Brian Yang

👉️ https://lnkd.in/gAUt3gZB 🧠 Writing PyTorch extensions? Your memory allocator might be the hidden bottleneck. When you're building custom operations—or integrating specialized hardware—PyTorch's default memory management won't always cut it. This blog dives deep into how to write efficient, debuggable, and performance-tuned allocators for your extensions. Key insights include: 🚀 When (and why) to use custom allocators 🎯 Memory alignment, pooling, and fragmentation strategies 🔍 Debugging tools and sanity checks you wish you had sooner 🤝 How to coexist with PyTorch’s built-in allocator without reinventing the wheel 💬 Building custom extensions or facing memory headaches? Let’s trade strategies—what’s worked for you? #PyTorch #AIEngineering #DeepLearning #MemoryManagement #CUDA #PerformanceEngineering #PythonDev #MachineLearning

1

A developer in San Francisco selects MiniMax as her inference backend. The request crosses the Pacific, gets processed in a Chinese data center, returns in a second. She pays $0.30 per million tokens. Claude Opus would cost $5.00 for the same volume. No customs declaration. No tariffs. No entry in any trade database. Chinese tech media calls this Token 出海: Token Export. This month’s OpenRouter data showed it crossing a threshold: Chinese models surpassed American ones in total token consumption for the first time. This did not happen overnight. I spent February tracing the structural forces behind it: efficiency architectures, energy cost gaps, and a competitive intensity that pushes pricing below rational margins. Nvidia’s record quarter and falling stock price are part of the same story. China’s Compute Bet explains the system. https://lnkd.in/ghEBwk-s The Export That Tariffs Can’t Touch explains what happens when that system meets the global market. https://lnkd.in/gh_ajmmZ

3

DealWire.TECH Deal of the Week - Feb. 20, 2026. Waymo raises $16B to scale robotaxi infrastructure Deal: $16B late-stage round, Valuation: ~$126B Lead investors: Dragoneer, DST Global, Sequoia, Alphabet Waymo, the autonomous driving unit of Alphabet, has raised a massive $16 billion late-stage round, valuing the company at approximately $126 billion. The financing brings together leading global venture and crossover investors, with Alphabet continuing as the strategic anchor. The raise comes as Waymo transitions from years of R&D and pilots into full-scale commercial deployment. The company already operates fully driverless ride-hailing services in multiple U.S. cities and is targeting rapid geographic expansion, with ambitions to reach 1 million weekly rides by the end of 2026. The new capital will accelerate fleet deployment, expand into new urban markets, and build the infrastructure required to support large-scale autonomous mobility networks. Why this matters: • Autonomous driving becomes infrastructure Waymo’s $16B round reframes robotaxi networks as long-term infrastructure assets rather than experimental tech. • Part of the trillion-dollar AI infrastructure cycle Unlike foundation model raises, this capital is flowing into AI deployed in the physical world—vehicles, sensors, and city-scale operations. • Winner-take-most dynamics emerging Only a few companies globally can fund autonomy at scale, making capital access a key determinant of long-term leadership. • Capital intensity becomes the moat Success will depend not just on algorithms but on financing and operating large fleets and urban logistics networks. • Structural advantage for Waymo More than a decade of development, billions of miles of data, and deep integration with Alphabet’s AI and compute stack give Waymo a meaningful edge as competitors like Tesla and Cruise continue to push forward. • Start of the AI transport wave Robotaxis are likely just the first step toward broader AI-driven autonomy across trucking, logistics, maritime, air cargo, and drone fleets. DealWire Takeaway: Waymo’s $16B financing marks a turning point: autonomous driving is shifting from frontier technology to city-scale AI infrastructure. As AI moves into the physical economy, the winners will be those who combine advanced algorithms with capital, data, and operational scale. Waymo’s latest round positions it among the frontrunners in the emerging autonomous transportation stack. #DealWire #Waymo #AutonomousVehicles #AIInfrastructure #Robotaxi #VentureCapital #FutureOfMobility #ArtificialIntelligence

