Introduction:
Mira Murati, CTO of OpenAI, joined a fireside chat at MIT’s inaugural CERC event, discussing the company’s advancements in AI and its ethical and societal implications.

GPT Builder Interface:
OpenAI unveiled the GPT Builder interface, a platform enabling users with no coding knowledge to create their own customized versions of ChatGPT. This platform aims to empower creators and foster innovation in AI applications.

Monetization for Creators:
Mira highlighted the importance of recognizing and compensating creators for their contributions to the platform. OpenAI is exploring various monetization models to ensure creators are fairly rewarded for their work.

Aligning AI with Human Intentions and Values:
Mira emphasized OpenAI’s commitment to developing AI systems that align with human intentions and values. The company strives to create safe and beneficial AI technologies that positively impact society.

Mira’s Career Path:
Mira’s diverse career experiences in product management, engineering, and design, spanning industries such as aerospace, automotive, and AI, were briefly highlighted.

Questions from the Audience:
The audience was invited to submit questions via Slido, which would be discussed live during the event.

Introduction:
The presenter, Mira Murati, expresses excitement about being at the event and announces new advancements in APIs and products, including Inkabase and 2K Kids. The goal is to develop AI agents that collaborate and teach humans how to collaborate with AI systems in various spaces, such as creative and workspaces. Murati emphasizes the ubiquity and importance of AI agents in the future, but also acknowledges the risks and trepidations associated with their widespread use.

Iterative Deployment Strategy:
Murati discusses the need for a cautious and iterative approach to deploying AI agents to understand their implications for society. The first iteration of AI agents has been rolled out, and early observations show creative uses, including tools like browsing and DALI.

Human-GPT Collaboration:
Dan Huttenlocher spontaneously raises the topic of human-GPT collaboration, expressing MIT’s interest in this area. Huttenlocher highlights the common perception of AI replacing humans, tracing it back to the 1960s.

AI as an Augmentation Tool:
Murati agrees with Huttenlocher’s perspective and emphasizes the importance of viewing AI as an augmentation tool rather than a replacement for humans. She believes that AI can amplify human abilities and enhance collaboration, leading to greater creativity and productivity.

The Need for Ethical Considerations:
Murati stresses the significance of ethical considerations in AI development and deployment. She underscores the responsibility of researchers and developers to ensure that AI systems align with societal values and minimize potential harms.

Conclusion:
Murati reiterates the importance of continuous adaptation and learning as AI technology evolves. She emphasizes the need for ongoing dialogue and collaboration among researchers, developers, and policymakers to shape the future of AI in a responsible and beneficial manner.

AI Collaboration and Augmentation:
Mira Murati emphasizes the significance of AI collaboration, viewing AI systems as tools that elevate human capabilities in various aspects, including creativity and productivity.

Tools and Accessibility:
Murati advocates for user-friendly AI tools that can be easily utilized by individuals with or without programming expertise. Natural language programming is highlighted as a means to make AI more accessible and inclusive.

Rapid Prototyping and Iteration:
Murati envisions AI collaboration enabling rapid prototyping and iteration, facilitating effective communication and collaboration among team members.

Current Examples of AI Collaboration:
ChatGPT, Code Interpreter, and Copilot are cited as examples of existing AI systems that demonstrate the feasibility of human-AI collaboration.

Potential Future Questions:
Murati acknowledges the question of how much AI delegation may occur over time and the possibility of tasks becoming more abstract. She identifies the automation of repetitive tasks and assistance in brainstorming as immediate benefits of AI collaboration.

Overall Leadership Approach:
Murati expresses optimism about the potential of AI collaboration and its impact on society. Her optimism guides her leadership at OpenAI, where she focuses on developing AI tools that empower humans and enhance their capabilities.

Optimism and Preparation:
Mira Murati emphasizes the significance of optimism and faith in the successful development of AI technologies. However, she stresses the need for thorough preparation and a clear-eyed assessment of the challenges ahead.

Challenges in AI Deployment:
Immediate challenges include making AI accessible while predicting capabilities and risks as more powerful models are developed. Medium-term challenges involve building the science of prediction to understand and mitigate risks. Long-term challenges involve catastrophic risks and self-improving systems.

