Oumi

They built a smaller model. It exceeded expectations. Murali Minnah, Co-founder of Wired Informatics, needed something he could actually productionize — deployed on-premise into health systems, with latency constraints a frontier API could never meet. The smaller model kept latency low. It beat GPT-5.4 by 4.5% accuracy. And it runs inside client infrastructure, not a managed cloud. He joins us live July 1 to walk through exactly how they built it. 🗓️ July 1 | 10am PDT / 1pm EDT 📍 Live and then on demand at same reg link Register: https://lnkd.in/eg7pQZQb

"Bigger models were not feasible for us." That's Murali Minnah, Co-founder of Wired Informatics. They needed a model that understood their clinical terminologies — one they could productionize inside client infrastructure. So they built a word sense disambiguation model with Oumi instead. It beat GPT-5.4 by 4.5% accuracy. At 10x lower cost. Murali joins us live July 1 to show exactly how they did it. 🗓️ July 1 | 10am PDT / 1pm EDT 📍 Live webinar and on demand after at same registration link Attend: https://lnkd.in/eSQcX7X7

A word-sense disambiguation model in the domain of healthcare services outperformed frontier models by 4.5% accuracy at substantially lower cost. On July 1, Murali Minnah from Wired Informatics joins us live to show exactly how they built it — and what it compounds to in production. We'll also build two models live: 🏥 Financial statement sentiment analyzer — built in hours, outperforming Opus 4.8 📋 Customer support ticket triage — precision routing with full explainability Why this matters: frontier models don't learn your business. Every run with your own model does. 🗓️ July 1 | 10am PDT / 1pm EDT 📍 Live webcast, streaming live on socials, and on demand at same link after Register here → https://lnkd.in/e_v6RunT

I've been advising Oumi and partnering with Emmanouil (Manos) for a few months. Here's why I made the bet → My thesis: Every company needs their own in-house frontier lab. Their own specialized intelligence - built for their task, trained on their data, owned as their IP. Not rented from a vendor who can change pricing, deprecate your model, or improve every competitor's product at the same pace. Not generic. Not on someone else's roadmap. Specialize. Own. Compound your intelligence. That's the moat. We are building a new category: "Specialized Intelligence" The capability that only frontier labs had — now available to every enterprise. 1) Today, most enterprises are renting their AI. Why that's the wrong bet: → Performance: A model built for everything is optimized for nothing. Tasks are constrained. Domains are specific. The accuracy bar isn't "good enough" - it's deterministic and auditable. Generic models don't clear it. → Cost: Uber burned through their entire 2026 AI budget in four months. 73% of enterprises exceeded AI cost projections. Renting gets more expensive as usage scales - and still underperforms. → Access: Frontier labs control who gets access to their intelligence. We recently saw it with Mythos. Companies that build on top of them can't own their future. 2) The companies that figured this out aren't renting. They're building: → Cognition and Cursor - their own coding models → Intercom - Fin, built on their own intelligence → Airbnb - custom models for pricing, search, and fraud detection → Harvey - legal AI tuned to how law actually works The trend continues - and it's accelerating. 3) Owned intelligence compounds. Each training loop sharpens it - an asset that improves on your terms while competitors wait for the same generic update. As Satya Nadella wrote last week: "The real opportunity is not in picking the best model but instead in building a learning loop on top of models where human and token capital compound." ─── So, Why Oumi? Oumi gives every enterprise a frontier lab in-house, without the billion-dollar research budget. Build your own specialized intelligence. Own it as your IP. The team built Google PaLM and Gemini, Apple's Health foundation models. Backed by 14 leading research institutions. ─── I've placed early bets before - on foundational models before ChatGPT, on AI infrastructure and open source before neo clouds were a thing, on agentic coding before Claude Code. The pattern: a structural shift that's been building finally has the tooling to reach mainstream. Looking forward to continuing to partner with Manos, Oussama, Matthew, Ethan Batraski, Ganesh Srinivasan, Erik, Stefan, Damon, and the Oumi team. And also I get to team up with Bill MacCartney again 😃

