Harry Stebbings@HarryStebbings15h ago

I first met @matanSF following a kind intro from @MattEvantic.   We went for a walk in Hyde Park. I was in my trusty short shorts and it was kinda awkward as a walk.  It was awkward because after 5 mins, he was clearly a genius and for the next 55 mins, I just had to pretend like I had more questions to ask before saying with intense eagerness, “can I invest”.  Thank the lord he let me. And through that I got to spend time with his co-founder, @EnoReyes.  Eno is this insane combination of a truly brilliant technologist with an intense awareness of what it takes to build an insanely high-margin, efficient business in AI.  I sat down with Eno when he was in London recently (episode in comments) and have added my handwritten notes below.  Special thanks to @rabois @shaunmmaguire @byersblake @Sabina_Smith_ @laurenmhreeder for some amazing question suggestions. -- 1. The Frontier in AI Right Now The true frontier of AI is building verification systems where no benchmarks exist. Creating systematic frameworks for what “good” looks like allows AI to reliably execute complex, high-friction human tasks that were previously difficult to automate. Is this the true beauty of Instinct @noahrshinn @saranormous 2. The Two Ways That Large Language Models Will Win Frontier model providers face two paths: dominate infrastructure through high-volume inference or move up the stack into high-margin applications. But model-locked applications can conflict with what enterprises actually want: the best possible outcome across multiple models. Love to hear your thoughts on this specifically @AnjneyMidha @mmurph 3. What Is Required to Get the Most Out of Models? Simple gateway routing outside the execution layer delivers only basic cost savings. Maximizing agentic performance requires operating statefully inside the workflow itself, dynamically understanding context, execution history, and what needs to happen next. How do you think about this @alexatallah @shensi @ThibaultJaigu @rauchg @koblovinamerica 4. Why 80% of Neo Labs Will Die and What Separates the Winners From the Losers The winners will anchor themselves to durable enterprise workflows that do not disappear as underlying frontier models improve. Single biggest advice to VCs on investing in neolabs today @LiamFedus 5. Why the Harness Is So Valuable and Who Ultimately Is the Sovereign of Your Intelligence Continuous learning and workflow optimization happen at the harness layer, not inside closed model APIs. True enterprise sovereignty requires owning your harnesses and learning loops so critical intelligence remains proprietary rather than being ceded to third-party labs. 6. Why We Should Not Be Scared to Use Chinese Open-Source Models Labeling open-source weights as dangerous “Chinese models” can obscure the distinction between model provenance and actual security risk. Open models can reflect creator biases, but their risks should be evaluated technically rather than by origin alone. As open weights improve, they could power a growing share of standard enterprise workflows.