Prashanth Ramdas

Super excited to introduce Amazon Nova Sonic, our new foundation model for building real-time voice applications, now available in Amazon Bedrock. Nova Sonic generates speech in a fundamentally different way by handling voice inputs and response generation through a single model. This means it can maintain every nuance of a conversation—context, tone, pauses, and even interruptions. Plus, it's built to support real-world applications and agentic workflows that let it interact with customer data and APIs to get things done. We look forward to seeing all the incredible things you'll build with Nova Sonic! https://lnkd.in/gRYmRr8R

4,419

124 Comments

Interested in how we test our systems at Amazon? In our new article in Communications of the ACM, Ankush Desai and I write about the evolution of systems correctness and testing practices at AWS. We cover a mix of classic approaches, formal approaches, and new techniques like deterministic simulation. https://lnkd.in/gwdncKM4

401

10 Comments

We released an improved speech-to-speech model and new API capabilities to help build production grade voice agents. Highlights: - Realtime API is now GA - gpt‑realtime: our most advanced speech-to-speech model is more natural, expressive, and better at following complex instructions (even mid-sentence language switches) and tool calling - New integrations: support for remote MCP servers, image input, plus phone calling via SIP - Two fresh voices: Cedar and Marin - 20% price reduction https://lnkd.in/gnzHiNjd

866

23 Comments

Organizations need AI models that deliver both performance and cost-efficiency at scale. Today, we’re adding Claude Haiku 4.5 from Anthropic to Amazon Bedrock, giving customers another powerful option that delivers near-frontier intelligence at low cost and fast speeds. Claude Haiku 4.5 matches Claude Sonnet 4's capabilities on coding, computer use, and agent tasks. It’s perfect for scaling AI across your organization. This means you can now more affordably deploy intelligent agents in customer-facing applications, and in free product offerings without compromising on quality. With Opus 4.1 for the most advanced intelligence, Sonnet 4.5 for advanced coding and agentic tasks, and now Haiku 4.5 for efficient scaling, AWS customers have a wide variety of Claude models to choose from, all with the security, compliance, and capabilities of Amazon Bedrock. https://lnkd.in/gPgrF8jV

1,424

40 Comments

NVIDIA just hired WSJ's tech editor as their new managing editor. RH's Editor in Chief is the former EIC at Architectural Digest Robinhood's Editor in Chief is the former EIC at The Outline and BDG The Trade Desk's Editor in Chief is the former EIC at AdWeek Jerry's Lead Data Journalist is former Bloomberg/Mother Jones Data Reporter Mastercard's VP of Editorial Content is former cnet/wsj reporter CARFAX's EIC is the former editor at Consumer Reports IBM's Editorial Lead is a former AFP reporter BlackRock's VP Content is former editor @ Fortune SoFi's Managing Editor is former cnet managing editor Resy/Amex's Managing Editor is former SF Chronicle Wine Editor Redfin's Chief Economist came over from Boston Fed ... and so many more These companies are not replacing journalists with AI. They are hiring great talent and putting them to work on great coverage. A real reason to be excited about the reporting coming out of these orgs you would not traditionally think of as media companies. And a reason to be excited about job prospects for talented editorial folks. Slowly curating the best of the best of the output from this effort - if your team is following suit, or you're looking for roles in this space... say hi//follow along. Also - a follow up conversation here with examples of the work these orgs are doing (and tradeoffs they face) - https://lnkd.in/egkmumhM

1,557

139 Comments

Last week in PyTorch Conference 2025, we announced we are open sourcing torchcomms and NCCLX/CTran. Today, we share more details about NCCLX/CTran design and how we used them in production GenAI training and inference. Check out our white paper on this topic - https://lnkd.in/gySuXi6Y Some features we covered in this paper - * Host-driven collectives * Zero-copy data transfer * CTran/Network co-design (DQPLB) * Zero-copy and SM-free Send/Receive for PP * RMA Put for TP * Fault tolerant AllReduce * GPU-resident collectives for EP * Low-latency optimization * Scalable Initialization in Training * GPU Memory Management for comms * Fault localization and Performance Observability * CPU emulation This paper covers years of innovation and production experience from GPU communication teams at Meta, in supporting generations of LLAMA models training and inference. Hope you enjoy reading it!

