a
Where We Go From Here with OpenAI's Mira Murati
@ a16z
09/25/23
Takeaways
MM
Mira Murati
09/25/23
@ a16z
Building good products on top of AI models is incredibly difficult.
MM
Mira Murati
09/25/23
@ a16z
I was born in Albania just after the fall of Communism, which greatly influenced my education and focus on math and sciences.
MM
Mira Murati
09/25/23
@ a16z
In communist Albania, there was a strong emphasis on math and physics, while humanities subjects were often viewed with skepticism due to questionable sources of information.
MM
Mira Murati
09/25/23
@ a16z
I pursued my interests in math and sciences relentlessly, eventually studying mechanical engineering and working in aerospace before joining Tesla.
MM
Mira Murati
09/25/23
@ a16z
At Tesla, I became very interested in AI applications, particularly in autopilot technology, which led me to explore AI's potential in different domains.
MM
Mira Murati
09/25/23
@ a16z
After Tesla, I worked on augmented and virtual reality to gain experience in different domains, although it was too early for practical applications at that time.
MM
Mira Murati
09/25/23
@ a16z
I was drawn to OpenAI because of its mission to develop artificial general intelligence (AGI), which I believe is the most important technology we can build.
MM
Mira Murati
09/25/23
@ a16z
The shift in AI development has seen a growing number of influential figures in the field coming from physics and math backgrounds, unlike 15 years ago when it was more about traditional engineering.
MM
Mira Murati
09/25/23
@ a16z
Theoretical math requires a different way of thinking, where you need to sit with a problem for a long time, building intuition and discipline to solve it.
MM
Mira Murati
09/25/23
@ a16z
As we deploy AI technologies, we face massive systems and engineering challenges to make them efficient and accessible to users without deep knowledge of machine learning.
MM
Mira Murati
09/25/23
@ a16z
The difference in user engagement between making AI models available through an API versus a user-friendly interface like ChatGPT is significant; the latter captures people's imagination more effectively.
MM
Mira Murati
09/25/23
@ a16z
The design of AI interfaces is evolving towards natural language interactions, allowing users to collaborate with AI as if it were a co-worker rather than just a tool.
MM
Mira Murati
09/25/23
@ a16z
We are at an inflection point in redefining how we interact with digital information through AI systems, which may evolve to understand our context and goals in a more personalized way.
MM
Mira Murati
09/25/23
@ a16z
The development of ChatGPT was rooted in a long-term vision of creating safe AI systems that align with human values, utilizing reinforcement learning with human feedback.
MM
Mira Murati
09/25/23
@ a16z
AI safety has transitioned from a theoretical concept to practical applications, integrating safety measures into AI systems to ensure they follow user intent effectively.
MM
Mira Murati
09/25/23
@ a16z
Hallucinations are a significant challenge for AI models, and we have been studying how to improve their truthfulness and express uncertainty.
MM
Mira Murati
09/25/23
@ a16z
The project that led to ChatGPT began as Web GPT, which utilized retrieval to provide information and cite sources, ultimately evolving into ChatGPT due to its unique dialogue capabilities.
MM
Mira Murati
09/25/23
@ a16z
We focused on alignment and safety for GPT-4, aiming to gather feedback from researchers to enhance its reliability and robustness.
MM
Mira Murati
09/25/23
@ a16z
Alignment means ensuring the model acts according to the user's intent, while safety also involves preventing misuse that could lead to harmful outputs.
MM
Mira Murati
09/25/23
@ a16z
We are exploring reinforcement learning with human feedback to address the issue of hallucinations, which is a complex problem but potentially solvable with this approach.
MM
Mira Murati
09/25/23
@ a16z
There is no grand plan; our progress towards AGI is a series of incremental steps based on what we need to achieve at each stage.
MM
Mira Murati
09/25/23
@ a16z
The trajectory of AI systems indicates they will increasingly perform tasks autonomously, but human oversight will remain essential for direction and guidance.
MM
Mira Murati
09/25/23
@ a16z
We aim to provide a platform that allows users to build on our models, offering tools for customization and control to create effective products.
MM
Mira Murati
09/25/23
@ a16z
The foundation models today represent the world in text, and we are expanding to include images and video to create a more comprehensive understanding of our environment.
MM
Mira Murati
09/25/23
@ a16z
The challenge of super alignment is critical as we develop powerful models, and we have dedicated a team at OpenAI to focus on this complex technical issue.