We’re sharing new results showing that frontier AI models are able to outperform human experts on financial predictions for the first time. We built an environment for the challenging task of Earnings Predictions, finding that the most recent frontier models (Opus 5.5, Fable 5.1, Astra) with Samaya’s harness, meaningfully outperform expert consensus predictions on commonly reported earnings metrics such as company revenue, gross margin and more. We built this environment to have a “point-in-time” gate in the harness to ensure no information leakage, and created financial retrieval and data tools that would respect this gate and give the models access to the same information available to experts. We also adjusted for bias in expert consensus to create a hard, bias-corrected baseline, where outperformance has real signal. Only the most recent frontier AI models are able to beat this hard baseline, pointing to an important inflection point in AI capabilities in finance. Like with FrontierFinance, we see substantial headroom for further RL training in this environment — the models show large variations in error across different rollouts. We’re continuing research on using earnings prediction and other predictive tasks to post-train models to learn from their mistakes. Get in touch if you’d like to collaborate or try out an alpha version. Link to the full research below.

Seeing a lot of mixed takes on what LLMs do to vertical software moats. Most are framed as if the biggest threat to software is "an LLM in a chat window." But the real threat is the north star: "AI Agents working with 100% reliability, enterprise context, and at institutional scale" that is slowly becoming reality. Some quick thoughts on the biggest misses I'm seeing: Business logic and institutional context are THE most important value add. Firms want expert vertical AI Agents not because they want another chatbot, but because they want something that can deliver outcomes with deep, firm-specific context. Similarly, software companies that are deeply embedded in business context have a meaningful moat. But AI's advantage is that it can be built to adapt to different types of business logic while using a shared platform. Thoughtful UI matters. UI alone is not a moat. But assuming that sophisticated AI Agents can be reduced to a chat box is a huge oversight. Users need to understand how to build Agents that best represent their workflows, how to collaborate and provide feedback, how to get them to reliably interface with other tools. Overloading all of that into a chat box just doesn't work — our lived experience at Samaya — and the need for guided, structured interaction is real. Strong engineering is still not "trivially accessible." While AI coding tools offer a real increase in productivity, building reliable, fast, scalable, and secure systems — the foundation for enterprise-grade AI Agents — is still a substantial lift. And that's not even mentioning model training work that requires deep technical understanding. AI and software teams that maintain a bar for technical excellence continue to have an important edge. With the coding tools, it's technical excellence coupled with focus and velocity. (See e.g. xitter.cf/ibab/status/1983356398… from @ibab ) Nailing the "long heavy tail." When I was at Google, I spent a lot of time training models to be accurate on the "long heavy tail" — a large set of less common but important use cases. Fast forward to now: what's in the long heavy tail has changed, but not its existence. We consistently find small errors in our AI systems (e.g., company tickers, mixing up metrics) that have to be corrected with urgency so they don't compound as the Agent executes. I expect we will always have a changing "long heavy tail" that needs custom development. Proprietary data is not necessarily a moat. On the surface, proprietary data seems like a strong moat, especially for software incumbents. But in practice, a lot of proprietary data is not truly proprietary. It may be a data product with revenues and pressures tied to licensing it out, aggregated across different primary sources — now easier with AI — or tied to other third parties. I expect we'll see a trend towards data becoming more broadly accessible as the value accrues to what you do with the data.