Building ChatGPT for Finance — Joseph Kim, OpenAI's Head of Financial Services

Latent Space · 2026-10-31

Building ChatGPT for Finance — Joseph Kim, OpenAI's Head of Financial ServicesLatent Space ▶ Play the video

Why it earned a slot

The finance angle forces a hard look at how much domain-specific reasoning you can bake into a shared LLM before you hit the regulatory wall. Joseph Kim’s focus on “training financial reasoning into a shared model” suggests a shift from pure prompting to a hybrid fine-tuning regime, a move that most fintech startups are still debating. Compared with the more sandboxed approaches we saw in early 2025, this episode promises concrete trade-offs between model alignment, data privacy, and latency that will shape the next wave of AI-driven financial products.

The short version

Finance is the ultimate testbed for LLM safety and value.

Why it matters

Regulators are tightening around AI-generated advice, and every new model that claims financial reasoning must survive both compliance audits and real-world risk. The decisions made at OpenAI will set a de-facto benchmark for the entire industry, from robo-advisors to fraud detection. Practitioners need to understand whether a shared model can ever be trusted with high-stakes capital flows.

My take

I think the moment we start embedding sector-specific logic into a general-purpose LLM, we cross a threshold where model governance becomes inseparable from product design. The trade-off between a monolithic model and a modular, plug-in architecture will dictate how quickly firms can iterate without re-training from scratch. Expect the next six months to be a litmus test for whether this hybrid approach scales or collapses under compliance pressure.

How it connects

• Hybrid fine-tuning vs. prompt-only pipelines – a recurring theme in recent agentic product talks.
• Regulatory sandboxes are beginning to require explainable model outputs, which this episode hints will be baked into the training loop.
• The finance use-case will likely drive new tooling for risk-aware inference, echoing trends in AI-driven security.

Bottom line

Start mapping your compliance requirements to model components now; you’ll need a clear split between core LLM and finance-specific modules before the next product cycle.