OpenAI President Greg Brockman on AI's Impact on Cybersecurity, Drug Development, and Workflows

OpenAI is committed to ensuring that Artificial General Intelligence (AGI) benefits all of humanity, with a core belief that humans must remain in control of the future, using technology to amplify their agency. According to Greg Brockman, this involves developing technology that is not only reliable but also secure, especially as the barrier to entry for software becomes incredibly low.

Brockman suggests that security spending should 'vastly outstrip' software spending, as the fundamental dynamics of cybersecurity are undergoing a significant shift. He believes there's an opportunity to create a much more secure world by internalizing how much these 'physics' are changing.

Greg Brockman noted that OpenAI operates on the principle that progress in AI 'never gets easier, you just go faster.' This sentiment reflects the constant and increasing pace of activity within the AI field, where every perceived peak of impact reveals yet another level of possibility.

Brockman emphasized that this rapid progress isn't confined to a single company or technology but is a collective societal effort to adopt AI, leverage its full potential, and improve everyone's lives. He views recent advancements, such as the math results, as significant milestones in this continuous acceleration.

Greg Brockman highlighted that technological advancement and scientific discovery have been the primary drivers of improved human quality of life over the past thousand years. He asserted that AI now definitively possesses the ability to massively accelerate scientific discovery, a capability that was once questioned.

Brockman stated that it's now 'impossible to deny' that AI can generate new knowledge. While AI may not yet be formulating entirely new theories, it is capable of resolving century-old mathematical problems.

The focus now shifts to what humans can achieve when equipped with these powerful AI tools. Brockman believes that AI significantly amplifies human leverage, making previously unthinkable results aligned with human goals and desires accessible to everyone, which he finds incredibly exciting.

OpenAI has resolved nearly 400 significant open math problems, and the mathematical community is currently analyzing these solutions for deeper insights. Greg Brockman noted that domain experts are finding the AI's approaches surprising and insightful, with unexpected applications of ideas.

Brockman contrasted current AI outputs with previous 'incomprehensible proofs' in mathematics, such as the Four Color Theorem, which relied on mechanical case-checking without providing elegant insights. He emphasized that the AI's work offers 'beautiful insights' that are human-understandable.

He expressed optimism that as AI models develop a better 'theory of mind'—the ability to understand what humans know—they will guide humans toward new understanding. This evolution could make complex mathematical concepts more accessible, echoing how tools like ChatGPT have helped individuals learn new math.

AI is proving to be a valuable tool in helping people understand complex domains, connect information, and improve their comprehension of the world and each other. Greg Brockman pointed out that technical mechanisms can enable AI to solve problems and identify connections across different fields.

To ensure human oversight and safety, OpenAI has been discussing techniques like 'debate' and 'iterative amplification' since 2017. These methods are designed for scenarios where AI solves problems too complex for humans to easily check.

Brockman provided an example of an AI summarizing a book, a task difficult for humans to verify without reading the entire book. With a 'debate' technique, a second AI could attempt to prove the summary is flawed by identifying a crucial plot point missed in the summary. This allows humans to exercise oversight without performing the entire task themselves, aligning with OpenAI's goal of keeping humans in the driver's seat.

Greg Brockman stated that if the reasons for AI's capability in one dimension are understood, similar results can be expected in other fields of human endeavor with similar properties. He cited mathematics as a domain where AI has achieved incredible results due to high model intelligence and available data.

He clarified that while early AI applications focused on easily verifiable arithmetic problems, current AI is tackling more complex 'natural language proofs,' which are much harder to check for accuracy compared to formalized proofs. However, the core insight remains data availability.

Brockman noted that mathematics and theoretical physics benefit from abundant data. The challenge for biology, which often lacks such extensive datasets, is whether it can generate or access the necessary data to leverage similar AI capabilities, despite AI techniques that can work with less data.

Biology faces significant data bottlenecks compared to fields like mathematics and theoretical physics, where extensive data is readily available. Greg Brockman noted that despite advancements in techniques for working with less data, the fundamental physics of deep learning (compute + data + algorithms = capability/results) still apply.

AI can automate non-biological tasks within the drug discovery process, such as writing regulatory filings, which can take up to a year. This automation can significantly reduce bottlenecks that are not directly related to deep biological theory.