313

9 Comments

Building on top of large language models is fun, but getting consistent performance and reliability is extremely challenging. I love this post from Kimberly Patron about how Sierra uses health and performance-based traffic routing and request hedging to get meaningfully better performance and tail latency from high scale LLMs https://lnkd.in/g4qWRCb5

284

13 Comments

Excited to share my conversation with Marc Baselga and Ben Erez on the Supra Insider podcast. We went deep on one of the biggest shifts happening in product leadership right now: why many experienced PMs and product execs are returning to individual contributor (IC) roles in the AI era, and why that’s not a step back, but a strategic leap forward. Other topics we discussed: -- Why the leverage equation has changed—and how managing AI agents now rivals managing people -- The critical importance of hands-on credibility in AI-native companies -- Why great PMs in the future will blend product, design, analytics, and frontend execution -- What it takes to deliver real customer outcomes (not just launch features) -- How to evaluate breakout companies—and what founders often get wrong when hiring product leaders I also shared my thoughts on: -- Joining the right company: mission, trajectory, talent density, and founder courage -- Crushing it as a “super IC”: the power of speed, ownership, and infectious energy -- Transitioning from product leader to GM or founder with the right mindset If you’re navigating your next chapter in product (or just trying to stay sharp in the AI-native world), I hope you’ll find this useful. Episode links: Spotify: https://lnkd.in/gK48PFWK Apple: https://lnkd.in/gmT265BY YouTube: https://lnkd.in/gtr4RAKd Substack: https://lnkd.in/gM38FSWc

96

10 Comments

I met one AI development agency founder, and they said our 50% RAG apps failed in production last year. Because teams don’t know how to handle: chunking + embeddings + vector DB + reranking + hallucinations + evaluation + monitoring. So I'm sharing a RAG Playbook that shows exactly how to build production-ready RAG systems, with practical examples. ✅ What you’ll learn from this playbook This playbook helps you go from: “RAG demo is working” → “RAG system is reliable in production” You’ll learn how to: - reduce hallucinations using prompting techniques - use advanced chunking methods (not random splits) - select the right embedding model for your use case - choose the best vector database (based on real needs) - add reranking to improve relevance - follow clear steps to build an enterprise-grade RAG system - run evaluation before production using key scenarios - monitor and optimize performance after deployment - track RAG quality using 4 powerful metrics If you’re building a RAG-based feature for your product, this playbook will save you weeks of trial and error. ✅ Save this post (you’ll need it later) 🔁 Repost for your team/network ➕ Follow Shubham Vora to learn more on AI agents.

139

77 Comments

This Scale AI story might matter more than you think. Meta just invested $14.3B to take a 49% stake in Scale AI — and brought Scale’s CEO into its internal “superintelligence” unit. On the surface, it looks like a major bet on data. But look a little deeper, and it’s a shift in how control is consolidating across the AI value chain. Why it matters: – Scale was one of the few trusted, neutral data providers for labeling and evaluation – Now that neutrality is in question, contracts are reportedly getting pulled by OpenAI and Google – Startups and enterprises alike are suddenly reconsidering where their training signal comes from What comes next: – A surge of boutique data providers — decentralized, domain-specific, and more transparent – Startups reframing data control as part of their moat: “We don’t just use models — we own the context” – Infra teams re-evaluating dependencies — either proactively, or reactively when vendor dynamics shift As training data usage becomes more regulated and controlled — moving away from the wild west of early foundation model training — the pressure to secure differentiated data pipelines only increases. When we talk about defensibility in AI, it’s tempting to focus on model IP or feature differentiation. But data leverage — how it’s sourced, labeled, evaluated, and governed — is becoming a more important part of the strategic equation. It’s not quite infrastructure. But it’s no longer a commodity either. It’s a layer of increasing consequence — and one worth watching closely. Are you rethinking your data leverage yet? https://lnkd.in/gx9GPefp

11

1 Comment

Zibra AI is looking for partners to build the next generation of cloud rendering infrastructure. Have you ever tried using cloud rendering infrastructure for VFX production? In theory, it offers compelling advantages: on-demand compute and storage scaling, centralized resource management, and high availability of processed data. In practice, however, it remains impractical for many high-end production pipelines. The main technical barrier? Volumetric VFX file sizes. No one wants to wait hours for a terabyte-sized VDB project to download from the cloud—or be hit with a massive data transfer bill at the end of the month. Even today, we still hear about teams physically transferring data between locations using hard drives. But what if your volumetric data took up only 5% of its original size—without compromising quality? With that kind of efficiency, you can simulate, render, and distribute content seamlessly across any location. This is now possible thanks to Zibra AI’s compression technology: ZibraVDB. It replaces OpenVDB as the new standard for volumetric visual effects, offering state-of-the-art compression rates, fast real-time decompression, and real-time rendering. ZibraVDB supports all the functionality required by render farms and integrates smoothly with major VFX production tools. Adding it to your pipeline is as simple as inserting a caching node in your Houdini project. If data storage and transfer have been holding you back from building a cloud rendering infrastructure, send me a direct message. With ZibraVDB, cloud rendering finally becomes practical.

27