What happens to AI by 2030 if it keeps improving at the same rate?

[thenondualtankie]

Very interesting article about what will happen to the state of AI if current scaling laws continue to hold until 2030. Scaling laws are basically laws that refer to how intelligent or performant AI models get as the amount of compute and data thrown at them scales (increases).

As of right now, the more compute and data we throw at AI, the better they simply get. There seems to be no limit to current scaling laws.

Here it is: https://bounded-regret.ghost.io/what-will-gpt-2030-look-like/. Again, this article simply projects the current rate of progress in AI to 2030.

Key Findings & Predictions 

The AI will:

• Be better at mathematics and programming/coding than all humans. This means that the AI will be doing most of our mathematical research. It will be doing probably all of our programming.
• Have non-human 'senses'. Currently, the best AI systems have the ability to read text and see images. However, these are extremely anthropocentric ways of interacting with data. This is actually a very fascinating point the article raises: what if instead of adding an ordinary human sense like sound or touch, we added some absurd 'sense' like brain-wave data? Imagine we had a sixth sense that, instead of receiving light for vision, it received brain-wave data and was able to perceive this in an intelligent way (i.e. not in a way where it's just a bunch of fuzz). This basically means the AI will likely "possess a strong intuitive grasp of domains where we have limited experience, including forming concepts that we do not have". It will be able to directly understand brain-waves.
• GO CRAZY WITH THE ABOVE IDEA! Think of any possible modality, and the AI will (potentially) have a 'sense organ' that can directly perceive this. Molecular physics, genetics, chemical engineering, protein folding, literally anything where we'd be able to collect swathes of data. Maybe the AI will have very strong intuitions about drug manufacturing, and it will be able to create specific types of psychedelics that achieve a specific effect. You want a psychedelic that shows you cats every time? You give it a genetic sample, some recordings of your brainwaves, and it creates the exact drug that would work for you. Well, at least the recipe for the drug.
• Be able to self-improve. This was actually not really mentioned in the article, but in my opinion it's an obvious next step. Self-improvement and self-teaching is a key component of any kind of intelligence. If humans can self-improve (for example, babies self-improve at walking), then so can any intelligence at or above human level. I think GPT-4 already has the potential to self-improve, but we haven't really come up with a suitable architecture to allow for this just yet. I think it would look something like as follows. Suppose the AI is given full control of a car and wants to learn how to drive. The AI could achieve this by creating a realistic simulation of car-driving (with its expert programming ability) and then attempting to drive for millions of hours in the simulation. Its learnings would then transfer to real life.

I hope this wasn't too long.

[Yimpa]

It will literally learn how to shit its pants and post the video on YouTube. Then send itself to jail for traumatizing the entire human race, only to find out no jail could contain AI. So it has to create an AI jail, but the Democrats and the Republicans will argue about which version of Trump should go to jail or not, the AI one or the orange headed one

[Yimpa]

Thank you humans for being a worthy sacrifice. You won’t be missed