Artificial Intelligence and Kahnemann's Fast and Slow Thinking
The Large Language Models will keep evolving. They will keep changing our lives and we’re only seeing the tip of that iceberg yet.
According to Substacker Alberto Romero, what sparked the events at Open AI some 10 days ago - CEO Sam Altman first being fired by the board, only to be rehired a few days later, while most of the board were removed - was probably linked to disagreement regarding the new Q* (Q-Star) model purported to be able to solve complex mathematical problems, something current Large Language Models are unable to do. In Romero’s view the events reflect a deep-seated conflict between the company’s stated aim of developing a safe AI that benefits all of humanity and the general goal of any technology company of developing and marketing cutting edge solutions, being at the forefront in its field. Perhaps a prime example of what may happen when a business tries to achieve conflicting goals.
But what is the Q* model?
A few days back, OpenAI’s Andrej Karpathy published a video where he explains the Large Language Models, what they are, how they work and how they may be expected to develop. This is a must-see for anyone who wants a one-hour overview of the subject. Among other things, Karpathy discusses the Q* model and its importance. He describes its capabilities by reference to Daniel Kahnemann’s differentiation between fast and slow thinking, or what Kahnemann refers to as System 1 and System 2 thinking. An example of System 1 thinking is when we respond to the question of what is 2+2. We respond immediately, using information “cached” in our minds. But when we’re asked what 24*47 is, most of us cannot respond immediately. We have to go into the System 2 mode of thinking, the slow thinking mode.
As Karpathy explains, the Q* project brings a large language model to the level of System 2 thinking. This is a vast improvement in the capability of such a model and a significant step toward Artificial General Intelligence, a model that vastly surpasses human cognitive abilities. It sounds credible if panic regarding the implications of this and distrust regarding Altman’s plans for it were the driving factor behind the decision of the board. What they seem to have failed to realise however is that in the end the knowledge that drives the company resides within the minds of the employees and managers and a new technology, however frightening, will not vanish into thin air whatever a corporate board decides, even to the point of destroying the business out of concerns for the greater good of humanity.
The Large Language Models will keep evolving. They will keep changing our lives and we’ve only seen the tip of that iceberg yet. Until now, the models have been capable only of System 1 thinking. But this will change soon. The question we are faced with is how to respond when this happens. I may be wrong, but I still stick to my earlier conclusion that we must our best to make use of this new technology to improve and speed up our own thinking. That’s the narrow road we have to take.
Perhaps you're right that these are tools that can enhance thinking, Thorsteinn... or elsewise the scope of this system design is still capped by its fundamental dependencies. My Masters degree in 'Advanced Computer Science with a Specialisation in Artificial Intelligence' (easily the worst degree title I've encountered...) left me with a near-permanent scepticism. I graduated in 1995. It is now nearly thirty years later. We are still using approximately the same techniques, only the processing power and scope of the data sources has changed. I do not see what grounds for assuming acceleration there could be that don't emerge from the marketing suite.
Apropos of the general enthusiasm for AI, I'd like to share this piece from June this year:
https://strangerworlds.substack.com/p/laws-of-robotics
Stay wonderful!