Can GPT‑4 Really Threaten Humanity? Inside Sam Altman’s Candid Chat with Lex Fridman

GPT‑4内幕大曝光

Altman explains that GPT‑4 was completed last summer and has since been aligned to better meet human needs. Training still follows the pre‑training plus RLHF pipeline, but with far fewer RLHF examples that prove decisive. Data came from public datasets, partner contributions, and a small amount of Reddit memes.

Even the OpenAI team cannot fully interpret GPT‑4; they continue to probe its reasoning by asking it questions and analyzing its responses. Altman notes that GPT‑4, like ChatGPT, has begun to exhibit reasoning abilities, though the origin of this capability remains unknown.

“Even with a large model, we can predict its overall performance by training only a part of it, similar to forecasting whether a one‑year‑old could pass the SAT.”

To address differing user values, OpenAI opened system‑message permissions, allowing users to steer GPT‑4’s persona, such as making it answer in JSON or adopt a Socratic style.

最会打太极的CEO

When pressed about GPT‑4’s size, Altman gives vague answers, mentioning public data, partner‑provided code, and internet content, while dismissing the rumored 1 trillion‑parameter claim as false.

He acknowledges that GPT‑4 is the most complex software humanity has built, yet emphasizes that practical utility matters more than raw parameter counts.

“People are no longer fascinated by CPU clock speeds; they care about what the technology can do for them.”

Altman also discusses the infamous GPT‑4 jailbreak, comparing it to early iPhone jailbreaks, and notes that OpenAI’s response is to make the system robust enough that extensive permissions are unnecessary.

GPT的“偏见”不会消失

Altman admits bias is inevitable in GPT models and will never disappear, as the models reflect the diverse perspectives of the external data they are trained on.

“Bias is a result of personalized control; iteration refines that control.”

He stresses that external contributions are essential for improving AI and that the bias issue must be openly addressed.

One More Thing

In the closing segment, Altman offers a contrarian piece of advice to young people, warning against the allure of “success‑hacking” posts and encouraging them to ignore most advice.