💡 𝗘𝘃𝗲𝗿𝘆 𝗰𝗼𝗺𝗽𝗮𝗻𝘆 𝗶𝘀 𝗯𝗲𝗰𝗼𝗺𝗶𝗻𝗴 𝗮𝗻 𝗔𝗜 𝗰𝗼𝗺𝗽𝗮𝗻𝘆. But how do we build with AI in a way that avoids token cost blowouts, loss of IP or direct access to customers, and performs optimally for your specific task? 🤕 It's clear that the current way of building with AI, defaulting to frontier models from OpenAI, Anthropic, and others, for every task is not working... In this webinar, I'll explain what other options exist and how to take back control! ✅ Live demo of 2 small, specialized models: a financial document sentiment analyzer, and a customer support ticket triager, both with explanations accompanying the predictions. Built in < 1 day for $5 and outperforming Opus 4.8 ✅ Real-world case-study from Murali Minnah of Wired Informatics of how they productionized a model for word-sense disambiguation of medical terms that outperformed frontier models Join us for the fun, stay for the learnings - registration link in the comments:

AI hallucinations have now become a legal liability. A Munich court ruled last week that Google is liable for false statements in AI Overviews. The case: Overviews falsely told searchers that two publishers were running subscription scams — citing a source that never mentioned them at all. The court's reasoning matters more than the fine. It held that an AI Overview is not a search result, it's Google's own statement — generated by their AI, structured by their algorithms — and Google is responsible for it. Earlier this year, at the request of the New York Times, Oumi analyzed thousands of AI Overviews using HallOumi, our specialized hallucination-detection model. What we found: 📉 Only 39% of AI Overviews were fully trustworthy — correct and supported by their cited sources.  📉 1 in 3 individual claims wasn't backed by the citations at all. 📉 Better answers, weaker receipts: the newest model hallucinated sources more often, not less. Every enterprise shipping LLM outputs to customers should read this ruling carefully. Hallucinations can cause reputational damage, regulatory exposure, and now legal liability — using a frontier model is not a defense. The fix starts with models you own and can stand behind — specialized AI purpose-built for the task that can be more accurate, reliable and possible to fix. Not rented generic AI from a black box that you can’t control. That's what Oumi makes possible: any enterprise can build its own custom models in hours, from just a prompt. Models that compound intelligence with every interaction in production. Munich is the first ruling like this, but it won't be the last. The enterprises that win the next era of AI will be the ones who can deliver AI solutions with higher quality and reliability at the right token economics.

Tokens to Profits Last night, we hosted a dinner with a group of AI founders, infrastructure builders, and technical leaders to discuss the future of building AI companies and tokenomics. What struck me was how much the conversation has changed. A year ago, most discussions centered on model capabilities. Now they're increasingly about economics and defensible moats. Not whether AI works, but what it costs to run at scale. A few observations that stuck with me: - The industry has been conditioned to believe every task requires the biggest, smartest, fastest model. In reality, many workloads are being solved with the AI equivalent of an 18-wheeler when a scooter would do. - Change management is a huge challenge in enterprise AI deployments. - When balancing performance, latency, and cost in the short term, performance remains the top priority due to the current immaturity of these models and their outcomes. -Emerging trust deficit between the model providers and builders. - We are going to move in the medium term towards deploying autonomous AI workflows leveraging various models all designed around meeting certain defined outcomes. - Competitive advantage won't come from using the same models as everyone else. It'll come from proprietary data, models and customization. Vertical AI will thrive. - In the long term, opportunity exists to make the overall stack 100-1000X more efficient. We ended the night with my favorite question: What widely held belief about AI will prove wrong over the next five years? The answers ranged from "compute won't be scarce" to "today's model architectures will look primitive" to concerns that we're underestimating AI's impact on mental health. One thing feels increasingly clear: the next phase of AI won't just be about intelligence. It'll be about the economics of intelligence. Thank you to Steven Hong, PhD and Emmanouil (Manos) Koukoumidis for co-hosting with me, and everyone who joined us for a thoughtful discussion! Ankit Jain, Eric Nguyen, Faraz Siddiqi, Gert Lanckriet, James Joaquin, 朱浩然 Justin Zhu, Kahini Shah, Marco M., Mike Henry, Mo Islam, Rohan Ganesh, Shara Balakrishnan, Ph.D., Tim Weingarten, and Varun Palivela. What should we discuss next? cc: Obvious Ventures, Amanda Denney, McKenzie Quinn

Narrative violation: according to Stanford University research, local models can answer 71.3% of real-world chat and reasoning queries accurately, up from 23.2% in 2023. Obviously at a fraction of the cost and energy consumption of frontier APIs. The obvious conclusion: you don't need a frontier model for most tasks. The future is multi-model: local, open-source, smaller and cheaper for the majority of workloads, frontier APIs when no other choices!