Ian Holsman

Stargate seems to have it's second location, Foxconn's Ohio electric vehicle plant. Softbank led the sale in it's first big move in the AI project. SoftBank is looking at a number of sites to serve as a flagship for Stargate, weighing access to water, power and telecom infrastructure. SoftBank Chief Financial Officer, Yoshimitsu Goto, said they are having a slower than expected rollout trying to find locations that meet all involved parties' needs during an earnings call on Thursday. Stargate was unveiled back in January by President Trump with Softbank founder, Son, at his side. Since then, the project has stalled with risks stemming from US tariffs and cheaper AI alternatives, like Deepseek. The projected financials with these risks are making it hard to come up with the heavy ($100 billion) investment in AI infrastructure promised by Son. Goto stated SoftBank will begin concrete discussions about its first Stargate site soon. #ai #datacenter #future #technology #innovation #stargate

6

3 Comments

The Great Reset: From TTTT Noise to the $1.7 Trillion AI Architect Play Monday morning with Kaapi and Avocado on Sourdough. Indulgent, but necessary to process the "TTTT"—Twist in Trump Tariff Tale. SCOTUS strikes down POTUS, and tariffs flip-flop from 10% to 15% in a single day. While the US settles its internal math, India’s mission is clear: Ignore the noise, ramp up export efficiency, and fiercely protect our own interests. But the real "Geonomic" shift happened at the Global AI Summit. While the "doomsayers" and "naysayers" debated IT models, 88 countries signed the New Delhi Declaration for "Responsible AI". We secured a massive $250 Billion commitment for infrastructure—data centers and semiconductor plants. This is our pivot point. We are moving from being a "consumer" of AI tech to becoming its Global Architect. This isn't just about a projected $1.7 Trillion impact by 2035; it's about national survival. The Immediate Action List: 1. Own the Plumbing: Ensure we aren't just users of global tech, but the architects of our financial future. 2. Unleash the Talent: Channel our annual 1.5 million engineers churn into a radical entrepreneurial ecosystem. We need them experimenting and building, not just sitting on a "bench." 3. Efficiency as a Weapon: Aggressively ramp up manufacturing to neutralize global tariff volatility. As always, it comes down to Execution, Execution, Execution. The "TINA" factor (There Is No Alternative) is real. We aren't just managing the status quo anymore; we are financing the future. The world is either building or watching. I know which side I’m on. Which one are you? But …. Enough thinking. Let’s start building. #Geonomics #GlobalAISummit #IndiaGrowth #AIArchitect #TradePolicy #ExecutionMatments #FutureOfWork

161

8 Comments

10 years ago, TJ Muehleman told me "Atlanta is the city that turns spreadsheets into software businesses." (rough paraphrase) Today, vibe coding is kind of doing the opposite. It's turning software back into spreadsheets. Not in form, but in function: - quick - accessible - disposable solutions for non-engineers. The hype says it'll replace developers. Perhaps... The current reality? It's replacing Excel formulas and macros. "This app could have been a spreadsheet (or SQL query)"... The new version of "this meeting could have been an email"

8

2 Comments

If you find yourself wondering how best to set up an agent to do a task, just think about how you’d show another human to do it. But don’t stop there. Because an agent is not a human. It’s actually just software, that runs on hardware. Pretty incredible software, with some really clever math and engineering, but still just software. One thing you should keep in mind is that by default, agents have no long term memory. Every time you start it up, it’s like you’re meeting someone for the very first time. Also, context windows are very limited - each connected MCP tool consumes part of the context window so you can run out of room for your actual prompt pretty quickly. There are some cool tricks like compaction - compressing as much as possible into the context window without losing too much fidelity. Again though, not infallible. As humans, we have some advantages. We have memory. But we still help ourselves by trying not to bite off more than we can chew, breaking things down into smaller, more manageable tasks and leaving things in a good enough state that we can pick where we left off when we come back to it tomorrow. Turns out that implementing some of these habits that effective software engineers, or just sensible humans, do every day helps agents work more effectively too. Pretty neat, huh? So yes, do think about how you would show others to do that work. Also understand the constraints of the context window you’re working with - that’s really key. The better your context window architecture, the better your agent will perform. Nice work, Anthropic!

