Great piece, Richard. Yes, ontologies are a plus today — but things are moving fast. I expect we'll soon have super-capable agentic LLMs that can plan, strategize, pause, invent, and use tools on their own. As an agent developer myself, I know this isn't the side I'm expected to take, but it's how things are advancing from where I sit.
Congrats on the piece — and on Latent Space continuing to be the place this kind of long-view thinking gets published. Always a pleasure to read.
On second thought, even if we get there with standalone LLMs, agents may still win in practice. If a 10T model can do what a 1T model can achieve with good agentic tooling and ontology guardrails, most companies will likely choose the cheaper, faster, more controllable setup for a long time.
Great piece, Richard. Yes, ontologies are a plus today — but things are moving fast. I expect we'll soon have super-capable agentic LLMs that can plan, strategize, pause, invent, and use tools on their own. As an agent developer myself, I know this isn't the side I'm expected to take, but it's how things are advancing from where I sit.
Congrats on the piece — and on Latent Space continuing to be the place this kind of long-view thinking gets published. Always a pleasure to read.
On second thought, even if we get there with standalone LLMs, agents may still win in practice. If a 10T model can do what a 1T model can achieve with good agentic tooling and ontology guardrails, most companies will likely choose the cheaper, faster, more controllable setup for a long time.