Hume AI Boosts 10 GB/s Upload Speed with Backblaze B2

Ever tried moving petabytes of audio data on a tight deadline? It's enough to make even the calmest engineer need a hard reboot. 🤖 The brilliant minds at Hume AI train emotionally intelligent speech models, which means their #MachineLearning pipelines are incredibly data-hungry. When their old fragmented setup started acting like a digital traffic jam, they consolidated their #CloudStorage on Backblaze B2. The result? A unified object store hitting 10 GB/s upload speed! 🚀 Now their compute clusters stay perfectly fed without skipping a beat. Plus, with $0 egress fees, they can spend their budget on cutting-edge #AI research instead of paying tolls just to move their own bytes around. Check out the full case study to see how Hume AI banished storage silos and kept their data pipelines flowing: https://lnkd.in/eMXZRy7b

Users don't interpret a long AI pause as "a distributed system doing distributed system things." They interpret it as incompetence. Unmanaged latency kills AI adoption faster than almost anything else. To build trust, we have to look past the model itself and optimize the entire traffic delivery path. 🎯 🔗 Learn how we're helping enterprises conquer AI latency: https://go.f5.net/l8a12s5z

Building AI: Lesson #1 Latency percentiles lie if you only look at the average Average latency hides your worst experiences. p50 says nothing about p99. A model that "feels fast" on average can still time out for 1 in 100 users, and at scale, that's thousands of angry sessions a day. We stopped reporting mean latency internally. Now every dashboard leads with p95/p99, split by input length. It changed which optimizations we prioritized, tail latency, not average throughput. If you're only tracking averages, you're optimizing for the wrong metric. #ArtificialIntelligence #GenerativeAI #LLM #AIInfrastructure #MLOps #Inference #AIEngineering #MachineLearning #PerformanceEngineering #Latency #Observability #CloudComputing #CTO #CIO #EngineeringLeadership

Agentic AI doesn't just use more compute. It leaves a bigger data trail behind it, and most telemetry stacks weren't built for that kind of volume. Every agent decision, tool call, and retry generates its own signal. Multiply that across a fleet of agents and the pipeline becomes the real bottleneck. Apica Flow controls what actually needs to move downstream. Link in comments. #ApicaFlow #TelemetryPipeline #Observability

What if the secret to the next generation of AI isn't just building bigger data centers, but fundamentally changing how models think—and how we train them across the globe? We are unpacking a massive drop from the AI frontier: the newly released Inkling model card from Thinking Machines Lab. Cost-Efficient AI: Why Inkling's 975B MoE Changes Everything. SWE-1.7, Condensed Chain-of-Thought. #SWE #Inkling #MoE At 975 billion parameters, Inkling isn't just another massive Mixture-of-Experts. It is a natively multimodal engine built from the ground up for a new era of cost-efficient intelligence. We’re going to look at how this beast acts as the foundation for SWE-1.7, pushing the absolute limits of agentic software engineering and long-horizon tasks. But the real story here is the engineering underneath. We are breaking down the wild logistics of their fault-tolerant, distributed reinforcement learning pipeline spanning three continents. We’ll explore the math behind their gradient norm stabilization, the use of the Muon optimizer to keep the training run alive, and a fascinating new capability called condensed chain-of-thought—where the AI literally self-compacts its own reasoning to slash compute costs. Whether you're building autonomous agents or just trying to keep up with the bleeding edge of AI architecture, strap in. Let's dive into the Inkling framework. https://lnkd.in/gr87mf8N

"It is expensive to be a poor programmer" Current frontier AI subscription tiers, widen the barriers of entry to new programmers and computational scientists. You get 20x performance by paying 5x normal subscription. You barely learn what is multi-agents/workflows/loops with the normal subscription. You get the scraps with free models.

