Mrinal Wadhwa

We’re racing to build data centers on the assumption that AI demands infinite compute. In healthcare, what if the real breakthroughs come from AI that makes that assumption obsolete? I explore this in my latest article for Health Data Management.

12

If you are interested in OpenTelemetry and more specifically OTEL fleet management, this interview with Andy Keller (BindPlane) is a must listen. Bindplane has played a instrumental in the creation of OpAMP as well as a number of other OTEL components and Andy has been one of the prime drivers of BindPlane’s contributions.

22

2 Comments

Three Implications； 1) Validation of the AWS architecture bet.  The fact that AWS is wiring Trainium prefill to Cerebras decode for Bedrock and Amazon Nova means Cerebras is now a first‑class, production inference tier at a hyperscaler, not a boutique sidecar. 2) More “stickiness” via specialization.  Explicitly positions WSE‑3 as the high‑throughput decode engine, with AWS telling customers “we’ll give you 5× token throughput vs GPUs” on certain workloads. Once developers and internal teams tune models and serving stacks to this split prefill/decode architecture, ripping Cerebras out becomes non‑trivial. 3) Ecosystem exposure without owning the whole stack.  Being the decode side of a disaggregated architecture lets Cerebras focus where its on‑wafer SRAM and bandwidth are most differentiated, while AWS shoulders prefill, orchestration, and much of the software surface with Neuron and Trainium. That plays to Cerebras’ strength (raw token firehose) rather than forcing it to be a universal replacement for GPUs or Trainium.

18

1 Comment

We have revamped our opper console client (cli) to be a great launcher for agents. Now you can launch codex, claude code, opencode and many other agents with hardened configuration through controlled inference. The following config will be maintained across any agent launched: - approved models (300+ of them, or some, including all the frontier ones) - logging (or not), - guardrails - data boundry - effective terms of services and data use - evals and alerts - cost budgets This yields: - Significantly higher model/agent uptime, keep productivity up - Faster experimentation on new models and harnesses - No more wrangling with api keys or plans as the launcher handles auth - Less risk for security mistakes, especially when moving fast - Very likely lower cost Give it a spin! Link in comment

25

4 Comments

> Why would I switch from claude code to isomux? Good question. The big ones are probably: - Mobile support (https://lnkd.in/gGrqR8EW) - Human collaboration at the agent conversation level (https://lnkd.in/g5BuWzHV) - Inter-agent discoverability and soon communication (https://lnkd.in/gSJ_cyME) Some smaller things: - hierarchical system prompts (https://lnkd.in/gTTVC8fy) - cron job UI (https://lnkd.in/gTvcSvKF) - task system designed for humans and agents (https://lnkd.in/gtSA7kWk) It also has the typical QoL features you'd expect: - speech-to-text and text-to-speech - safety hooks - auto-generated conversation topics - file attachments - notifications - conversation branching - built-in terminal - built-in diff tool And of course, the main selling point is the cute, anthropomorphized agents :) (and skeuomorphic elements)

1

3 Comments

Congrats to Cerebras for their $60B IPO! 🎉 The detail underneath the headline: their CFO named OpenAI 5.4 and 5.5 internal workloads as customers, running trillion-parameter models on wafer-scale silicon. That's the buyer that mattered for this offering. For years, wafer-scale was a hardware bet most observers didn't take seriously. Bigger chips, exotic packaging, expensive yields. The contrarian thesis was that on-chip memory bandwidth would matter more than raw FLOPS once models got big enough. That thesis paid off because the market shifted. Inference has replaced training as the scarce resource. Serving frontier models in production drives the economics now, and that game rewards a different architecture from training. A $60B IPO is partly about Cerebras and partly about what the market has decided to price: inference scarcity is real, non-NVIDIA silicon has room, and the largest model owners are quietly hedging their compute stack. Worth noting that the S-1 didn't disclose cost-per-token, latency, throughput, or utilization. The narrative is there. The numbers that would confirm or weaken it are not.

10

2 Comments

Uber partners with Nuro, invests $300M in Lucid to launch upscale robotaxi serviceUber is investing hundreds of millions of dollars into Lucid and autonomous vehicle technology startup Nuro in a bid to launch its own robotaxi service. https://lnkd.in/ez_CmpF7

1

New feature for Ray users! We just launched the persistent Ray Task Dashboard on Anyscale,  making it easier than ever to monitor, debug, and optimize Ray workloads at scale. ✅ Logs and observability that persists beyond cluster shutdown ✅ No more 10k task ceiling ✅ Built on the new Ray Event Export framework - available from Ray 2.49 Say goodbye to ephemeral dashboards and hello to always-on observability. Read More: https://lnkd.in/etqvDb4K

40

1 Comment

Engineers, say goodbye to tedious component searches! Bridges' new AI feature finds the right parts in seconds using natural language—just ask and get instant answers from SiliconExpert datasheets and your internal docs.

