AI Radar

Noteworthy AI conversations and writing, summarized from publisher notes.

OpenAI Enterprise, GPT-5, and Why Digital Autonomy Lags Physical

· BG2Pod with Brad Gerstner and Bill Gurley · BG2Pod

Sherwin Wu and Olivier Godement, who lead engineering and product for the OpenAI Platform, walk through OpenAI's enterprise business and the work of its Forward Deployed Engineering (FDE) team. They detail three flagship deployments: T-Mobile, where OpenAI models power both text and natural-sounding voice support; Amgen, where GPT-5 accelerates drug R&D and the regulatory/document workflow; and Los Alamos National Labs, where the O3 reasoning model was deployed on-prem on an air-gapped supercomputer called Venado, with model weights physically transported into the facility. They frame the platform as essential to OpenAI's mission of distributing AGI's benefits, since many high-impact use cases in healthcare, telecom, and government only flow through B2B channels. On the model side, they discuss GPT-5 as OpenAI's first launch framed as a full system rather than a single model, with deliberate focus on behavior, style, and tone alongside capability. Key tradeoffs include reasoning depth vs. latency, and the team describes customer-driven feedback that made GPT-5 very literal in instruction following, sometimes requiring prompt rewrites. They cover the recent GA of the real-time voice API (which unifies STT, reasoning, and TTS rather than stitching them), and Reinforcement Fine Tuning (RFT) as a more powerful successor to supervised fine-tuning, with early wins from Rogo (financial docs) and Accordance (tax). On the broader question of why digital autonomy lags self-driving despite a lower safety bar, they point to two factors: decades of accumulated work and real-world scaffolding (roads, signs, laws) versus AI agents being dropped into unstructured environments with no standard interfaces. They also address the MIT report that 95% of AI deployments fail, attributing failures to missing scaffolding, lack of eval discipline, and absence of a cross-functional tiger team with top-down sponsorship. The episode closes with rapid-fire picks: Sherwin is long esports and short the AI tooling/evals category, Olivier is long healthcare/pharma AI and short memorization-based education, and both share AGI-pilled moments and career advice emphasizing critical thinking and AI-native fluency.

AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.