Super Alignment:
OpenAI invests in super alignment research to understand how to align highly capable models with human values. This research addresses the technical difficulty of aligning models to prevent catastrophic risks.

Examples of Risk Prediction:
Mira Murati shares examples of risk prediction successes and failures in deploying AI models. She mentions the unexpected prevalence of spamming content rather than misinformation as a risk for GPT-3. The impact of AI technologies on education is still being studied due to its complexity and wide-ranging effects.

Collaboration and Engagement:
Mira Murati recognizes the responsibility of AI developers in understanding the societal impact of these technologies. OpenAI collaborates with organizations like Code.org and Khan Academy to empower educators and gather feedback. Partnerships with universities help OpenAI understand how AI affects classrooms.

Preventing a Race to the Bottom on Safety:
The Bletchley Park meeting was an example of good intentions without clear mechanisms for ensuring safety. OpenAI promotes voluntary commitments by AI developers to prioritize safety. The company believes that transparency, auditing, and collaboration are key to preventing a race to the bottom on safety concerns.

Structure and Governance:
OpenAI was restructured as a limited partnership to raise funds for research and development while ensuring that its non-profit mission remains the primary objective. Mechanisms and incentives play a critical role in aligning the organization’s actions with its mission.

Importance of Governance:
Regulation is essential for frontier AI systems to prevent unintended consequences and ensure responsible deployment. Voluntary standard setting by industry leaders and government initiatives can help promote safe and ethical practices.

OpenAI’s Approach to Safe Deployment:
OpenAI employs a comprehensive set of safeguards and measures to ensure safe and responsible deployment. These measures include safety systems, red teaming, and continuous feedback loops to improve the technology.

Industry Standards:
The industry should establish standards for the deployment of AI technologies, particularly for frontier systems. Voluntary standard setting and government initiatives can help prevent a “race to the bottom” and promote responsible AI practices.

Focus on Frontier Systems:
Regulation and collaboration are crucial for managing the risks associated with frontier AI systems. Governments and AI developers should work together to regulate frontier models and address potential dual-use applications.

Executive Order Implications:
Meta is supportive of regulation around frontier models to define capabilities and determine thresholds for regulation. Practicality and alignment with scientific advancements are key considerations.

AI and Existing Laws:
Using AI does not excuse breaking existing regulations and laws. Guardrails in general models to prevent misuse are essential, like preventing AI from pretending to be a lawyer or a doctor.

Domain-Specific Guardrails:
Common sense guidelines and guardrails exist, but nuanced decisions depend on the use case and domain. Human oversight and collaboration with trusted partners help understand risks and establish specific guardrails. Incentive alignment with enterprise customers promotes robust and reliable technology.

Respecting Rights of Creators and Authors:
Meta works with licensed data and collaborates with creators to develop useful products. Research is ongoing to attribute value to specific data contributed to model training. An upcoming marketplace will enable GPT publishers to be compensated for their work.

Data Attribution and Compensation:
OpenAI is working closely with creators to understand their needs and preferences regarding data attribution and compensation. There are technical challenges in developing economic models and attributing data usage to creators. OpenAI aims to empower creators and allow them to choose whether their data can be used for training models.

Data Privacy and User Controls:
OpenAI has invested in data privacy and user controls to give individuals control over how their data is used. Users can opt out of having their data used for training purposes.

Engineering Challenges and Opportunities:
OpenAI faces numerous engineering challenges, including frontier problems in technology development. These challenges often involve the intersection of multiple domains such as society, technology, safety, and legal frameworks. OpenAI is committed to addressing these challenges and shaping the future of AI in a responsible and beneficial manner.

OpenAI’s Role in Society:
Mira Murati emphasizes the need for experts to consider how to evolve and modify various domains to create an infrastructure that enables society to adopt and utilize AI technology beneficially. Murati believes that this is a crucial period in history to engage in significant and influential work that will shape future living standards and quality of life.

OpenAI’s Dominance in Natural Language Processing and Image Synthesis:
Dan Huttenlocher acknowledges OpenAI’s leadership in natural language processing and image synthesis.

