Brad Lightcap on OpenAI’s Early Years, Agents and the Future of Software

OpenAI COO Brad Lightcap joins Jack Altman to look back on the company’s early research years and discuss agents, software jobs, startups and his use of Codex. The discussion was recorded in early 2026; forecasts, business figures and product descriptions are the speakers’ contemporaneous claims.

Altman asks Lightcap to divide his time at OpenAI into chapters. Lightcap joined in 2018, when the organization was much less widely known and its work was followed by small technical communities. He recalls people asking what the company did.

Lightcap had worked with Sam Altman at Y Combinator, focusing on companies outside conventional software, including fusion, satellites and biotechnology. He says Sam asked him to help find someone to run non-research work at OpenAI; he eventually began helping himself.

Spending time with researchers convinced him that scaling general architectures might lead to more capable systems. He stresses that at the time he was not qualified to judge whether the idea would hold, but thought its implications would be important if it did.

He describes an intensely research-centered early culture. His operational work ranged from securing capital and building supercomputers with partners to sourcing parts for broken robots. He says the common aim was to accelerate research.

Before ChatGPT, Lightcap says, there were “sparks”: models were beginning to support conversation, and users of the completions API were already trying to interact with them in turns. Early DALL-E adoption also suggested that people wanted a direct, prompt-based interface.

The team expected ChatGPT to matter but underestimated its scale, he says. His own planning guess was a peak of one million concurrent users. That anecdote describes his recollection of internal expectations, not a published usage measurement.

Lightcap divides the technology’s development into a scaling period before mass adoption, a chatbot period from roughly 2022 to 2024, and an emerging agent period in which systems can use tools and work asynchronously. These are his categories, not fixed industry boundaries.

He believes diffusion into the economy can take much longer than a model release: even if technical progress stopped, he imagines years of adoption and innovation. He treats the mismatch between capability development and adoption as a question to watch.

Asked where agents’ capabilities might end, Lightcap says he does not know. He points to unresolved problems such as retaining context and memory over long tasks, and speculates about systems coordinating with other agents. Neither speaker presents a demonstrated endpoint.

Altman contrasts earlier science-fiction discussions with the practical experience of using powerful software. Lightcap sees parallel conversations about near-term productivity and longer-term implications. Both note how quickly people grow accustomed to capabilities that initially seemed surprising.

Lightcap cites a reported case of an owner using AI and working with a lab on an experimental treatment for a dog’s cancer. He presents it as an example of individual empowerment. The interview does not establish efficacy or a cure, and an anecdote is not evidence that such a treatment works generally.

Altman asks about the gap between Silicon Valley enthusiasm and skepticism elsewhere. Lightcap acknowledges that the future is uncertain and says the industry has failed to communicate a compelling picture of potential benefits to many people.

Lightcap expects lower costs and shorter paths from an idea to an implementation to help individuals create products, conduct research and serve customers. He also recognizes potential harm and says institutions must address downsides; his confidence about the eventual balance is an optimistic forecast.

The speakers turn to coding and employment. Altman mentions data he has seen about engineering job postings, but the interview does not provide a source or establish a general labor-market trend. He asks whether better tools will expand the work or replace workers.

Lightcap argues that when a service becomes cheaper, demand can increase, and he expects software engineers’ work to change rather than simply disappear. Altman notes that engineering costs are lower, not zero, and that companies may use savings to build better products. Neither prediction guarantees employment outcomes.

Lightcap sees many businesses and public institutions using inadequate or outdated software. He cites hospitals, power systems and hotels as areas where modernization and security matter. He argues that AI-assisted development could extend software to underserved problems and help improve vulnerable systems.

He expects people to remain responsible for designing, implementing and maintaining a potentially much larger amount of software. His estimate of how much software might be created is speculative; the discussion does not demonstrate that generated code is automatically safe.

Asked about Codex’s recent progress, Lightcap credits the focus of the product and model teams and faster iteration. He gives early-2026 figures for GPT-5.4 usage and revenue run rate; these are his contemporaneous company claims, not independently audited results or current figures.

Altman asks how OpenAI manages work across many areas. Lightcap says the company does not want to confine itself to a single consumer, enterprise, hardware or software category; it experiments, expands efforts that work and moves people away from those that do not.

He describes research as an expansion-and-contraction cycle: many projects explore possibilities, a smaller number receive more resources, and new bets later emerge. He thinks product development and deployment can follow a similar pattern.

Altman imagines a more unified experience across home and work. Lightcap says users currently have to make too many choices about models and modes; he hopes systems will eventually allocate the work more automatically. This is a stated direction rather than a promise about a particular release.

Altman asks where startups can build durable businesses if a model provider might add a competing feature. Lightcap says founders should look for specific hard problems newly made feasible by improving models, rather than sitting directly where general model capabilities are concentrated.

He describes renewed energy among founders and uses a ripple metaphor: as a model improves, opportunities spread outward into areas that were previously impractical. Examples he imagines include modernizing old code and serving overlooked industries.

Altman worries that the next model improvement could overtake an earlier opportunity. Lightcap says close knowledge of a user, a problem and how an industry currently serves it remains valuable. He returns to the basic advice to talk to users.

The speakers observe that a large market can support multiple products built on similar capabilities. Altman says some founders are willing to discard an existing product while retaining the team’s knowledge and customer relationships if assumptions change.

Lightcap says some established founders are effectively restarting part of their companies to explore a new technical foundation. Existing teams and customer relationships can be a head start, even though abandoning prior work is difficult.

Asked about a sell-off in public software companies, Lightcap avoids a specific market call. He says the companies he sees have customer relationships and domain knowledge and are actively reconsidering their products. His positive assessment is a view of their prospects, not investment advice or a guarantee.

Altman asks how people used to pre-AI workflows can adapt. Lightcap recommends direct experimentation and says Codex has become a daily tool for him, even though he does not work as a software engineer. He values its ability to create a first version of a task.

As an example, Lightcap describes recruiting for a forward-deployed engineering organization. He says he asked Codex to examine public profiles and writing from a list of candidates and rank them against technical requirements, surfacing several people for closer review.

He says the result helped him decide whose work to inspect, rather than replacing his own examination. The interview does not evaluate the accuracy or fairness of such rankings; public online presence is only a partial view of a candidate.

Lightcap describes forward-deployed engineers as a way to solve specific business problems that historically did not justify custom software. He expects AI to reduce the time and cost of producing tailored solutions, while acknowledging that the role of the engineers will differ from earlier versions.

Altman asks about working with his brother Sam Altman. Lightcap says they have worked together for about ten years, beginning at Y Combinator. He describes Sam as more comfortable in small technical discussions than in a public role and as someone focused on long time horizons.

The two say differences in time horizon can make discussion of future technology disorienting. Lightcap says customers, friends and family are still trying to understand current tools while company leaders are thinking about what might come next.

Lightcap says OpenAI’s mission remains a way for the team to evaluate which projects to pursue. He describes the work as highly incomplete and says the organization repeatedly asks whether an activity advances its intended outcome. Altman thanks him for the discussion.