Podcast Summary: BG2Pod with Brad Gerstner and Bill Gurley
Episode: Inside OpenAI Enterprise: Forward Deployed Engineering, GPT-5, and More | BG2 Guest Interview
Date: September 11, 2025
Guests:
- Sherwin Wu (Head of Engineering, OpenAI Platform)
- Olivier Godman (Head of Product, OpenAI Platform)
Host: Apoor Vagrawal
Overview
This BG2Pod episode offers a deep dive into OpenAI’s push into enterprise AI adoption, the evolution of their B2B platform solutions, real-world customer deployments across industries (including telecom, healthcare, and government), the technical/artistic philosophy behind GPT-5, forward deployed engineering, the nuances of AI model customization, and what’s next for enterprise AI. The guests, Sherwin Wu and Olivier Godman, bring both insider perspectives and a candid, occasionally nerdy vibe to the discussion, sharing behind-the-scenes details, notable successes, challenges, and even spicy opinions on AI and tech trends.
1. OpenAI’s Enterprise Mission & Platform Fundamentals
Timestamps: [02:49]–[04:55]
- OpenAI’s transition from developer-focused APIs to broad enterprise (B2B) solutions predates ChatGPT.
- Sherwin Wu underscores the mission: distributing AGI benefits globally, with the API/platform as the vehicle for “reaching everyone in every corner of the world.”
- API adoption includes startups, digital native companies, Fortune 500s, governments, and direct enterprise solutions.
Quote:
“Our mission, obviously, is to build AGI ... but also to distribute the benefits of it to everyone in the world, to all of humanity.” – Sherwin Wu [03:12]
2. Deep Dive: Real-World Enterprise Deployments
Timestamps: [06:03]–[16:46]
T-Mobile (Telecom)
- Customer support automation: OpenAI models now handle significant chunks of text and voice support, reducing latency and making interactions human-like.
- Heavy integration work required—forward deployed engineers embed with T-Mobile, connecting models to CRMs and legacy tools, often without APIs.
- The collaboration directly improved OpenAI’s core models through “real world” feedback loops.
Quote:
“The features in the T Mobile app ... are actually handled by OpenAI models behind the scenes and it does sound super natural, human sounding latency quality wise.” – Olivier Godman [06:54]
Amgen (Healthcare)
- Use cases split between R&D/speeding drug development and overwhelming regulatory/admin documentation tasks.
- OpenAI’s latest models (GPT-5) already creating tangible efficiency, with potential to affect hundreds of millions of lives via faster drug time-to-market.
Quote:
“If a new drug is developed faster ... this could be hundreds of millions of lives.” – Apoor Vagrawal [13:02] “Once you realize what’s at stake ... it is probably one of the biggest bottlenecks to human progress.” – Olivier Godman [49:53]
Los Alamos National Labs (Government/Defense)
- Ultra-secure, air-gapped on-prem deployment on the “Bonado” supercomputer.
- OpenAI engineers physically transported model weights; no remote access.
- Used for data analysis, experiment planning, and as a “thought partner” – a setup not possible with previous models.
Quote:
“We literally had to bring the weights of the model physically into their supercomputer ... you’re not allowed to have cell phones or any other electronics.” – Sherwin Wu [15:10]
3. Anatomy of a Successful (or Failed) Enterprise AI Deployment
Timestamps: [16:46]–[20:17]
- Most deployments fail (e.g., headline: "95% of AI deployments don’t work" – MIT report).
- Success = Combination of top-down buy-in and bottom-up tiger teams (mix of technical + institutional knowledge).
- Crucial role of evals (evaluation metrics): tightly defining and measuring “what good looks like.”
- Hill climbing: iteratively tuning for performance, often relying on hands-on wisdom/artistry more than pure science.
Quotes:
“The vast majority of the knowledge is in people’s heads ... standard operating procedures are largely people said.” – Olivier Godman [18:14]
“When you define good evals, that gives a clear common goal ... when you fail to, it's a moving target.” – Olivier Godman [19:16]
4. Technical Nuances: Autonomy, Scaffolding, and the Digital/Physical Gap
Timestamps: [20:17]–[25:38]
- Physical autonomy (e.g., self-driving cars) arguably surpassed digital autonomy by 2025 because physical environments have existing, shared “scaffolding” (roads, laws, norms).
- Digital systems lack this; AI agents dropped into complex, inconsistent corporate environments.
- AI agent “autonomy” is early—steep improvement, but still limited by lack of digital “roads.”
Quote:
“Self-driving cars have a good amount of scaffolding ... AI agents are just kind of dropped in the middle of nowhere.” – Sherwin Wu [23:55]
5. GPT-5: Not Just Benchmarks – Craft, Style, and Real-World Feedback
Timestamps: [25:38]–[32:39]
- GPT-5’s big leap: not just in intelligence/capability, but also focus on “craft” (behavior, style, user experience).
- Extensive customer feedback loop. Trade-off: intelligence vs. latency, “reasoning tokens” (longer thinking, slower replies).
- GPT-5 excels at coding, robust reasoning, near-zero hallucinations, superior instruction-following—sometimes almost too literal.
- Prompt engineering remains essential—old prompts can backfire due to improved literalness.
Quotes:
“We have worked so closely with a bunch of customers for ... months to better understand the concrete locks, blockers of the model.” – Olivier Godman [26:22]
“You throw it at GPT-5 Pro and it just one-shots it ... but you might wait 10 minutes.” – Sherwin Wu [28:18]
6. Multimodal AI Progress: Text Leading, Audio and Video Advancing
Timestamps: [32:54]–[38:08]
- Major step: Real-time API for voice—directly speech-to-speech without “stitching” text intermediary.
