I believe LLM’s already have greater general intelligence than any human.

I believe LLM’s will continue to improve their general intelligence until it marginal improvements become unprofitable or a different technology supersedes it.

But you don’t want a generalist to do a specific task, you want a specialist, and for that reason I believe it is still possible to beat LLM’s in a specific task.

When the costs and benefits of a specific task are extremely large, the value created by completing the task optimally can be significant.

So for these specific tasks, LLM’s will not be able to beat specialised intelligence in the long run, therefore making the pursuit of specialised intelligence for these specific tasks a worthwhile investment.

Where are costs and benefits of a specific task extreme?

Ones that result in death, disease, damage, disaster.

Ones that result in irreversible change.

Ones that result in wealth transfer.

Ones that result in the reallocation of scarce resources.

In the western world a good example of death, disease, irreversible change would be the health sector. And a good example of reallocation of scarce resources and wealth transfer would be all financial services or business strategy.

I would not trust any LLM for medical advice, when it recommends a fatal dose of the incorrect medicine and challenged afterwards it will respond nonchalantly “Oh yes, great catch.”

This level of uncertainty is so unacceptable it’s not even debatable.

So the way I would beat an LLM is to find an extremely valuable business problem.

Filter out any problem which cannot be broken down into basic inputs and outputs.

I would likely create a simulation or reinforcement learning model to find the “optimal” strategy for maximising the outputs using minimal inputs.

If the problem was valuable enough to warrant this level of precision, I would break it down and apply the laws of physics as my primary inputs.

“Physics is the law, everything else is a recommendation” — Elon Musk

This strategy would yield results which are not beyond the comprehension of LLM’s, but they are not accessible to you, after all you are a nobody. It is important to remember that LLM’s like ChatGPT are a product, you are the customer not the owner. If your task requires significant energy to solve ChatGPT actively refuses.

Don’t believe me, ask ChatGPT to “count every number from 1 to 1,000,000" and see if it is willing to comply with this simple task.

Don’t write software to count to 1,000,000.

But creating a solution which requires millions of simulations to identify optimal solutions is one way to beat LLMs.