Activity
15K followers
-
Peter DeSantis shared thisFun weekend at Berkeley RDI’s Agentic AI Summit. Talked about how constraints drive innovation – the hard tradeoffs in silicon, memory, software and systems that push you toward more creative architecture. Especially in silicon, where many design decisions are zero sum and prioritizing generalizability leaves gains on the table for more specific workloads. Enjoyed the panel afterward too. We came at it from very different angles and I learned something from everyone up there. So many cool things going on. A fun time to be in technology…it’s really Day 1 in AI!Peter DeSantis shared thisAI's full potential is bottlenecked by an efficiency problem that spans the entire stack. At Berkeley RDI's Agentic AI Summit, Amazon SVP Peter DeSantis traced how the predictable memory and compute flow of AI models led Amazon to build Trainium on a systolic array architecture, stripping out flexibility those workloads don't need while preserving what matters. But no single chip will power the next decade. As AI workloads keep evolving, Peter sees a growing and more diverse hardware ecosystem ahead. Annapurna Labs' John Liu followed with a deep dive on using agentic looping to optimize models on Trainium. Key takeaways: test agents on data they've never seen, verify visibility scope before fixing multi-agent failures, and recognize when existing rules are causing the very failures you're trying to patch.
-
Peter DeSantis shared thisWith all the buzz around quantum computing lately, this New Scientist conversation with Peter Shor is worth a read. Breaking encryption attracts most of the headlines, but Shor himself (who invented the very algorithm that could theoretically break encryption) isn't very worried. A quantum computer capable of running his algorithm would have to be very, very large, and it’s unlikely to exist in the near term. And we already have good methods for post-quantum cryptography. The hard part is implementing them. That's why Amazon has been investing in quantum-safe encryption algorithms for the better part of a decade. The migration is long and complex, but it's cheap insurance relative to the alternative. The more interesting question is what quantum computers WILL be good at. Shor believes the list will be fairly narrow. A quantum computer isn't going to be a faster version of the computer on your desk. My mental model goes back to Richard Feynman, who said that if you want to simulate the quantum world, you'd better build a quantum machine. That's what these systems will excel at — simulating the chemistry and physics that classical computers can’t. Even with all the computing power available today, we can't accurately simulate important molecules in science and engineering. Take nitrogen fixation. The industrial processes we use to fix nitrogen for fertilizer emit enormous amounts of CO2. Nature does it far more efficiently, but the molecule at the heart of that process (FeMoCo) is too complex to fully simulate classically. A moderately sized quantum computer could simulate that chemistry far better and help us find cleaner processes. The same goes for materials science, superconductors, and drug discovery. While the list of useful applications may be narrow, the impact will be enormous. And the list may be longer than we think. Shor mentions optimization algorithms, which he believes have been dismissed too quickly. We don't yet know whether quantum computers will meaningfully accelerate optimization problems, but if Peter Shor thinks it's worth a second look, it's probably worth a second look. So when people ask me whether quantum computers will live up to the hype, my answer is that it depends on what you're hoping for. If you're waiting for a machine that does everything faster, you'll likely be disappointed. But if you care about some of the hardest, most important problems in chemistry and physics, there's a lot to be excited about. You can count me amongst the excited! https://lnkd.in/gBCv55WdPeter Shor’s algorithm could break the internet – but he's not worried | New ScientistPeter Shor’s algorithm could break the internet – but he's not worried | New Scientist
-
Peter DeSantis posted thisOne of the things I find myself saying a lot lately — to our teams, to customers, to anyone who'll listen — is that we're really at the beginning of this. Not the middle. The beginning. The pace of progress can make it feel like the big architectural questions are settled, and that what's left is scaling what we already have. But I think that confuses scaling the current paradigm with scaling AI itself. And when you look at the full stack — really look at it — you realize the current paradigm is still being invented at every layer. Take the models themselves. Transformers have been remarkably durable, and there's still a tremendous amount of innovation happening within them. But they're only part of the story. There's a lot of work happening across the industry on world models, real-time multimodal video, and entirely new approaches to how models learn and reason. That's new science. And the workloads of a few years from now will look quite different from the workloads of today, which means we're going to need new approaches. The chip layer is just as interesting. For a while, the right answer was to build big, general-purpose AI platforms, because the workloads were too uncertain to specialize around. That's still true for a lot of what we do. But we're starting to understand the shape of these workloads well enough that thoughtful specialization is becoming possible. How memory and compute get packaged together, how we handle the very different demands of prefill versus token generation, how we make long context efficient. There's enormous room to innovate here. Some of the best engineering problems I've seen in my career live in this space right now. And underneath all of this is efficiency. The cost of intelligence has come down by about an order of magnitude in the last couple of years. I think it needs to come down by a few more before AI becomes what we all believe it can be. What I find most interesting is that lowering cost isn't just about making the same capability cheaper. It's what unlocks the next set of capabilities. Drive down the cost of inference, and you can afford far more reinforcement learning, which makes models meaningfully better, which drives more demand, which justifies driving cost down again. It's a genuine flywheel, and one we're still early in spinning up. That kind of progress doesn't come from any single layer. Models, chips, systems, and software advance together, or they advance slowly. That's something you learn pretty quickly looking at the full stack. It's also a big part of why I remain deeply optimistic about what's ahead. The most interesting work in this space isn't behind us. It's in front of us. And that's what makes this such a great time to be building.
