Reid Hoffman interviewed Sam Altman a few days ago. Having been blown away by GPT-3 and now DALL-E, I had a listen on the train and took some notes. Enjoy!
- Right now you could do a great copywriting business, or education service
- Biggest short term opportunity is for creatives, being an assistant to them
- We haven’t yet seen people go after the $1T opportunities. Expects a serious challenge to Google in the next few years based on a large language model (LLM)
- Will finally see good chatbots, with great medical/educational services through those
Given large language models are delivered via an API service, how to create an enduring differentiated business on top?
- Middle-layer becomes really important: tuning an existing LLM
- eg tuning a model for medecine, education, friendship
- with unique flywheel which improves over time
- “We will not be doing prompt engineering in 5 years”
- Doesn’t think we will see many new base models
What will some of the new surprises be?
- For example people were surprised OpenAI could train from the internet data, but then be used to produce code (Copilot)
- Doesn’t really answer the question, but does say that LLMs will create net new knowledge, it’s not just about indexing existing stuff
- Science-dedicated products like AlphaFold. Will see a lot more of these
- Tools that make engineers/scientists more productive. Will have an outsized impact
- Will come later: AI scientist that can self-improve
- Thinks the AI can help us align it… for instance you could ask it “don’t be racist” once it grasps these higher level concepts
Upcoming AI moonshots?
- Language models will go a lot further than what people think. A lot of algorithmic progress to come
- Multi-modal models coming. Not just texts and images, other media and all integrated
- Models that continuously learn. For instance GPT-3 right now does not improve based on how much it is used
- “Marginal cost of intelligence and marginal cost of energy will trend towards 0”
- “AI will seep into everything” (See Andrew Ng AI is the new Electricity on this)
Metaverse and AI?
- Expects that in the upside case the metaverse is on the order of a new compute platform like the iPhone
- Whereas AI is like a technological revolution
Life sciences and current LLMs
- So far models are not good enough
- Some parts of the cycle time cannot be sped up, eg testing
- “Simulators are still very bad”
Best utopian science-fiction?
- Star Trek
- The Last Question
- Ian M Banks on Culture series (really need to try again to read this…)
Main societal issues in next 20-30 years as AI continues to grow?
- Will need to work on wealth redistribution, how to govern it, and how to ensure fair access
- OpenAI run the largest UBI research project in the world. 4 of 5 years in
- Looking at how AI can help reskilling
- The narrative was AI will come for blue collars, white collars, then artists and creative fields
- But it seems to be going exactly the opposite way
Will leave you with…