How do we define Artificial General Intelligence (AGI)?

Defining AGI is challenging because we don't fully understand human intelligence. A common definition involves machine intelligence that can perform any intellectual task a human can, at least as well as the best human. However, some argue current AI is far from consciousness or intent, relying on mass language modeling.

Can current AI like ChatGPT be considered AGI?

Most speakers believe current AI, including models like ChatGPT or GPT-4, are not true AGI. While they can exhibit impressive logic and language skills, they primarily perform mass language modeling and pattern matching, not true consciousness or understanding.

How will AGI impact society?

The impact of AGI is uncertain and could be positive or negative. Its development and deployment rely on agreeing on core principles, ensuring adherence to human values, and implementing proper governance for beneficial use.

How will we know when we have achieved true AGI?

Knowing when AGI is achieved is a hard question, with no consensus. One perspective is that it will be when we can no longer identify tasks that humans can do significantly better than AI, effectively crossing a threshold in collective consciousness.

Is AGI already here, or is it still far off?

Opinions are sharply divided. Some believe AGI is already here or very close, either as a continuum of current tools or that GPT-3 was already AGI. Others feel we are very far from AGI, considering current AI to be pattern matching rather than true intelligence.

Key Moments

AI Startup Founders Debate the Creation of Artificial General Intelligence

Y Combinator

Science & Technology5 min read6 min video

Nov 21, 2023|47,236 views|564|51

YC Y Combinator yt:cc=on

Save to Pod

Key Moments

TL;DR

AI founders predict AGI is decades away or already here, disagreeing vehemently on its arrival and definition, raising questions about whether current 'intelligence' is mere pattern matching.

Key Insights

While some founders believe AGI could arrive in 4-5 years, the majority predict it will take at least a decade or two, with some suggesting it's still very far off.

A significant portion of founders question the definition of AGI, with disagreement on whether current models like GPT-4 exhibit true reasoning or are simply sophisticated pattern predictors.

Some argue that the tools available today, even if not complete, can be assembled to solve arbitrarily complex problems, hinting that AGI might be an emergent property rather than a distinct breakthrough.

The ability to know when AGI has been achieved is identified as a major challenge, with the idea that it might be a spectrum rather than a binary state.

Concerns are raised about AI confidently making factual errors (e.g., 50 million > 100 million), highlighting a lack of genuine reasoning compared to human capabilities.

The impact of AGI is seen as dependent on human ability to agree on core principles for its development and deployment, emphasizing the need for ethical considerations and governance.

Varying timelines for AGI creation highlight fundamental disagreements

The consensus among Y Combinator-backed AI founders regarding the creation of Artificial General Intelligence (AGI) is far from uniform. While a minority suggest it could be here within the next four to five years, many express a more conservative outlook, placing AGI's arrival at least a decade or two away. This divergence in timelines underscores a deeper disagreement about what AGI truly signifies. For instance, one perspective defines AGI as machine intelligence capable of performing any task a human can, at least as well as the best human. However, even with such a definition, the timeline remains uncertain. Robert and a colleague held opposing views, one predicting it within their lifetime but not in the next 10-20 years, while the other believed it could happen in a couple of years. This spectrum of predictions, from imminent to distant, suggests that the path to AGI is not perceived as linear or predictable, with unpredictable breakthroughs potentially accelerating or delaying its arrival.

The definition of AGI remains a significant point of contention

A core issue shaping the AGI debate is the lack of a clear, universally accepted definition. Respondents frequently wrestled with how to quantify 'general intelligence' when human intelligence itself is not fully understood. The concept of an autonomous agent within foundation models, possessing thoughts and beliefs akin to humans, is dismissed by some as 'BS.' Instead, current AI like ChatGPT is characterized as performing 'causal mass language modeling,' primarily focused on predicting the next token. This predictive capability, while powerful, is seen by some as a far cry from consciousness or intent. The difficulty in defining human intelligence—how, why, and what it means to be intelligent—makes it even harder to set a benchmark for AGI. If AGI merely mimics what we perceive as human intelligence, is that sufficient? This philosophical quandary complicates any timeline or assessment of progress.

Evidence for current AI exhibiting 'human-like' traits is debated

The perception of AI's current capabilities is another area of stark disagreement. Some founders feel that AGI, to some extent, is already here, arguing that today's tools, though incomplete, can be combined to solve highly complex problems. One respondent even believed GPT-3 was already AGI, viewing current advancements as a continuous evolution rather than a discrete leap. This viewpoint suggests that AGI might not be a singular invention but an emergent property of increasingly sophisticated systems. Others, however, strongly disagree, stating that current AI is 'definitely not intelligent.' They point to basic errors, such as AI confidently asserting that '50 million is greater than 100 million,' as evidence that the systems are not truly reasoning but rather matching patterns and making statistical predictions. The confidence with which AI presents information, often mimicking human language, can be deceptive, masking a fundamental lack of understanding or logical deduction.

The challenge of recognizing AGI's arrival

Beyond predicting when AGI will be created, a significant challenge lies in knowing when it has actually been achieved. One founder admitted they haven't reconciled what 'true AGI' means. The idea that AGI exists on a spectrum is considered plausible, with different people potentially recognizing certain capabilities as AGI at different times. For some, moments where advanced models like GPT-4 exhibit logical inference—where pasting information allows the AI to infer intent—are signals that AGI is getting close. The 'Turing Test' is mentioned as a potential benchmark, but the goalposts may shift. The moment when we can no longer think of specific tasks that humans can still perform better than AI might be the unofficial marker for AGI's crossover into the collective consciousness.

The potential impact of AGI hinges on human guidance

The potential societal impact of AGI, whether positive or negative, is viewed not as predetermined but as contingent on human actions. Founders emphasized that the outcome depends heavily on our collective ability to agree on core principles for both the development and deployment of these advanced models. Establishing governance and control mechanisms is crucial to ensure AGI benefits society while staying within acceptable boundaries. Moral and philosophical questions are expected to intensify as we approach AGI. The ideal scenario envisions AGI operating based on human values and backed by robust governance to help people lead more fruitful and connected lives. This perspective places significant responsibility on humanity to shape AGI's integration into society ethically and beneficially.

AGI could be closer than experts believe, or further than the public imagines

The perception of AGI's proximity is starkly divided. Public discourse, often fueled by interactions with advanced tools like GPT-4, can create a sense that AGI is 'right around the corner,' feeling incredibly real and human-like. However, founders involved in the deep technical development suggest this perception is misleading. Building capabilities beyond language, such as the complex 'brains' required for video game characters, involves many elements that current AI lacks. This suggests that the leap from sophisticated language models to true general intelligence might be underestimated by the public. Conversely, some experienced researchers believe AGI is 'probably a lot further away than a lot of people think,' implying that the current trajectory might not lead directly to AGI without significant paradigm shifts or fundamental discoveries.

Mentioned in This Episode

●Software & Apps

●Concepts

Common Questions

There's a wide range of predictions, with some believing it could happen within our lifetime, possibly in 4-5 years, while others think it's at least 10-20 years away, or even further. Some believe it might be a spectrum and that certain capabilities already exist today.