Ian Ferreira

Voice AI is quietly hitting an inflection point 🔊 NVIDIA just published a great deep dive on scaling speech recognition for real-world voice agents - latency, streaming accuracy, robustness, and why “ASR quality” isn’t just about WER anymore. 👉 https://lnkd.in/enDUKHvG A few things that stood out: - Streaming + low latency is non-negotiable if you want natural conversations - ASR needs to handle interruptions, background noise, accents, and partial utterances, not clean demo audio - Scaling voice agents isn’t just model quality, it’s systems engineering If you want to actually try building or running voice agents instead of just reading about them, Pipecat is a great place to start: 👉 https://pipecat.ai This is the kind of progress that makes voice agents finally feel… practical. Not demos. Not prototypes. Production. Curious to see where the ecosystem lands this year 👀

149

10 Comments

🧠 How do LLMs "remember"? In this Medium post, I break down Episodic, Procedural, and Semantic memory in LLMs—with real-world analogies and practical examples. A quick, clear read for anyone curious about how AI thinks. 👉 [https://lnkd.in/gpfBxdMy] #AI #LLM #Memory #MachineLearning #ArtificialIntelligence

28

3 Comments

Just read the MIT Technology Review piece on Google releasing energy data for Gemini. It’s interesting to finally see concrete numbers behind what an AI prompt actually consumes, rather than broad estimates or assumptions. As AI use scales, this kind of transparency feels important for understanding the real infrastructure and sustainability tradeoffs behind everyday interactions. Worth a quick read if you’re thinking about where large-scale AI is headed. #AI #AISustainability #AIInfrastructure

13

Lately I’ve been jumping between a few AI and data-intelligence projects at Snowflake, and one thing keeps hitting me over and over again: the hardest part of AI has very little to do with the model itself. Models are cool. They look great in slides and polished demos. But the second you plug them into real workflows, real products, and real customers… things get interesting really fast. Here are a few things I keep running into: Data looks clean until the moment you actually rely on it Every team has its own idea of how something “should work” A perfect solution on paper turns into something completely different in production Integrations sound easy, right up until you start doing them for real And my personal favorite: “We’ll fix that later,” which usually translates to “we’ll rebuild this twice” What’s funny is that these challenges barely change from company to company. Whether it was Snowflake, ServiceNow, Cisco, or even earlier in my research days, the pattern is pretty much identical: AI only becomes useful when teams communicate well and work through the messy parts together. It’s not glamorous. It’s not magic. It’s just alignment, transparency, and adjusting when reality doesn’t match the plan. And honestly, that’s what I like about this field. The tech is fun, but the real progress comes from people actually building things together. Just wanted to jot this down before I forget. Curious what others are seeing in your teams lately. For you, what’s been harder: the model or the people part? By the way, here’s a video that showed up in my feed and fits this topic pretty well 👇 https://lnkd.in/g3wTuzb6

2

We’re bringing customers enhanced model choice today with the addition of Claude Opus 4.5 in Amazon Bedrock https://lnkd.in/g9CdMFWj Claude Opus 4.5 is Anthropic’s most advanced model and it’s excellent for building AI agents that handle complex, long-running tasks across thousands of workflow steps. I think builders will really enjoy how the model handles sophisticated coding—turning hours of work into minutes. Claude Opus 4.5 is already proving its value to devs, with a new state-of-the-art 80.9% on SWE-bench. The model will also be available today in Kiro. This new model joins the existing hybrid reasoning models, Claude Sonnet 4.5 and Claude Haiku 4.5, in Amazon Bedrock, giving customers a range of models that toggle between near-instant responses and extended thinking. With Amazon Bedrock's enterprise-grade security and responsible AI controls, the future of agentic AI is here.

1,034

19 Comments

Integrating large language models and powerful algorithms into your workflows is a massive undertaking. All that hard work can be undermined if the outputs fall short, quickly eroding stakeholder interest and trust. That's why effectively evaluating LLMs and algorithms is crucial. From initial prototypes to production, and through continuous experimentation and benchmarking, mastering evaluation ensures these advanced tools truly deliver on their promise. #LLM #Algorithms #AI

8

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

AI Agents: The Definitive Guide to Agentic Systems and How to Build Them for Production https://bit.ly/3NvNUjs #AI #MachineLearning #DeepLearning #LLMs #DataScience

6

A.I. Clouds are not all the same. Justin Gage of Amplify Partners shares, in explicit detail, how Modal is just built different, and why it's a mistake to think of them as an "inference provider". https://lnkd.in/gmaYQZtT

55

LLMs have been trained on hundreds of millions of documents, yet we evaluate these models using benchmarks containing only 1K to 650K Q&A pairs. Even the largest benchmark, TriviaQA (2.5 GB), represents less than 1% of the total data used to train a large LLM—for example, Gemini, which was trained on 750 GB of data. This is comparable to deploying software with less than 1% tested. So why are we surprised when things go wrong?

29

3 Comments

Current State v. Future State. Business system MCPs plugging-in to an orchestration layer— all powered by the electricity of AI 💡 AI is the apex creature 🦈 (global knowledge, smarter than us) and everything else is critical lower apex (proprietary knowledge unlocked from silos)🐟 🐙 🐛. They all feed user natural language interfaces—or better yet—automation. One of my engineering heroes (Jeff “JJ” Johnson) used to tell me, “Technology is best when you don't know it’s there.”

