A compressed version of Scottish Mortgage’s conversation with the Chinese AI lab MiniMax’s CEO and founder, Dr Yan Junjie (IO).
Listen to the full episode here.
If you’ve ever watched AI-generated videos on YouTube, there’s a good chance they were made using one of MiniMax’s models. The firm has also become popular outside China for allowing others to run and modify the ‘weights’ that underpin its technology at no charge. In this rare English-language interview, IO discusses MiniMax’s business model with investment manager Lawrence Burns and explains how Chinese AI companies differ from Silicon Valley’s.
Timecodes:
00:00 Introduction
01:28 Dr Yan Junjie (IO) interview begins
01:42 Minimising inference cost, maximising intelligence
02:27 Defining AGI
02:59 MiniMax’s M- and H-series AI models
04:33 Low inference costs as a ‘moat’
05:55 Open source as a catalyst for monetisation
08:38 AI ecosystems: China v US
10:24 ‘The largest post-training system’
11:11 A deeper understanding of the universe
Glossary (in order of mention):
Inference and inference cost: Inference is the process of running a trained AI model to produce an answer or other output. Inference cost is the computing expense of doing so.
Frontier model and frontier intelligence: A frontier model is among the most capable AI systems available at a given time; frontier intelligence refers to that leading edge of capability.
Artificial general intelligence (AGI): A proposed form of AI able to perform a wide range of intellectual tasks rather than being limited to one narrow use. There is no universally agreed-upon definition. However, IO suggests it will be attained at the point where new scientific theories and breakthroughs “mostly depend on AI”.
Large language model (LLM): An AI model trained on large amounts of text to understand and generate language.
AI agent: An AI agent can plan and carry out tasks on a user’s behalf.
Open-weight model: An AI model whose learned numerical parameters are available for others to download, run and often adapt; this does not necessarily mean that all its code or training data is open.
Model scaling: Expanding the data, computing power, model size or training activity used to build or improve an AI system.
Closed model: An AI model whose provider keeps its weights private, so users normally access it through that company’s product or an ‘application programming interface’.
Fine-tuning: Further training of an existing AI model on targeted examples so that it performs particular tasks or behaves in a desired way.
Feedback loop: A repeated cycle in which information from users feeds into model improvements that shape later results.
Post-training: Additional model development after initial training, used to improve how it follows instructions or handles particular tasks.
Self-hosting an AI model: Running a model on computing infrastructure controlled by the customer rather than relying on the model provider to operate it.
AI sovereignty: The ability of an organisation or government to control the AI models, computing infrastructure and data it relies on, reducing dependence on an external provider.
Compute: Shorthand for the processing power and hardware resources used to train or run AI models.
Return on investment (ROI): A measure of the gain or benefit from an investment relative to its cost.
Hedge fund: An investment fund that can use a broad range of strategies, typically offered to professional or wealthy investors.
Reinforcement learning: A machine-learning method in which a system improves through feedback framed as rewards or penalties for its actions.
Data synthesis: The creation of artificial data that imitates patterns in real examples and can be used to train or test an AI model.
Check the podcast description to ensure this content is suitable for you. Your capital is at risk.
Presenter: Claire Shaw
Executive Producer: Leo Kelion
Line producer: Jessica Rooney
Broadcast Technician: Samuel O’Hare
Editor: Jody Black
","author_name":"Scottish Mortgage"}