10:00AM PDT SEPTEMBER 10, 2026
Previously Author of 10+ books on LLMs and agentic AI, founder of Kylie.ai (early agentic AI, YC-backed, acquired 2019), AI Lecturer/Author for Pearson + O'Reilly, and your friendly neighborhood ML engineer.
Becoming AI native means owning the specialized intelligence that defines your business - your taxonomy, your style, your customers' quirks - instead of renting a generic version of it by the token. Teams get pushed toward that leap by a handful of forces: fear that a frontier lab will move into their domain, a quality bar the general models don't clear for their vertical, or an AI bill that's outgrown the intelligence they actually need. This session is about the road that gets you there.
Most teams move through four stages, and knowing which one you're in is half the battle.
Almost everyone starts here. It's powerful, someone else pays to train and serve it, and you pay by the token. But the model was molded by someone else, for everyone else, and you don't control weights, price, latency, deprecation, or in most cases, your data.
Prompt engineering, then context engineering, then harness engineering - MCP, RAG, tool optimization, compaction. Humans picking up the slack for a model that cannot change. It carries you a long way, but every model in your harness is still the model you rented off the shelf, and every request re-explains your business in tokens you pay for.
Frontier open models close most of the quality gap at a fraction of the cost, and this is where your own evals play a massive role. Benchmarks and leaderboards give you a handy model shortlist but it’s your evals that tell you where each model actually wins on your task. Route to open models for categories where they win, keep the rest on closed, and you end up with a stronger and cheaper system than either model alone.
At some point your taxonomy, your style, your customers' quirks are your value proposition. Training is how you stop needing to re-specify them on every request or worse, ship off the logic to a closed AI lab, and move them into weights you own.
We'll walk through some common patterns where this pays off and show a worked example of tuned open models beating closed ones on computer use.
The real payoff isn't one trained model. It's the loop of eval → training → monitoring → update eval → eval.
Rented intelligence can only ever catch up to you; your owned version will stay ahead of the curve. Come find out which stage you're in and what the next step actually looks like.
Ask us anything about training, Fireworks, or just AI in general.