11

2 Comments

Customers are asking for AI that goes beyond assistance to true collaboration. Claude Opus 4.5 from Anthropic delivers on that vision: understanding objectives, factoring in constraints, and ready to power complex workflows. Opus 4.5 joins the full set of Claude models now in Microsoft Foundry, alongside models from OpenAI, Mistral AI, Cohere, Meta, and others, so you can easily compare and find the best-fit models for your needs. With Microsoft Foundry, Azure customers get the broadest selection of model providers on any cloud, plus the governance, security, and interoperability to accelerate innovation with confidence. 🔗 https://lnkd.in/gdRxNB_q

107

3 Comments

The future of AI scaling will prioritize efficient, low-cost inference. As of late 2025, the industry has shifted decisively into an inference dominant era, where inference revenue has surpassed training revenue for the first time, and ongoing operational costs now dwarf one-time training expenses. This transition is driven by exploding real world usage, billions of daily queries, real time agents, and edge deployments making inference the primary bottleneck for scalability, energy efficiency, and profitability. Training remains essential for frontier models but is episodic and increasingly commoditized, with costs dropping dramatically (e.g., GPT-4-level models trained for under $6 million). Specialized inference hardware is emerging as the key enabler. Google’s TPUs (now in their sixth-generation Trillium, with strong efficiency gains for large-scale cloud inference) continue to power hyperscalers’ internal workloads. Meanwhile, Groq’s purpose built LPUs excelled in low latency, high throughput performance at lower cost until Nvidia’s $20 billion licensing deal and talent acquisition in December 2025 effectively integrated Groq’s technology into Nvidia’s ecosystem, consolidating leadership in hybrid GPU/LPU architectures. Ultimately, winners in the next phase will optimize for tokens-per-dollar and milliseconds of latency, enabling widespread adoption of advanced AI agents and real-time applications. Nvidia now positioned to dominate both training and inference, while custom ASICs like TPUs remain vital for hyperscale efficiency. https://lnkd.in/gV6KrTRy

30

1 Comment

The biggest AI performance unlocks are coming from algorithm-system co-design. DeepSeek's recent papers focus on closing the gap between open and closed LLMs—not by throwing more compute at the problem, but by designing smarter architectures around hardware limits. 🧠 Engram Think of it as a separate 'filing cabinet' for facts. ✅ Stores facts and named entities in a static lookup table ✅ Cheaper than running everything through attention ✅ Frees up the model to focus on reasoning ❌ Problem: Adds memory pressure, slows inference ✔️ Solution: Pre-fetch from cheaper memory before it's needed Result: Better reasoning, minimal latency hit. 🧠 mHC Expands the 'highways' that move information between model layers ✅ Better signal flow = better performance ✅ No increase in model size ✅ Training stays stable ❌ Problem: More GPU-to-GPU comms needed ✔️ Solution: Smarter GPU scheduling Result: More capability, same footprint. A lot of AI progress assumes unlimited hardware access. DeepSeek is optimizing for constraints—memory bandwidth, interconnect limits, etc. Will it work in production-grade environments? DeepSeek V4 will be the real test. More here: https://lnkd.in/d4raD2GR

15

Yesterday, our sister company Canal+ announced multi-year agreements with Google Cloud and OpenAI to incorporate generative artificial intelligence into its production workflows and streaming service. This collaboration aims to enhance content creation and improve user experience through advanced AI tech. We're just getting started. https://lnkd.in/dF35xJvj

25

1 Comment

Over the holidays, I read this paper by DeepSeek called Manifold-Constrained Hyper-Connections, focused on improving how information moves inside AI models. It turns out that AI models have been using narrow “express lanes” to pass information between their internal layers but the paper talks about turning that single lane into a multi-lane highway. Think of training a model like mixing a song: - Each track (layer) adds an instrument, a vocal, or an effect. - Residual connections are like keeping a reference track Some approaches let every track influence all others, like adding every instrument to every channel. This can make richer music, but the mix can also become distorted. Manifold-Constrained Hyper-Connections (mHC) is like a mastering engineer applying rules to keep the mix balanced: layers influence each other, but no part dominates. The result is an LLM that is capable of handling complex details, but costs a lot less to build. https://lnkd.in/gxsZvZcy

17

1 Comment

In the recent Quarterly report Google signals another year of elevated AI CapEx. https://lnkd.in/e--b3nhJ But there are two universes playing out. Universe #1: Narrative • AI demand is strong and expanding across Cloud. • Infrastructure investment is framed as a growth enabler. Universe #2: Constraint • Capital has already been poured in aggressively, ahead of proven, repeatable ROI. • Depreciation and fixed costs are real now; unit economics are still forming. Same moment. Same company / market. Two very different signals. Insight: AI infrastructure is being financed like a utility, but monetized like speculative software. The real bottleneck isn’t demand or belief—it’s the absence of decision-grade ROI math that can sustain capital intensity under scrutiny. That gap explains more than the headline does. --------------------------------------------------------------------------------- Ayesha Saeed - Haq is an enterprise AI, Analytics, and Decision Intelligence leader with Fortune-20 experience building production decision systems that are designed to scale, governed to survive scrutiny, and measured in real business outcomes. #EnterpriseAI #AIEconomics #CloudInfrastructure #CapitalAllocation #DecisionIntelligence