Kevin Rosendahl

🚀 Excited to share a fresh Roblox Studio update: Studio MCP Server improvements + external LLM support for Assistant. We’ve added tools like start/stop Play, run scripts in Play, and richer output logs, all aimed at tightening the edit → test → fix loop for agentic workflows. And creators can now power Assistant with external LLMs (e.g., Anthropic / OpenAI / Gemini), depending on what works best for them. DevForum details: https://lnkd.in/gEEag2b3 Can’t wait to see what the community builds with this 🙌 #Roblox #RobloxStudio #AI #LLM

96

4 Comments

In the world of AI, where performance and efficiency often clash, #MoonshotAI has cracked the code with their new Kimi Linear model. This isn't just another upgrade - it's a complete rethink of how attention mechanisms work. At its core is the #Kimi Delta #Attention (KDA) module, which replaces traditional quadratic-cost attention with a linear-complexity approach. The result? Faster processing and significant memory savings. Imagine being able to handle 64k tokens without breaking a sweat - this model does that while using almost half the memory of its predecessors. KDA uses a fine-grained diagonal gate to manage memory decay and positional awareness, making it not just efficient but also smarter about how it uses resources. The benefits are clear: up to 75% reduction in key-value cache size and decoding throughput that's six times faster than before. This means you can process longer sequences without the usual slowdowns or memory bloat. It's like having a supercharged engine for your AI workloads. What makes this even cooler is how it balances speed and performance. By using KDA in most layers and keeping only a few global attention layers, Kimi Linear maintains strong results while delivering massive efficiency gains. This isn't just about being fast but also about doing more with less. If you're into AI research or building models that need to handle long contexts, this is worth checking out. Read More: https://lnkd.in/geUBd_wA #AI #Efficiency #MachineLearning #Innovation

5

From code reviews ➝ train stations 😄 Seeing this on a BART platform is surreal. Qodo is now running across the digital boards at Embarcadero & Montgomery in SF. We’re building AI code review that enforces engineering standards and catches more real issues - with less noise. Because in 2026, the winning teams won’t just write more code. They’ll ship higher-quality code, at a higher pace.

153

12 Comments

This is wild. Anthropic let multiple AI agents work in parallel to build a C compiler from scratch - no human compiler engineers involved. In about two weeks, they produced ~100k lines of Rust code and a compiler that can compile the Linux kernel and even Doom. Cost: around $20k in API credits. What’s exciting isn’t just that it worked. It’s the shift from “AI helps write code” to “AI systems can autonomously build complex software by collaborating.” Feels like the early days of distributed systems - but for intelligence. Definitely worth a read: https://lnkd.in/eu4SJVSw #AI #Engineering #AutonomousAgents #SoftwareDevelopment

36

2 Comments

OpenAI's CFO Sarah Friar published a piece laying out how the company thinks about delivering more intelligence at lower cost, and the framing is worth sitting with for a minute. The argument is that gains compound across the entire stack: chip efficiency, compute infrastructure, model improvements, and product delivery all reinforce each other. It's not one lever pulling harder, it's multiple layers improving in parallel. That compounding logic is actually what makes the economics of AI hard to predict from the outside. If you only track model benchmarks or headline API pricing, you might miss that a quieter efficiency gain two layers down is what actually changes the unit economics for builders and enterprises. Having a CFO make this case publicly is also a bit of a signal in itself. The conversation is shifting from "look what the model can do" toward "here's why the cost curve keeps moving" which is the kind of narrative that matters when you're trying to close large enterprise deals or justify continued infrastructure investment. Whether the compounding holds at the scale OpenAI is targeting is still an open question. But the framework they're describing, full-stack efficiency rather than model-only improvements, is a reasonable way to think about where durable cost reductions actually come from. https://lnkd.in/gUQBTAan

1

New blog post: A Fork in the Road: Deciding Kafka's Diskless Future. Kafka is getting serious about S3 and finds itself at an architectural crossroads that will shape its next decade. Several new KIPs (1150, 1176, 1183) aim to reduce replication costs across cloud availability zones, but the implications go far beyond networking cost. It’s a mistake to think of S3 as simply a cheaper disk or a networking cheat. Building on object storage opens the door to operational benefits such as elastic, stateless compute — something many modern analytics systems already exploit. In my latest post, I outline two competing future paths: 🔹 Evolutionary path: Reuse large parts of existing Kafka components to reduce code changes and long-term maintenance. 🔹 Revolutionary path: Separate stateless and stateful layers to realize the full operational benefits of disaggregated storage. The post examines the trade-offs of each path, how the current KIPs map to these two paths and poses a broader question: what should Kafka become? And what will keep it relevant in the decade ahead? https://lnkd.in/d4E2BAe4

149

12 Comments