By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.

Inside OpenAI Enterprise: Forward Deployed Engineering, GPT-5, and More | BG2 Guest Interview

Open Source bi-weekly convo w/ Bill Gurley and Brad Gerstner on all things tech, markets, investing & capitalism.

Sherwin WuguestApoorv AgrawalhostOlivier Godementguest

Sep 11, 20251h 8mWatch on YouTube ↗

Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.

CHAPTERS

Jump to publisher timestamps

Chapter summaries

Setting the stage: OpenAI Platform leaders and the enterprise focus
The host introduces the conversation at the OpenAI office with Sherwin Wu and Olivier Godement, framing the episode as a deep dive into OpenAI’s enterprise and platform work. The focus is positioned as the “other side” of OpenAI beyond ChatGPT, including customer deployments and what’s next in AI.
- •The host introduces Sherwin Wu (engineering lead for OpenAI Platform) and Olivier Godement (product lead for OpenAI Platform).
- •The conversation is framed around OpenAI’s enterprise work, not just consumer ChatGPT.
- •The host previews discussion of customer examples, industry impact, and forward-looking views.
Why the API and B2B are core to OpenAI’s mission
Sherwin explains that OpenAI’s original product was the API, predating ChatGPT, and argues that enterprise distribution is central to spreading AI’s benefits broadly. He outlines what “the platform” includes and why developers and enterprises help extend reach into many real-world use cases.
- •Sherwin says the API was OpenAI’s initial product and was built for developers.
- •He connects the platform strategy to OpenAI’s mission of distributing AI’s benefits widely.
- •He describes platform components: developer API, public sector/government offerings, and emerging enterprise products beyond the core API.
- •He argues many societal-scale outcomes (like more medicines and better public services) require enabling enterprises.
T-Mobile case study: automating customer support across text and voice
Olivier describes OpenAI’s work with T-Mobile to automate and improve customer support, including voice calls. He emphasizes the customer support load and the importance of natural-sounding, low-latency voice experiences that can operate inside consumer-facing app workflows.
- •Olivier describes T-Mobile’s heavy customer support demand, including voice calls.
- •OpenAI and T-Mobile worked on automating both text support and voice support.
- •He says some T-Mobile app call flows are handled by OpenAI models behind the scenes.
- •Naturalness, latency, and voice quality are highlighted as crucial for adoption.
Forward-deployed engineering: integrations, orchestration, and evals for enterprise reliability
The discussion expands from models to the system work required to make enterprise deployments succeed. Sherwin and Olivier describe forward-deployed engineers embedding with customers, connecting models to internal tools, building interfaces, and defining “what good looks like” via evals—especially challenging for audio.
- •Olivier and Sherwin describe forward-deployed engineers embedding deeply with customers and building for specific systems.
- •Sherwin says enterprises often lack clean APIs, requiring gateways and custom integrations to connect internal tools.
- •They stress orchestration work: models need access to CRM and internal context to solve real tasks.
- •They emphasize the difficulty of defining robust evals, including audio evals for long calls and transfers.
- •Olivier highlights the Realtime API as engineering to reduce interruptions and lag for natural voice experiences.
Closing the loop: customer deployments improving OpenAI’s models
Sherwin explains that deep enterprise engagement doesn’t just ship a solution—it feeds improvements back into OpenAI’s models. Using T-Mobile as an example, he says learnings from being embedded helped inform snapshot improvements in the Realtime GA release.
- •Sherwin says working closely with T-Mobile helped improve OpenAI’s models directly.
- •He ties improvements in a Realtime GA snapshot to learnings from T-Mobile’s needs.
- •The host summarizes the “before/after” impact as improvements in tech support and customer support experiences.
Amgen case study: speeding drug development and reducing administrative burden
Olivier describes work with Amgen focused on accelerating parts of the drug development and consolidation process. He splits healthcare needs into two buckets—R&D and heavy administrative/document workflows—and argues models can help across both, noting Amgen as a major GPT-5 customer.
- •Olivier says OpenAI has been working with healthcare companies, including Amgen.
- •He describes two buckets: R&D work with large volumes of data, and administrative/document-heavy work to get medications to market.
- •He highlights regulatory submissions and review processes as major workload areas where automation/augmentation can help.
- •He notes Amgen has been a top customer of GPT-5.
Los Alamos National Lab case study: air-gapped, on-prem deployment on a supercomputer
Sherwin shares a high-security deployment with Los Alamos National Labs requiring a bespoke on-prem installation. He describes deploying a reasoning model into an air-gapped environment on their supercomputer and the unique constraints—including physically transporting model weights due to security restrictions.
- •Sherwin describes Los Alamos National Labs as a high-security government research lab with defense-related work.