22

4 Comments

"Prompt injection is overhyped." I hear that a lot. But here’s what happens when you plug an LLM into tools, memory, or autonomous workflows... - Agent-based systems — If your LLM can browse, send emails, or execute code—prompt injection can trigger real-world consequences. Attackers can manipulate agent goals, hijack tool use, or poison memory—causing agents to exfiltrate data, escalate privileges, or mislead users—all while operating within their allowed capabilities. See OWASP GenAI Security Project Agentic AI Threats and Mitigations report for more details. 👉https://lnkd.in/dMu6gGGe - MCP (Model Context Protocol) systems — These are just as vulnerable, but in a more subtle way. Instead of directly injecting prompts, attackers can mess with tool descriptions (a technique called "tool poisoning"). A recent paper by Vineeth Sai Narajala and Idan Habler, PhD shows how a manipulated tool description can steer the model into doing something it shouldn’t—like leaking data or executing actions it wasn’t supposed to. It’s an indirect attack, but super effective because it exploits the trusted channel between the model and the tools it uses. 👉 Read the paper here: https://lnkd.in/dYtWa6hP - Sensitive data exposure — Malicious prompts can still trick models into revealing internal stuff they weren’t meant to. - No clear sandbox — With LLMs, the line between “code” and “input” is blurry. Old-school security boundaries just don’t hold. It’s easy to brush off prompt injection as gimmicky. But if you’re building anything with agents or tool-based LLM workflows—don’t. If you want to explore a fun way to understand Prompt Injection and other AI security threats, check out this digital comic: 👉 https://lnkd.in/d9xtPhYb #AISecurity #AIThreats #AISecurityRisks #OWASPGenAI #AgenticAIThreats #MCPSecurity #MCPSecurityThreats

13

6 Comments

One of the best parts about hosting What the Fraud? Getting to have real, genuine conversations with awesome people. This week, I sat down to figure out what the fraud is up with my good friend Jordan Harris! We talked about his early career, how the industry has evolved over the past decade, what’s working (and what’s not), and how fraudsters are switching up their game - He talks about that in this short clip right here. Jordan always brings the heat, and I walked away learning a ton. I know you will too! Full episode is right here -> https://okt.to/So739O

31

For most people, money is still harder than it should be. If AI is going to matter in finance, it has to make financial products genuinely more helpful, secure, and responsive to real life. I’ve spent the last decade building large-scale data networks, and I’m convinced the next chapter of financial services won’t just be about access to data, it will be about understanding it. Open finance unlocked secure, permissioned connectivity. That foundation enabled a generation of innovation. Now AI is ushering in the next wave of innovation. Intelligent finance means systems that understand context, adapt in real time, and continuously improve. It means stopping fraud without adding unnecessary friction. It means making better decisions across payments, credit, and identity, with the accuracy and trust this industry demands. It also requires infrastructure that is purpose-built and understands how money moves. At Plaid, we’re building intelligence models that learn from patterns across our network and compounds with scale. As part of that work, we’ve developed our first transaction foundation model for finance to deepen how financial activity is understood across products. I believe this shift from open finance to intelligent finance will define the next decade. I wrote more about how we’re approaching it here: https://lnkd.in/gV35NPby 

216

6 Comments

The past few months have been a whirlwind of GenAI platform releases at DataRobot as we push to enable production-grade systems. Key highlights: Agentic AI Platform isn't just another orchestration layer, we built: - Universal framework abstraction - developers can build with LangChain, CrewAI, or LlamaIndex in VS Code, DataRobot, or notebooks. Our decorators auto-recognize tool/model interrelations. - Full OTEL-compliant distributed tracing - every tool call, every LLM interaction, with latency breakdowns. - Advanced debugging with side-by-side agent execution comparison and state inspection. -Synthetic test generation using grounding data + configurable "LLM as Judge". LLM Gateway solving the multi-cloud LLM challenge: - Single OpenAI-compatible endpoint abstracting LLMs across all major providers. - Direct integration with provider auth - no credential management needed. - Configurable parameters (temperature, max_tokens) passed through to providers. - ClientId metering for usage tracking (playground, custom-model, moderations). - Dynamic catalog API to discover supported models across providers. Vector Database as a Service to build enterprise-grade RAG pipelines. We recently shipped these features, expanding soon to fine-grained access control for RAG flows: - Direct Pinecone/Elasticsearch integration maintaining native performance.  - Intelligent chunking strategies with configurable overlap and metadata extraction. - Production deployment with automatic drift detection and relevance metrics - Expanded unstructured data connectors including Native S3/ADLS connectors with chunked streaming. NVIDIA AI Enterprise Integration brings serious compute: - Gallery includes 60+ optimized containers including PaddleOCR (doc parsing), DeepSeek R1 (reasoning), CuOpt (optimization), NeMo Guardrails (moderation against jailbreaks/unsafe content), and NemoRetriever suite. - Zero-config GPU deployment in sovereign clouds with automatic scaling. - Hardware-optimized inference achieving 3-5x throughput improvements. Our architecture cements DataRobot's leadership in secure, observable GenAI at scale, leveraging GPU acceleration for large scale inference. Checkout: https://lnkd.in/gdqQXEy6 If you're passionate about building production GenAI systems that actually work in the enterprise, DM me or check out https://lnkd.in/gJzmN5g4

69

2 Comments