6

Congratulations to Yann LeCun and the AMI Labs team. Following Fei-Fei Li’s World Labs, this is another major capital signal that serious money is moving into 𝐧𝐨𝐧-𝐋𝐋𝐌 paradigm. AMI Labs, just raised $1.03B at a $3.5B pre-money valuation to build AI around world models rather than LLMs. And it changed something deeper: the AI race is no longer just an LLM race. It has split into three camps. 1) 𝐓𝐡𝐞 𝐋𝐋𝐌 𝐂𝐚𝐦𝐩 The bet is simple: scale text prediction. LLMs demand massive data, compute, and energy as an ongoing requirement. Hallucinations are architectural. Periodic retraining is mandatory. The weaknesses are structural: lacks reliability, adaptability, predictability, and accountability. These systems are frozen after training, cannot learn continuously and brittle in the real world. That is why so much of today’s AI industry looks like scaffolding built around static models: wrappers, guardrails, orchestration layers, patches, and human babysitting. In many cases it is productivity theater with no real intelligence. 2) 𝐓𝐡𝐞 𝐖𝐨𝐫𝐥𝐝-𝐌𝐨𝐝𝐞𝐥 𝐂𝐚𝐦𝐩 This is where Yann LeCun’s AMI Labs and Fei-Fei Li’s World Labs sit. The bet here is to move beyond text into multimodal world prediction: learning from video, audio, space, and sensor data to support reasoning, planning, and eventually robotics and other physical-world systems. This is a real break from traditional LLM thinking. But it still appears to remain within the pre-trained-model paradigm: train → deploy → retrain Smarter than text prediction, yes. But not yet a full break from model-centric AI. This camp simulates a better model of the world by improving the representation of reality, but they do not create a system that truly learns. 3) 𝐓𝐡𝐞 𝐒𝐞𝐥𝐟-𝐋𝐞𝐚𝐫𝐧𝐢𝐧𝐠 𝐀𝐈 𝐂𝐚𝐦𝐩 This camp starts from a different premise. Real intelligence is the ability to continuously learn, adapt, and apply what is learned across changing contexts. That means persistent memory, reasoning, understanding, dynamic world models, and autonomous adaptation through experience. These systems self-improve with experience. The result is compounding intelligence, not just faster or more fluent answers. This is developmental intelligence that starts with foundational knowledge, then develops upward through experience, without consuming planetary-scale data, compute, and energy. None of this can be brute-forced, it's the hardest and the least crowded path. This is the real paradigm shift: compounding ROI, no endless retraining, far lower resource requirements, and a cleaner path to real intelligence. AMI Labs is more ambitious than LLMs, but real intelligence comes from continuous learning and compouding capabilities through use in production. 𝐋𝐋𝐌𝐬 𝐠𝐞𝐧𝐞𝐫𝐚𝐭𝐞. 𝐖𝐨𝐫𝐥𝐝 𝐦𝐨𝐝𝐞𝐥𝐬 𝐬𝐢𝐦𝐮𝐥𝐚𝐭𝐞. 𝐒𝐞𝐥𝐟-𝐥𝐞𝐚𝐫𝐧𝐢𝐧𝐠 𝐀𝐈 𝐬𝐲𝐬𝐭𝐞𝐦𝐬 𝐜𝐨𝐦𝐩𝐨𝐮𝐧𝐝. That is the emerging stack. The real AI race has only just begun.

54

9 Comments

Ownership with NO off switch 🔌 At NVIDIA, responsibility isn’t episodic. For Jensen Huang, problems don’t get delegated and forgotten. They live rent-free in his head until reality resolves them. That choice creates a very specific leadership model: ▫️ Decisions stay exposed to consequences ▫️ Weak signals are carried forward, not ignored ▫️ Thinking doesn’t end when the meeting ends ▫️ Feedback loops stay tight, even at massive scale The cost is personal bandwidth. The payoff is clarity most leaders never reach. This is what long-term ownership actually looks like. Uncomfortable. Demanding. Compounding. I broke down why Jensen Huang chose to lead this way, and the price he paid to build NVIDIA to last. Full breakdown here 👇 https://lnkd.in/ePbQi9ig

18

3 Comments

Three frontier models dropped this week. From three different companies. Google shipped Gemini 3 Deep Think. Competitive programmer-level reasoning. OpenAI launched GPT-5.3-Codex-Spark on Cerebras hardware. Near-instant coding feedback. MiniMax released M2.5. Open source. Frontier-level coding at a fraction of the cost. Meanwhile, Anthropic just raised $30B at a $380B valuation. Here's what matters for your team: competition is compressing prices faster than anyone predicted. MiniMax M2.5 can run agentic tasks around the clock for pennies compared to what GPT-4 cost 18 months ago. The winners won't be companies that pick the "right" model. The winners will be companies that build workflows flexible enough to swap between them. Model loyalty is a liability now. The best engineering teams I see treat LLMs like cloud providers. Abstract the interface. Benchmark on your actual workloads. Switch when the math changes. And the math is changing every week. What's your team's model switching strategy?