- •OpenAI could not use a normal API deployment; they built a custom on-prem deployment onto a supercomputer called Venado.
- •He says they deployed a reasoning model, o3, into an air-gapped environment and adapted it to the lab’s hardware/networking stack.
- •He recounts physically bringing the model weights into the secure environment with strict device restrictions.
- •He notes limited visibility into usage, but feedback indicates applications in experiments, data analysis, and as a thought partner for research.
Why many AI deployments fail: leadership buy-in, tiger teams, evals, and iteration
Prompted by a claim that many deployments don’t work, Olivier and Sherwin lay out what they’ve seen as leading indicators of success in enterprise rollouts. They emphasize leadership alignment plus empowered execution, capturing institutional knowledge, defining evals early, and iteratively “hill climbing” performance.
- •Olivier describes patterns from working with many enterprises and calls out top-down buy-in plus an empowered execution team.
- •Sherwin adds the need for a tiger team combining technical skills and institutional/subject-matter knowledge.
- •Sherwin says enterprise knowledge is often undocumented and “in people’s heads,” making adoption hard without the right team.
- •They stress “evals first” to avoid a moving target; Sherwin says good evals are harder than they look.
- •Sherwin describes iterative improvement (“hill climbing”) from a baseline score toward high performance, sometimes requiring fine-tuning and patience.
Physical vs digital autonomy: timelines, scaffolding, and enterprise infrastructure gaps
The conversation compares self-driving progress to AI agents in the digital world. Sherwin argues agent timelines are much shorter so far, and that self-driving benefits from standardized real-world scaffolding (roads, traffic laws), while enterprise digital environments often lack the connectors and structure agents need.
- •Sherwin argues self-driving has had a long timeline to mature, while AI agents are early and accelerating.
- •He dates meaningful agent robustness to the reasoning paradigm and its recent development.
- •He introduces “scaffolding” as a key difference: roads and traffic laws standardize the physical world for autonomy.
- •He suggests many failed enterprise deployments lack sufficient infrastructure/connectors for agents to interact with systems and data.
- •He says forward-deployed work often involves building platforms, connectors, and organizing data to create that scaffolding.
GPT-5 overview: focusing on behavior, craft, and customer-driven blockers
Olivier describes GPT-5 as not only a leap in benchmark capability but also a step forward in behavior—style, tone, and instruction adherence. He emphasizes a close, months-long feedback loop with customers to identify concrete blockers and improve how the system behaves in real usage.
- •Olivier distinguishes intelligence/benchmarks from model behavior (style, tone, and overall feel).
- •He says GPT-5 benefited from unusually close collaboration with customers over months.
- •He lists behavioral goals: faster experiences, better instruction following, and being more likely to say no when it doesn’t know.
- •He suggests builders can feel the difference versus models that are only strong in an academic sense.
GPT-5 trade-offs and early feedback: latency vs thinking, hallucinations, and prompt sensitivity
Sherwin discusses a core trade-off: more reasoning tokens can improve results but increase latency, which is hard for product use cases. He and Olivier share platform/customer feedback, including strong coding performance, reduced hallucinations, very literal instruction following, and the need to adjust older prompts to fit GPT-5’s behavior.
- •Sherwin highlights the trade-off between longer “thinking” and response latency/performance for builders.
- •He reports positive feedback on coding strength and ability to solve hard problems with more thinking time.
- •Sherwin says customers observed improved robustness and reduced hallucinations, while noting it’s not perfect.
- •They discuss instruction-following becoming very literal, requiring developers to write clearer prompts.
- •Sherwin gives an example where reusing older “be concise” prompts caused overly terse outputs, fixed by removing redundant constraints.
- •He notes areas of constructive feedback including code quality/idioms and making thinking more dynamic for simpler tasks.
Multimodality progress: closing the intelligence gap between text and voice
Olivier describes ongoing work to make voice, image, and video experiences more capable, noting that voice can feel less intelligent than text today. He also highlights the practical requirements of “economically valuable” voice settings, where accuracy on details and appropriate repetition matter in real calls.
- •Olivier says multimodality is progressing across voice, image, and video, with recent generations unlocking new use cases.
- •He notes feedback that text still leads on perceived intelligence, and voice needs to catch up.
- •He describes voice strengths in casual conversation but identifies gaps in work settings (e.g., handling precise identifiers and not guessing).
- •He emphasizes that real customer calls and deployments expose the gaps needed to improve models.
Realtime API vs stitched pipelines: latency, signal loss, and unifying orchestration
