By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.
Bloomberg OriginalsInside OpenAI, the Architect of ChatGPT, featuring Mira Murati | The Circuit with Emily Chang
Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.
CHAPTERS
Chapter summaries
Entering OpenAI and setting up the big questions around AI
Emily Chang frames OpenAI as an under-the-radar startup that beat tech giants to market with ChatGPT and DALL·E, igniting an industry-wide race. She sets up the episode’s core tension: whether what OpenAI is doing is “magic,” just algorithms, or something that could help or harm society.
- •OpenAI is presented as a surprisingly low-profile company behind major AI breakthroughs
- •ChatGPT and DALL·E are described as products that forced big tech to respond quickly
- •The episode tees up questions about AI’s promise versus potential danger
- •Emily sets the expectation that the visit is aimed at getting answers from inside OpenAI
Tour moments and meeting Mira Murati as a key strategist
Emily and Mira Murati walk through the space and react to its atmosphere and design. Emily positions Murati as a central, reflective leader shaping OpenAI’s strategy, beyond the more publicly known CEO figure.
- •The setting is described as futuristic and calming, shaping the tone of the conversation
- •Mira Murati is introduced as a “chief architect” behind OpenAI’s strategy
- •Their interaction establishes a behind-the-scenes access feel
- •The visit transitions from ambiance into deeper discussion of OpenAI’s culture and choices
OpenAI’s internal culture: symbolism, alignment focus, and managing disagreement
A discussion of artwork leads into Murati’s view of OpenAI’s priorities around safety and alignment. Emily asks about internal disagreement, and Murati acknowledges some people have left over deployment strategy debates, emphasizing structured dialogue to address concerns.
- •An OpenAI-logo painting is discussed, with Murati interpreting it as AI that “loves humanity”
- •Murati lists challenges OpenAI focuses on: hallucination, truthfulness, reliability, and alignment
- •Murati says some people have left to start other organizations due to disagreements on deployment strategy
- •Murati describes the need for constant dialogue and ways to systematize concerns
What a CTO does at OpenAI: strategy, gaps, and enabling teams
Emily asks Murati to define the CTO role. Murati describes it as hands-on team guidance combined with long-term strategy work, identifying gaps, and ensuring teams have what they need to succeed.
- •The CTO role is framed as both operational and strategic
- •Murati emphasizes guiding teams on the ground while planning longer-term direction
- •Identifying organizational and technical gaps is part of the job
- •Success depends on supporting teams to tackle very hard problems
Why OpenAI released ChatGPT: learning from real-world use and feedback
Emily asks about the decision point to release ChatGPT publicly. Murati says OpenAI released it because they needed broader feedback to understand usage, risks, and limitations, and to bring the technology into public awareness—while being surprised by the scale of adoption.
- •Murati says ChatGPT existed internally “for a while” before release
- •Release was motivated by the need for feedback on use cases, risks, and limitations
- •Murati says they wanted to learn more about what they had created through public interaction
- •They were surprised by how strongly ChatGPT captured public imagination
How ChatGPT works at a high level: scaling models and investing in safety
Murati explains ChatGPT as a neural network trained on massive data and compute to predict the next word. She describes how increasing model size and data increases capabilities, and argues that alignment and safety investment can improve reliability over time.
- •ChatGPT is described as a neural network trained on huge data using a massive supercomputer
- •Training objective is predicting the next word in a sentence
- •Murati says larger models and more data lead to stronger capabilities
- •Murati links more alignment and safety work with improved reliability and safety
Search disruption and the limits behind confident-sounding answers
Emily presses Murati on whether OpenAI is coming for Google Search. Murati says the goal was not to dominate search, but to offer a more intuitive way to understand information—while acknowledging that producing correct answers consistently is difficult for systems built around next-word prediction.
- •Murati says OpenAI didn’t set out to dominate search market share
- •She contrasts keyword search with a more intuitive, conversational way of understanding information
- •Emily questions why systems don’t simply say “I don’t know” more often
- •Murati explains that general capabilities make it difficult to handle limitations like correctness
Misinformation and hallucinations: what worries Murati most
Emily raises concerns that biased or incorrect training data could accelerate misinformation. Murati calls misinformation complex and highlights hallucinations—models convincingly making things up—as a top concern, warning that users cannot blindly rely on outputs.
- •Emily points to bias and incorrect data as a driver of misinformation risk
- •Murati calls misinformation a complex, hard problem
- •Murati identifies hallucinations as a major concern, especially with advanced models
- •Murati emphasizes the need for awareness that outputs can be wrong even when convincing
Why call it “hallucination” and whether AI should identify itself
Emily questions using a human term for AI errors. Murati argues the behavior can resemble human conversation, noting people also fabricate answers when uncertain, and says it’s important to distinguish machine output from human output as society moves toward more hybrid collaboration.
