Ian Horrocks argues that AI’s next big step is hybrid AI: combining statistical learning systems with knowledge-based systems such as knowledge graphs and rules to reduce hallucinations, improve explainability, and make AI more useful in high-stakes settings.
Watch on YouTube ›Get the market thesis, key claims, assets, contradictions, and follow-up questions from any financial video — then unlock a version personalized to your portfolio, watchlist, and favorite speakers.
The talk opens by noting the major successes of modern learning-based AI in chatbots, image, text, speech, and drug discovery, but also points to embarrassing failures and growing concerns about scaling limits. Horrocks says recent model gains look more incremental, while power consumption and the scarcity of remaining human-generated training data raise sustainability questions. He contrasts learning-based AI with knowledge-based AI. Learning-based systems rely on large, noisy datasets and produce statistically likely outputs, but their reasoning is opaque and hard to audit. Knowledge-based AI, by contrast, uses curated facts and explicit structural knowledge, often represented as knowledge graphs and rules. …
Tactically, the near-term trade is around increased attention to hybrid AI tooling, especially knowledge-graph and RAG stacks that promise fewer hallucinations and better traceability. The immediate risk is that the concept remains promising but not yet proven as a broad production standard.
Over the next few months, the base case is incremental adoption in domains that need structured reasoning, with proof coming from narrower wins rather than a single breakthrough. If deployments start showing measurable gains in accuracy, counting, and auditability, the narrative around AI architecture could shift toward hybrid systems.
The long-run implication is that AI may settle into a two-layer regime: probabilistic models for scale and pattern recognition, plus symbolic knowledge systems for reasoning and verification. If that happens, explainability and domain-specific trust become core design constraints for the next generation of AI products.
Modern AI has been highly successful in chatbots, image recognition, text processing, speech processing, and drug discovery.
The speaker lists multiple areas where AI has clearly worked well.
Recent large-model improvements appear to be tapering and may be approaching practical limits due to power, electricity, and training-data constraints.
He ties slower incremental gains to compute and data bottlenecks.
Learning-based AI is powerful but statistically driven, noisy, and opaque in how it produces answers.
He argues the model is not guaranteed truth and is hard to inspect.
Unlock the full claims, asset map, scores, related transcripts, follow-up questions, and AI chat — shaped around your portfolio, watchlist, favorite speakers, and risks.
