Why a New Class of AI “Judgment Models” Could Have Big Business Implications
Wednesday, 16 September 2026 · 2 min read · Listen to the episode ↗
In this episode, the emergence of AI judgment models is explored, highlighting their ability to deliver rapid and cost-effective decision-making capabilities for businesses. Mark Zuckerberg emphasizes the need for responsible AI development, while Salesforce introduces its first in-house model, Koa, aimed at improving sales management. The discussion also touches on the challenges companies face in deriving meaningful value from AI tools, underscoring the transformative potential of these new models in various operational contexts.
A new class of AI models known as judgment models is emerging, which focuses on producing probabilities for specific questions rather than generating lengthy text outputs. These models promise to deliver judgments more quickly and at a lower cost than traditional AI models, potentially transforming business operations significantly.
Mark Zuckerberg highlighted the importance of responsible AI development, stating that individual labs must align their models with user expectations to mitigate risks associated with liability. He noted that Meta chose to delay the release of its AI model Muse to prioritize safety concerns.
The call for human-centric AI regulations has gained traction, with figures like Bernie Sanders and Steve Bannon expressing concerns about the concentration of technological power among a few oligarchs. The overarching fear regarding AI revolves around losing control and agency over its future, rather than specific existential threats.
Salesforce has entered the AI landscape with its first in-house model, Koa, aimed at enhancing sales management. Additionally, the company launched AIforce, an initiative allowing third-party agents to access Salesforce data. However, a study revealed that only 12% of companies utilizing AI tools are able to extract meaningful business value from them, underscoring the difficulties associated with legacy modernization projects.
Diogo Almeida introduced a novel training method called RLCD, claiming that his model, JEV, operates 20 to 200 times faster and is 40 to 400 times cheaper than existing models. JEV is specifically designed for calibrated decision-making, making it particularly beneficial for customer support, sales, and compliance tasks.
JEV's efficiency is notable, with one instance producing 777 judgments in under 0.7 seconds at an estimated cost of a quarter of a cent. This model has the potential to function as a code linter for knowledge work, analyzing documents and flagging issues, thereby enhancing teamwork by facilitating small judgments about responsibilities.
While judgment models are not intended to replace all generative AI functions, their integration into automated systems could have profound implications. The development team has dedicated two years to this concept, believing it to be both crucial and self-evident. They anticipate that once judgment models become established, their absence will be surprising to many.
This summary was generated from the episode transcript and can contain mistakes.