OpenAI’s Platform Bet: Where AI Value Accrues, Why Open Source Trails, and When Superintelligence Arrives

Jacob Effron asks where value will ultimately accrue across the model, infrastructure, and application layers. Peter Welinder says the simple answer is that value will accrue in many areas, but he suspects most of it will land at the application layer. He frames OpenAI's mission as making sure AI is safe and benefits all humanity, with a central part of that being enabling as many builders as possible rather than extracting a lot of value at the model and infrastructure layer.

Welinder points to OpenAI's dramatic price cuts as evidence of that approach, citing an earlier cut of roughly 70% and the release of GPT-3.5 Turbo, which was about 10x cheaper than the original 3.5 model and much faster. He says a lot of research and engineering is aimed at bringing prices down because lower prices broaden access and make the models applicable to more applications. Tooling also matters, but he expects much of it to come from the developer ecosystem: OpenAI had about 400 people, while millions were already building on the technology.

He expects value capture at the application layer to follow familiar business patterns such as network effects and branding. He cites the Bill Gates idea of a platform where the value generated on top is much larger than the value captured by the platform provider, and says that is OpenAI's aspiration. In that sense, he treats AI as another technology where many of the same principles apply.

OpenAI wants to stay at the platform level rather than building many vertical first-party applications. Welinder describes ChatGPT as a very general consumer product, not an attempt to be the best AI teacher, doctor, or lawyer; those applications will be built by companies that add proprietary content, domain expertise, and integrations. He points to plugins as the first step in connecting ChatGPT to external services and says OpenAI wants to be a platform at multiple levels, providing both reasoning and a route for companies to reach consumers.

Jacob Effron notes the pace of OpenAI's shipping—plugins, web browsing, and mobile—and asks how the team prioritizes. Welinder says OpenAI made a big, top-down bet on how general language models would become, concentrating most people, compute, and GPUs on that direction instead of spreading across too many bets. He notes that OpenAI used to have other bets, such as robotics and beating world champions in Dota 2, which pushed technology forward and revealed viability; as confidence in language models grew, most effort moved there.

The flexibility of language models is what lets OpenAI build features like browsing and plugins quickly, according to Welinder. He says that once you have that platform layer, the other requirement is smart, driven people—researchers and engineers who can take ideas from research all the way to product. He argues that many research-focused people at OpenAI are motivated to put work in users' hands and learn from it, and that browsing and plugins came from researchers' visions for building on language models. Plugins, in particular, unified several efforts around browsing, code interpreter, and connecting to external APIs so research teams could ship faster.

Effron asks about autonomous agents as the next step for large language models. Welinder says OpenAI already has something agent-like in plugins: a plugin is a little mini-agent that takes a task and connects to a sequence of APIs—for example, finding a recipe and putting the ingredients on an Instacart shopping list. He expects more complex tasks to become possible where the agent only checks in when it is uncertain about what it is supposed to do.

Welinder says the API side of OpenAI's product lets developers push agents much further and explore faster than OpenAI can. He points to open-source agent projects as inspiring examples of what people are doing. At the same time, he describes plugins as early and rough around the edges; they need to be more robust and smarter, and he expects generational model improvements over the next year or two to make many of these behaviors work.

OpenAI tries to evaluate its work through the lens of what the product should look like once AGI exists. Welinder's analogy is hiring an employee: you give big directions, the employee goes off to work, comes back with questions, shows you something, and incorporates feedback. He says that if OpenAI is building AI, that employee-like interaction is roughly what the product should evolve toward in a couple of years.

Rob Toews asks how open-source and open models fit into the broader LLM ecosystem, especially after the explosion of activity around Llama and its derivatives. Welinder says he may hold an unpopular opinion: open-source models may catch up in the very long run, but he expects that to play out on a longer horizon. He compares it to desktop Linux, which has existed for about 30 years but has not caught up to macOS, iOS, and Windows because there is never enough investment in all the details. He also argues that the capital and engineering required for training and inference at scale are hard to replicate openly, and that incentives and safety concerns make companies investing heavily unlikely to open-source their best models.

Welinder is not dismissive of open source: he says it is extremely useful for pushing research forward, trying new training and application approaches, and serving on-device, edge, on-prem, or smaller-model use cases. But for applications that need reliability and the smartest model, he thinks the rational choice will be a proprietary model, which will likely stay a couple of years ahead for the foreseeable future. He adds a nuance: narrow tasks such as summarization may not need the smartest model, but products tend to become more general over time, so a specialized open model can hit a wall at edge cases. Customer service is an example where handling all the edge cases requires the smartest model.

Toews asks whether open source represents a meaningful competitive risk to OpenAI from a business and addressable-market perspective. Welinder says no, because he thinks most value will be in the smartest models; he compares it to hiring, where you always try to hire the best people. A company choosing a dumber model would be at a competitive disadvantage to one choosing a smarter model, so OpenAI's bet is that value accrues where the most economically valuable problems can be tackled. He expects that to start with co-pilots for professions such as law and medicine and eventually extend to AI scientists working on new drugs or climate solutions.

Smaller models still have a role, especially for embeddings, retrieval, and grounding, where latency and control matter and recomputing embeddings can be expensive. Welinder says OpenAI looks for open-source investments that enable broader use of smarter models; Whisper is an example, because accurate transcription lets people feed audio into language models and build applications that were previously impractical. He says OpenAI does not expect to make meaningful money on speech recognition but wants to enable more use cases.

Toews asks what gaps are preventing enterprises from going full steam with these models. Welinder says the biggest gap right now is hallucination: the models can invent facts, so they cannot be entirely trusted, and this comes up repeatedly. He describes hallucination as an open research problem at the frontier and says OpenAI puts a lot of effort into making its models more robust.

