Video · 2025-09-11 · 1h 9m · 6 moments

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

✦ AI generated

timeline · colored by role

Example

Healthcare and pharma companies like Amgen have two distinct categories of AI need: accelerating pure R&D work involving massive scientific datasets, and automating the heavy administrative and document-authoring burden required to get an approved drug to market.

Olivier Godement describes working with Amgen and explains that healthcare/pharma AI opportunities fall into two buckets: R&D acceleration and automating the massive admin and regulatory documentation workload.

transcript

Olivier Godement: I feel like they are two big buckets of needs. One is, like, pure R&D. It's like, you know, you're seeing, like, a massive amount of data and, like, you have super smart scientists who are trying to, you know, come by, test out things, you know. A second bucket is, like, you know, much more, like, you know, common across other industries. It's, like, pure, like, you know, admin, document authoring, document-scribing work.

Anecdote◆

Because Los Alamos National Laboratory required extremely high security and clearance for national security research, OpenAI could not deploy its models through standard APIs and instead built a custom on-prem deployment, physically installing the o3 model's weights onto the lab's air-gapped Venado supercomputer.

Sherwin Wu recounts how strict government security clearance requirements forced OpenAI to physically install its o3 reasoning model onto Los Alamos's air-gapped Venado supercomputer rather than provide standard API access.

transcript

Sherwin Wu: Because they are so, they're a government lab, they're so, you know, restrictive and high security and high clearance with a lot of their things, we couldn't just do a normal deployment with them. They couldn't, you know, you can't have people doing national security research just hitting our APIs. And so we actually did a custom on-prem deployment with them onto one of their supercomputers called Venado.

Claim

The clearest leading indicator of a successful enterprise AI deployment is the combination of top-down executive buy-in with an empowered bottom-up 'tiger team' that mixes technical skills with institutional/subject-matter knowledge, allowed to start small and scale.

Olivier Godement, drawing on experience with hundreds of enterprise deployments, says the top success pattern is pairing top-down leadership buy-in with a bottom-up tiger team of technical and institutional experts that starts small and scales.

transcript

Olivier Godement: What I've seen being like clear leading indicator of success. Number one is like the interesting combination of like top down like buy in and like enabling like, you know, very clear group of like a tiger team, essentially, like, you know, the enterprise which sometimes a mix of like OpenAI, like, you know, enterprise employee.

Mechanism◆

Self-driving cars have advanced further than AI agents in the real world not because digital tasks are technically easier, but because roads, stoplights, and traffic laws already give self-driving cars standardized scaffolding to operate within — scaffolding AI agents dropped into arbitrary digital environments don't have.

Sherwin Wu argues physical autonomy (self-driving cars) has outpaced digital autonomy (AI agents) partly because roads and traffic laws provide standardized scaffolding that AI agents, dropped into unstructured digital environments, still lack.

transcript

Sherwin Wu: I actually do think self-driving cars have a good amount of scaffolding in the world for them to operate in. Like not completely unlimited. You have roads, roads exist, they're pretty standardized. You have stoplights. People generally operate in pretty normal ways. And there are all these traffic laws that you can learn.

supports · 1

Mechanism

A core, still-unsolved trade-off in building GPT-5 is between how long the model reasons (reasoning tokens/thinking time) and response latency versus intelligence — models get dramatically smarter given more inference time, which conflicts with product needs for fast responses.

Sherwin Wu identifies the reasoning-time-versus-latency trade-off as the hardest and still-unresolved design decision behind GPT-5, since more thinking time makes the model dramatically smarter but users won't always wait minutes for an answer.

transcript

Sherwin Wu: I actually think a very clear trade-off, which I honestly think we are still iterating on, is the trade-off between the reasoning tokens and how long it thinks versus performance. And honestly, this is something that I think we've been working on with our customers since the launch of the reasoning models, which is these models are so, so smart, especially if you give it all this thinking time.

gives example · 1provides context · 1supports · 1

Definition

Reinforcement fine-tuning (RFT) is fundamentally different from supervised fine-tuning (SFT) because it introduces reinforcement learning into model customization, rather than relying on labeled prompt-completion pairs that directly supervise the model's outputs.

Sherwin Wu explains that OpenAI's reinforcement fine-tuning (RFT) product differs from the original supervised fine-tuning (SFT) API by introducing reinforcement learning rather than requiring labeled prompt-completion pairs, making it more complex but far more powerful.

transcript

Sherwin Wu: So it really, it's introducing reinforcement learning into the fine-tuning process. So the original fine-tuning API does something called supervised fine-tuning, we call it SFT. It is not using reinforcement learning. It is, it's using supervised learning. And so what that usually means is you need a bunch of data, a bunch of prompt completion pairs. You need to really supervise and tell exactly the model how it should be acting.

Highlight slides

Why OpenAI Built an On-Prem Deployment for Los Alamos✦ from: Because Los Alamos National Laboratory required extremely high security and clearance for national security research, OpenAI could not deploy its models through standard APIs and instead built a custom on-prem deployment, physically installing the o3 model's weights onto the lab's air-gapped Venado supercomputer.Physical autonomy outpaces digital — not by difficulty, but by structure✦ from: Self-driving cars have advanced further than AI agents in the real world not because digital tasks are technically easier, but because roads, stoplights, and traffic laws already give self-driving cars standardized scaffolding to operate within — scaffolding AI agents dropped into arbitrary digital environments don't have.The scaffolding advantage for self-driving✦ from: Self-driving cars have advanced further than AI agents in the real world not because digital tasks are technically easier, but because roads, stoplights, and traffic laws already give self-driving cars standardized scaffolding to operate within — scaffolding AI agents dropped into arbitrary digital environments don't have.

Related episodes

Video · The Cognitive Revolution

Healthcare and pharma companies like Amgen have two distinct categories of AI need: accelerating pure R&D work involving massive scientific datasets, and automating the heavy administrative and document-authoring burden required to get an approved drug to market.

The clearest leading indicator of a successful enterprise AI deployment is the combination of top-down executive buy-in with an empowered bottom-up 'tiger team' that mixes technical skills with institutional/subject-matter knowledge, allowed to start small and scale.

A core, still-unsolved trade-off in building GPT-5 is between how long the model reasons (reasoning tokens/thinking time) and response latency versus intelligence — models get dramatically smarter given more inference time, which conflicts with product needs for fast responses.

Reinforcement fine-tuning (RFT) is fundamentally different from supervised fine-tuning (SFT) because it introduces reinforcement learning into model customization, rather than relying on labeled prompt-completion pairs that directly supervise the model's outputs.

Fable's Back, AI Engineer Recap, & SambaNova

Is Defense the Next Trillion-Dollar Category? | a16z American Dynamism Summit

Kavak's Playbook for Rebuilding a Company Around AI