Jeremy Barnes

Wild Moose is now consistently hitting >80% accuracy in identifying root causes - within just 3 weeks of deployment. The idea of an AI SRE has been floating around for years now. Investors put their money on it. But customers remained (rightfully) skeptical. Why? Because when something breaks in production, every minute counts. And an AI agent that points you in the wrong direction isn’t just useless - it's harmful. The truth is: building a generative AI agent that understands the intimate dynamics of your system is incredibly hard. That’s why despite all the hype, most tools in this space never made it past demo stage. For us, something shifted earlier this year. Technical bets we made early on - around how to express tribal knowledge, and how to gradually increase the Moose’s autonomy through real-time reasoning - started paying off. Everything started clicking. And with that, something else happened: Our customers became our biggest advocates. It’s hard to describe how energizing that feels as a founder. We’re feeling unstoppable. More to come.

165

17 Comments

We teamed up with The Product Folks on mosAIc 2025 - a virtual build a thon for AI builders. As part of the buildathon, there are several hands-on workshop that covers everything from evals, prompting, and agent design to security, infra, and shipping production-ready flows. This is what it actually takes to build a working agentic application. Alongside speakers such as Arvind, Ji Huang, Madhu, Akash, Rohit Khanna, Rahul, Vaibhavi, Eashan, I'll be doing a hands-on workshop as well, to answer questions like: → What truly goes into building an AI agent that can do real meaningful job - layers involved in building it → How do agents work securely on behalf of users and connect with external tools to get the job done. → How should builders think about auth where agents are the primary actors If you’re working on AI infra or products, this is probably the most useful 2 weeks you’ll spend this year. Last few spots left. Link in the comments.

70

5 Comments

Why LLMs freak out over the seahorse emoji? (+ why is it a big deal & how to fix it) Ask an LLM: “Is there a seahorse emoji?” It will confidently say: Yes. & then return 🐟 🪼 🦄 or something else This small illusion is a window into a big challenge in how LLMs “believe” things. The model builds an internal representation of “seahorse + emoji” despite no such token existing. At the decoding stage, that fuzzy vector is force-matched to the closest known token, usually 🐠 or 🐴. Technically, this is a latent alignment failure. The model’s residual stream encodes associations that the vocabulary cannot express, and the projection layer tries to collapse them into something that looks valid. Now scale this behavior to enterprise use-cases 👇 ➝ In insurance, a model might generate a “policy clause” that doesn’t exist in any regulator’s handbook. ➝ In legal AI, it might hallucinate a “precedent” that was never filed. ➝ In healthcare, it might suggest an “ICD code” that doesn’t map to any billing standard. The model’s latent beliefs don’t always map cleanly to real-world tokens or entities, but they still surface as confident outputs. In my experience, fixing this requires system-level engineering, not just prompt tricks: 1️⃣ Vocabulary-constrained decoding Restrict generation to valid ontologies — e.g., only recognized ICD codes, policy IDs, or SKUs. This eliminates “phantom tokens.” 2️⃣ Intermediate validation layers Don’t just post-validate outputs; intercept at generation. We build middleware that checks candidate tokens against a domain dictionary before they surface. 3️⃣ Activation audits Use probing tools (like logit lens) during fine-tuning to detect latent ghost concepts. If your model repeatedly encodes “nonexistent entities,” you can retrain or regularize before deployment. 4️⃣ Fallback strategies When the model ventures outside the allowed distribution, the safest answer is “no such entity exists.” Teaching the system to abstain is as valuable as teaching it to generate 👉 The seahorse emoji might be harmless. But in real systems, the same failure mode can break compliance, mislead doctors, or fabricate contracts. If you’re building mission-critical AI: Don’t just audit the outputs. Audit the beliefs. #llms #chatgpt #genai

139

21 Comments