OpenAI’s Plans for Expansion into Other AI Domains:
Huttenlocher inquires about OpenAI’s aspirations to lead in other AI domains, specifically video generation, biology applications, and generative AI for scientific discovery.

OpenAI’s Research Agenda and Path to Safe and Beneficial AGI:
Murati explains that OpenAI’s research agenda is guided by the goal of achieving safe and beneficial artificial general intelligence (AGI). A significant strategic decision a few years ago was to focus on scale and scaling large English models, which represented a major paradigm shift. The current focus is on integrating multiple modalities into these large models, requiring substantial data, diverse modalities, and continued scalability.

Modalities and Applications:
Mira Murati discusses the different modalities, such as text-to-speech, images, voice, and video, that OpenAI is working on. The company aims to bring these technologies into its platform so that other people can build with them as well.

Domain Expertise:
OpenAI does not typically invest in specific domains like biology unless there is a strategic reason to do so, as it lacks expertise in those areas. The company prefers to collaborate with organizations that have expertise in other domains to explore new possibilities at the frontier.

Scientific Advancements:
Mira Murati expresses excitement about how AI systems can help us discover new things and advance research. OpenAI has been using Copilot-like and Wave systems internally to assist with tasks such as safety monitoring, leveraging GPT-4’s capabilities. The company is working on developing an AI scientist that can run on GPUs to advance research and alignment.

Human Out-of-the-Loop Systems:
Mira Murati acknowledges the potential benefits of fully autonomous multi-agent generative AI systems, but emphasizes the challenge of ensuring alignment and following intent. OpenAI has made progress in this area with RLHF and instruction following models, but there is still much research to be done.

Alignment and Scalability:
Mira Murati highlights the importance of alignment research and capability prediction to ensure the robustness and safety of autonomous agents. The company is exploring reinforcement learning with AI feedback to address the challenges of scaling human feedback. The quality of the data used for RLHF is crucial and requires specific data collection campaigns.

Education and Coding:
Mira Murati believes that the ability to code in natural language can make learning and building things easier. However, she emphasizes that there is still a lot of code to be written in the world and it is a great skill to learn. The technologies developed by OpenAI can help people learn how to code more easily and cater to their individual learning styles.

Accessible Learning and Creation with AI:
AI systems like GPT are becoming more accessible and user-friendly, enabling individuals to learn and create without extensive coding knowledge. Users can provide high-level guidance to AI systems, allowing them to build prototypes and products with minimal coding. The availability of tools and resources makes it easier than ever to explore AI and its potential applications.

Potential Partnership between OpenAI and MIT:
There is growing interest within the MIT community to explore a partnership with OpenAI. Discussions have been ongoing between MIT faculty and research groups and OpenAI researchers. The level of student interest in AI highlights the need for such a partnership to foster expertise in this rapidly evolving field.

Mechanisms for Fairness, Combating Bias, and Ethical Development:
Despite broader discussions on ethical AI development, there is a demand for concrete examples and mechanisms to address fairness, bias, and ethical concerns. OpenAI is actively working on these issues across various domains, including text, images, and multimodal combinations. Ongoing research and initiatives aim to develop effective mechanisms that promote fair and ethical AI development.

Balancing Metrics and Holistic View:
When optimizing for one metric, unexpected degradation in other areas can occur. A holistic view is necessary, especially for complex issues like harmful bias. Constant auditing is needed when tweaking the system to prevent unintended consequences.

Example of DALI Gender Ratio Tweak:
Adjusting the gendered image ratio in DALI 2 led to unintended hyper-sexualization of women’s images. Intervention was necessary to achieve a more balanced result and mitigate unintended consequences.

Diverse Talent and Red Teaming:
Diverse talent in the team is crucial for handling complex issues like harmful bias. Red teaming with experts from different fields provides valuable insights and learning opportunities.

Post-Deployment Monitoring and Tweaking:
Monitoring and tweaking are essential at all stages, from data to deployment. Usage policies and monitoring systems help ensure compliance with guidelines. Fine-tuning is necessary to address over-conservatism and optimize for user experience.