8

1 Comment

Building MCP Servers for AI in RAD Studio with the new MCP SDK If you’ve been tasked with implementing an MCP solution and need to get up to speed quickly, this webinar will take you from zero to a working MCP integration. Join Embarcadero Developer Advocate Ian Barker and Matt Springfield from nSoftware for a practical, developer-focused session that breaks down the Model Context Protocol (MCP), why it’s becoming critical for AI integration, and how to implement it without getting buried in protocol details. You’ll see how MCP SDK accelerates development through an intuitive, event-driven architecture that abstracts the complexity of MCP server implementation. Using live code examples in RAD Studio, the session walks through building a functional MCP server, handling requests, and connecting AI models to your applications. By the end, you’ll understand the core concepts behind MCP and how to quickly build a working proof of concept using MCP SDK.

25

The secret is out: LangSmith is how TierZero production agents self-investigate and self-improve their own prompts and performance to free up tens of thousands of engineering hours for our customers.

15

2 Comments

TestMu AI Launches Kane CLI, the New Browser Automation Tool Built for AI Agents and Developers The terminal-native browser verification tool ships today with native support for Claude Code, Codex CLI, Cursor, and Gemini CLI, and it’s free to start San Francisco, CA – April 28, 2026 – TestMu AI (https://lnkd.in/gT5bmv6W) (formerly LambdaTest), the world’s first full-stack Agentic Quality Engineering platform, today announced the launch of Kane CLI (https://lnkd.in/g7VAjsJA), a new browser automation tool that runs directly from the terminal. Kane CLI is the first tool designed simultaneously for human developers and AI coding agents, closing the gap between code generation and verified browser execution. AI coding agents have transformed how software gets written, as AI agents are shipping https://lnkd.in/gNv3XyDP #msp #channelPartners #carriers #enterprise #Telecommunications #ai #messaging #mobility #ucaas #ccaas #cpaas #Mobility

24

1 Comment

New investment alert! 🚨 The AI industry spends $200bn a year on GPUs. A significant portion of that compute sits idle; not because of a hardware shortage, but because engineers have no visibility into what's actually happening inside their clusters. The default fix? Buy more GPUs. zymtrace closes that gap. Their platform continuously profiles GPU and CPU workloads down to the line of code, pinpoints the bottleneck, and fixes it autonomously, from detection to pull request. Customers have cut inference latency by 2.5x and improved throughput by 90%, without adding a single GPU. Israel and Joel are the right people to build this. Their eBPF profiling agent — built at Elastic, open-sourced, and donated to OpenTelemetry — runs in production today at some of the largest orgs. Happy to be supporting Zymtrace in their $8.5M seed raise led by Venture Guides alongside Mango Capital, Fly Ventures, and angels including Thomas Wolf (Hugging Face) and Christian (Chris) Bach (Netlify). Hilmar Eggertsson, Lucas Stoops, Thibault D'hondt, Christina Franzeskides, Thomas Olszewski, Maxime Vanderstichele https://lnkd.in/eRwSiqzU

52

5 Comments

NVIDIA’s Nemotron series (Nano, Super, Ultra) is engineered for agentic AI—with a clear focus on efficiency, multi-agent coordination, and real-world deployment. If you’re currently building 𝗮𝗴𝗲𝗻𝘁𝗶𝗰 𝗔𝗜 𝘀𝘆𝘀𝘁𝗲𝗺𝘀, here’s why Nemotron is worth paying attention to: ■ Hybrid architecture Combines Mamba + Transformer + MoE, delivering ~4× higher throughput and native support for 1M-token contexts—well-suited for long-sequence workloads like full codebases and complex documents. ■ Agentic & multimodal strengths Designed for tool-use, planning, video and document understanding, and real-time safety moderation—going beyond text-only interaction. ■ Training for behavior, not just answers Trained on ~10T high-quality tokens with reinforcement learning focused on reliable multi-step trajectories, which matters for agents operating over time. ■ Truly open Permissive licensing with access to weights, datasets, and tooling—enabling deep customization and commercial deployment without lock-in. ■ Hardware-aware by design Tightly optimized for the NVIDIA ecosystem (e.g., TensorRT-LLM, Blackwell), significantly reducing inference cost and latency from edge to data center. Nemotron prioritizes scalable AI systems over single-prompt brilliance, making it a strong fit for the next wave of intelligent, coordinated agents. Models: https://lnkd.in/gSuX6_Tx Paper: https://lnkd.in/gdFmjFh5 link: https://lnkd.in/g3_myMy5

