Ahmed Aly

I didn’t stay and grow at Amazon for 5 years just because I could write good code. Plenty of smart engineers can do that every single day. Writing clean code was just the entry ticket. What truly set me apart was taking ownership, even when it wasn’t asked. If I was assigned a bug to fix, I could’ve stopped after the code review. But I would follow it to production, sit with support teams, talk to impacted customers, and make sure the bug was truly gone, and people felt the difference. If I was asked to build a new API, I could’ve stopped after passing the unit tests. But I’d stress-test it with real traffic, set up alerting, write docs, and hop on calls with downstream teams to make sure their integration was smooth. If I noticed on-call alerts going off every night, I could’ve shrugged and marked them as “infra issues.” But I dug into root causes, ran post-mortems, and sometimes pulled late nights re-architecting fragile parts, so the next person on call wouldn’t have to lose sleep. If I wrapped up my sprint tasks ahead of schedule, I could’ve clocked out early. But I’d sit with teammates struggling with blockers, offer code reviews, and even pick up “thankless” chores like improving internal tools or onboarding docs. Ownership isn’t about heroics or working 14-hour days. It’s about caring enough to see things through, to fix the root, not just the symptoms. To ask, “What else could go wrong?” and, “How can I make this easier for the next person?” That’s how you get noticed. That’s how you build trust. That’s how you grow. Don’t just be the engineer who writes code. Be the one who owns the outcome, end to end. That’s what really sets you apart.

297

10 Comments

The biggest difference between Google and Waymo, almost one year in: Waymo has largely adopted the Amazonian practice of presenting information in documents, not slides. Meetings start with a few minutes of pre-read / frantic commenting on a Google Doc, followed by a live burn-down of the comments that require discussion. It's efficient and information-dense. I love it.

2,183

63 Comments

My fellow Googlers recently have published a white paper on AI Agents. Many believe that AI Agent is the application layer of LLMs, and it seems like we are stepping closer to having useful and reliable AI Agents. This is a recommended must-read.

93

6 Comments

Update: Google is going back to in-person onsites for SWE roles in the next 2-3 months. I published a piece a few weeks ago where I summarized AI's effect on eng interview processes after surveying ~50 FAANG interviewers on interviewing.io, and in that piece I said there was no indication that any of the FAANGs were going back to in-person. (Piece linked in the comments.) Then I got this update from one of our interviewers last week. Things change fast. I'll share them as I learn more!

451

45 Comments

Updates from #llamacon2025 Last week, we open-sourced two new tools to help address major pain points developers face when building with Llama. llama-prompt-ops A tool to optimize or migrate prompts from other LLMs to Llama — systematically and at scale. https://lnkd.in/g-D9vwaB synthetic-data-kit A modular toolkit to generate high-quality post-training data (Q&A pairs, reasoning samples, DPO datasets, and more) to fine-tune Llama models and help close the gap with closed-source models. https://lnkd.in/g2CdKh4i Avialble on #unsloth now -- https://lnkd.in/gGZ-RDUM We’re just getting started. Let’s build together — contribute, fork, test, and let us know what features you’d love to see next. #Llama #OpenSourceAI #LLM #PromptEngineering #SyntheticData #LlamaEcosystem

109

3 Comments

Late last week, I had the pleasure of sitting down with Michael Nuñez from VentureBeat to discuss my team's latest work on building and open sourcing SWE-PolyBench, the first industry benchmark to evaluate AI coding agents' ability to navigate and understand complex codebases, introducing rich metrics to advance AI performance in real-world scenarios. In the interview I discuss the importance of building fine grained metrics to track, measure and improve upon agents' reasoning, decision making and their ability to understand (very) large context spaces. https://lnkd.in/gVJKaxt7

81

3 Comments

The teams that learn fastest win. As a Sr Principal Technologist at Amazon Search, I focus on accelerating innovation through AI and large-scale experimentation. Traditionally, improving digital experiences meant slow, expensive A/B testing with live customer traffic. It could take weeks or months to gather results. Many great ideas stalled — or never launched at all. Thanks to advances in AI, that’s changing. New research — like the AgentA/B system from Northeastern University and Amazon’s Search team [https://lnkd.in/eKeddkV3] — shows that LLM-powered virtual customers can simulate realistic web interactions at scale. We can now test hundreds of ideas before using a single real customer session. Virtual customers aren't here to replace real feedback — they're here to supercharge early learning: * Explore more ideas safely and quickly * Spot weak designs before they reach production * Protect live traffic for validating the strongest innovations The payoff is clear: * Speed: Ideas tested in days, not months * Scope: Bolder, more ambitious exploration * Quality: Evidence-driven decisions earlier in the cycle This is one of the biggest shifts happening in how we innovate for customers: A faster, safer, smarter feedback loop — powered by AI. And it’s just getting started. #AI #Experimentation #CustomerExperience #SearchInnovation #GenerativeAI #LLMAgents

