: model-neutral, data-vendor-neutral, and integrated across public, licensed, and internal data in a compliant workflow. This is where we believe vertical AI wins: by owning the workflow, trust, and data layer of a specific industry. Congrats to the Jacob Chanyeol Choi, Hojun Choi and the LinqAlpha team. And to Josh Jung & Julie Seo who led the investment at Z Venture Capital. Check out why we invested in the comments below.

37

1 Comment

Is agentic AI bringing about the revenge of the CPU? I’ve been thinking about that question a lot lately, and NVIDIA gave us another interesting data point at Hot Chips last week. Vera is NVIDIA’s first CPU built around its own custom core, and they designed it specifically for agentic AI. Why? Because an AI agent is doing much more than running matrix multiplication. It is browsing, retrieving, reasoning, compiling code, calling tools and orchestrating increasingly complex workflows. NVIDIA calls agentic AI “the most complex computing workload in history.” I think that complexity is going to change how we think about the architecture of AI systems. The last wave of AI put enormous focus on the GPU. The next wave will demand the right compute for every part of the workload. CPU. GPU. Accelerators. Memory. Networking. Heterogeneous computing is having its moment. And perhaps the “general purpose” CPU is too (again). https://lnkd.in/gaUUypeB

526

39 Comments

Introducing Solar Open 2. An open-weight foundation model built for agents that do real work. 250B MoE, 15B active params per token. 1M-token context. Top MMLU-Pro and LiveCodeBench scores in its class. #1 on the Korean benchmark average. We built this for office work: reports, spreadsheets, presentations, the stuff that actually fills a workday, run end to end by an agent. On Ko-GDPval (58 professions, real workplace deliverables), Solar Open 2 matches a 1.6T model at a third of the active compute per token. The weights are available on Hugging Face today under a commercial-use license, plus a full Technical Report on architecture, training, and ablations. Go verify it. Deploy it on your own infra. Build on it. 📑 Tech report & model download: https://lnkd.in/gHNEWqXn 📌 Blog: https://lnkd.in/gwKXG5SQ #SolarOpen2 #UpstageAI #SolarLLM #OpenWeights #AgenticAI #SovereignAI

81

1 Comment

Big news for deep-tech builders 🚀 TESORO VC × Parallel Works: Unlocking Scalable HPC & GPU Power for Founders We’re excited to share that Tesoro VC has entered a strategic ecosystem partnership with Parallel Works, bringing scalable HPC and GPU compute directly into the hands of our AI + Semiconductor Accelerator startups. This is a major unlock for the kinds of companies we support. ⚡ Why this matters for Tesoro founders Many of our startups aren’t building lightweight software — they’re: • Designing next-gen semiconductor architectures • Running advanced chiplet and packaging simulations • Training compute-intensive AI models • Performing large-scale physics, materials, and systems simulations These workloads demand serious compute — and managing that infrastructure can easily distract from core product development. Through Parallel Works’ ACTIVATE platform, our startups now get: ✅ Access to HPC + GPU resources across hybrid cloud and on-prem ✅ A ready-to-use compute framework with best practices built in ✅ Cost controls and budgeting tools (huge for early-stage teams) ✅ Repeatable workflows that scale from prototype → production Net effect: founders spend more time innovating and less time wrestling with infrastructure. 🧩 Why this strengthens the Tesoro ecosystem This partnership doesn’t just help individual startups — it reinforces the entire Tesoro platform. For industry partners Startups can design, simulate, and validate technologies in environments that more closely resemble real enterprise and manufacturing conditions. For semiconductor & manufacturing stakeholders Founders can run higher-fidelity simulations earlier, reducing technical risk before fabrication and scale-up. For AI and national security–aligned innovation Secure, scalable, portable compute workflows are essential in regulated and mission-critical environments. For investors This improves technical de-risking. Companies can validate performance and scalability earlier — leading to stronger milestones and clearer paths to commercialization. At Tesoro, we’re building more than an accelerator — we’re building an execution layer for the AI + Semiconductor economy. Pairing founder support, industry access, and capital pathways with advanced compute infrastructure creates a powerful bridge from: Idea → Simulation → Validation → Scale Proud of this step forward and excited to see what our founders build with this new level of computational firepower. 📰 Read the full announcement here: https://lnkd.in/giXf9_vZ

47

2 Comments

LB Investment recently led the Series A investment in ROAI, a startup building optimized robotic production lines for next-generation manufacturing. Built on deep experience from Hyundai Motor’s manufacturing operations, ROAI is developing autonomous robotic planning and intelligent production infrastructure for industrial automation(mainly led by our professional, KiHeon Kim, who previously worked on engineering and development at Hyundai Motor). Backed by our KRW 303B AI-focused fund, myself and LB are actively investing in Physical AI, Vertical AI, AI chips, and AI-native infrastructure. (now, leading 5+ investment rounds of AI startups, with total investment commitments KRW 50B(USD 35M)). 로아이, 130억 규모 시리즈A 투자 유치 - 전자신문 https://lnkd.in/gDmz9uYs #ROAI #PhysicalAI #VerticalAI #ManufacturingAI #IndustrialAutomation #Robotics #VentureCapital #LBInvestment

39

Some of you already know what I’ve been working on, but I haven’t shared much about it here. I’ve been building robots with our team at khameleon, and I’m happy to finally share what we’ve been up to! We’re developing robots to help with commercial cleaning, starting with hotels. There’s a lot of work beyond keeping the floors clean, and we’re building a robot that can switch tools to take on different tasks throughout the day. We’ve spent the past several months building and testing, including in real hotel environments. Getting to see how the staff work and where they could use a hand has helped us figure out where to focus. I’m proud of our team and excited to share a first look. Would love to hear what you think!

59

5 Comments

👏👏 We’re excited to join NVIDIA’s Inception Program as we continue building multimodal data infrastructure for the next frontier of AI

76

5 Comments