Shotaro Kohama

So I'm at this meetup - and it's all about chatbots. I'm Evan Lin, by the way, a Technical Evangelist at LINE Taiwan. Got invited to share some updates on LINE API and my own dev experience. It was a blast. You can learn a ton about the latest platform updates - like Flex Message, which is pretty cool. It's powerful: think clothing display cards, transportation itinerary planning cards... the works. LINE's got a lot of useful features, too - like LINE Login and Messaging API. These are total game-changers for e-commerce websites and client apps. The meetup itself? Super community-driven, with a bunch of lightning talks. I mean, we're talking Chatbot UX design, voice input... all the good stuff. Key takeaways: LINE bot can remind users, no problem. And chatbots should be the main thing - not some annoying side feature. Reduce user resistance, avoid blocking... and demand satisfaction is key to a pleasant experience. Simple. And if you want more info, check out the collaborative notes: https://lnkd.in/gTymun-T Optional learning community: https://chatbots.kktix.cc/ Source: https://lnkd.in/gh7msDBM #Chatbots #Innovation #LINEAPI #TechEvangelist #CommunityDriven

1

We (w/ Shansan Gong, Ruixiang ZHANG, Huangjie Zheng, Jiatao Gu, Navdeep Jaitly, Lingpeng Kong) released a family of 7B diffusion language models, DiffuCoder, that specializes on code generation, with a focus on understanding and improving masked diffusion models. A core analysis of DiffuCoder is the autoregressiveness (AR-ness) score, a novel metric that quantifies the causal patterns in decoding, revealing how diffusion models break from strict left-to-right generation for more flexible, non-linear code planning. Recent advances in autoregressive (AR) models dominate code generation, but diffusion-based LLMs (dLLMs) like DiffuCoder offer a promising alternative, especially for complex programming tasks. DiffuCoder explores how these models decode differently—showing less global AR-ness in code tasks compared to math—and how temperature affects both token selection and generation order, unlike traditional AR models. We also introduce coupled-GRPO, a post-training RL method with a coupled-sampling scheme, to reduce performance drops during accelerated decoding, boosting parallelism and efficiency. We use a self-improvement pipeline that leverages AR-ness analysis, coupled-GRPO optimization, and evaluation on benchmarks like AceCode-89k to refine decoding strategies. This approach enables DiffuCoder to navigate diverse code generation pathways and enhance performance with modest computational overhead. Looking ahead, we aim to further leverage Reinforcement Learning to steer code generation through these decoding patterns, with the discrete nature of AR-ness scores providing a foundation for search-based strategies—ideal for the sparse rewards of optimizing complex code structures. Check out our full paper and code for a deeper dive! Paper: https://lnkd.in/gVWU3BDJ Code: https://lnkd.in/gmXTZ_6n Models: https://lnkd.in/gTcKCDr9 #MachineLearning #AI #CodeGeneration #DiffusionModels #NLP

220

5 Comments

📣 Simular has raised a $21.5M Series A to build AI-powered computer-use agents! When SPC Founder Fellows Ang Li and Jiachen Yang (former DeepMind scientists) joined SPC in 2023, they knew there was a way to make workflows completely autonomous. Now at Simular, they're building AI agents that handle tasks the way a knowledge worker would: autonomously. The agents navigate interfaces, complete workflows, and operate across all applications. Their breakthrough came with Agent S, an open agentic framework that hit 69.9% success on OSWorld, the industry's toughest benchmark for computer task completion. That's nearly human-level performance (72%). Now they're launching Simular 1.0: a native desktop agent that browses & operates in the background, across applications. It repeats successful tasks, learning from user corrections like a real teammate. Users can redirect the agent mid-task using natural language, and contextual triggers launch workflows automatically based on real-time activity. The technology is already processing millions of steps across insurance, recruiting, and travel workflows. Their hybrid system combines neural exploration with symbolic execution and continual learning to deliver enterprise-ready autonomy. The $21.5M Series A was led by Felicis Ventures, with participation from nVentures, Basis Set Ventures, Flying Fish Partners, SPC, and angel investor Lenny Rachitsky. Congratulations Ang Li, Jiachen Yang, and the Simular team!

