The Hacker Mindset
The most important difference between human and other intelligence is perhaps our unmatched ability to hack.
In previous blog posts, I talked about the limitations of large scale parallel training, which is the foundation of large language models, and the challenges for LLMs to achieve human level reasoning capabilities (note I never said LLMs can’t reason :-). But the most important difference between human and other intelligence is perhaps our unmatched ability to hack - the ability to quickly understand a system to a “just-enough” level, and to utilize it in novel ways to achieve our own goal, be it virtuous or evil.
Our superpower of hacking is why neural networks and LLMs were invented and widely applied to our daily life, even though we never quite figure out how our brain exactly works. It is also why people keep finding examples where LLMs make silly mistakes, and why there is one benchmark after the other to show that LLMs are getting better and better, but never quite there yet.
Hacks in Nature
Nature is full of hacks among living things. Cuckoo birds are known for laying their eggs in the nests of other bird species. Their eggs usually hatch sooner than the host birds’ eggs, and the baby bird will even push other eggs out of the nests. So the host birds, often smaller birds, tirelessly feed the rapidly growing cuckoo chick, impacting their own reproductive success.
Another fascinating example is beaded lacewing larva. These larvae are much smaller than the termites, yet they somehow deceive the termites into accepting them into their nests. Living comfortably among its prey, they use a potent chemical released from their anus to subdue and kill termites.
New hacks in nature depend on evolution, which usually takes millions of years to happen. Humans however, can create new hacks at a much faster speed.
Self Hacking
Humans hack nature to invent useful tools and we hack each other to form a dynamic, complex society. But we can also hack ourselves as well. In the book Tiny Habits for example, BJ Fogg detailed an “algorithm” which you can use to form new desirable habits and get rid of existing undesirable habits. To be able to execute the self-hacking algorithm though, it will depend on human’s self consciousness - our mysterious ability that seems to dynamically create hierarchies in our cognition systems, and to have the higher level system monitoring and reflecting the lower level ones.
Self hacking is one way we defend against undesirable hacks on us. Banner ads, when they were first invented, had an astonishing click through rate of over 40%. Now their click through rate has dropped to below 0.5% because we have learned the clickbaity nature of these ads, and hacked ourselves not to click on them.
Turing Test is an Interactive Hacking Game
Some people claim that LLMs have passed the Turing Test as they have become so good at mimicking human’s thought process and use of language. These claims however, treat Turing Test as a static test based on average human conversations, which is what today’s machine learning benchmarks are already doing. It would be better to think of Turing Test as an interactive hacking game between human and machine, where:
Human tries to hack the computer to show traits that are unlikely human;
Machine tries to hack the human into believing it is human.
With this definition of Turing Test as the standard for AGI, machines have an enormous gap yet to close. It is very easy to construct prompts on which LLMs make basic reasoning mistakes, or self contradicting conclusions.
The Hacker Mindset
Acknowledging hacking as the superpower of human beings, it is a good idea to embrace the hacker mindset - a mindset of always being curious about the new things in the world, of being aware of your goals, and of being proactive in connecting the two. A hacker’s mindset would allow us to extend our ability with powerful tools, and to hack ourselves to guard against undesirable hacks targeting us.