Mikhail Rudenko

We just released Nemotron Speech ASR: 🤖0.6B streaming cache-aware transducer 📉low latency (down to 80ms) 📈high throughput (up to 900 concurrent streams on H100) 🎮adjustable latency-throughput-accuracy trade-off without re-training 🌎English ASR 🔗https://lnkd.in/eFeEQR2F

538

20 Comments

The critical trade-off between throughput vs. sample off-policyness has always been central in RL, from robotics to video games, I have seen it myself multiple times as a key design driver. But at the current scale of large foundation models, the centrality of its role in RL-based post-training efficiency has become even more evident. The high level architecture in the picture describes the custom system adopted by MiniMax in training their M2.5 model released a few days ago (link in the comments). They claim that this element alone provided an approximate 40x(!) training speedup. There are so many variables at play in these applications, and so much space for improvement, I cannot think of a more exciting space to work in. 

27

1 Comment

🚘 100 million miles. Zero drivers. Waymo just crossed a milestone that would take a human 7,000 years to match. While much of the world debates AI’s risks in theory, Waymo is racking up real-world progress. It has doubled its autonomous mileage in just six months, expanded to cities like Atlanta and Austin, and is now delivering more than 250,000 rides per week without a driver. Not test rides. Not demos. Actual paid rides. In real cities. With real people. 🛣️ What’s the headline here? • 100 million fully autonomous miles now logged • Over 10 million trips delivered with no one behind the wheel • New city launches coming in Miami, D.C., and Tokyo • First snowy city deployment planned for Washington D.C. • Live competition from Tesla, Zoox, Nuro, and policymakers 💥 This isn’t hype. It’s rollout. Cruise is out. Tesla still uses safety monitors. Waymo is quietly becoming the U.S. leader in driverless mobility. 🧠 The strategic implications go beyond autonomous vehicles: • AI agents are entering the physical world • Trust is being built ride by ride, not just with performance benchmarks • Urban AI infrastructure is being shaped through real deployment, not just policy 🔮 Here’s my take: The companies that move beyond pilots and focus on scale with responsibility are writing the real playbook for AI in the real world. 💬 What’s yours? Will autonomy look more like Waymo or Tesla? And when will you be ready to trust a ride with no one in the front seat? #AI #AutonomousVehicles #Waymo #Tesla #Robotaxi #FutureOfMobility #AVs #AIadoption #LeadershipInTheLoop

5

Finally finished going through the much-discussed paper “AlphaGo Moment for Model Architecture Discovery.” I’m thoroughly impressed, but not fully converted yet. What’s great: automated loops that ideate → implement → evaluate at scale are real productivity multipliers, and they aren’t weighed down by “we’ve always done it this way” thinking. Where I push back: building AI architectures isn’t the same as winning AlphaGo. Unlike AlphaGo, AI architectures may not always have a perfect, rule-based verifier. Move 37 was one of the finite possible moves (however big that finite number maybe doesn't matter at scale). Real problems on other hand are multi-objective, noisy and not within a finite set of possibilities. Brute force will happily optimize for the benchmark, even if that doesn’t translate to real-world value. The biggest leverage still lies in problem selection, framing, and evaluation design, which are deeply human decisions. If research were purely compute-bound, the only winner would be Nvidia, and my “research roadmap” would just be new purchase orders for GB200s. :) Humans set the questions, shape the rewards, design the guardrails, and extract the theory. At least for now, this automation is a force multiplier for that work and not a substitute. #AI https://lnkd.in/gZf8ZyEf #ArtificialIntelligence

11

1 Comment