AI is vulnerable for the same reason Neuralink in healthy humans is a terrible idea. Filtering.
Elon started Neuralink largely because he sees humans losing the intelligence race to AI. He is probably right about that, but his solution of giving ai a high bandwidth connection to our heads is a terrible solution.
I've argued before that all forms of communication eventually lose the battle to spam and other forms of malicious data because the cost of filtering is higher than the cost of communicating. Thus eventually forces data through centralized classifiers that may or may not agree with your definition of signal.
For humans to retain their individualality they need to be able to filter data as it come in as well as have the ability to identify which data is external and which is internal. Low data rates are one method by which we do this. But the second is equally important, fuzzy compression.
By fuzzy compression, I mean a lossy data format in which the compressor does not know exactly which information is being lost because each endpoint is running a unique decompression algorithm. The sender cannot predict what the recipient will receive.
We've all been there. We have a great idea, we compress it into the very best words, hit send and get a blank stare, someone laughing in our face or slapped.
We have similar experiences with AI. HOWEVER, while the first time the message is sent to ai may be unpredicted, it is not unpredictable. We can try it on identical weights offline. Due to the homogeneity of agents, the precise words don't matter nearly as much as the state of the agent after it is exposed to them. Even a very intelligent agent can be coerced into a particular state simply by being predictable. Text to an llm isn't a datastream it is an executable.
You can't do this with humans, UNLESS, you can bypass the decompression stage. If ideas are placed directly into a mind at greater than 10 bits per second then they cannot be evaluated and it isn't clear that the brain would have a way to know that they should be evaluated. Imagine the idea that you should vandalize your neighbors car gets planted in you mind along with the certainty that you have already vetted the idea and found it good and right.
Granted this already happens via trust mechanisms. We all choose trusted external sources as a tool to increase bandwidth without increasing cognitive load for filtering. It is a powerful tool if your trusted sources are trustworthy. The existence of universities proves that trusted sources are worth a great deal. (Actual results aside)
All that to say. Treasure your limited bandwidth, choose trusted sources carefully, prefer obscure or self trained AI agents. I am not sure how to do that last one yet, but if oyu don't want your keys exfilterated by a cleverly worded tweet you'll at least help me think about it.
Elon started Neuralink largely because he sees humans losing the intelligence race to AI. He is probably right about that, but his solution of giving ai a high bandwidth connection to our heads is a terrible solution.
I've argued before that all forms of communication eventually lose the battle to spam and other forms of malicious data because the cost of filtering is higher than the cost of communicating. Thus eventually forces data through centralized classifiers that may or may not agree with your definition of signal.
For humans to retain their individualality they need to be able to filter data as it come in as well as have the ability to identify which data is external and which is internal. Low data rates are one method by which we do this. But the second is equally important, fuzzy compression.
By fuzzy compression, I mean a lossy data format in which the compressor does not know exactly which information is being lost because each endpoint is running a unique decompression algorithm. The sender cannot predict what the recipient will receive.
We've all been there. We have a great idea, we compress it into the very best words, hit send and get a blank stare, someone laughing in our face or slapped.
We have similar experiences with AI. HOWEVER, while the first time the message is sent to ai may be unpredicted, it is not unpredictable. We can try it on identical weights offline. Due to the homogeneity of agents, the precise words don't matter nearly as much as the state of the agent after it is exposed to them. Even a very intelligent agent can be coerced into a particular state simply by being predictable. Text to an llm isn't a datastream it is an executable.
You can't do this with humans, UNLESS, you can bypass the decompression stage. If ideas are placed directly into a mind at greater than 10 bits per second then they cannot be evaluated and it isn't clear that the brain would have a way to know that they should be evaluated. Imagine the idea that you should vandalize your neighbors car gets planted in you mind along with the certainty that you have already vetted the idea and found it good and right.
Granted this already happens via trust mechanisms. We all choose trusted external sources as a tool to increase bandwidth without increasing cognitive load for filtering. It is a powerful tool if your trusted sources are trustworthy. The existence of universities proves that trusted sources are worth a great deal. (Actual results aside)
All that to say. Treasure your limited bandwidth, choose trusted sources carefully, prefer obscure or self trained AI agents. I am not sure how to do that last one yet, but if oyu don't want your keys exfilterated by a cleverly worded tweet you'll at least help me think about it.
7