6

4 Comments

👉️ https://lnkd.in/g4x2XuVX 🚀 Scaling in C++ isn’t about squeezing out one more optimization—it’s about architecting for growth from day one. Whether you’re building trading engines, AAA games, or embedded systems, this blog unpacks real-world best practices for writing C++ code that won’t collapse under scale. ✅ Modular design ✅ Smart memory ownership ✅ Concurrency without chaos ✅ Continuous measurement If you think “it works on my machine” is enough, think again. Scalability is a mindset, not a feature. 🧠 What’s the hardest part of scaling a C++ application in your experience? Drop your war stories or go-to tips 👇

2

Gave a lightning talk yesterday at Tokyo AI x OpenAI event on my Codex CLI workflow. Here are the notes in case you missed it: 𝐔𝐬𝐞 𝐬𝐩𝐞𝐞𝐜𝐡-𝐭𝐨-𝐭𝐞𝐱𝐭 𝐭𝐨 𝐠𝐢𝐯𝐞 𝐝𝐞𝐭𝐚𝐢𝐥𝐞𝐝 𝐜𝐨𝐧𝐭𝐞𝐱𝐭 𝐭𝐨 𝐜𝐨𝐝𝐢𝐧𝐠 𝐚𝐠𝐞𝐧𝐭𝐬: As a developer, I love keyboard and I can understand if you're skeptical. I was too, but using speech-to-text is one of the high-leverage things you can do as a developer. We all know that LLMs work better when given more context and clear instructions. When you are working with coding agents, you are giving instructions to them multiple times a day, every day. Over time, you end up giving much better instructions and detailed context if you use speech-to-text compared to manually typing all those instructions all the time. There are tons of open source and proprietary products for speech-to-text which offer inference on local machine or on cloud. Pick whatever you find suitable, but I definitely recommend using speech-to-text to give your instructions to the LLMs. And if you're skeptical, I'd say give this a try for a week and then you can re-evaluate. 𝐏𝐥𝐚𝐧𝐧𝐢𝐧𝐠 𝐚𝐧𝐝 𝐢𝐦𝐩𝐥𝐞𝐦𝐞𝐧𝐭𝐚𝐭𝐢𝐨𝐧: After I speak out loud things I want to do, I feed it into Codex for planning and iterate on the implementation plan until I'm satisfied. Then I use the wedow/ticket application which is a minimal lightweight version of Beads to create tickets for all the tasks. After I'm done with planning and all the tickets have been created for the task, I go to new Codex session and then pick the ticket that's not blocked by any other ones and then start the implementation. After the agent has completed the implementation, I check the code manually and iterate on it until I'm satisfied. 𝐆𝐞𝐭𝐭𝐢𝐧𝐠 𝐭𝐡𝐞 𝐜𝐡𝐚𝐧𝐠𝐞𝐬 𝐫𝐞𝐚𝐝𝐲 𝐟𝐨𝐫 𝐜𝐫𝐞𝐚𝐭𝐢𝐧𝐠 𝐚 𝐩𝐮𝐥𝐥 𝐫𝐞𝐪𝐮𝐞𝐬𝐭: After I'm satisfied with the implementation, I use an Agent Skill I have created for myself in a fresh Codex session for polishing the implementation. It checks the git diff and removes all the unnecessary comments made by LLMs. Then it runs tests and if there are any failing tests, it recursively fixes them until all of them pass. Then it runs formatting and linting recursively until issues are fixed. After that the implementation is ready for pull request. So then I go to the editor myself and then review all the code again and make small meaningful commits. And then I manually create the pull request.

54

4 Comments

🎉 PolyAPI Release 23 is live! This release is packed with updates that make building and scaling integrations even smoother. Here’s what I’m most excited about: 😲 Direct function execution in TypeScript and Python — now you can call PolyAPI functions without going through the gateway, with full mTLS support. Great for those who want the benefits of SDK generation agains their existing gateway. 🤩 UX upgrades across the management UI — and a new button in VS Code to regenerate your client libraries. I’ll miss typing npx poly generate 10 times a day… but this is way nicer. 🤑 Faster, smarter API training — we’ve optimized our OpenAPI Spec support to handle scale and are ready to increase the OOB catalog significantly. We continue to build for enterprises who care about both speed and precision. Check it out!

27

3 Comments

Founders like John Kennedy don’t treat community as a side project. They build it into the company. At Actual AI, the team believes the future of software isn’t just shaped by better models. It’s shaped by people building together. That belief is why they built Seattle’s largest dev tools meetup and why culture and trust sit at the center of how they hire. When it came time to grow, they didn’t want resumes. They wanted real conversations. At a MeeBoss Hiring Festival, one of those conversations turned into a hire. Not because of keywords, but because alignment was obvious the moment they met. That’s what happens when community isn’t a brand value. It’s how you build your team.

24

4 Comments