The drug development process is broadly divided into early-stage molecular discovery and late-stage clinical trials. Brockman emphasized that innovation in 'wet labs' (experimental biology) and simulation will become incredibly important in areas where fundamental data currently does not exist.

He stressed the need to examine every step of the drug discovery process. Some areas will see rapid acceleration with AI, making previously painful tasks incredibly easy, but other bottlenecks will emerge, requiring new solutions beyond just AI, especially in data generation and experimental design.

Greg Brockman highlighted a fundamental difference in human aptitude: while humans are naturally well-suited for natural language, biology remains 'totally foreign' to them due to its complexity. He suggested that for a neural network, there's no inherent reason it should be better at natural language than at understanding biology.

This perspective implies that neural networks, given their success in natural language processing, should also be capable of decoding the 'language of biology' from first principles. This opens up massive potential for AI to overcome human limitations in biological research.

AI can significantly increase the parallelism of drug discovery processes, reducing the duration of many steps. Brockman also pointed out that many existing cures might be 'hidden' within scientific literature; AI could rapidly identify these relevant papers, enabling treatments for specific conditions and massively scaling up such operations in the short term.

Greg Brockman discussed the lengthy drug development timeline, which can span up to 15 years, largely due to successive, increasingly expensive trials and stringent safety regulations. He argued that if AI can demonstrably prove a fundamental shift in the 'physics of drug discovery' by delivering safer or more efficient results, regulatory expectations must change.

He drew a parallel with self-driving cars, suggesting that if AI-driven vehicles become 10 or 100 times safer than human drivers, the entire transportation system, including its regulations, would need a complete rethinking. This could lead to new regulations, such as restricting human-driven cars in certain areas.

Brockman noted that regulatory environments are not uniform globally, leading some researchers to conduct trials in jurisdictions with more flexible frameworks. This highlights the dynamic interplay between technological advancement and the need for adaptable regulatory systems.

The core argument is that unambiguous evidence of AI's superior performance in critical domains like drug discovery or transportation will inevitably force a re-evaluation and overhaul of existing regulatory frameworks, trickling through every part of the system.

Greg Brockman highlighted 'Dots' as a 'proactive agent' that has significantly changed internal workflows at OpenAI. Unlike traditional tools, Dots is 'always on, always working' and continues tasks even when the user stops, signifying a shift in the 'form factor' of workplace AI rather than just raw capability.

His personal 'Dot' tracks his pending work, reminds him of unfulfilled commitments, and even proactively identifies calendar conflicts or suggests leaving meetings early to make other appointments. This demonstrates its deep contextual awareness of his daily activities, projects, inbox, and Slack communications.

The agent integrates across various platforms and understands the user's top-of-mind priorities. Brockman views this as a 'game changer' in a workplace setting, leveraging the intelligence of advanced models to anticipate needs and manage tasks autonomously.

Greg Brockman asserted that consumer and enterprise AI are destined to converge, despite historical distinctions. OpenAI currently operates research, a consumer platform (ChatGPT), and an enterprise/developer business platform.

The primary difference between these segments lies in cost tolerance. Mass-market consumer products demand high efficiency to provide free or very low-cost access, requiring substantial engineering efforts for scalability.

In contrast, enterprise users prioritize 'peak intelligence.' They are more willing to invest in higher compute usage as long as they see proportional returns in productivity, making cost less of a limiting factor for premium features.

OpenAI possesses a unique advantage, according to Greg Brockman, by having both 1.2 billion ChatGPT consumers and a leading enterprise coding business. Their strategy is to unify these fronts rather than compete separately, believing this is the general direction of AI to blur consumer and enterprise distinctions.

Brockman emphasized that AI is significantly lowering barriers to entry for entrepreneurship. He noted observing people quitting their jobs to start companies, explicitly stating they wouldn't have done so without reliable AI to manage tasks that were previously professional hurdles.

Greg Brockman believes the ultimate evolution of user interface (UI) in AI is its disappearance, stating, 'the best UI is no UI.' He envisions a future where users simply converse with AI agents like 'Dots' through verbal commands or messaging apps like Slack and iMessage, or even by calling them directly.