What if the secret to the next generation of AI isn't just building bigger data centers, but fundamentally changing how models think—and how we train them across the globe? We are unpacking a massive drop from the AI frontier: the newly released Inkling model card from Thinking Machines Lab. At 975 billion parameters, Inkling isn't just another massive Mixture-of-Experts. It is a natively multimodal engine built from the ground up for a new era of cost-efficient intelligence. We’re going to look at how this beast acts as the foundation for SWE-1.7, pushing the absolute limits of agentic software engineering and long-horizon tasks. #ThinkingMachinesLab But the real story here is the engineering underneath. We are breaking down the wild logistics of their fault-tolerant, distributed reinforcement learning pipeline spanning three continents. We’ll explore the math behind their gradient norm stabilization, the use of the Muon optimizer to keep the training run alive, and a fascinating new capability called condensed chain-of-thought—where the AI literally self-compacts its own reasoning to slash compute costs. #InklingModel Whether you're building autonomous agents or just trying to keep up with the bleeding edge of AI architecture, strap in. Let's dive into the Inkling framework. https://lnkd.in/gr87mf8N

Episode 13 of my Learn AI series is live. Running an AI model for yourself is one problem. Running it for a thousand users at the same time is a completely different one. A GPU processes requests one at a time by default. If a hundred users send a message at the same second, 99 of them are waiting in line. Batching solves this by grouping multiple requests together and processing them in one GPU pass. Instead of handling requests one after another, the model handles them together. The GPU stays fully utilized and users stop waiting. Concurrency is the layer on top. It controls how many requests your system accepts, queues, and routes before they ever reach the model. Without it, a traffic spike does not just slow things down. It crashes the whole system. This is the difference between a demo that works and a product that holds up under real load. Every inference engine you use, vLLM, TGI, Ollama, has batching and concurrency settings baked in. Most people never touch them. The ones building production AI systems have to. Episode 13 breaks down how both work and why they matter when you are serving real users. https://lnkd.in/eSQ4ixhz #AI #AIInfrastructure #LLM #MachineLearning #LearnAI #MLOps #DataEngineering #LocalAI #GenerativeAI #vLLM

𝗔𝗜'𝘀 𝗘𝗻𝗲𝗿𝗴𝘆 𝗣𝗿𝗼𝗯𝗹𝗲𝗺 𝗗𝗼𝗲𝘀𝗻'𝘁 𝗦𝘁𝗮𝗿𝘁 𝗪𝗵𝗲𝗿𝗲 𝗬𝗼𝘂 𝗧𝗵𝗶𝗻𝗸 It’s been a while since I posted anything technical here. I've been heads down on something for the past few months and I think it's finally ready to talk about. So here's where it started. Every time AI energy comes up in conversation, it goes the same way. Someone mentions GPUs. Someone mentions how much it costs to train a model. Someone throws out a number about parameters. That's where most of the conversation tends to stop. But there's a whole part of the AI pipeline that almost nobody brings up. Before a model ever touches your data, that data already went through a lot. It gets loaded, cleaned, joined, filtered, transformed, aggregated, sometimes at a scale that's hard to picture. The model isn't really where the computation starts. It's just where people start paying attention. That gap is what got me curious. What's actually happening underneath the model, in the software choices nobody questions. So I started measuring it. I'll be sharing what I found over the next few weeks, bit by bit. Stay with me on this one. It's about the layer nobody's really optimizing. If we actually care about sustainable AI, should we only be looking at the model? Or the whole pipeline feeding it? #AI #GreenAI #SoftwareEngineering #DataEngineering #SustainableAI

Independent industry coverage continues to highlight Penguin Solutions’ role in the next phase of AI infrastructure. As production-scale inference and agentic AI accelerate, token economics and AI factory efficiency are becoming increasingly critical. A recent Yahoo Finance article features CEO Kash Shaikh discussing how persistent, context-rich AI inference workloads are reshaping infrastructure requirements and how Penguin’s AI Factory Platform helps customers build and operate AI factories at scale with best token economics and measurable ROI. Read more: https://lnkd.in/gkd2tr5u