Speeding Up Science (with link)
AI for DC
Hi
With corrected link for OpenAI for Science — apologies!
Welcome (back) to The Prompt. AI isn’t well understood, but we learn a lot in our work that can help. In this newsletter, we share some of these learnings with you. If you find them helpful, make sure you’re signed up for the next issue.
[News] Doing science faster, from our VP of science
We’re at the beginning of something new, where AI systems are starting to meaningfully contribute to science itself.
In the past year, we’ve seen GPT-5 help researchers make measurable progress on real problems. Mathematicians have used it to produce correct proofs in minutes, physicists to guide complex simulations, and biologists to explore protein designs that would have taken months to develop on their own. In some cases, GPT-5 has surfaced connections between ideas that scientists hadn’t made in decades. It’s not “solving” science, but in the hands of experts, GPT-5 is helping scientists move faster, test more ideas, and discover what was once out of reach.
While earlier AI systems worked within the boundaries of known information, we’re seeing GPT-5 start to think beyond them, generating creative and novel – though sometimes imperfect – ideas. We expect this trajectory to continue, expanding what scientists can explore and discover.
That’s why we’re launching OpenAI for Science: to build the next great scientific instrument by combining frontier AI models with the tools and workflows of science. Our goal is to help scientists push the frontier of knowledge further and faster – compressing decades of discovery into a few years. About 3 in 5 Americans are concerned about how long it takes for scientific and medical breakthroughs to reach them. This is one answer.
OpenAI for Science is built in part on the foundation of our partnerships with the US National Laboratories, where researchers are testing how AI can accelerate progress across energy, materials science, astrophysics, national security research, and more. “Advanced materials” include new classes of engineered substances that are lighter, stronger, more conductive, or more durable, enabling breakthroughs in batteries, semiconductors, superconductors, and catalysts.
Across labs such as Argonne, Berkeley, Los Alamos, and Oak Ridge, scientists are using AI to model complex systems, analyze vast datasets, and identify promising directions for real-world experiments. These efforts have shown that when the best of American science meets the best of American technology, AI can be a force multiplier for innovation and help the US stay at the fore of scientific leadership in a world where discovery is key to both prosperity and security.
This work is grounded in our mission to ensure that artificial general intelligence benefits all of humanity. Science is one of the clearest, most direct ways that mission comes to life. The results of scientific progress are visible, testable, and shared. And when discovery moves faster, everyone benefits. Most Americans see AI in science and medicine as just common sense: 73% say we need new and innovative solutions to accelerate breakthroughs, and 69% call US leadership in science a top national priority.
If we can help compress decades of progress into years, the impacts will reach across health, energy, climate, and security, strengthening the foundations of science and the institutions that depend on it. We couldn’t imagine a more exciting or meaningful challenge. We’re just getting started, but we are deeply optimistic about what’s to come. – Kevin Weil, Vice President, OpenAI for Science
Learn more about OpenAI for Science here. We’re also inviting scientists who are using AI to help tackle hard problems to reach out here.
[Data] The common sense of AI in science
[Policy] Proposals for faster discovery
OpenAI sees federal government investment in AI-enabled scientific discovery going hand-in-hand with investment in AI-driven advanced manufacturing and supply chains to usher in America’s next great industrial era and widespread economic opportunity. Among policy ideas and proposals for how the federal government can support faster science:
Think bigger than any one bill. The American Science Acceleration Project (ASAP), led by Sens. Mike Rounds (R-SD) and Martin Heinrich (D-NM), is a national campaign seeking to bring together government, industry, academia, philanthropy and other policy thinkers to make American science 10x faster by 2030. The Project calls for “an equally ambitious investment in data, compute and AI to build a ‘superhighway for science,’” enable broader collaboration; and streamline the deployment process by reducing unnecessary obstacles to innovation.” OpenAI is a proud supporter of ASAP.
Identify metrics to measure the speed of scientific innovation. Metrics to track inputs, uptake, and outputs could help clarify where science is accelerating and where further support may be needed:
Deployment and access to advanced AI infrastructure: Track how many institutions have operational access to large-scale compute resources and the extent to which those resources are used in active research projects. Wider access and higher use indicate that cutting-edge infrastructure is reaching more scientists.
Uptake of AI modeling in federally funded research: Measure the proportion of federal research grants that budget for, or report using, AI models. Growth in this share suggests that AI methods are moving from pilots to routine practice across scientific disciplines.
Adoption of AI tools within government science agencies: Track the agencies that have piloted or embedded advanced AI systems in their core scientific workflows. Increasing adoption shows that the federal research enterprise is itself becoming an early user and validator of emerging capabilities.
Leverage shared compute platforms and open datasets: Monitor the number of researchers who log into compute resources enabled by public-private partnerships, as well as the volume of open scientific data they access. Increasing participation and data consumption reflect democratization of key inputs to innovation.
Patent activity for AI-enabled innovations: Track the annual number, and growth rate, of patent applications and grants that involve AI technologies – showing how scientific advances translate into protectable inventions.
Peer-reviewed publications on AI: Measure the count and year-over-year change in AI-related papers appearing in scholarly journals and on pre-print archives. A sustained rise indicates robust research activity and distribution.
Researcher-to-discovery efficiency: Calculate the ratio of active researchers (or total R&D spending) to quantifiable discoveries, such as new compounds, materials, or algorithms.
Digitize government data currently in analog form. Another way to unlock faster scientific discovery is to unlock government data, much of which is in the public domain but in analog form. Making it more accessible or machine-readable could help American scientists discover new insights and discoveries that help advance public health and preserve US scientific leadership. It also would democratize access to a taxpayer-funded asset for everyone’s benefit. This is an idea we submitted to the Trump Administration for consideration for its AI policy agenda.
[Feedback] Tell us your AI story
Now that you’ve gotten this far in the newsletter, we want to hear from you. Tell us your own personal story about how you see AI making a difference in the world – whether it’s through science, healthcare, education, the workplace, or your home. Submit your stories and examples to our email address ga@openai.com, or post them as comments.
[About] OpenAI Academy
The Academy is OpenAI’s free online and in-person AI literacy trainings for beginners through experts.
OpenAI has called for a nationwide AI education strategy – rooted in local communities in partnership with American companies – to help our current workforce and students become AI-ready, bolster the economy, and secure America’s continued leadership on innovation.
12:00 PM – 1:00 PM EDT on Nov 13
[Disclosure]
Graphics created by Base Three using ChatGPT.