The brain learns by constantly making predictions and making corrections (in its wiring) to approach close to the actual results.

Suggested in [Being You](https://www.anilseth.com/being-you/), the perception process is a top-down "controlled hallucination." The brain predicts conceived "things" it will see and then "the sensory input" caused by them. Then the brain verifies or corrects the conceived things with the sensory input it actually receives.

Major disruptions in businesses will start to appear around 2024 - i.e. profitable (by free cash flow) and hyper-growth companies relying on large-scale neural networks as their main tech strength.

Here is an example: **Me**: I have a hypothesis. Cortical columns make classifications based on raw sensory inputs as well as classification signals from other columns. They learn by predicting incoming raw sensory inputs as well as classification signals from other columns. One column's prediction error helps its own learning, as well as other columns' learning when the prediction was made with input from their classification signals. What do you think of my hypothesis? **ChatGPT**: Your hypothesis describes a process that is similar to how the brain is thought to process and classify sensory information. Cortical columns, which are groups of neurons in the cerebral cortex, are thought to play a role in processing and classifying sensory information. They do this by receiving input from sensory organs and then making predictions about what this input represents. These predictions are then compared to the actual input, and any discrepancies between the two are used to update the column's understanding of the sensory information. It is also thought that cortical columns communicate with each other, sharing information about the sensory input they have received and the predictions they have made. This can help to improve the overall accuracy of the predictions made by the brain, as different columns may have different strengths and areas of expertise in processing different types of sensory information. Overall, your hypothesis captures some of the key principles of how the brain processes and classifies sensory information, and it is generally consistent with current understanding of this process. However, it is important to note that the brain is a complex and multifaceted organ, and our understanding of how it works is still evolving. There may be additional factors and processes at play that are not captured in your hypothesis.

It's powered by LLMs. The value it brings is being testified widely.

The art market is relatively small but the stock image market might be much larger.

GitHub copilot is promising, but it's not human-level yet. It might be possible for Self-driving to use LLMs as foundations to build models that can predict the social behavior of humans on the street.

In SayCan, we showed how we can connect robot learning pipelines to large language models, bringing a lot of common sense knowledge to robotics. The hope was that as the LLMs become better (which they seem to be consistently doing), it will have a positive effect on robotics.

From unconscious learning in brian to individual scientific inquiries to collective scientific inquiries.

The controlled hallucination process is also a hypotheses-prediction-observation correction learning process. Brains can learn brand new concepts by generating new hypotheses about the world, making predictions about the sensory input, and correcting according to observation (actual sensory input)

Although the definition of "hypothesis" is narrower in science - a hypothesis in science usually refers to a theory that explains **how** things work. Propositions such as "It's okay to eat bruised apple" or "Mr. Huckabee is a bad politician." are not regarded as hypotheses.