Rocketship.vc

Today, we are proud to announce that JACKAL-0004 and Mosaic have achieved all Mission X-3 test objectives. Jackal has been fully commissioned and is prepared for its next phase of mission. We could not be more excited for this moment. True Anomaly is ready for what comes next. Check out our blog for more details: https://lnkd.in/g8p2vjv6

I’ve been teasing a little side project for a few weeks now. Then life intervened — I had to step away for ~2–3 weeks, my laptop rebooted for a security update, all my shells and apps restarted… and when I finally came back to the project yesterday, I ran into something interesting about how coding agents / copilots work. Like many of you, I’ve started to think of these agents as “junior engineers” who can help me pick up where I left off. I even tried to set things up the “right” way: - Kept an agents.md file with notes for the agent - Documented the architecture in architecture.md Tried to follow good practices so future-me (and the agent) wouldn’t be lost And yet, when I returned, I still struggled to rebuild my mental model of the codebase. Worse, even with those files, the agent also failed to fully recover the context we had previously built up together. My takeaway: today’s coding agents mostly know what was done, but not why. The design rationale, trade-offs, and prior lines of thinking — essentially the “story” of the project — were gone. The agent had solid short-term memory (great within a live session or a saved checkpoint) but almost no durable long-term memory. Coming back after a few weeks felt uncomfortably close to starting from scratch. I’m increasingly convinced that if we want agents to truly act like long-term collaborators, we need a much richer memory model, including: - Factual memory – what artifacts/classes/files exist, and what they do - Rational memory – why certain decisions were made, what we considered, and what we ruled out - Process memory – the plan, the preconditions/postconditions, and which steps are done vs. pending Coding is essentially planning + methodical execution, with interruptions guaranteed in real life (vacations, new projects, forced reboots, etc.). Until agents can reliably retain both the facts and the reasoning across those interruptions, we’re still in the early innings of “AI pair programmer.” Curious how others are handling long breaks with AI copilots today — any workflows or tools that actually preserve this “why” over weeks or months? #AIAgents #SoftwareEngineering #DeveloperExperience #GenerativeAI #FutureOfCoding

Recently, I wrote about the ways criminals can misuse AI—from identity theft and fraud to a wide range of other abuses. That post sparked a lot of questions, and one theme kept coming up: “Is this just another version of the broader fear of AI?” For many people, the biggest fear is AGI: systems that become so capable they end up controlling key parts of human life. That’s not a fear I share. Working closely with a wide range of today’s models, my takeaway is: - They are incredible productivity enhancers and powerful co-pilots. - They are nowhere close to having the kind of deep understanding or autonomy required to “take over.” Yes, the pace of progress feels exponential—as Jensen Huang put it, we’re living through “three overlapping exponentials” in compute, AI, and data centers. It’s natural to conclude that AGI must be right around the corner. But I don’t believe our current architectures get us there: - These models still don’t show fundamental understanding of the knowledge they’re trained on. - I haven’t yet seen true intuition or original conceptual breakthroughs that would signal early hallmarks of AGI. Where I am very optimistic is here: We’re at the beginning of a golden age of AI-assisted human discovery. With the right guardrails and strong human collaboration, these tools can dramatically accelerate innovation—without replacing the human judgment, creativity, and curiosity that actually drive it. Curious how you’re thinking about this: Are your primary AI concerns about misuse by bad actors today, or about longer-term AGI scenarios? #AI #ArtificialIntelligence #Innovation

Yesterday I had the chance to share what we’re seeing in fintech + AI at the Zenskar CFO Exchange: https://lnkd.in/gWPpehZB One topic that resonated: Treasury Management for startups. It’s rarely the headline in a CFO job description, yet it can quietly add (or subtract) months of runway. The setup: when a startup raises a round, it receives a couple of years of capital upfront. Most teams park it in a high-yield account or short-term deposits. That’s safe—and simple—but it’s also leaving value on the table. Why this matters now - Rates are non-zero: idle cash has a cost-of-inaction. - Cash needs are lumpy: headcount ramps, vendor prepaids, and GTM bursts don’t follow neat curves. - Counterparty risk is real: bank concentration and policy drift need monitoring. Boards are asking: “Are we being prudent and proactive with cash?” What great startup treasury should optimize for - Safety first (policy- and limit-aware) - Always-on liquidity (cash where you need it, when you need it) - Consistent yield (without adding operational burden) - Governance & auditability (clear controls, clean trails) Where GenAI can change the game - Live cash forecasting from billing, payroll, and ops systems—no manual spreadsheets. - Policy-aware allocation engines that propose (and execute) safe ladders across T-Bills, MMFs, and term deposits, within board-approved limits. - Counterparty & concentration monitoring that flags exposure drift in real time. - Scenario planning (new hire plan slips, enterprise deal closes late, cloud spend spikes) with instant treasury rebalancing suggestions. - Narrative, board-ready reporting with full audit trails and approvals baked in. The opportunity for founders: Build AI-native treasury tools that are: * Plug-and-play with banks, brokers, and ERPs; * Guardrail-heavy (policy, limits, segregation of duties); * Low-latency (decisions and moves in hours, not quarters); * CFO-friendly (explainable recommendations, one-click approvals). If you’re building in AI-powered treasury or cash management, or you’re a CFO feeling this pain, I’d love to connect. There’s a real chance to turn “money sitting in the corner” into smart, governed, liquid runway—without taking on new risk. #Fintech #AI #CFO #StartupFinance #TreasuryManagement #CashManagement #GenerativeAI

