Meanwhile, the actual research community tells a different story. A 2025 survey by the Association for the Advancement of Artificial Intelligence (AAAI), surveying 475 AI researchers, found that 76% believe scaling up current AI approaches to achieve AGI is “unlikely” or “very unlikely” to succeed. The researchers cited specific limitations: difficulties in long-term planning and reasoning, generalization beyond training data, causal and counterfactual reasoning, and embodiment and real-world interaction.
I am not at their level yet, but this is my take too.
IMO until we truly understand human intelligence / consciousness, we don’t have a benchmark for whether the machine has achieved AGI.
Not to say the current approach of brute-forcing would never work. IMO more work can be done in areas beyond vision and natural language. Personally, I am interested in somatosensation.
Another subfield of AI that looks promising is reinforcement learning. Not sure if these are the correct terms, but all these models do “offline” learning. Yeah yeah, there’s RLHF and whatnot, but my understanding is that it has always been split into a training phase and an inference phase. I wonder if it’s possible to do “online” learning, in which the model actually incorporate new information into its weights in real-time, and use said info right away.
I am not at their level yet, but this is my take too.
IMO until we truly understand human intelligence / consciousness, we don’t have a benchmark for whether the machine has achieved AGI.
Not to say the current approach of brute-forcing would never work. IMO more work can be done in areas beyond vision and natural language. Personally, I am interested in somatosensation.
Another subfield of AI that looks promising is reinforcement learning. Not sure if these are the correct terms, but all these models do “offline” learning. Yeah yeah, there’s RLHF and whatnot, but my understanding is that it has always been split into a training phase and an inference phase. I wonder if it’s possible to do “online” learning, in which the model actually incorporate new information into its weights in real-time, and use said info right away.