An AI model can be capable yet lack the knowledge needed to work inside a particular company. Databricks' chief executive argues in the interview that models miss the institutional context windowThe maximum amount of text a model can consider at once — its working memory for the current conversation or task.Full definition → employees acquire through meetings and experience, including how decisions actually get made. His proposed response is to organize that knowledge so AI systems can use it. The adoption problem, in this account, includes making company-specific understanding available to the model, alongside improving the model's general abilities.
Key takeaways · AI-distilled
Ghodsi argues today's models could already automate far more work than companies use them for; what they lack is context, since they haven't been in every meeting or learned how decisions actually get made.
His proposed fix is an organizational 'ontology', defined more broadly than the Palantir version, which Databricks has built internally.
He separates speculative superintelligence risk from the more immediate challenge of AI-powered cyberattacks, and the panel discusses why a large-scale AI cyberattack wave hasn't appeared yet.
The hosts also press on what would count as meaningful recursive self-improvement, framed as a 'black box test' for whether RSI is actually happening.
Terms in this piece · Glossary
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.