Google apparently in recognition that they lost LLM game (first to OpenAI chatgpt and then to Chinese deepseek) started drum rolling for AGI. Given how people are afraid of AGI, they brought out a paper on responsible AGI (somewhat similar to earlier responsible AI ). Read the Google post here.
Does it mean that LLM ≠ AGI
There were posts earlier where OpenAI said chatgpt is almost AGI. Even some in Google Gemini team said Gemini is almost sentient. But if people are talking about AGI seperately from LLM, perhaps that is acceptance of that fact that LLM may not ever reach human capability of intelligence. In fact there was a recent study that asserted that LLM could not match human ingenuity of ‘zero shot abstract thinking’.
Martha Lewis, a coauthor of the study, tells LiveScience, – while we can abstract from specific patterns to more general rules, LLMs don’t have that capability. “They’re good at identifying and matching patterns, but not at generalizing from those patterns.” Read the full LiveScience post.
Can the AGI be responsible?
AGI has to evolve from current generative AI framework. Google Deepmind has categorized the challenge into four baskets:
- Misalignment
- Misuse
- Mistake
- Structural Risk
Misalignment refers to AI model doing something that developers did not intend it to do. Misuse refers to the model being misused by a human controller/user to work as adversary to humanity. Mistake refers to AI model doing something bad without triggering internal checks and balances. Structural Risk ensues from multi-agent dynamics without any fault from individual model. Example for this can be, a new path opens up due to complex interlinking of activities between AI models which developers didn’t intend to.
Of all the four types, Misalignment and Structural Risk are the most difficult to address because they are the most complex and difficult to uncover. We will limit ourselves to the issue of Misalignment for this post. Marcus Arvan explained in a LiveScience post, ‘If any AI became ‘misaligned’ then the system would hide it just long enough to cause harm’. He has given real life examples where LLM shocked the user with the answers.
‘The basic issue is one of scale. Consider a game of chess. Although a chessboard has only 64 squares, there are 1040 possible legal chess moves and between 10111 to 10123 total possible moves — which is more than the total number of atoms in the universe. This is why chess is so difficult: combinatorial complexity is exponential.
LLMs are vastly more complex than chess. ChatGPT appears to consist of around 100 billion simulated neurons with around 1.75 trillion tunable variables called parameters. Those 1.75 trillion parameters are in turn trained on vast amounts of data — roughly, most of the Internet. So how many functions can an LLM learn? Because users could give ChatGPT an uncountably large number of possible prompts — basically, anything that anyone can think up — and because an LLM can be placed into an uncountably large number of possible situations, the number of functions an LLM can learn is, for all intents and purposes, infinite.’ He argues.
Google’s approach to tackle misalignment
Google plans to use second AI model to validate a model’s answer. It’s sort of AI police to police AI model. Can this work? To find answer, we should ask, does policing work for human citizens? You could say mostly. But we should not miss that mostly it’s willingness of human citizens to follow rule and respect police that makes the job of police manageable. We surely cannot assert that for AI citizens. But given that entire AI paradigm works on goal-seeking principle, meticulous control of goal dynamics may provide an understanding of AI motivation.
Demis Hassabis, the Nobel Laureate CEO of DeepMind has a sobering thought about the approach to AGI. In a lecture in Cambridge University, he said, “Lot of Silicon Valley Companies work with the principle of ‘Move Fast, Break Things’ but I think it is not appropriate, in my opinion, for this type of transformative technology. I think instead we should be trying to use scientific method and approach with humility and respect this kind of technology deserves. We don’t know lot of things. Lot of unknowns about how this technology is going to develop. With exceptional sort of care and foresight we can get all the benefits and minimize the downside of it.” He invites people to focus on research and debate now and be mindful that the technology does not really go out of hand.
We are with him on that thought.