In practice, Welinder says more companies are getting around hallucinations by grounding models in external data. A common pattern is to embed the question and company documents, use a vector search database to find relevant documents, give those to the language model, and have it summarize the relevant information into an answer. He says instructing the model to say 'I don't know' if it cannot find the answer works really well, and he calls hallucination the number one issue he has seen in applications.

Effron asks whether OpenAI waits for the ecosystem to build tooling around problems like this or solves them itself. Welinder says OpenAI listens to developers about obstacles and may build common tooling to some extent so teams can get off the ground. But OpenAI prioritizes what it is best in the world at—training models and making inference really good—because lower prices, higher reliability, and lower latency are the fundamentals developers ask for. Plenty of other companies are building good tooling on top, but without those fundamentals the tooling does not matter.

Toews lists a range of concerns: misinformation and deepfakes, bias, job displacement, and the possibility of a runaway powerful AI posing an existential risk. Welinder says the first few risks are fairly surmountable; misinformation becomes a problem at scale and depends on distribution channels such as social media or email, where infrastructure already exists to protect against it. On bias, he says it is impossible not to have bias in models, so OpenAI wants to give developers and users tools to instruct the model toward the behavior they want within broad bounds, rather than imposing a particular political orientation.

The risk he thinks is receiving too little attention is superintelligence. He says the field is moving toward AGI, but the next step is models that are much smarter than humans, and the timeline is unknown. He finds it striking that there is surprisingly little research on how to ensure that outcome is good—there are teams at OpenAI, but no superintelligence safety departments at universities—and he says the outcome could be existential for humanity. He argues for both technical work on alignment and control and regulation that helps governments understand who is training such models and how much compute they are using.

Effron asks for a speculative prediction about when superintelligence might arrive. Welinder defines AGI as autonomous systems that can do economically valuable work at the level of humans, and says OpenAI has a shot at that before 2030. He acknowledges unknown blockers and external events, such as a GPU shortage, but says the field seems to be moving almost automatically because of the pace of innovation. He adds that he has been in the field for about 15 years and something feels very different now; people who have been in the field much longer say the same.

Welinder would not be surprised if something resembling AGI exists by the end of this decade. He notes that superintelligence can be defined along different dimensions, such as running much faster than humans or running many more experiments in parallel, and some of those aspects may arrive before others. He also raises the possibility that a superintelligent system is too expensive to run economically at first, requiring another five to ten years, but his speculative bet is that around 2030 we will start seeing early signs of superintelligence.

Effron asks about Geoffrey Hinton's public warnings and how worried Welinder is about an extreme downside where humanity's survival is at risk. Welinder says he is probably a little more optimistic that humanity will figure it out together, comparing it to the fact that there has not been a nuclear war because of some level of self-preservation. He expects that as we get closer to AGI and superintelligence, we will tread much more carefully.

OpenAI's strategy, as Welinder describes it, is to build the right organizational processes and frameworks before superintelligence arrives. He says it is important to release models when the stakes are low so the field can learn from risks such as misinformation and bias, because it would be wrong to wait until superintelligence exists and then try to make it safe. He notes that OpenAI held back GPT-4 for almost half a year to gain clarity on potential downsides. He also argues that a leading lab setting an example adds accountability for others.

Welinder emphasizes the upside as well: if superintelligence can solve problems such as climate change, cancer, and aging, it could create more abundance and a higher standard of living for everyone. He says that is why it is important to get the technology out, but to do it carefully. He frames safety and beneficial outcomes as linked rather than opposing goals.

Effron asks what Welinder would focus on if he ran a well-funded university department for superintelligence safety. Welinder names interpretability—understanding what is happening inside still-black-box models—and points to work by people such as Chris Olah, now at Anthropic, as well as the possibility of moving faster than computational neuroscience because these models can be studied directly. He also wants much crisper work on alignment and specification: how goals and guardrails are defined, with collaboration among technical people, social scientists, and philosophers. On technical approaches, he mentions shaping reward functions for reinforcement learning and having one model oversee another model's actions, while saying he does not claim to know which approach will prevail. He argues society should invest much more in getting many more smart people to work on these problems.

Effron asks how OpenAI uses ChatGPT internally. Welinder says ChatGPT has a broad set of use cases, and users usually have a dozen or more places in their workflows where they turn to it for answering questions, summarizing, or helping with writing. Internally, he sees a lot of use for coding, including debugging by pasting in stack traces. He personally uses it to get past writer's block, improve his writing, and draft emails from a rough description of the point he wants to make.

Asked for a belief he holds strongly that most people would disagree with, Welinder returns to superintelligence. He thinks we should be thinking much more about its implications, and he does not think it is generally accepted that we are heading toward it. He says people keep identifying things models cannot do, and then the models can do them within a short amount of time, so he believes we should act on a potentially much shorter timeline and have a plan for it.

Effron asks what would most likely explain OpenAI losing its leadership position five years from now. Welinder says his biggest fear is losing touch with users and developers. He describes a tension in which developers build something good and then a new OpenAI model generally improves enough to replace some of what they built, even though that is not intentional. He worries that rapid growth makes it harder to attend to every developer—he remembers knowing early API customers by first name, while now there are millions of customers—and says scaling that customer experience is critical to the mission.

For follow-up, Welinder points listeners to OpenAI's blog, which he says many ChatGPT users may not know about. He says OpenAI uses it to demystify the technology, explain how the company thinks about safety and the approach to superintelligence, and describe the technical approaches underpinning the models. He encourages people to read it and says the posts are meant to be approachable for a more general technical audience.