Sep 25, 2023 · 25m · a16z
Where We Go From Here with OpenAI's Mira Murati
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In an interview hosted by Andreessen Horowitz's Martin Casado, OpenAI CTO Mira Murati shares insights on AI research mindsets, the real-world deployment strategy behind ChatGPT, scaling laws, and the roadmap toward safe Artificial General Intelligence. She highlights how natural language interfaces, product engineering, and empirical safety feedback loops are driving the next wave of AI innovation.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
The host as informed peer 3.5 Guest teaching 2.8 Guest disagreement 1.2 The host pushing back 2.0
speaking balance: gold is the host, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 25:23 Rejecting Doomer/Accelerationist Binary
The guest resists the host's forced binary choice between being a doomer or accelerationist, stating 'Let me say something else.'
Hardest push from the host ▶ 15:27 Challenging Alignment vs Safety TerminologyThe host interrupts to demand clarity on whether the guest's use of safety means functional correctness or protection against active harm.
Biggest teaching moment ▶ 15:40 Defining Technical Distinction of AlignmentThe guest clearly educates the host by breaking down alignment as meeting user intent versus safety as preventing malicious misuse and hallucinations.
The host holds their own ▶ 20:24 90s Silicon Industry Historical AnalogyThe host demonstrates technical background by comparing AI model consolidation to 1990s CPU coprocessor integration.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The host as informed peer | Guest teaching | Guest disagreement | The host pushing back | Why |
|---|---|---|---|---|---|---|
| Mira Murati's Background and Journey to OpenAI | 1 | 1 | 0 | 0 | The host asks an open background question about the guest's path to OpenAI. The guest provides a friendly personal monologue about her upbringing in Albania, engineering background at Tesla, and transition to OpenAI. | |
| The Role of Math and Physics Mindsets in AI Development | 4 | 3 | 1 | 2 | The host shares personal experience programming against models, describing it as wrapping a supercomputer with an abacus. The guest elaborates on how interfaces are evolving toward natural language collaboration rather than standard programming. | |
| The Research Roots of ChatGPT, RLHF, and Safety | 4 | 4 | 1 | 3 | The guest explains the research history of RLHF, WebGPT, and ChatGPT's role in gathering alignment feedback for GPT-4. The host interjects to press for precision on the conceptual boundary between alignment and safety. | |
| Product Strategy and Real-World Deployment Feedback | 5 | 4 | 2 | 3 | The host probes scaling laws and cites the Nishihara cat-to-Einstein reasoning continuum. The guest grounds the debate in practical capabilities, describing models as currently at intern-level reliability. | |
| Platform Economics and Building Products on Top of Models | 5 | 3 | 1 | 2 | The host draws an informed historical analogy to the 1990s CPU integration of co-processors to ask about platform economics. The guest responds by explaining API tiered economics and the difficulty of product building above the model layer. | |
| Multimodal Future, Collaborative Agents, and Superalignment | 2 | 2 | 2 | 2 | The guest outlines OpenAI's roadmap around multimodal pre-training, collaborative agents, and superalignment. When the host asks if she is a doomer or accelerationist, the guest gently rejects the binary choice. |
Statements from this episode (19)
Murati: OpenAI and DeepMind were the only AGI-focused labs
“There were two places at the time that were laser focused on this issue and OpenAI and DeepMind.”
Murati: AGI will be the most important technology humanity builds
“There's not going to be a more important technology that we all build than AGI.”
Murati: ChatGPT uses same core model as OpenAI API plus RLHF
“It's fundamentally the same technology, maybe with a small difference, with reinforcement learning, with human feedback. For ChatGPT, but it's fundamentally the same technology, and the reaction, and the ability to grab people's imagination, and to get them to…”
Murati: OpenAI feared ChatGPT was not good enough a week before launch
“It's a strategy that we, we've been using from the beginning and also with ChatGPT where, you know, the week before we were worried that it wasn't good enough. And we also, what happened, you know, we put it out there and then people Told us it is good enough …”
Murati: WebGPT was the direct retrieval-based precursor to ChatGPT
“The precursor to ChatGPT was actually another project that we called WebGPT, and it used, ah, retrieval. To be able to get information and cite sources.”
Murati: OpenAI had already trained GPT-4 before releasing ChatGPT
“One thing that people forget is that actually at this time we had already trained GPT-IV. And so internally at OpenAI, we were very excited about GPT-IV and sort of Put Chagipiti in the rearview mirror.”
Murati: ChatGPT was originally released to gather research feedback for GPT-4
“One of the main things was actually to put chat GPT in the hands of researchers out there that could give us feedback since we had this dialogue modality. And so this was the original intent to actually get feedback from researchers and use it to make GPT-IV m…”
Murati: Scaling RLHF alone might be sufficient to solve AI hallucinations
“And we also wanted to figure out the issue of hallucinations, which is always an extremely hard problem. But I do think that with this method of tree enforcement learning with human feedback, maybe that is all we need if we push this hard enough.”
Murati: Developing AGI in a lab vacuum without user feedback is impossible
“It would not be possible to just, you know, sit in a lab and develop this thing in a vacuum without feedback from users from the real world.”
Murati: No evidence that scaling compute and data is slowing AI progress
“So there isn't any evidence that we will not get much better, much more capable models as we continue to scale them across the axis of data and compute.”
Murati: Reaching AGI will likely require breakthroughs beyond pure model scaling
“Whether that takes you all the way to AGI or not, that's a different question. There are probably some other breakthroughs and advancements needed along the way, but I think there's still a long way to go in In the scaling laws and to really gather a lot of be…”
Murati: OpenAI charter defines AGI as autonomous execution of most intellectual work
“In our chart, OpenAI Charter, we define it as a computer system, basically, that is able to perform autonomously the majority of intellectual work.”
Murati: Current AI models perform most tasks at a human intern level
“With fine-tuning, you can get a lot, obviously but in general, I think we're sort of And most tasks kind of like intern level, I would say that's what I generally say.”
Murati: AI models could become reliable very quickly
“I think though that it's important to Pay attention to this emerging capabilities, even if they're highly unreliable. And especially for people that, you know, are building companies today, you really want to think about, okay, what, what's somewhat possible t…”
Murati: AI systems will perform more human work autonomously
“The trajectory is one where these AI systems will be doing more and more of the work that we're doing, and they'll be able to operate autonomously, but we will need to provide direction and guidance and oversee.”
Murati: AI market will feature a diverse range of models
“You can see even today, you know, we make a lot of models available through our API and from the various, from the very small models to our frontier models, and people don't always need to use The most powerful, the most capable model. Sometimes they just need…”
Mira Murati: Building good products on AI models is incredibly difficult
“There is a lot of focus right now on building more models, but, you know, building good products on top of these models is incredibly difficult.”
Murati: Future AI models will expand pre-training into multimodal inputs
“The world is not just in text, it's also in images, so I think That will certainly expand in, in that direction, and we'll have these bigger models that will have all these modalities and that's kind of the pre-training part of the work, where we really want t…”
Murati: Solving AI hallucinations with browsing and citation is achievable
“And there is a ton of work that needs to happen here and maybe introducing browsing so you can get fresh information and you can cite information and solve hallucinations. I don't think that's impossible. I think that's achievable.”