Demis Hassabis Says Today’s AI Models Still Lack Key Human-Like Abilities
Demis Hassabis, the head of Google DeepMind, was the one to declare that the present-day large language models (LLMs), no matter how sophisticated, still got some areas to improve on that are very crucial. He acknowledges that AI systems at present can articulate their answers very well and even carry out many tasks that seem to be very impressive on the surface but they still lack the overwriting capabilities that are essential to true intelligence.
Hassabis mentions that long-range planning is one of the significant drawbacks. The current models are capable of excellent instant feedback but they cannot reliably handle multiple-step planning over time. This becomes an issue as real-world problem-solving scenarios usually involve coming up with a strategy, revising it each time, and keeping the end goal in sight.
Another point that needs attention is the ability to continuously learn. Hassabis observed that the vast majority of the AI models nowadays do not imitate the human learning process for adoption and sustainability. Instead of gaining new knowledge from everyday encounters, these systems literally remain dormant until they are re-trained in another development cycle. This impedes the AI’s natural ability to adapt to shifting environments, updating information and changing situations.
He also pointed out that there are limitations to the reasoning ability. The contemporary artificial intelligence sounds self-assured and eloquent, but it still might have difficulties with profound reasoning, particularly in cases where the use of logic, making up one’s mind consistently, or discerning intricate cause-and-effect relationships are required. It was unmistakable that Hassabis was equating language fluency with weak thinking capability.
Hassabis is still optimistic about the future of AI, though, and expresses his concerns. He has talked about AI in a good way regarding the drug discovery process, for instance, arguing that the technology could still provide benefits to the real world even if it has not yet reached the level of complete intelligence.
His comments are a reflection of a larger discussion within the AI circle. The majority of the leading figures admit that AI is evolving faster. However, at the same time, they caution that the current systems might not progress any further unless researchers tackle fundamental issues like memory, flexibility, and reasoning. According to Hassabis, enriching these lacking qualities will represent a significant leap towards the creation of AI systems that are more trustworthy, more beneficial, and more akin to human-level comprehension.