However, Brockman acknowledged fundamental differences between personal and workplace AI settings. Enterprise environments require additional layers such as admin controls, security measures, and consideration for multiple stakeholders, meaning that a truly 'no UI' approach might be more complex to implement.

OpenAI is actively developing 'generative UI,' which allows the AI to create custom visual layouts on the fly, tailoring the interface to the user's immediate needs. This addresses the challenge of surfacing functionality or communicating limitations in a 'no UI' environment.

He also mentioned the concept of a 'thinking' model, where AI can provide a quick initial answer while continuing to process for a deeper, more comprehensive solution. This dynamic and adaptable approach to interfaces aims to integrate AI seamlessly into user workflows, despite the inherent complexities of diverse use cases.

Greg Brockman explained that AI is moving towards a model where it 'molds itself much more to you,' meaning user interfaces are becoming precisely what is needed at the exact moment it's required. This allows for the on-the-fly generation of bespoke applications.

He provided an example of an AI creating a one-off app to help assemble a bike—a task for which a user would never hire a developer. AI can perform such custom tasks in real-time, demonstrating its adaptability.

OpenAI is focused on integrating this adaptable and user-centric approach into every aspect of its product development, ensuring that interfaces are no longer static but dynamic and responsive to individual needs.

OpenAI is making a substantial commitment of one billion dollars to empower 'frontline defenders' by providing access to their advanced AI models for cybersecurity. This initiative, part of their Daybreak program, aims to help identify and fix vulnerabilities in systems, leveraging 'peak intelligence' for defense.

The company prioritizes coordination with governments and the broader community to enhance incident response capabilities. Brockman highlighted that AI offers new opportunities for 'formal verification of software,' a previously impractical method that models are now adept at, allowing for the proof of software security properties.

He believes that the 'physics around cybersecurity' are changing, and there's an opportunity for a much more secure world. This proactive stance reflects OpenAI's view that leading in cyber security is a core priority, especially given their frontier model capabilities.

Greg Brockman explained OpenAI's internal security protocol: 'every time a new model drops, you need to apply it to secure yourself, find your vulnerabilities and then saturate.' This continuous process involves using advanced AI models to scan and stress-test their own infrastructure.

As an example, OpenAI temporarily reallocated approximately 25% of their production engineers to solely focus on security. They tasked their peak intelligence model at the time to probe their infrastructure, identify issues, and build tools for repeatable scanning. After fixing the identified problems, they reached a saturation point where that specific model could no longer find new vulnerabilities.

Brockman acknowledged that this process is ongoing and will be repeated with each new model release. Furthermore, he highlighted a long-term strategy: training AI models to write secure code from the outset. This aims to create fundamentally more secure systems through the correct application of AI models.

Greg Brockman stated that for defenders to outstrip attackers in cybersecurity, organizations must fundamentally rethink their operational approach. This includes a collective re-evaluation of critical infrastructure and how those lacking resources can be supported.

He highlighted 'internet-wide tech debt' as a significant barrier to effective security. Brockman pointed out that proactive measures, like reassigning 25% of engineers to security or implementing feature freezes for an entire quarter to address vulnerabilities, are rarely adopted but are necessary.

Brockman strongly believes that cybersecurity spending should 'vastly outstrip' software spending. He argued that as software development becomes increasingly inexpensive, the value shifts towards ensuring reliability and security. This means security investment should reflect the foundational importance of secure and trustworthy software.

Greg Brockman addressed the unsustainable 'XKCD' model, where critical infrastructure often relies on a single, obscure open-source maintainer. He explained that this situation arose because writing reliable software is difficult, and few companies audit their extensive software supply chains.

Human developers lack the capacity and expertise to manually audit the myriad of obscure third-party packages for vulnerabilities. This limitation has historically made it impossible to prevent supply chain attacks where malicious packages are uploaded and then used by numerous companies.

Brockman highlighted emerging tools, such as Socket, that use AI to automate the scanning of open-source supply chains for vulnerabilities. This development represents a crucial shift, enabling comprehensive auditing that was previously infeasible, and is part of a coordinated response involving companies like CrowdStrike to secure the ecosystem.