Slowly but surely we're moving towards general AI. There is a marked split across general society and even ML/AI specialists between those who think that we can achieve AGI using current methods and those who dismiss the possibility. This has always been the case, but what is remarkable about today's environment is that researchers keep making progress contrary to the doubter's predictions. Each time this happens, the AGI pessimists raise the bar (a little) for what constitutes AGI.
Just in the last five years, here are some categories of pessimistic predictions that have been falsified:
- "AI/ML can't solve scientifically useful problems" - then AlphaFold changed the protein folding field
- "We're entering an AI winter" [0] - then transformers continued to show promise across multiple domains
- "ML models can't perform creative work" - then came GANs, large language models, DALL-E, and more.
- "Generative ML models are just memorizing the dataset!" - then came multiple studies showing this to be false for well trained GANs, diffusion models and other types of generative models. Take a look at DALL-E 2 generated images of "a bear putting on a shirt in H&M".
- "AGI is impossible - look at language models, they have no understanding of the world and make silly mistakes" - the second part is true, large language models are artificially limited due to being language-focused. Nowadays there are approaches such as Gato and other multi-modal models. Humans develop intuition through multiple sources of information: sight, sound, smell, and touch. Given enough multi-modal context I'm confident multi-modal models will be able to show human-like intuition.
I'm not anti-skeptic. Skepticism is essential to all good science. I think the danger of skepticism with respect to AGI is that we're being complacent. Given the trajectory of improvements in machine learning, we should start preparing for a world where AI is indistinguishable, or far superior, to human intelligence.
[0] - https://www.bbc.com/news/technology-51064369