Latent Space

Joseph Kim on the 20% the base model won't build for you

Joseph Kim· Head of ChatGPT Finance at OpenAI
·~27 min·English·Latent Space
LLMAgentsAI CompanyBusiness StrategyAI Infrastructure
TL;DR

OpenAI's ChatGPT-for-finance lead on why the base model is only the easy 80% of a product, and why the hard, durable 20% is in-house data, an org-chart safety model, taste trained from human judgment, and evals that predict the job.

01Core Mental Model

The 20% Is the Whole Game

<strong>The platform hands you 80% of the product; the missing 20% is the only part that wins.</strong> Codex gives everyone the same primitives, so your data, UI, and integrations are what set you apart.

Codex will get you about 80% of the way to a product that would be successful in the market, but, you know, um, the 20% is missing. And so, we need to find some way for my team to go build that 20% to actually go win

— Joseph Kim, Latent Space
Key Insight
Kim frames his own team as a startup building on OpenAI, which quietly concedes that the base model is a commodity. If the 80% is the same for everyone, the defensible business is never the model, it is the sliver of judgment and integration the platform cannot ship for you.
02How OpenAI Builds

Research First, Then Deployment

<strong>Finance-specific training does not make a finance-only model, it folds back into the one main run everyone builds on.</strong> Research comes first on purpose, deployment second.

Internally, we call ourselves a research and deployment company. And so, what that means and we kind of structure that in order on purpose, like research comes first, and then deployment comes second

— Joseph Kim, Latent Space
Key Insight
Refusing to silo the finance intelligence is itself the platform strategy. By pushing that data into the shared model instead of a private one, OpenAI keeps its own finance team on the same footing as any outside startup, which is exactly why section one's 20% is where everyone, inside and out, has to compete.
03Enterprise Reality

First Principles Lie; the Data Decides

<strong>On paper the data bundle looked redundant; testing found some big-name connectors with 40%+ error rates,</strong> so OpenAI brought the data in-house, and it is now the number-one thing customers call out.

in enterprise, it's, you know, you have to deal with the reality of the industry that you're in and go build around that

— Joseph Kim, Latent Space
Key Insight
The lesson is not really about data licensing, it is about trusting measurement over a clean argument. A first-principles case can be internally airtight and still wrong because it never looked at the error rate; the industry's messy reality, not the logic, is what set the product decision.
04Safety by Analogy

Give the Agent an Org Chart

<strong>Shape the agent's safety profile to look like a person, and the review structures that already catch human mistakes can be leveraged instead of rebuilt.</strong> Read-write permissions and the junior-to-MD chain are his examples.

if models behave more like people, then you can leverage those systems

— Joseph Kim, Latent Space
Key Insight
This reframes alignment as an interface problem. Banks spent a century building hierarchies that assume every worker is fallible; by making an agent legible as a junior analyst rather than a black box, Kim gets to reuse those controls instead of inventing agent-specific ones that customers would not trust.
05The Capability Frontier

Factual Mistakes Recede, Taste Is the Frontier

<strong>The dangerous mistake, a wrong fact that flips the conclusion, is the one models are quickly getting good at; what is left is taste.</strong> Judgment is comparative, not right-or-wrong.

the career ending mistake is something where you get a piece of information factually incorrect. And that piece of factual information changes the conclusion

— Joseph Kim, Latent Space
Key Insight
The reassuring half, that the career-ending factual error is receding, hides the harder half. Taste has no answer key: a slide is not wrong, it just looks sloppy, and softer objective functions like that are precisely the ones scaling alone does not fix. The frontier moved from correctness to judgment.
06Training Judgment

Taste Trains Like It Always Did

<strong>You train taste with a reward model over pairwise human judgments, the same signal that trained the earliest models.</strong> Humans still beat models at the aesthetic call.

the practical answer is you just train a reward model, right? You just train a reward model. You get a bunch of humans to look at pairs of things and say which one is better

— Joseph Kim, Latent Space
Key Insight
The field moved to verifiable rewards and forgot that early RLHF was pure taste, so Kim's point lands as a quiet correction: the tools for judgment already exist. Coding is the precedent he leans on, part verifiable (does it compile) and part taste (would you merge it), and finance has the same split.
07Where It's Going

Climb the Ladder, Lose the Answer Key

<strong>Today's AI does junior-analyst work because that work has a verifiable answer; moving up to MD-level judgment means the objectives get much fuzzier.</strong> An investment call can take years to grade.

the objectives become much fuzzier. It goes from this is clearly a good or bad slide, or something that we'd put in front of a client or not, to well, I don't know if this investment idea is good

— Joseph Kim, Latent Space
Key Insight
The hierarchy doubles as a difficulty curve for AI. Low rungs pair tedious, well-specified tasks with instant feedback, which is why they fell first; the top rungs trade a clear answer key for a verdict that arrives years late, so progress there is gated by research patience, not by model size.
08The Real Bottleneck

A Good Eval Predicts the Job

<strong>The bottleneck is not the model, it is the eval; most finance benchmarks are saturated or broken, hiding real progress.</strong> The goal is an eval whose score predicts whether you can automate a profession.

one of the things that my team is working on currently is to try to build a much better eval that actually predicts the ability to automate the work of that specific profession

— Joseph Kim, Latent Space
Key Insight
Kim names a bias most labs live with: good numbers get believed and only bad numbers get investigated, so a saturated eval quietly stops measuring anything. Tying the metric to automating a real profession is an attempt to make the score fail loudly when the model is not actually ready.