7

This section on how dbt Labs transitioned from a purely PLG motion to layering on enterprise is a fascinating one. Two instructive passages that I have bolded (below from First Round Review's path to PMF series featuring dbt Labs) TLDR: GTM comprises your ICP, channel, and message. When you transition from PLG / bottom up to Enterprise / top down motion, naturally your ICP and channel also changes, but your messaging / proposition needs to change too, e.g., the enterprise may be multiple personas and are buying assurance as much as solution. --- "Handy found product-market fit organically for dbt as an open-source tool mostly used by data practitioners and developers. But a few years into running a commercial business, he realized he had to build a growth curve all over again with C-suite data leaders. “Even though we had an unbelievable amount of market pull, as we initially commercialized, it wasn’t easy for us to transform this open-source command line tool into a product that enterprises would pay a million dollars for,” says Handy. “When you have enough product-market fit, sometimes it allows you to get away with not being super tight on product marketing or sales motions. So around 2022, we went from this gigantic acceleration curve and overnight we realized, we have to sit at the adults’ table and figure things out real fast,” says Handy. After the PLG flame started to fizzle, Handy turned his attention to layering on a sales-led motion for the cloud platform. “We had to focus our efforts on telling cohesive stories to senior data leaders. We had to have a very clear, explainable answer to the question, ‘Why should I use the commercial product and not the open-source product? And it had to be digestible by someone with a C in their title,” he says. Handy’s answer: dbt Cloud can handle complex data for companies of every size. “The longer people used dbt, the more complex their code became,” he says. “It was a problem for the most sophisticated dbt users, who were often at the largest companies. So there was a real opportunity for us to step in and solve that for them with dbt Cloud.” To tell that story to enterprise customers, Handy relied on data, naturally. “At a user conference we presented a chart that showed the number of dbt projects that had a certain number of models in them — over 100, over 1,000, et cetera,” he says. “We watched that number climb and we knew as users ourselves, ‘Oh my God, trying to work in a dbt project with 5,000 models in it is challenging.’ So we started with that quantitative data point and asked folks in our community about their experiences with these very large, complex dbt projects, and validated that this was a pain in the ass without a cloud platform.”

23

4 Comments

Andrew Ng is right: vertical AI is where the value is. Andrew Ng said something on last week's 20VC podcast that captures what we've learned building Diffusely: the real value in AI isn't in foundation models - it's in vertical applications that transform specific industries. As Ng put it: "Moats tend to be a function of the industry rather than a function of the technology." This isn't about having better models than OpenAI or Google. It's about industry-specific data, workflows and integrations that horizontal tools can't replicate. CarCutter integrates with DMS systems and understands dealership workflows. AutoRetouch handles brand consistency across thousands of SKUs. Both leverage our 200M image library of professional photography. Horizontal AI tools can't do this - not because their models aren't good enough, but because they're not built for it. Ng emphasizes that AI's value isn't saving 20% on one task. It's redesigning workflows for transformation. One automotive client reduced photography time from 28 to 5 minutes per vehicle. €600k saved annually. Fashion brands achieve 90% faster time-to-market. This is 10x gains - fundamental changes to operations, not incremental improvements. The current narrative is that foundation model companies will dominate. Ng's thesis - and our experience - suggests otherwise. Foundation models are infrastructure. Valuable, but not where economic value gets captured. The value is in vertical applications built on proprietary data that integrate deeply with industry workflows. Vertical AI requires understanding the industry as deeply as you understand AI. Professional services, not just software. Patience to integrate with legacy systems. Most AI companies optimize for horizontal scale because it's faster. Building vertical AI is slower, requires more expertise. 𝐓𝐡𝐚𝐭'𝐬 𝐰𝐡𝐲 𝐢𝐭'𝐬 𝐝𝐞𝐟𝐞𝐧𝐬𝐢𝐛𝐥𝐞. Three years building vertical AI convinced us Ng is right. The companies that win won't just have the best models. They'll have the deepest vertical expertise, specialized data, and strongest integration into industry workflows. #verticalai #ai #moat #entrepreneurship #enterprisesaas

35

1 Comment

Waymo collected ~20M miles over 15 years before launching L4 in SF. That’s years of data collection, billions in capital, and a scale most companies simply can’t afford. And in many industries — especially defense — you don’t even have the luxury of commercial-scale deployment in the first place. The reality? Autonomy doesn’t fail because of models. It fails because of missing edge cases, weak coverage, and slow data iteration cycles. That’s exactly the problem we’re solving at DiffuseDrive. In this Forbes piece, Sabbir Rangwala breaks this topic down. Synthetic data isn’t about replacing reality. It’s about completing it. https://lnkd.in/dqNSA4t7 Roland PinterJordan KretchmerOutlander VCPresto Tech Horizonse2vc #PhysicalAI #Autonomy #DefenseTech #SyntheticData #Robotics #GenerativeAI

91

2 Comments