65

4 Comments

What exactly is cloud computing and why is everyone doing it? In the latest No Dumb Questions, Phoebe Sajor interviews Josh Zhang, the technical lead of Stack Overflow's infrastructure team, to ask all her burning questions about the cloud and how AI is changing the compute landscape. https://lnkd.in/epJvrEry

5

1 Comment

[Call for Engineering Leaders & Hiring Managers] Thank you to everyone that joined us for our mock behavioral night earlier this month! Our next technical interview night with our AI Fellows is on April 9th. These events are designed to mirror real-world hiring processes while creating a space for supportive, collaborative engineering environments.​ Here’s what Johnnie Tabron Jr., Solutions Engineer at Cloudflare, had to say about his experience: “I had the chance to connect with two engineers with backgrounds in QA and web development. We had engaging discussions around system design and practical applications of AI, and what stood out was their ability to reason through tradeoffs, approach ambiguous problems, and apply product thinking in ways that go beyond just coding. I also shared how they could build on tools within the Cloudflare developer platform, which offers a powerful, accessible stack for bringing ideas to life. It was a genuinely great exchange.” We’re incredibly grateful to Johnnie and the many industry leaders who give their time, not only to support our talent community, but to strengthen the broader tech ecosystem. If you’re interested in getting involved at our upcoming event on April 9th, we’d love to have you join us: https://hubs.ly/Q047Zlbr0

6

🚀 C++ / Rust Systems Engineers – AI Inference Working with a team building a next-generation AI inference system purpose-built for transformers, delivering 10x+ performance and dramatically lower latency than current GPU systems. This is a true inference systems role - focused on the runtime and distributed execution layer that runs large transformer models on custom hardware. You’d be working on: • Distributed inference execution (multi-node & intra-node) • Runtime systems powering large-scale model inference • Porting state-of-the-art transformer models onto new accelerator hardware Great fit for engineers with backgrounds in performance-critical C++ systems, GPUs/accelerators, compilers or high-speed networking. If you enjoy building low-level systems that make frontier models run faster at inference, drop me a message ❗ #Inference #SystemsEngineering #DistributedSystems #GPUComputing #AIInfrastructure

21

Lablup is now a Python Software Foundation Participating Sponsor 🎉 From our very first PyCon Korea to managing 16,000+ GPUs across 110+ sites today, the Python ecosystem has been foundational to everything we build. The AI industry is built on open source. More companies should be investing in the foundations that sustain it. As a company that has long been part of this ecosystem, this PSF sponsorship is how we're broadening that commitment. 🔗 https://lnkd.in/gncjYYeS #Lablup #BackendDotAI #Python #PSF #OpenSource

33

Honored to have contributed to this exciting collaboration with Zhiyu (Albert) Zeng and Nathan Ai at Cadence, along with our academic and industry partners from UC Santa Barbara (Prof. Zheng Zhang’s group) and Intel Corporation (Hai Li & Ian Young). Our paper “DeepOHeat-v1: Efficient Operator Learning for Fast and Trustworthy Thermal Simulation and Optimization in 3D-IC Design” has just been published in IEEE Transactions on Components, Packaging and Manufacturing Technology. This work leverages AI to bridge the gap between physical modeling and design optimization—towards fast, generalizable, and trustworthy thermal analysis for 3D-ICs. 🔗 Link to the paper: https://lnkd.in/g2WAHbGA #thermal #operatorlearning #3DIC #AI

23

1 Comment

We at Haladir (YC W26) released a report on RLFR: Reinforcement Learning from Formally-Defined Rewards for Code Generation (https://lnkd.in/guCH7w7C). The core idea is simple: we present a simple offline and single-turn RL framework that uses formal verification as the primary reward signal for program synthesis in LLMs rather than unit test-based rewards (RLVR), showing improvements over a swath of different coding benchmarks. We fine-tune both Qwen2.5-Coder-7B-Instruct and Qwen3-8B using RLFR and Dafny, a specification language that utilizes Boogie (IVL) and Z3 (SMT solver) to check verification conditions. Despite solely training on Dafny, we show improvements on Python-native coding benchmarks, suggesting generalization past the formal verification domain.

90

8 Comments