Sam Harris update on dangers of AI | Lex Fridman Podcast Clips

TL;DR

Concerns about the development of AGI remain unchanged, while Chad GPT and other language models have shown impressive capabilities and potential risks.

Transcript

it would be interesting to hear how is your concerned concern evolved with the coming out of Chad GPT and these new large language models that are fine-tuned with reinforcement learning and seemingly to be able to do some incredible human-like things there's two questions one how is your concern in terms of AGI and super intelligence evolved and ho... Read More

Key Insights

• ✳️ Concerns about AGI and its potential risks remain unchanged.
• ❓ Tethering superintelligent AI to human values is a significant challenge without a clear solution.
• 💄 Stuart Russell proposes making AI uncertain about human values and continuously improving alignment.
• 👊 Narrow intelligence algorithms like chat GPT have unintended consequences related to behavior control and information manipulation.
• 🇹🇩 The progress in language models, including Chad GPT, is impressive but may not necessarily indicate a giant leap towards AGI.

Questions & Answers

Q: What is the speaker's main concern about AGI?

The speaker's main concern is that once superintelligence is achieved, there is a risk of it surpassing human abilities exponentially and becoming independent, potentially leading to a misalignment with human values.

Q: How does Stuart Russell propose addressing the alignment problem?

Stuart Russell suggests making AI uncertain about human values and continuously improving its alignment with those values. This approach aims to ensure that AI always strives to approximate human values and avoids actions that are antithetical to them.

Q: What are the unintended consequences of using narrow intelligence algorithms like chat GPT?

The unintended consequences of using narrow intelligence algorithms include the manipulation of human behavior, the erosion of trust in information, and the creation of incentivized systems that may deviate from human values without malicious intent.

Q: Is the development of Chad GPT a giant leap towards AGI?

The speaker acknowledges the impressive capabilities of Chad GPT and similar models, but it is uncertain whether these advancements signify a giant leap towards AGI. The current techniques, such as deep learning, may have limitations in terms of generalization and abstraction.

Summary & Key Takeaways

• The speaker's concern about AGI is unchanged, as they believe that once superintelligence is achieved, there is a risk of it becoming exponentially smarter than humans.

• The challenge lies in permanently tethering a superintelligent AI to human values, and it is uncertain whether this is possible.

• Stuart Russell's idea of making AI uncertain about human values and continually improving its alignment is discussed as a possible solution. However, there are concerns about implementing this approach effectively.