3 ms·

The term "AI" hypes the capabilities and dangers of LLMs by invoking lay people's thoughts about AGI and sentience. I think firms like OpenAI play up this idea

The term "AI" hypes the capabilities and dangers of LLMs by invoking lay people's thoughts about AGI and sentience. I think firms like OpenAI play up this idea to get people interested in the tech; and some firms seem to be working toward legislation to cement their position and limit competition (since just creating an LLM is not a sufficient moat).. and this AGI angle appears to play well in lobbying. OpenAI--one of the most vocal about the danger--is particularly vulnerable to open competition since they lack the device control that Microsoft, Google, and Apple have. And while Microsoft has funded a lot of OpenAI, they actually have a parallel AI team. The moment they no longer need OpenAI, Microsoft can pull out of their deal. OpenAIs position is particularly precarious (similar to Anthropics position--another vocal AI-danger proponent). Ilya says he thinks LLMs are capable of achieving AGI all by themselves, but there's no evidence of that and it's open for debate. LLMs simply generate the next token based on the scoring of previous tokens.. and then does that repeatedly. I'm very skeptical of the idea that the human mind is equivalent to an LLM. Don't get me wrong.. this tech is all very impressive. But the idea that there's no limit to this tech is unfounded. And if people understood the way LLMs generate tokens, I doubt there would be as much discussion about the dangers of LLMs.

benlivengood 2y ago
Gato and RoboCat[0] are examples of what transformer architectures with environment/action tokens can accomplish. These are pretrained with visual and experiential data, but ultimately are "simply generating the next token". This is why it's likely that transformers are capable of AGI, and the top models are no longer strictly LLMs since they are multimodal. [0] https://deepmind.google/research/publications/35829/ https://deepmind.google/research/publications/35829/
bunderbunder 2y ago
> Ilya says he thinks LLMs are capable of achieving AGI all by themselves, but there's no evidence of that and it's open for debate. This debate is probably mostly just an argument over semantics. What exactly do we mean by "Large Language Model"? What exactly do we mean by "Artifical General Intelligence"? On the strict end of things, if "LLM" the current state-of-the-art generative transformer architecture, and if we think that "general intelligence" requires something like sentience, or at least a certain capacity for flexible, iterative symbolic reasoning, then I think that most people would agree that LLMs clearly aren't any more capable of achieving artificial intelligence than a stock 1989 Buick LeSabre is of achieving sustained flight. But the more flexible you're willing to be about how you define "large language model" and "artificial general intelligence", the more room there is for it to be possible. But also, the less practical value there is in making such a statement. Beyond some hand-waviness event horizon, you can probably just declare that any model capable of handling natural language inputs and outputs is a language model, take a nice slow toke, and begin pontificating about the Church-Turing thesis.

benlivengood 2y ago

Gato and RoboCat[0] are examples of what transformer architectures with environment/action tokens can accomplish. These are pretrained with visual and experiential data, but ultimately are "simply generating the next token". This is why it's likely that transformers are capable of AGI, and the top models are no longer strictly LLMs since they are multimodal. [0] https://deepmind.google/research/publications/35829/ https://deepmind.google/research/publications/35829/

bunderbunder 2y ago

> Ilya says he thinks LLMs are capable of achieving AGI all by themselves, but there's no evidence of that and it's open for debate. This debate is probably mostly just an argument over semantics. What exactly do we mean by "Large Language Model"? What exactly do we mean by "Artifical General Intelligence"? On the strict end of things, if "LLM" the current state-of-the-art generative transformer architecture, and if we think that "general intelligence" requires something like sentience, or at least a certain capacity for flexible, iterative symbolic reasoning, then I think that most people would agree that LLMs clearly aren't any more capable of achieving artificial intelligence than a stock 1989 Buick LeSabre is of achieving sustained flight. But the more flexible you're willing to be about how you define "large language model" and "artificial general intelligence", the more room there is for it to be possible. But also, the less practical value there is in making such a statement. Beyond some hand-waviness event horizon, you can probably just declare that any model capable of handling natural language inputs and outputs is a language model, take a nice slow toke, and begin pontificating about the Church-Turing thesis.