Steven Kester Yuwono

🚀 We are excited to announce the release of MobileLLM-R1 — a series of sub-billion parameter reasoning model. Despite the common belief that reasoning only emerges after training on massive amounts of data, we prove otherwise. With just 4.2T pre-training tokens and a small amount of post-training, MobileLLM-R1 demonstrates state-of-the-art reasoning ability. Here is some highlights: 🎉 100% Open Source: model, training recipe, and data sources. https://lnkd.in/gG-WmYgz 🌟 2x-5x performance boost over other fully open-source models: MobileLLM-R1-950M model scores ~5× higher on MATH vs. Olmo 1.24B and ~2× higher vs. SmolLM2 1.7B—despite being smaller. ⚡ 1/10 of the pre-training token count compared to Qwen: Trained only on 4.2T Tokens total (11.7% compared to Qwen’s 36T tokens), MobileLLM-R1-950M matches or surpasses Qwen3-0.6B on multiple reasoning benchmarks. Collaborating with Ernie, Changsheng, Chia-Jung, Wei, Yuandong, et al. And thanks for the support from Raghuraman, Yangyang, Vikas

542

6 Comments

If you’ve been heads down building, here’s your catch-up of everything that happened in This Week of AI👇 🏗️ OpenAI & SoftBank Group Corp. kick off the $500B Stargate Initiative with compact, energy-efficient AI data centers 🧠 Yahoo Japan mandates daily AI usage for all 11,000 employees to double productivity by 2030 💻 Alibaba Group drops Qwen3-Coder — a 480B MoE model for agentic programming and long-context coding ⚔️ Microsoft, Google & OpenAI intensify the AI agent arms race with new launches and talent wars 🚨 UK police roll out real-time AI camera surveillance to detect driver violations nationwide Want the full breakdown + my predictions? 👉 Read the latest AI with Aish blog here: https://lnkd.in/d5vCxWDA PS: I do the reading, so you don’t have to 💙 Co-author: Harini Anand + AI 🤖 〰️〰️〰️〰️ ♻️ Share this to keep your network in the loop 👩‍💻 Follow me (Aishwarya Srinivasan) for weekly AI updates, product breakdowns, and industry signals that matter

188

43 Comments

Andrew Ng is right: the bottleneck in AI isn’t coding anymore—it’s product management. At Foot Locker and in my grad work in AI, I’ve seen how quickly AI tools can accelerate prototyping and delivery. The challenge isn’t building faster—it’s aligning on what to build, how to validate user needs quickly, and ensuring the right feedback loops are in place. The role of product managers has never been more critical. We’re not just writing specs; we’re orchestrating speed, empathy, and strategy in a world where execution is no longer the slowest part.

539

72 Comments

Vibe-coding is not the same as AI-Assisted engineering. My latest write-up: https://lnkd.in/gGyHwUUR ✍ A recent Reddit post described how a FAANG team uses AI and it sparked an important conversation about semantics: "vibe coding" and professional "AI-assisted engineering". While the post was framed as an example of the former, the process it detailed - complete with technical design documents, stringent code reviews, and test-driven development - is a clear example of the latter imo. This distinction is critical because conflating the two risks both devaluing the discipline of engineering and giving newcomers a dangerously incomplete picture of what it takes to build robust, production-ready software. As a reminder: "vibe coding" is about fully giving in to the creative flow with an AI (high-level prompting), essentially forgetting the code exists. It involves accepting AI suggestions without deep review and focusing on rapid, iterative experimentation, making it ideal for prototypes, MVPs, learning, and what Karpathy calls "throwaway weekend projects." This approach is a powerful way for developers to build intuition and for beginners to flatten the steep learning curve of programming. It prioritizes speed and exploration over the correctness and maintainability required for professional applications. There is a spectrum between vibe coding and doing it with a little more planning, spec-driven development, including enough context etc and what is AI-assisted engineering across the software development lifecycle. In stark contrast to the post, the process described in the Reddit post is a methodical integration of AI into a mature software development lifecycle. This is "AI-assisted engineering," where AI acts as a powerful collaborator, not a replacement for engineering principles. In this model, developers use AI as a "force multiplier" to handle tasks like generating boilerplate code or writing initial test cases, but always within a structured framework. Crucially, the big difference here is the human engineer remains in control, responsible for the architecture, reviewing and understanding every line of AI-generated code, and ensuring the final product is secure, scalable, and maintainable. The 30% increase in development speed mentioned in the post is a result of augmenting a solid process, not abandoning it. For engineers, labeling disciplined, AI-augmented workflows as "vibe coding" misrepresents the skill and rigor involved. For those new to the field, it creates the false and risky impression that one can simply prompt their way to a viable product without understanding the underlying code or engineering fundamentals. If you're looking to do this right, start with a solid design, subject everything to rigorous human review, and treat AI as an incredibly powerful tool in your engineering toolkit - not as a magic wand that replaces the craft itself. #ai #programming #softwareengineering

1,490

78 Comments

𝗛𝗼𝘄 𝗠𝗖𝗣 𝗦𝗲𝗿𝘃𝗲𝗿𝘀 𝗠𝗮𝗻𝗮𝗴𝗲 𝗦𝗲𝘀𝘀𝗶𝗼𝗻𝘀 𝘄𝗶𝘁𝗵 𝗟𝗟𝗠𝘀 𝗮𝗻𝗱 𝗖𝗼𝗻𝘁𝗲𝘅𝘁 𝗪𝗶𝗻𝗱𝗼𝘄𝘀 Managing sessions in LLMs isn’t just about keeping track of prompts. It is about orchestrating memory, context, and tools seamlessly across multiple agents. That’s where MCP servers come in. Here’s the flow I captured visually → Context Window as Working Memory: Holds system prompt, user inputs, assistant responses, tool calls, and rolling context → MCP Servers: Each server maintains its own session state, ensuring consistency across interactions → System Prompt + Tools + Code Loops: Work together to power agentic behavior while grounding outputs in real tasks → LLM Integration: The context is constantly summarized and updated before being sent to the model This structure ensures scalable, multi-session orchestration, which is critical for enterprise AI where multiple tasks and tools interact simultaneously.

149

14 Comments