Reid Hoffman and OpenAI CFO Sarah Friar on where AI goes from here

Sarah Friar described AI as being at an early stage comparable to the Industrial Revolution, with use shifting from answering questions toward completing tasks through agents and acting as a proactive virtual coworker. ChatGPT has 1.2 billion weekly users, which she said is about one-fifth of the world’s population; 90% can access it for free. About one-third of users ask health or wellness questions each week, rising to 80–90% in some health deserts. OpenAI has also announced 100,000 advertisers across 89 countries.

The shift from experimentation to routine work is already visible in companies’ use of agents. Frontier firms use about eight times as many agentic tokens as average firms, up from roughly twice as many in January, while OpenAI’s internal use is 33 times the frontier-firm level. Friar expects frontier firms to reach OpenAI’s current level within one to two years, with traditional firms following. Companies are extending AI from coding into cybersecurity and knowledge work, then adapting it to particular industries and functions. She expects jobs to change in both their responsibilities and the tasks within them, rather than whole occupations simply disappearing. On ChatGPT, users with finance job titles spend about a quarter of their activity on financial tasks, a quarter on engineering and a quarter on marketing.

Friar said natural-language instructions let people build without coding directly, citing Linda.ai’s work with a dental receptionist named Linda: the system contacted 600 people and scheduled 250 new patients in three days. In healthcare, a proof of concept with HCA reduced patients’ time in hospital by 25% without patients getting sicker. OpenAI acquired Orchestra Bio, a platform for AI agents and scientists intended to accelerate drug discovery and delivery. Friar also cited work with Boston Children’s Hospital, which provided data and scientific literature on more than 300 unresolved diseases; OpenAI has reached answers for 40. She said a held-back frontier model solved the Navier–Stokes Millennium Math Problem and advanced roughly 350 other theorems, with possible applications including more efficient artificial-heart valves and data-center cooling.

Friar presented restraint as part of OpenAI’s approach to safety, saying the company is “pacing the frontier” and has held back its most powerful model partly to give researchers time to assess safety and alignment. She also said highly cyber-capable models have been withheld until OpenAI is confident they are safe, despite the short-term business cost. OpenAI worked with PHP to fix vulnerabilities before attackers found them; Friar said the flaws could have affected about 70% of websites. She cautioned that the security work is unfinished and threats are accelerating, while arguing that frontier cyber capabilities ultimately need to be safe to deploy for defense. She said trust is earned through visible outcomes over time, not assurances, and that workable regulation should be developed with institutions able to keep pace. In OpenAI’s research, only 10% of identified frontier firms are based in Europe.

Friar said OpenAI’s mission to benefit humanity informs decisions on access and risk, while maintaining strong finances is part of her job. OpenAI is now live on Google Cloud Platform, joining Amazon, Azure, GCP and Oracle as cloud providers through which it can serve customers; she called this a tailwind into 2027. She estimated AI is adding about one percentage point to GDP, depending on how it is measured, and expects that growth to improve outcomes nationally and globally. Data centers can bring communities skilled work, including electrician and HVAC jobs, and OpenAI helps train people for those roles. The company gave a Georgia community an $80 million grant to deploy over five years, with residents choosing its use; Friar approved $4.5 million for free school lunches, after which uptake rose 50%. She said developers build trust by listening to concerns, explaining mitigation and asking residents what they want. On water, she cited a closed-loop-cooled data center using about as much in a year as 300 households, and also cited about two Olympic swimming pools’ worth of water for a data center.