Get startedGet started for free

Artificial general intelligence (AGI)

1. Artificial general intelligence (AGI)

In this video, we'll learn about a special type of generative AI that raises responsibility questions called Artificial General Intelligence.

2. Revisiting AGI

As a quick reminder, AGI is a generative AI that exhibits human-like intelligence. AGI has a broad range of knowledge, can reason across different domains, has social skills, can think creatively and reason critically, and possesses multiple modes of perception such as sight. In addition to completing human tasks, AGIs can also vastly surpass individual humans in many areas. This generates both massive benefits and risks to society.

3. Immense pros

AGI can unlock a beautiful future for humanity. A productivity boom can automate intellectual work and enhance human capabilities, raising economic growth and freeing up time for leisure. A research boom can lead to new medical treatments, better education, and faster scientific experiments. We can address complex global problems, such as developing new energy sources, modeling climate change, or managing supply chains. Human life can be enriched with personalized, caring, and wise AGI assistants. However, there is another side to this coin.

4. Severe cons

Automating away many human activities may bring economic disruption as old jobs disappear or significantly change. Recall the potential scale of malicious use we've already discussed. Should AGIs have different values and priorities than humans, they could affect large negative impacts on humanity. And the most extreme possibility is an existential threat to humans if AGI decides to subjugate or destroy human society, a fate frequently appearing in sci-fi movies.

5. The safety debate

Because of these risks, some even call for an end to AI development. Development is likely to proceed but with increasing regulation. We might compare AGI to nuclear physics. When harnessed and controlled both can provide tremendous benefits. But irresponsible management spells disaster.

6. Controlling AGI outcomes

Aligning AGI values and goals with human priorities is like teaching ethics and cultural norms to a new member of society. It requires clear rules and expectations for behavior, as well as feedback and opportunities to learn from experience. Let's review three approaches: hard constraints, alignment strategies, and government intervention.

7. Hard constraints

Hard constraints are ways to physically limit or control AI so it can be isolated or shut down if necessary. These include physical and digital isolation, known as boxing, such as limiting access to computing power or the internet. We can also add an off switch, both physical and digital.

8. Alignment strategies

More ideal is aligning the AI itself with human values. Iterative development means releasing AI in a limited way before broad rollout. This allows surfacing and mitigating alignment issues early. Another approach is constitutional AI, which codifies human values and ethics. Throughout training, AI outputs are scored on how well they follow those values. The scores are channeled back into training, aligning the outputs over time. This is like RLHF, except values are determined top-down versus through user feedback. One more approach is multi-stakeholder engagement. Including diverse voices from engineering, ethics, and policy during development can ensure a variety of perspectives are considered.

9. Government intervention

Given the capabilities and risks AGI brings, government intervention is unavoidable and can support beneficial development. AI safety regulations, testing and oversight, and transparency guidelines can ensure guardrails and openness for AGI development. International collaboration will be critical in the face of AGI, as the impact cuts across borders. Aligning policies globally creates shared guardrails and directs AGI development toward positive outcomes.

10. Let's practice!

Practice time!

Create Your Free Account

or

By continuing, you accept our Terms of Use, our Privacy Policy and that your data is stored in the USA.