- Still gaps: Model must learn subtle conversational/workplace cues (ex: handling social security numbers, accents, call pauses).
- Feedback from actual phone-call deployments (e.g., T-Mobile) help close these gaps.
- Some customers still use stitched (speech-to-text/to-speech) solutions, but shift is toward native real-time.
Quote:
“It’s pretty mind blowing ... [that] the speech-to-speech setup actually works correctly ... accents, tone, pauses.” – Sherwin Wu [36:47]
7. Model Customization for Enterprises – From Supervised to Reinforcement Fine-Tuning
Timestamps: [38:08]–[42:51]
- OpenAI invested early in supervised fine-tuning, but reinforcement fine-tuning (RFT) is the new paradigm.
- RFT enables fully customized models tuned for gradable, objective tasks (e.g., parsing financial documents, handling tax tasks).
- Data requirements: precise tasks, high-quality grading, deep bottom-up knowledge.
Quote:
“It allows you to ... use your own data and actually crank the RL ... to create a best-in-class model for your own particular use case.” – Sherwin Wu [40:58]
8. Rapid-Fire Section: “Long/Short” Game & Hot Takes
Timestamps: [42:51]–[50:18]
- Sherwin Wu:
- Long: Esports & pro gaming—massive cultural/tailwind yet undervalued in U.S.
- Short: Most AI tooling—frameworks, evals tools; field is too dynamic for current tools to last.
- Olivier Godman:
- Short: Education models reliant on memorization—LLMs excel here; human learning must pivot to critical thinking.
- Long: Healthcare & life sciences — perfect storm for AI breakthroughs due to structured data, regulatory/document burden, technical readiness.
Quotes:
“I’m short on the entire category of tooling around AI products.” – Sherwin Wu [45:28]
“I’m pretty short on any education which basically emphasizes human memorization at that point.” – Olivier Godman [47:38]
9. Favorite Underrated AI Tools
Timestamps: [50:21]–[53:39]
- Granola: Both guests highlight it; excellent Google Calendar integration, quality transcription/summarization.
- Codex CLI: Massive recent improvements, now “mind-meld” feeling for dev productivity, especially with GPT-5.
Quote:
“You get this weird bionic feeling where you’re like—I feel so mind melded with the model right now and [it] perfectly understands what I’m doing.” – Sherwin Wu [52:01]
10. AI & the Future of Software Engineering
Timestamps: [54:09]–[57:02]
- There will be dramatically more “software engineering”—maybe not all full-time jobs, but more code will be written by more people thanks to AI-augmented tools.
- Non-engineers (inc. product managers, students) are coding prototypes with LLMs/Codex faster than ever.
Quote:
“There is a massive software shortage in the world ... the goal of software was never to be that rigid, hard-to-build artifact ... I expect we’ll see way more of a reconfiguration of people’s jobs and skillsets.” – Olivier Godman [55:50]
11. Advice for Students & Early Career Entrants
Timestamps: [57:04]–[59:43]
- Critical thinking > memorization; fields requiring critical thinking (math, physics, philosophy) best for futureproofing.
- Leverage your “AI native” youth advantage—today’s students/interns are far ahead in exploiting new tools, transforming workplaces.
Quotes:
“Prioritize critical thinking above anything else ... if you go in a field that turns down that thing and gets back to memorization ... you will probably be less future proof.” – Olivier Godman [57:09]
“Don’t underestimate how much of an advantage you have right now because of how AI native you might be ... you’ll have a huge leg up ... leverage that.” – Sherwin Wu [58:06]
12. Roses, Thorns, and AGI-Pilled Moments
Timestamps: [60:04]–[68:31]
Thorns
- The “blip” (board coup) tough on OpenAI’s team—yet ultimately increased organizational resilience.
- Major December 2024 outage: a wake-up call to improve reliability.
Roses
- Sprint to GPT-5 launch showed OpenAI at its best—cutting-edge science, serious customer focus, massive scaling, no outages.
- The first Dev Day (2023) as a “coming of age” for OpenAI: celebrating developer community, successful demos, emotional highs.
AGI Awakening
- Both guests “AGI-pilled.”
- Aha moments: never having to code manually again, major GPT-4 jumps, and astonishing multimodal (especially voice) capabilities.
Quotes:
“The first one was the realization in 2023 that I would never need to code manually ever, ever again.” – Olivier Godman [65:15]
“The blip ... was a really tough moment ... made OpenAI stronger for real ... company built a thicker skin and ability to recover like way quicker.” – Olivier Godman [60:44]
Notable & Memorable Moments
- Physically transporting model weights into a government air gap supercomputer:
“We literally had to bring the weights of the model physically into their supercomputer in San Francisco.” – Sherwin Wu [00:00, 15:10]
- Day-1 status of digital agents compared to self-driving cars:
“AI agents are just kind of dropped in the middle of nowhere.” – Sherwin Wu [00:16, 23:55]
- The “monkey paw” of literal instruction-following:
“It’s almost like the monkey paw—developers ask for better instruction following. Yes ... it follows it almost to a T ... (then) the response is too terse.” – Sherwin Wu [30:49]
- Emotional pride in OpenAI’s progress:
“All the love that GPT5 has been getting in the past couple of weeks ... once you see it, it’s really hard to come back to a model which is extremely intelligent, but an exclusively academic way.” – Olivier Godman [27:02]
13. Closing Thoughts
- OpenAI’s platform focus is to reach everyone, from global enterprises to individual developers.
- Success requires deep customer integration, technical artistry, and evolving service models.
- GPT-5’s arrival is more than a technical leap; it’s a user experience revolution.
- AI is fundamentally reshaping industries—healthcare, customer support, and R&D first, with others soon to follow.
- The future: critical thinking, adaptability, and an “AI native” instinct will shape the next-generation workforce.