By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.
BG2 PodInside OpenAI Enterprise: Forward Deployed Engineering, GPT-5, and More | BG2 Guest Interview
Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.
At a glance
WHAT IT’S REALLY ABOUT
How OpenAI makes enterprise AI work: deployment craft over demos
- OpenAI’s platform and enterprise work are presented as central to its mission by enabling developers, enterprises, and governments to embed AI into real-world services beyond ChatGPT.
- Customer case studies highlighted three different deployment patterns: voice automation and orchestration at T-Mobile, drug development and document-heavy workflows at Amgen, and a bespoke air-gapped on-prem deployment for Los Alamos National Lab.
- The guests argued many enterprise AI efforts fail due to missing prerequisites like leadership buy-in paired with a capable tiger team, clear evals that define success, and the patience to iteratively improve performance to high reliability.
- GPT-5 development was described as focusing not only on intelligence and benchmarks but on model behavior—instruction following, reduced hallucinations, and practical usability—while managing latency versus ‘thinking time’ trade-offs.
- They described multimodality as progressing rapidly but requiring new infrastructure such as the Realtime API and new training emphasis to make voice interactions accurate and production-ready in high-stakes workflows.
IDEAS WORTH REMEMBERING
5 ideasOpenAI frames enterprise and the API as mission-critical distribution, not a side business.
Sherwin Wu explained that OpenAI began with a developer-facing API and still sees the platform as a core way to distribute AI’s benefits broadly by enabling companies and governments to build AI into services that reach end users.
Successful enterprise AI is as much integration and operational craft as it is model quality.
In the T-Mobile example, Olivier Godement described moving from “tokens in, tokens out” to a production voice support experience that depends on systems design, deep integrations, evaluation, and latency/interrupt handling; both speakers emphasized forward-deployed engineers embedding with customers to make it work.
Some high-security deployments require fundamentally different delivery models than cloud APIs.
Sherwin Wu described a custom on-prem, air-gapped deployment for Los Alamos National Lab on their Venado supercomputer, including physically bringing model weights in and adapting to their hardware and networking, with limited visibility into usage but feedback indicating research support and experiment acceleration.
The gap between pilots and working deployments is governance, evals, and iteration discipline.
Olivier Godement and Sherwin Wu said leading indicators for success include top-down buy-in plus a bottom-up “tiger team,” defining evals early, and iterative “hill climbing” from low initial performance to high reliability, sometimes including fine-tuning when models hit limitations.
GPT-5 optimization is framed as behavior and product usability, not just benchmark scores.
Olivier Godement said GPT-5’s impact isn’t only benchmark intelligence but also “craft” (style, tone, behavior), driven by months of close customer feedback; Sherwin Wu highlighted the trade-off between longer “thinking” for higher quality versus latency for product use cases.
WORDS WORTH SAVING
5 quotesWe literally had to bring the weights of the, the model, uh, physically into their supercomputer.
— Sherwin Wu
Physical autonomy is ahead of digital autonomy in 2025.
— Apoorv Agrawal
AI agents are just kind of dropped in the middle of nowhere.
— Sherwin Wu
I think at that point, I may have worked with like a couple of hundreds, I think, enterprises.
— Olivier Godement
It’s almost like the monkey paw where it’s like developers and, and platform customers ask for better instruction following, so we’re like, "Yes, we’ll give you really good instruction following," but it was like, you know, it follows it almost to a, to a T.
— Sherwin Wu
SPEAKERS
Sherwin Wu
guestOpenAI Platform engineering leader focused on the API, reliability, and enterprise/forward-deployed deployments.
Apoorv Agrawal
hostHost of the BG2 Pod who interviews guests about OpenAI Enterprise and platform strategy.
Olivier Godement
guestOpenAI Platform product leader working on enterprise deployments, customer feedback loops, and product direction.
High quality AI-generated summary created from speaker-labeled transcript.
