Devesh Rathod

Devesh RathodAug 31, 2026

Anthropic just introduced MHS (Model Hardware Standard) — think MCP, but for physical hardware. It gives AI agents a standardized interface to interact with sensors, robots, lab equipment and other programmable hardware. Imagine: PCB design → firmware → physical testing → measurements → AI-driven iteration. MCP connected agents to software. MHS might connect them to the physical world. AI × Hardware is getting interesting. https://lnkd.in/djqKnZdV

Devesh RathodSep 6, 2026

So yes, we can “vibe code” a PCB now. GPT Astra working with tools like KiCad definitely looks impressive. But I think there’s a more interesting question: At what cost? PCB routing is fundamentally a constrained optimization problem. We already have decades of specialized algorithms built to search that space efficiently. An LLM can understand intent, read datasheets, operate tools and iterate on a design. But can it actually beat specialized routing algorithms on: → compute cost → routing time → via count → signal integrity → constraint satisfaction → manufacturability That’s the benchmark I’d like to see. Because “AI can design a PCB” and “AI found a better way to design a PCB” are two very different claims. Vibe engineering is getting real. Now let’s benchmark it against the algorithms that have been solving these problems for decades.

Devesh RathodSep 19, 2026

We might be using LLMs for things that never needed language generation in the first place. Came across Jev by TypeSafe AI, and the idea behind it is pretty interesting. Most LLM-based systems work something like: input → transformer → autoregressive decoding → tokens → JSON/text → parsing → validation → action Which makes sense when you actually need the model to generate something. But imagine your agent just needs to decide: Should I call search_web, query_database, or send_email? Do we really need an LLM generating tokens just to answer that? That’s basically where Jev comes in. Jev is what TypeSafe calls a System One Model. Instead of generating text token-by-token, you give it some state + typed questions, and it returns typed decisions with probability distributions/confidence. So instead of: "I think the best action is search_web..." you get something closer to: search_web → 0.91 query_database → 0.07 send_email → 0.02 Your code can immediately branch on that. The technical part I found interesting is their RLCD — Reinforcement Learning for Calibrated Decisions approach. The goal isn't “generate the nicest response.” It’s closer to: make the correct decision and know how confident you are about it. And then there’s the cost. Jev is currently listed at $0.042 / 1M input tokens with free output tokens. That makes a lot of sense because… there really isn't much output to generate. TypeSafe also reports roughly 70–500ms end-to-end latency, although those performance numbers are currently their own benchmarks. I don't see this replacing LLMs. I see something more interesting: LLM → reason, write, code, explain Jev → classify, route, score, gate, decide An AI agent probably doesn't need a massive generative model for every single decision in its control loop. Sometimes you don't need another 500 tokens. You just need a reliable if statement with probabilities. That feels like a pretty interesting direction for agent architecture. https://lnkd.in/dbtWrwdw #AI #LLM #AIAgents #MachineLearning #Jev #SoftwareEngineering