-
Peter DeSantis shared thisAfter 27 years at Amazon, I figured it was time to end my quiet period on LinkedIn. Reading Andy's shareholder letter was all the nudge I needed. I never thought I’d be at the same company for 27 years… it just sneaks up on you when you’re having fun building things with smart people. One thing I love most about Amazon is our willingness to think big and be patient. When I joined, we were a bookstore with ambitions to deliver everything to everyone. That vision consumed my first decade. Then I helped launch EC2 – the start of a 20-year journey building AWS services and expanding our global infrastructure. A little over a decade ago, I started working with the Annapurna team on our chips business. When Andy wrote about it last week, I stopped mid-read – seeing it laid out so clearly reminded me of how far the team has come. Building our own chips wasn't an obvious call. There was real skepticism, and good reasons to avoid the complexity. But we believed that controlling things down to the silicon meant we could optimize AWS in ways that truly served our customers. That bet paid off. Every EC2 instance runs on our Nitro chip. Graviton is used by 98% of our top 1,000 EC2 customers. Trainium2 is essentially sold out. Trainium3 just started shipping and is nearly fully subscribed. None of this happened overnight, and none of it happened without a team willing to be patient and a little misunderstood along the way. I've watched this cycle play out often at Amazon. Bets that looked wrong, then right, then inevitable. It’s made possible by a culture that gives long-term thinking room to breathe while maintaining the urgency to move forward. Three months ago, I stepped into a new role bringing together our foundational AI models, custom silicon, and quantum computing. People often ask why bring these together, and why now. The question was never if but when. When model scientists know where infrastructure is heading, they explore techniques that wouldn’t otherwise be viable. When chip designers know where model architectures are going, they optimize differently. When you build models on your own chips, you train at greater scale and experiment faster. And as quantum matures, it will benefit from everything we've learned scaling the other two – especially our hard-won expertise designing some of the world's most complex chips. The industry is moving fast and there’s a lot of noise. But what I keep coming back to is how early we are in this massive transformation. Optimal AI infrastructure is still being invented by us and others. And today's models are genuinely amazing, but in a few years, we'll look back at them like dial-up internet... slow and expensive. We're working on problems that don't have obvious answers yet – and some of what we're doing will look strange for a while. At Amazon, that's usually a good sign. Read Andy's letter here: https://lnkd.in/gMFAYKHT
-
Peter DeSantis liked thisPeter DeSantis liked thisHad a great time speaking as well as meeting industry practitioners, researchers, and VCs at University of California, Berkeley's RDI Agentic AI Summit 2026. Peter DeSantis kicked off the session with his keynote on custom chips unlocking model efficiency and innovation from co-designing chips with model providers. Really set the tone for calibre of speakers and insights for the 2-day event. I shared an overview of using looping to optimize models on custom silicon (AWS Trainium) and five insights for all people using loops in production (link to video in comments). Almost every session / panel mentioned the OpenAI x Hugging Face agentic hack and agentic defenses. I just happened to have as my first insight to "design every part of your agentic loop as if it will be manipulated by the agent, especially the benchmarking / evaluation component". Until next year!
-
Peter DeSantis liked thisPeter DeSantis liked thisAgents don't break your model. They break your infrastructure. That was the through-line of the panel I moderated at the UC Berkeley RDI Agentic AI Summit — with Peter DeSantis (AWS), Saurabh Tiwary (Google Cloud), Jonathan Cohen (NVIDIA), and Chuan Li (Lambda). We built this stack for humans: short, bursty, request-response. Agents run at machine speed, for hours, spawning sub-agents and burning tokens overnight. They break things in new places. Chuan Li said it best: it's like routing machine-speed traffic onto roads built for human drivers. Capacity, traffic lights, and speed limits fail first. The whole system has to be rethought. And we're barely at the starting line. More in the comments 👇 #AgenticAI #AIInfrastructure #VentureCapital
-
Peter DeSantis liked thisPeter DeSantis liked thisAI's full potential is bottlenecked by an efficiency problem that spans the entire stack. At Berkeley RDI's Agentic AI Summit, Amazon SVP Peter DeSantis traced how the predictable memory and compute flow of AI models led Amazon to build Trainium on a systolic array architecture, stripping out flexibility those workloads don't need while preserving what matters. But no single chip will power the next decade. As AI workloads keep evolving, Peter sees a growing and more diverse hardware ecosystem ahead. Annapurna Labs' John Liu followed with a deep dive on using agentic looping to optimize models on Trainium. Key takeaways: test agents on data they've never seen, verify visibility scope before fixing multi-agent failures, and recognize when existing rules are causing the very failures you're trying to patch.
-
Peter DeSantis liked thisPeter DeSantis liked thisPersonal update: I’ve decided to join Amazon’s Annapurna Labs. After almost 20 years at Google, I look back with infinite gratitude at the projects we built, the lessons we learned, and the many brilliant folks I’ve had the privilege to work with. From the earliest days of the TPU program, we set out to do something hard and meaningful. I’ll always be proud of what we accomplished together. At the same time, I could not be more excited about this new challenge. New teams, new customers, new hardware, new challenges, and new ways to help. Annapurna’s hardware roadmap combined with Amazon’s reach is an incredible opportunity, and I can’t wait to be part of what’s next for Trainium.
Recommendations received
1 person has recommended Peter
Join now to viewOther similar profiles
Explore more posts
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top content