After the OpenAI Saga: Governance, Incentives, Diffusion Models and Why Your ASP Is My Opportunity

One host calls the outcome broadly positive: OpenAI is back in a stable position, still has what they regard as the leading model in GPT-4, and has reworked its board. They add that it was better for this to happen now than years later, and that governance of the nonprofit and the company should be more stable as a result. The other host's first lesson is that governance matters, and that entrepreneurs will think twice before placing their fate with groups that have explicitly mixed incentives.

The second lesson is that money matters. The factors of production are labour and capital, with compute as the AI-specific form of capital, which Microsoft held, while the team rebelled and the board underestimated the support Sam Altman and Greg Brockman had. The host adds that many OpenAI staff were upset by the destruction of the mission and also of the value they had been promised through the $86 billion tender offer, a reminder that stakeholders carry leverage.

The conversation then widens to boards and executive teams generally. The hosts say founders should ask who sits on their board, what expertise and strategic view each member brings, and what incentives they have, noting a "professional managerial class" whose loyalties lie with outside rewards such as conferences and awards. They cite Shopify as a company that resets this by being performance-based and mission-focused, and Munger's saying that every time you think you understand incentives you find you were underestimating them.

On the commercial side, the hosts expect more interest in owning models and in open source, and in reducing reliance on one vendor. They note that some builders shipped tools during the week, such as Braintrust's AI proxy that lets the OpenAI SDK query other models including Mistral and Llama via Perplexity, so teams can load-balance and compare performance. In their view such tools were coming anyway, and the episode only sped up the market.

The typical journey, one host says, is to prototype on GPT-4, then either stay for advanced reasoning or high throughput, move to GPT-3.5, or try fine-tuning, and few outside very large scale or proprietary data actually fine-tune. That implies an orchestration layer, and it clarifies the dimensions on which to judge a provider: reliability, latency, cost and control. OpenAI still leads on capabilities like code generation and GPT-4V, and is a great partner, they say, but the questions are now more obvious.

The other host adds that large enterprises have always wanted a second source for negotiating leverage, as in networking, where Juniper exists because buyers want to push on Cisco, or AMD against Intel. So there will always be a vendor or two beside the leader for advanced LLMs. If anything, OpenAI emerges more stable and more focused, since board-level instability has been removed.

Turning to video, one host recalls that a year and a half earlier Midjourney, Stable Diffusion and DALL-E 2 sparked a wave of image-generation startups, until ChatGPT pulled most entrepreneurs to LLMs. A handful of founders stayed with diffusion, and now products are appearing, such as Pika, started by two Stanford PhD students, a text-to-video engine, as well as avatar-cloning companies like HeyGen aimed at marketing and internal training. Audio tools are emerging too, such as music tools or adding voice to a track.

The host calls it a renaissance, partly because founders did not get distracted by LLMs, and expects the traction of these companies to draw talent back. They like the dynamic in which showing something is possible floods an area with talent and innovation, similar to Mistral's small open models. Diffusion is also cheaper: data is hard but easier to get than the whole internet, and these models cost in the millions rather than tens of millions to train, with the original Stable Diffusion model reportedly costing a few hundred thousand dollars of GPUs.

That suggests that in language, most people build on a few platforms, while in image and video more teams may grow their own model. Still, they advise trying Stable Diffusion first and building on someone else's LLM for most things initially. Both hosts stress that video generation is hard, because of short training clips, captioning, expense and temporal coherence, so there are many technical unknowns, yet teams like Pika's and early Midjourney's are small, which shows how few people you need.

One reason people moved from diffusion to LLMs, the hosts say, is that enterprises hold so much text and code, while image, video and audio felt more verticalised. They now think creative fields are quite commercial. Early critics of Midjourney asked how many people want to make images, but its scale has shown that was wrong.

They argue that tools like Pika, HeyGen and Midjourney make the pie bigger by serving all creators, not just film professionals, with uses in communication, marketing and advertising. One host points to the market capitalisation of incumbents such as Adobe as a sign that the creative world is not small, and says many past creator-economy companies failed for different reasons. Enterprises use enormous amounts of imagery and video externally and internally, for slides and communication.

The hosts note that markets like this expand and contract at once: more people can do the task, while the cost per task falls to a tenth or a hundredth. One recalls the saying that your margin is my opportunity, while the other says it is really about turning services revenue, which is people-intensive and low margin, into software revenue. Using rough numbers they say are partly made up, software spend may be around $500 billion a year against $3-5 trillion in services, and an 80% margin on $2 trillion beats 30% on $5 trillion.

As an illustration they describe Anduril in defence: a cost-plus model pays a small percentage on top of a million-dollar drone, while Anduril sells a similar or better drone for far less at a higher margin, earning comparable dollars. The host edits the quote to "your ASP is my opportunity", comparing it with SaaS democratising CRM from a hundred large enterprises to tens of thousands of buyers. For video generation, the price falls from hundreds or thousands of dollars an hour to single-digit dollars a minute, widening the audience.

They respond to a skeptic's joke that only five AI businesses have real traction, naming foundation models, Midjourney, Copilot and inference platforms, with a comparison to the internet in 1996-97, when the list of real use cases was short and much looked dumb. Early waves like the internet, mobile and cloud had easy early wins, then many ideas that failed, because it is hard to tell good ideas in a new market. One host says they would take the other side of the joke, since so much traction in such a short time is a very positive sign.