Trade-off between Truthfulness and Creativity:
Balancing creativity and factuality in AI models can be challenging. Different people have varying expectations and preferences. Giving users more control over the model’s personality and behavior within certain bounds can be a solution.

Background:
Mira Murati grew up in Albania and studied in Canada and the US. She has a strong interest in math, science, and building stuff.

Career Path:
Murati studied mechanical engineering and math in college. She worked in aerospace, automotive, and virtual reality/augmented reality fields. Currently, she is dedicated to artificial general intelligence (AGI).

Advice for Aspiring Entrepreneurs and Innovators:
Follow your curiosity and passion. Focus on learning and maximizing your knowledge in areas that excite you. Don’t be afraid to jump domains and start from scratch. Take risks and try new things, especially early in your career. Seek out mentors and role models who can guide and support you.

Advice for People from Minority Backgrounds:
Network with people from diverse backgrounds and experiences. Seek out opportunities to learn about different cultures and perspectives. Be persistent and resilient in the face of challenges.

Mira Murati’s Career Advice:
Build stuff and try as many things as possible. Don’t overthink it, especially early in your career. The best way to learn about what you want to do and what you’re good at is to just go do stuff.

OpenAI’s Environmental Efforts:
Making AI systems more efficient. Working with Microsoft to offset the environmental impact of AI systems. Using advanced AI systems to find new scientific solutions to climate change and develop more sustainable energy infrastructure.

Jobs in the Future:
Dan Huttenlocher believes that in the far future, jobs may become optional as AI takes on more tasks. However, he doesn’t see this happening in the short to medium term (next 10 years).

OpenAI’s Legacy:
Mira Murati wants OpenAI to be remembered for its positive impact on the future and humanity. She hopes that OpenAI’s AI systems will help solve some of the world’s biggest challenges, such as climate change and sustainable energy.

OpenAI’s Impact on Society:
Mira Murati hopes OpenAI will significantly impact society by helping build safe AI systems and distributing their benefits equitably. She also emphasizes the importance of global governance for AI and OpenAI’s role in driving innovative approaches to this challenge.

Alignment and Safety:
When considering open-sourcing systems, OpenAI prioritizes alignment, safety, and the potential consequences of releasing these systems. Guardrails, infrastructure, and careful deployment strategies are essential in ensuring the responsible use of large language models.

Open Academic Research:
OpenAI still engages in open research in areas where they can safely open-source systems and benefit the broader community. The debate surrounding open sourcing versus not centers around alignment, safety, and the potential risks of releasing these systems. OpenAI collaborates with experts to assess risks and capabilities of models, especially in specific domains like misinformation, bio-risk, and cybersecurity. They also fund safety research and promote new research directions through initiatives like the AI Safety Summit.

Open Source Concerns:
OpenAI acknowledges the challenge of preventing malicious modifications of open-sourced generative AI technology. Research is ongoing to address this issue and find ways to share these technologies while mitigating potential risks.

Embodied AI:
In the immediate term, OpenAI is not prioritizing robotics or other embodied AI initiatives. However, they are aware of public speculation and may consider exploring these areas in the future.

Explainability of AI Models:
Interest in explainability (or interpretability) of AI models has grown. “Explainability” implies the model can explain itself, while “interpretability” means humans can understand the model. Alexander Mandry from MIT joined OpenAI to lead research on interpretability and predicting capabilities.

LGBT Research:
Initial work focused on improving truthfulness and factuality in AI models. Research evolved into using Socratic dialogue to understand models’ capabilities and limitations. This approach provides insights into models’ behavior and helps identify hallucinations and other issues.

Super Alignment Team:
OpenAI’s super alignment team is conducting research on interpretability. New research in this direction is in its early stages.

Questions Unasked:
Dan Huttenlocher acknowledged that many questions remained unanswered due to time constraints. He invited Mira Murati to ask any questions she felt should have been asked.

Conclusion:
Dan Huttenlocher expressed appreciation for Mira Murati’s time and thoughtful answers. He emphasized the strong interest at MIT in OpenAI’s work and the desire for continued collaboration.