83

4 Comments

Failing the right customer-focused goal at Amazon is better than succeeding at the wrong one An engineering manager on my team once proposed an audacious goal "we will reduce server latency by 20%." We owned the navigation elements of the website shopping experience - rendered on every page so a 20% latency improvement would be huge! I was a new leader to the team and wanted to learn as much as possible, asking "How much of that 20% latency reduction will customers see?" "Customers should see about half the improvement." A 10% speed up for navigation elements for all pages at Amazon.com would be still be great! After discussion (we couldn't do more without resourcing trade-offs) I was onboard: "Great! Set the goal for a 10% customer improvement and get started!" Immediate push back. "Wait! We're only confident in the server side improvement - we might miss this goal!" They were afraid to commit even if it meant taking a server goal that might not matter to customers. This resistance meant we didn't take any latency goal and invested elsewhere. I've seen this pattern repeated since then - leaders preferring to commit to what they know they can deliver and being afraid to outline the goal impact in customer terms that would truly matter. In some cases this means the goal proposed was easily accomplished (we engineers can be a pessimistic lot). In some cases these less-inspiring goals were missed, or simply dropped. Either outcome isn't necessarily great, but missing a tightly defined goal that had unimpressive benefit reflects poorly on the judgment of a leader. Better to aim too high and miss (but get close!) than aim low and fail to impress or achieve. Within reason - teams and people will rise to higher expectations. I am guilty of over-optimism in setting goals, sometimes dramatically so. In 2021 I set a 1-year customer-obsessed goal based on "what would need to be true" hoping to inspire and exceed. We didn't make it - the goal was so audacious that I lost trust with the team "it's red from the start!" Our 2022 goal was more reasonable, but still audacious (yes - we missed that one too). We did ultimately succeed in June 2023, 18 months late -- but crucially I didn't lose the trust of my leaders along the way. We made measurable progress and I provided updates on where we were, the problems we faced, and (more importantly) how we continued to de-prioritize other goals to make the more important work happen. I learned to both educate and simplify in my updates so my leaders could understand, give helpful feedback, and support us when necessary. I helped them appreciate how hard the problems that were trying to solve were - not to excuse the misses, but to explain them. This helped me continue to have the privilege of leading my team. I was promoted in the midst of red goals - the most critical one across my org. Why? Because failing the right goal is better than succeeding on the wrong one.

138

1 Comment

This semester, I’ve been assisting the 5th-grade class at Palo Alto Unified School District (where my daughter is a student) in learning Scratch. Interestingly, several kids have expressed that Scratch doesn’t feel as “cool” as Python, which I taught them last year in 4th grade. Their main complaint? Scratch lacks actual code. While exploring alternatives, I came across an online teaching platform: Go+ Builder. It introduces a variant of the Go programming language, allowing students to drag and drop control flows into an editor window—bridging the gap between block-based and text-based coding in an intuitive way. https://lnkd.in/dBHN27YP If you’re involved in teaching programming to young learners, I’d love to hear your thoughts on similar tools! 🚀

66

3 Comments

Good interview with Jensen Huang and I could not agree more: "You’re not going to lose your job to an AI, but you’re going to lose your job to someone who uses AI." "In today's landscape, understanding how to use Al is no longer a luxury; it's a critical survival skill." "Mark Cuban's investment in free Al bootcamps for underserved high school students exemplifies this philosophy. Cuban urges learners to embrace Al early, absorbing its principles and applications to remain competitive in the evolving job market."

45

5 Comments

Scaling recommendation systems to thousands of GPUs isn't just about adding more hardware. As you distribute training, you often hit new bottlenecks like data imbalance and performance degradation. At Meta, my team is tackling this challenge head-on with 2D Sparse Parallelism in TorchRec. By strategically combining model and data parallelism, we can train terabyte-scale models much more efficiently. Curious about the technical details? We break it all down in our new post on the official PyTorch blog. https://lnkd.in/gZ89kb8z #PyTorch #RecSys #DistributedTraining

92