Unique Research · Insight / a16z conversation · October 17, 2023 · Full editorial rewrite of ~4,429 CJK — Q&A structure and numbered takeaways preserved
Conversation: OpenAI CTO Mira Murati × a16z GP Martin Casado on AI’s present and the future of human–computer interaction — plus the GPT-4 / ChatGPT story. Source: a16z — Where We Go From Here.
Ten takeaways
- Build intuition and discipline — believe you can solve the problem.
- We are at an inflection: redefining how humans interact with digital information — via AI.
- A week before ChatGPT shipped, OpenAI still worried it was not good enough; once live, people said it was enough and invented new uses — what happens when the tech is easy to get, use, and master.
- ChatGPT’s predecessor was WebGPT (retrieve + cite). Dialogue lets you ask, correct, and express uncertainty — so WebGPT became ChatGPT.
- Scaling data and compute still yields stronger models; AGI may need other breakthroughs.
- People do not always need the strongest model — often a cheaper fit for the use case. Many chase more models; building good products on top is hard.
- Foundation models are decent at text; next is multimodal (image, video) to understand the world more like humans.
- RLHF for reliable task completion; web access for fresh info and hallucinations.
- Integrate into a collection of agents users can collaborate with — and a platform for developers to build agents.
- As models strengthen, aligning them with intent gets harder — OpenAI has a team on superalignment.
Mira’s path
Born in Albania — math/physics emphasis when humanities sources were unreliable. Mechanical engineering → aerospace → Tesla (Model S dual-motor; led Model X launch). Interest in Autopilot and factory robots → Leap Motion AR → OpenAI in 2018 for DALL·E and ChatGPT, pushing public testing and human feedback. OpenAI’s mission beat DeepMind for her: nothing more important than AGI if intelligence is the core unit that lifts collective human capability.
Why so many AI leaders come from math/physics
Math’s theoretical space rewards long dwell time — sleep on a problem, return over days/weeks — not only rapid iteration. Build intuition for which problems matter. Casado: systems/engineering vs science? Murati: both. Systems matter for scale, efficiency, and accessibility so users need not know ML. API vs ChatGPT: similar RLHF tech underneath; imagination and daily use differ completely.
Natural-language interface
Programming is becoming less abstract — natural language to computers — and teaching us to collaborate with machines, not only program them. ChatGPT can feel like a colleague. Inflection: redefine interaction with digital information through AI collaborators — possibly several specialized systems, or one general companion that knows goals and helps. Strategy from the start: put tools in many hands and watch what happens.
OpenAI roadmap (not a grand plan)
ChatGPT was not the original plan. Roots >5 years earlier: safe AI without brittle hand-written objective functions → RLHF to align with human values. After GPT-3 on the API, instruction-following models trained on contractor feedback from customer prompts. Hallucination and uncertainty remained hard. WebGPT retrieved and cited; dialogue’s special properties produced ChatGPT.
Internally excitement shifted to GPT-4; then ~6 months of tuning/safety with ChatGPT as the research feedback channel. Alignment = match user intent; safety also covers deliberate misuse. Strategic bet: ship product, not only lab vacuum — that infrastructure enabled ChatGPT.
Scaling laws
No evidence yet that scaling data/compute stops yielding stronger models; AGI (per OpenAI charter: systems that autonomously do most economically valuable work) may need other leaps. Most tasks still “intern” reliability. Watch unreliable emergent capabilities — founders should ask what is possible today; reliability follows.
One model to rule them all?
Analogy: 1990s co-processors absorbed into CPUs. Future: giant general models vs many specialized ones. Murati: depends on the job. AI will do more of what we do; we still set direction and oversight. Platform already offers small → frontier models via API. Fit and economics matter; OpenAI wants builders with data, customization, and focus on the hard product layer above models.
Next 5–10 years
Multimodal understanding; RLHF for reliable outputs; browsing for freshness/hallucinations; a collaborative agent suite + builder platform; superalignment as models strengthen.