It's Too Late To Close Pandora's Box on AI
The plan to "pause" large language model development was never going to work
Those calling for a pause on large AI model development have a misunderstanding about both the nature of technological progress and the nature of the technology they are trying to “pause.”
While I agree that a misaligned superintelligent AI is indeed an existential risk, I don't think edicts to pause large model development are either possible or sufficient to stop more powerful models from being created.
While ChatGPT and Midjourney hold the spotlight, there has already been an explosion of new deep-learning models, and existing models are being used to create improved successors.
After the initial breakthrough, incremental technological progress is usually decentralized, and deep learning models are no exception. After the initial team makes a breakthrough (such as the Transformer model) and it becomes widely known that something is possible, the technology takes on a life of its own, outside the control of any one group. When anyone with access to fast hardware can create improved models, progress is impossible to control.
There are now language and image models that are nearly as good as GPT and Midjourney overall and superior in certain attributes, like efficiency. Furthermore, GPT-4 and Midjourney are being used to train their successors. GPT-4 is being used to train competitors (ex: GPT4 x Alpaca) and test the quality of competing models, as the team behind Vicuna did. Likewise, a number of models were initially trained on Midjourney images but have now exploded into thousands of competing models for different applications.
Moreover, large models have been getting more efficient at an incredible rate. Vicuna is 90% as good as GPT and can run on a MacBook CPU. In image models, additional network modules like LoRA are accelerating the training of large image models. There's now a huge and rapidly evolving ecosystem of models available on sites like huggingface.co and civitai.com for everything from text-to-video, robotics, to NSFW chatbot and video models. Restricting “large CPU clusters” incentivize thousands of teams to make more efficient models.
Given the infeasibility of pausing AI progress, our efforts should be channeled toward promoting AI safety research and fostering a consensus on safety standards. I recognize that the prospect of aligning a superintelligent AI with our values seems impossible at the moment, but we should try anyway:
We have a greater likelihood of success if we focus all our energy on alignment rather than waste it on useless advocacy or politics.
Failure has value: as a last resort, it can be used as evidence for policy changes.
We will use increasingly competent AI to help. There may be a short window where AI can help design control mechanisms for its successors.
It's possible that deep learning models will naturally align with human values, and our concern was misguided.