People keep asking me: “So… are we in an AI bubble?” My real answer: I’m feeling two things at once—excited and a little nervous. - Excited, because I think we’re about to jump to a new model paradigm beyond today’s giant transformer + RL stacks. - Nervous, because that shift will break a lot of assumptions (and a few business models). Here’s what I think changes next: 1) Data: quality > quantity We’ve spent years training on “the whole internet,” and a lot of it is noise. The next wave will be licensed, curated, high-quality datasets: think trusted archives, real world movements, specialized journals, proprietary corpora—data with real signal. That means smarter ingestion (active selection, filtering, feedback loops) instead of indiscriminate scraping. Teams that build repeatable pipelines for acquiring, clearing rights to, and continuously refreshing premium data will get compounding returns in model performance and defensibility. 2) Architecture: less blob, more brain Fully connected nets are sample-hungry and great at memorizing patterns—not always at understanding. I’m betting on dynamic, sparsely connected graphs that form/prune connections on the fly. Sparse attention, routing, and targeted updates reduce wasted compute, improve generalization, and speed training. Practically, this looks like conditional computation (only the relevant subgraph fires) and structured memory (relationships persist where they help, vanish where they don’t). 3) Specialists > one big generalist We’ll lean harder into orchestrated specialists—adjacent to Mixture-of-Experts and closer to the brain’s layout (vision, language, reasoning, affect). Different subnets can own distinct skills (retrieval, tool use, planning, multimodal perception) and coordinate through a lightweight controller. Upside: better accuracy, lower latency/cost per request, and easier lifecycle management (you can update a single expert without retraining the world). 4) Inference: engineered “aha” moments Today’s systems mostly follow the shortest path to the next token. There’s room for purposeful exploration at inference—sampling “distant” regions of the network or alternative experts to spark novel connections. Think of it as controlled divergence (explore) with disciplined convergence (verify and select). Done well, this yields fewer bland answers, more useful creativity, and still keeps reliability via self-checks, tools, and retrieval. So, bubble? I don’t think so. It feels more like a phase change. #AI #GenerativeAI #MachineLearning #LLM #DataQuality #DataCuration #AIArchitecture #SparseComputing #MixtureOfExperts #VentureCapital

Since my last post on why India must seize its demographic dividend in the age of AI, many people have asked what they can do for the country. My answer, in one word: *audacity*. For too long, we wore “developing nation” as a ceiling on our ambition. That ceiling is cracking. Look at ISRO’s Chandrayaan, the IPL—now the world’s second most valuable sports league—and the India Stack powering the largest real-time payments network on Earth. But audacity hasn’t yet reached everyone. Too many young Indians still hesitate to dream at national or global scale. To them I say: dream bigger than feels comfortable. Bigger than what seems “realistic.” Then build. Consider just a few who did: - From scrap to scale: A founder who began recycling tires in one city is now on track to build a national circular-economy champion. - India’s own rocketry: A team asked, “Why not our SpaceX?” and built a successful private rocket company. - Voice AI leaders: Founders who embedded voice AI into their platform now lead the category across the country. - Aviation made in India: Entrepreneurs who started with aircraft components are pushing toward full airframes—designed and manufactured at home. India doesn’t lack talent, resources, or opportunity. We sometimes lack the audacity to claim them. So if you’re young—and even if you’re not—choose boldness. Build the company, the product, the brand you wish existed. Let’s make audacity our national habit. #India #Audacity #Entrepreneurship #AI #MakeInIndia #GlobalAmbition

The “AI datacenters won’t pay back” math is a snapshot, not a movie. Yes, capex is massive. But what’s being built isn’t just rooms of GPUs—it’s a compute-native model economy where costs fall and utilization compounds. - Big Tech AI capex is tracking to ~$200B in 2025. - Meanwhile, OpenAI is already at ~$10–13B ARR (and growing). - Hyperscalers have extended server lives to ~6 years, and older GPUs cascade to inference and fine-tunes. - Inference costs keep collapsing (quantization, distillation, kernels). Each cohort gets cheaper to serve as demand explodes. If you assume all assets die in five years and revenue stands still, the spreadsheet screams doom. If you understand that the asset is not the facility but the model that was trained—the curve changes. The mistake isn’t the math; it’s the freeze-frame. AI infra is a flywheel: capex → better models → more usage → cheaper inference → higher utilization → better margins → next model.