The host asks why real-time voice changes the architecture compared to the older three-part pipeline (speech-to-text, reasoning, text-to-speech). Olivier explains the drawbacks of stitched systems—slowness and signal loss such as emotion and tone—and describes prioritizing natural speech-to-speech first, then working toward unified orchestration across modalities.
- •The host contrasts stitched pipelines with the integrated Realtime API approach.
- •Olivier cites stitched issues: higher latency/operational complexity and loss of speech signal across steps.
- •They highlight lost signals including emotion, accent, tone, and pauses—important in phone-call contexts.
- •Olivier says the team is actively working on architecture differences and unifying orchestration logic across modalities.
- •Sherwin notes many customers still use stitched approaches, but more are moving to Realtime for naturalness and low latency as intelligence improves.
- •Sherwin expresses that speech-to-speech working at all is surprising, given it streams audio in and generates audio out.
Model customization: supervised fine-tuning and reinforcement fine-tuning (RFT)
Sherwin explains OpenAI’s long-running investment in customization, starting with supervised fine-tuning and expanding into reinforcement fine-tuning. He describes RFT as adding reinforcement learning to fine-tuning, requiring gradable tasks and objective graders, and enabling customers to push toward best-in-class performance in narrow domains.
- •Sherwin says OpenAI offered supervised fine-tuning pre-ChatGPT and customers used it effectively.
- •He introduces reinforcement fine-tuning (RFT) as adding reinforcement learning to the fine-tuning loop.
- •He contrasts data needs: SFT uses prompt-completion pairs, while RFT needs gradable tasks plus an objective grader.
- •He says RFT is more complex but can be significantly more powerful and can aim for best-in-class domain performance.
- •He cites examples of customers using RFT: Rogo (financial document tasks) and Accordance (tax tasks using a tax-focused eval).
- •Olivier suggests that as base models improve at instruction following, RFT may become more common for pushing capability frontiers.
Rapid fire: longs/shorts, favorite tools, and how coding roles may shift
The conversation shifts to quick picks and opinions. Sherwin shares a long on esports and a short on certain AI tooling categories; Olivier shares a short on memorization-heavy education and a long on healthcare. They also discuss favorite tools (Granola, Codex CLI) and debate how the number and nature of software engineers may evolve.
- •Sherwin says he’s long esports, describing professional gaming’s growth and audience behavior.
- •Sherwin says he’s short broad AI tooling categories (including eval products/frameworks/vector stores previously, and now RL environment tooling), citing rapid stack evolution and competition.
- •Olivier says he’s short memorization-focused education and long healthcare/life sciences as a major AI beneficiary.
- •Both name Granola as an underrated AI tool; Sherwin also highlights Codex CLI paired with GPT-5 and mentions product polish and one-shot improvements.
- •They discuss whether there will be more software engineers, agreeing more coding will happen and more people (including PMs) may code.
- •Olivier describes having PMs build coding prototypes instead of writing PRDs, enabled by GPT-5 and Codex.
OpenAI highs and lows: the “blip,” outages, DevDay, and shipping GPT-5 at scale
Asked to reflect on their time at OpenAI, Olivier and Sherwin describe difficult and memorable moments. Olivier names the “blip” as the toughest, while Sherwin cites a major multi-hour outage; both also point to high points like DevDay and the GPT-5 sprint, including improved reliability during large launches.
- •Olivier identifies the “blip” (a board coup) as his hardest moment, saying it was difficult but ultimately strengthened camaraderie.
- •Sherwin highlights a multi-hour outage as a severe moment that underscored how essential the API had become, and describes postmortems with customers.
- •Olivier calls the GPT-5 sprint a standout positive moment, citing cross-functional excellence and the ability to scale quickly.
- •Olivier and Sherwin cite the first DevDay as a special “coming of age” moment for the developer community.
- •Sherwin recounts the stress of live demos and the relief afterward, ending the night riding home in a Waymo.
“AGI-pilled” moments: coding breakthroughs, GPT-4 shock, reasoning, and voice realism
The host asks whether they’re “AGI-pilled” and what convinced them. Olivier points to realizing he may never need to code manually and to voice/multimodality progress, while Sherwin highlights first using GPT-4 internally and later experiencing the reasoning paradigm through deep research-style tasks.
- •Olivier says he’s “AGI-pilled,” citing a 2023 realization that he would never need to code manually again.
- •Olivier adds that voice and multimodality made it feel more tangible, especially understanding tone and accents.
- •Sherwin says he’s “AGI-pilled,” describing first interacting with GPT-4 internally as a major shock in capability.
- •Sherwin identifies the reasoning paradigm as another breakthrough, pointing to deep research-like persistence and detailed outputs.
- •Both describe “feel the AGI” moments as asking something they thought the model couldn’t handle and seeing it succeed.

Publisher timestamps

Get more out of YouTube videos.

Read along on your phone, search it beside the video, or ask your favorite AI app what was said.