- •Emily challenges the human framing of the term “hallucination”
- •Murati argues some general capabilities appear human-like
- •Murati compares AI making things up to how humans sometimes respond when they don’t know
- •Murati says distinguishing machine-generated output from human output matters as hybrid outputs increase
Training data, value to creators, and job change (including prompt engineering)
Emily asks about using data from writers and artists and whether value flows back to creators amid job fears. Murati says she doesn’t know how attribution and accounting would work across the internet, anticipates job loss and transformation, and points to prompt engineering as an unexpected new job category.
- •Emily raises fairness and compensation concerns for writers and artists whose work is in training data
- •Murati says she doesn’t know how it would work to account for information created by everyone online
- •Murati predicts some jobs will be lost and many will change as AI integrates into work
- •Prompt engineering is cited as an example of a new job that wasn’t predicted
Prompting, fact-checking with browsing, and the role of contractors in safety work
Murati describes effective prompting as developing intuition, providing context, and working around hallucinations. Emily asks about reports of low-paid content moderation-style work, and Murati says OpenAI used contractors to scale, chose one vendor for safety standards, later stopped working with them, and maintains mental health and wellness standards for contractors.
- •Murati describes good prompting as intuition plus giving the model enough context
- •Hallucinations are discussed as a reason to add checks like fact-checking
- •Murati demonstrates using a browsing plugin to fact-check by browsing the web
- •Murati says contractors are used for scale; OpenAI later stopped working with the referenced contractor
- •Murati says OpenAI shares mental health and wellness standards with contractors
Kids, vulnerability, and what an “AI relationship” could mean
Emily asks about children potentially forming relationships with AI. Murati argues powerful systems should be deployed carefully around vulnerable populations because impacts are still not fully understood, and she says she could have a relationship with AI in the sense of a reliable tool that improves her life.
- •Emily frames concern about children having relationships with AI
- •Murati calls for caution placing powerful systems in front of vulnerable groups
- •Murati says checks and balances are important because it’s early and impacts aren’t fully understood
- •Murati says she could see a relationship with AI as a reliable tool that enhances life
Reid Hoffman on AI as a platform shift and the speed of integration
The episode shifts to Reid Hoffman, who argues AI builds on prior tech waves and may be even bigger. He introduces a hopeful framing (like smartphone tutors) and defines blitzscaling, predicting AI will be integrated faster than the iPhone and that “co-pilots” will emerge across professions.
- •Hoffman says AI is a crescendo that builds on internet, mobile, cloud, and data
- •He critiques discourse he sees as too fear-based versus hope-based
- •He describes a smartphone tutor for every child as within sight of current models
- •He defines blitzscaling as prioritizing speed over efficiency amid uncertainty
- •He predicts rapid adoption and “a co-pilot for every profession”
Origins of OpenAI, fear narratives, and debating whether the public is overestimating AI
Emily asks about Hoffman’s history with Elon Musk and OpenAI’s formation. Hoffman describes discussions with Elon and Sam, argues against fear terms like “robocalypse,” and says AI can appear credible while being wrong—especially on the newest edge of knowledge or specialized domains.
- •Hoffman recounts Elon telling him AI was coming, prompting deeper engagement
- •He says fear framing spreads easily and criticizes dramatic “robocalypse” rhetoric
- •He explains how OpenAI emerged as a counterweight to purely commercial development
- •He says AI can generate smart-sounding but incorrect answers, especially outside its depth
- •He emphasizes that seeming credibility can mislead people even when content is wrong
Investment frenzy, concentrated power, Microsoft ties, and what “open” should mean
Emily challenges whether AI is a bubble and whether power is concentrating among a few companies. Hoffman argues generative AI has broad impact and expects major new companies, addresses OpenAI’s structure and openness, and explains limiting open-source release as a safety choice—while acknowledging the need for transparency and broader social decision-making.
- •Hoffman argues generative AI has unusually broad reach, even amid investment uncertainty
- •He acknowledges AI development is shaped by an influential network of participants
- •He predicts disruption in search and a profusion of startups, with potential new giants emerging
- •He says he helped facilitate communications between Microsoft leadership and OpenAI leadership
- •He disputes claims that OpenAI is purely profit-maximizing, describing a nonprofit with commercial arrangements supporting the mission
- •He argues API access can be “open access” even when not open source, citing safety reasons for controlled release
- •He calls for more transparency but says AI behavior norms should be social decisions, not dictated by a few actors
Responsible innovation, extinction risk, and whether the AGI train should slow down
The conversation returns to Murati on acceleration risks and governance. She supports the idea of a trusted authority auditing systems under agreed principles, acknowledges extinction risk if goal-setting systems become misaligned, says it’s been hyped recently, and argues society should guide development rather than halt progress out of fear—while noting AGI with autonomous decision-making remains far away.
- •Murati says acceleration is a significant risk society must grapple with and that safe general AI is very hard
- •Murati supports a trusted authority that can audit systems based on agreed-upon principles
- •Murati acknowledges a potential human-extinction risk if AI systems set misaligned goals
- •She says the level of extinction risk hasn’t necessarily changed in recent months, though anxiety and hype have increased
- •Murati says systems are still far from autonomous decision-making and discovering new knowledge
- •She argues progress should be guided carefully rather than brought to a halt, agreeing with the framing that “the train has left the station”
