Luca Lin

Two articles I've been thinking about together this week: Amir Kabir at Overlook VC on "the defining tension of 2026: extraordinary concentration at the infrastructure layer, extraordinary democratization at the application layer" Take a read: https://lnkd.in/gyaFg-nB And Isaac Arnold's framework for what actually makes an application layer company durable: workflow entanglement, domain specific knowledge, and data that compounds from usage rather than just intelligence. Take a read for yourself: https://lnkd.in/gSDXA-vH Amir identifies where the opportunity is, Isaac identifies why most companies building there won't capture it durably and what separates the ones that will.

23

2件のコメント

Breaking: Composite Raises $5.6M to Make Browser Automation Accessible to All Professionals Exciting news in the AI productivity space! Composite just secured $5.6M in seed funding to build a cross-browser AI agent that automates tedious browser-based tasks for professionals. Unlike browser-specific AI tools, Composite works across any browser you prefer. Why this matters: Most professionals spend hours on repetitive browser tasks - from recruiters screening candidates to security engineers managing vulnerabilities. Composite's solution lets non-technical users automate these workflows without switching browsers or writing code. The most impressive part? The tool learns from your patterns and can handle complex multi-step tasks across different websites where you're already logged in. This could be a game-changer for enterprise productivity. What are your thoughts on AI automation tools? Would you trust an AI assistant to handle your routine browser tasks? Read more: https://lnkd.in/guQyyByJ #ArtificialIntelligence #ProductivityTools #WorkplaceAutomation #TechStartups #AIFunding #FutureOfWork #EnterpriseAI #BrowserTechnology

1

Latest from me and the team at Self Labs. We've closed $9M in funding from Greenfield Capital, SCVxSBI, Spearhead, Verda Ventures, Fireweed Ventures to expand our technical roadmap and grow our identity document support beyond biometric Passports, IDs, and Aadhar. Also announced today in our latest application v2.9 is Points a new way for our most active users to earn points for using their favorite apps: Aave Labs, Velodrome, ENS, Lemonade, Espresso, Karma, Talent Protocol, and soon Google Cloud Web3's Testnet faucets. https://lnkd.in/gtWRQguc

140

18件のコメント

Your agent has the answer. It's sitting right there in the context window. And it still gets it wrong. Stanford and Berkeley researchers found that LLM accuracy follows a U-shaped curve. Models perform well when relevant information is at the beginning or end of the context. Bury it in the middle? Accuracy drops 30-50%. Even with 200K+ tokens. They called it "lost in the middle." Here's what that looks like in production: Your agent has access to everything. The docs are loaded. The answer exists. But the model spends compute searching instead of reasoning. It finds something close enough. Confidently returns it. No error. No warning. Just a wrong answer dressed up as a right one. You don't find out until a customer does. This is why "just add more context" is a trap. You're paying for every token (whether the model uses it well or not). Teams shipping reliable agents aren't stuffing everything into one massive prompt. They're routing queries to specialized agents with bounded context. 10 relevant documents, not 500. Smaller context. Higher signal. Better answers. Multi-agent isn't just an architecture choice. It's a retrieval strategy. And it's the bet we're making at AgentShelf. Specialized agents. Bounded context. Answers you can trust and debug.

8

It's official. My co-founder Ahmad Shadid just announced what we've been building in stealth. The world's first Confidential Agentic Development Environment. During my time at Lockheed Martin we built analytics platforms for defense and intelligence customers. The hardest problem was never the technology. It was proving to security teams that sensitive data stayed where it belonged. Today, enterprises face the same wall with AI coding tools. Everyone knows they can 10x productivity. But if you work in defense, finance, healthcare, or anywhere with sensitive IP, your security team has to say no. The tools can't prove where your code goes or who sees it. That ends now. We built a platform with hardware-level encryption, cryptographic attestation, and verifiable isolation. Not privacy policies. Not promises. Proof. We're opening early access to a select group of engineers and enterprises. If your team has been stuck on the sidelines watching everyone else get AI superpowers, your turn is here. https://lnkd.in/dBd5iB-i

52

6件のコメント

2026 is the year venture capital stops funding stories, and starts underwriting systems. Capital hasn’t disappeared. It’s being redeployed with far more intent. Yes, private markets are sitting on >$3.5tn of dry powder, but in venture, that capital is no longer sprayed across themes. It’s concentrating around companies that can turn technical advantage into deployable systems, fast. Family offices are part of this shift, behaving less like opportunistic VCs and more like long-term underwriters. They’re anchoring fewer deals, leaning into structured equity, selective growth rounds, and businesses where cashflow visibility and control exist earlier than before. Not risk-off, just far more discriminating. Institutions are moving in parallel, pulling venture closer to infrastructure. Power, data centres, compute, energy transition, not as themes, but as constraints. AI alone is expected to drive hundreds of gigawatts of incremental power demand by the end of the decade, and venture returns will increasingly accrue to the companies that sit at those bottlenecks. This is where venture capital bifurcates. In 2025, 60%+ of global VC dollars flowed into AI, yet most of that capital clustered around a narrow set of platforms and infrastructure layers. In 2026, that concentration deepens, but the value starts moving downstream. Frontier and deep tech are back, not as moonshots, but as execution layers. Robotics, automation, and embodied AI are no longer optional adjacencies to software. They are becoming the physical interface of intelligence. As models mature and commoditise, venture value shifts toward: robotics integrated with real-world data, industrial and warehouse automation, defence and dual-use systems, healthcare and logistics robotics. By 2026, robotics investment is less about experimentation and more about deployment economics: labour substitution, safety, throughput, resilience. The winners won’t be the most novel machines, but the teams that can integrate hardware, software, data, and regulation into systems enterprises can actually run. Across VCs, family offices, and institutions, the filter is converging: What is defensible? Where does control sit? What scales without cheap capital? And who can execute when conditions aren’t perfect? The era of venture being impressed by possibility is over. The next cycle belongs to investors, and founders, who understand where value accrues, and how it is enforced in the real world. That’s the real movement heading into 2026. #VentureCapital #FrontierTech #DeepTech #Robotics #EmbodiedAI #AIInfrastructure #PrivateCapital #LongTermCapital

36

6件のコメント