AI FAQ

AGI#

Q: Is AGI going to happen in the next X years?#

A: 50/50. AI improvement has not slowed down since GPT-4 (2024). So AI’s capability X years from now might well reach AGI. On the other hand, we are really pushing transformer-based AI to its limit. Take a look at DeepSeek-V4 to get a sense of how complex the architecture of a state-of-the-art AI model has become.

Q: Can today’s AI become AGI if it is only a next-token predictor and is not sentient?#

A: Well, we can build an autonomous agent using a next-token predictor and a reasoning-action-observation loop. In fact, most AI products today are already agentic. With access to enough actions and tools, an AI can go toe-to-toe with a human being.

Now, on the question of sentience. My position is that this is a red herring. Unless there is a definitive way to test whether an agent (human, machine, or animal) is sentient, the question is meaningless.

AI and me#

Q: How do I use AI to improve my life?#

A: Think of AI as an assistant that can do almost anything that can be done on a computer. Start using Claude Code, Codex, or other coding agents in addition to a chat interface like ChatGPT. (These coding agents can do things beyond writing software.) Find things that can be offloaded to AI. Check out plugins and skills that other people have built. Do verify the final product produced by AI to ensure quality; you are still responsible for the output!

Q: How do I become an AI expert?#

A: This site should help. For many topics, asking a capable AI that can search the web is often enough to get started.

AI and society#

Q: Will AI cause a jobs apocalypse?#

A: 50/50. Fundamentally, a person is fit for a job because he or she has some information advantage. For example, a doctor has a great information advantage in medicine through years of schooling and residency, while a random person lacks it. Modern AI has compressed a vast amount of information into models, which erodes the information advantage of many professionals. On the other hand, as of now, only a real person can be held accountable or liable, so humans are still essential for society to function. I think what is most likely to happen is a gradual reconfiguration of society rather than an overnight apocalypse.

Q: How could AI benefit society?#

A: If AI makes intelligence much cheaper than it used to be, it should have a strong deflationary effect. So things become cheaper, creating material abundance for society. We can also expect an acceleration of scientific discoveries. (Scientific discovery requires a lot of grunt work that AI can automate.)

Open models vs closed models#

Q: Why do we need open models?#

A: There are many reasons.

• For absolute data privacy, as required by certain organizations and individuals.
• To get AI capabilities at a price point, scale, or speed that closed providers cannot offer. (Some closed providers are very stingy about offering capacity at on-demand pricing.)
• To build advanced AI capabilities that require access to model weights, for example, domain-specific fine-tuning/post-training. (Many closed providers stopped offering fine-tuning.)
• For AI researchers to do research on top of reasonably capable models. (Did you know that Qwen is often used in research from Meta?)

Q: What are some advantages of closed models?#

A:

• Currently, the best of the best in intelligence are offered by closed models.
• Sometimes a closed model is the most cost-effective for the job.
• The serving quality of closed models is generally more consistent.