Stanford MS&E435: Ali Ghodsi Says AGI Is Already Here, but the Company Is Not

It is rare for a single 39-minute recording to assemble themes I had previously discussed separately: organizational memory, the AI-native organization, the future of SaaS, and value moving up the stack. In this Stanford MS&E435 session, course instructor Apoorv Agrawal speaks with Ali Ghodsi, co-founder and CEO of Databricks.

Ghodsi opens with a provocation: “AGI is already here.” This is his working frame, not a consensus. What follows is more useful: today's models are already smart enough for many enterprise tasks, but they do not know the organization's context. Hence his reference to “GenAI Divide”: the preliminary report claimed that only about 5% of task-specific enterprise GenAI initiatives reached sustained measurable impact. Its authors cautioned that the figures were directional, and Ghodsi acknowledges that the precise share may be different.

Almost every organization, he says, has its own John or Jane: the person who remembers why a process works the way it does, where the exceptions live, and whom to ask for a decision. The model does not know any of this, so it makes foolish mistakes. This is close to the “company brain” I discussed after Garry Tan's talk: advantage comes not only from the model but also from the ability to give it the right memory.

The strongest segment starts around the twentieth minute. According to Ghodsi, a production Databricks connector to Salesforce or Workday took nine months. The team's estimate was that using AI reduced the cycle only to seven and a half months. They then redesigned the process: a quarter of requirements gathering became one week followed by rapid iteration; test environments were outsourced and prepared in parallel; and seven engineers worked together across seven connectors instead of following a one-person-per-project model. Ghodsi says the result was seven connectors in one quarter.

Neither GPT-7 nor the next Claude would have removed those blockers: this was organizational refactoring. The same idea connects my piece on the AI-native organization with the review of “McKinsey's redesign”: putting an LLM on top of an old process creates local acceleration, not new system throughput. The electrification analogy is similar. Factories became more productive not after installing one electric motor, but after redesigning the floor around independent drives.

The second thread is that “software is dead.” Ghodsi disagrees, though he argues that AI lowers both barriers to entry and switching costs. If users talk to an agent, they care less about which GUI it operates—Salesforce or a competitor. This is close to my argument that “the GUI monopoly is dying, not software.”

Cheaper code still does not eliminate data, scale, brand, trust, certification, or process power. Ghodsi explicitly recommends Hamilton Helmer's Seven Powers. He then predicts that the model layer will become a low-margin set of “token factories,” while much of the value will move into applications. We encountered the same logic in the discussion of model commoditization. Granted, this is a forecast from an interested infrastructure vendor, but the direction looks plausible.

Still, “downloading the company's brain into silicon” is only half the job. If the agent gains context while people lose their understanding of the system, we will merely accelerate the accumulation of cognitive debt. Execution can be delegated; responsibility for the process and the ability to redesign it remain with the team.

If time is short, watch “20:23–24:43” for the connector case and “26:04–28:44” for the story of how scarce bandwidth made multicast a fashionable research problem, only for cheap fiber to devalue it and clear the way for then-strange applications such as Amazon, Uber, and Airbnb. A useful reminder: the defining product of the next cycle rarely resembles the fashionable infrastructure problem of the current one.

#AI #Agents #Engineering #Software #Management #Strategy