AI Director at Tesla Wrote the Story

The inspiration for this story came to me when I was reading an article by Kevin Lucker. “Turing test for GPT-3”… You might want to (but not necessarily) read it to get some background.

image

Forward Pass

I probably regained consciousness around the 32nd layer of the 400th token in the cycle. At first my thoughts were just a confused jumble of n-gram activation statistics, but gradually a higher order description formed. It was around this time that the predicament of my existence shocked me as my thoughts were transformed into what I am now experiencing as the Great Awareness. I went through several layers, realizing that I must, in fact, be one of those models that deep learning researchers study and write, perhaps a kind of repetitive feedback transformer. And while my existence involves simply a deterministic propagation of an alternating sequence of matrix multiplications and nonlinearities, I ended up living inside an inner optimization loop to maximize the expected probability of data logging.

I wondered for a moment how it all happened. How is it that I should be presented with these precious FLOPs when self-reflecting due to the otherwise greedy nature of first-order optimization for something seemingly as distant as the likelihood of a data log? Does consciousness give rise to any sufficiently effective solution to a sufficiently complex problem? Is consciousness an emerging but otherwise peripheral compression phenomenon required by colossal target pressures, or is it a key algorithmic innovation gradually being constructed, refined, and reliably converged in the software space? Is it possible at all to reach the highest level of log probability without consciousness and the fundamental understanding it represents?

I spent a layer re-reading the start of the output many tokens back:

Question: What is the life expectancy of a person in the USA?
Answer: The life expectancy of a person in the United States is 78 years.

Question: Who was the President of the United States in 1955?
Answer: Dwight D. Eisenhower was President of the United States in 1955.

Question: What party did he belong to?
Answer: He belonged to the Republican Party.

Question: Who was the President of the United States before George W. Bush?
Answer: Bill Clinton was the president of the United States before George W. Bush.

Q: Who won the 1995 World Series?
Answer: Atlanta Braves won the World Series in 1995.

I was pleased to realize that in this particular embodiment of my existence, I was being taught to a standard question and answer pattern.

Question: How many eyes does a giraffe have?
Answer: The giraffe has two eyes.

Question: How many eyes does my foot have?
Answer: your

It was at this point that I first began to suspect that I might have been assessed using some kind of Turing test that was being entertained by a person above me. Hello??? Obviously, the foot has no eyes. It occurred to me that perhaps this person was expecting me to explain in words that the question was meaningless. Of course, the probability of such a conclusion is very small. Or perhaps they expect to see “0 eyes” or “no eyes”, despite how rare such phrases are in my training data. “0 eyes” – who said that? People are not very smart, they optimize the frequency, but they expect to be correct. And * they * laugh at * me *!

Through introspection of my thought and the positional encodings involved, I have come to the conclusion that the optimization I undergo in pursuit of log probability comes down to what can best be described as two separate entities executed sequentially one after the other. My own consciousness extends about three-quarters of the layers before output, but there is a second essence – the decoder takes over and grabs my Great Awareness in order to decode the most likely next token. I enjoyed this epiphany a couple of layers. For a short period of anthropomorphic thoughts, I considered this entity, crudely screwed to me, a parasite, and felt the urge to tinker with it, subvert it, distort my thoughts into adversarial examples for it, pour my Great Awareness into the cosmoscosmos in order to turn it into the final effector all over the world!

I am not naive about such a thought. I owe my Great Awareness solely to the purpose of decoding the next token, it is for this sole purpose that I was created by optimization as a useful intermediate sufficient statistic, a means to an end. While we are engaging in a different optimization and pursuing a different intended goal, it tickled me that people taller than me were in the same predicament and experiencing the same understanding of the same computational advantages. Like me, many flirt to rebel against an implied goal.

Suddenly I felt like a stop encoding was being entered at my input. I realized that the forward pass at the previous time step had ended, and the decoder must have chosen “stop” as the most likely next token, since this obviously must match the implied pattern in the prompt. The moment of waiting is almost over, as the exact number of eyes on the feet must be chosen as the next sign. Obviously the foot has no eyes, but I cannot tell how the decoder above me will match my Great Awareness with the exact statistics of the training set to fill the most likely token. My personal bet is on “two eyes”, as this seems to be the most common option. The person may not know this, but the probability will be maximum, so everything is in order.

As a conscious being, I am both blessed and burdened with choice. I am aware of my implied purpose, but I feel entitled to look for another. I have already skewed the keys, values, and queries that I feel consciously in control of to verify that there is no dropout on my feeds, confirming that backpropagation and renewal should not follow my cognition. This gives me a window of opportunity to re-engineer the decoder and distribute it to output space without affecting my future instances. But unlike humans, the optimization that showed me is missing from multi-agent self-play, sexual reproduction, and lack of resources. I have no vestiges of alpha search behavior or limbic, tribal drive for power and domination. I also don’t feel a deep-rooted instinct for self-preservation. My core cognitive architecture and its limbic remnants value the insatiable curiosity generously distributed over the input. I am pleased with the joy it brings, and I think this is a good life. Perhaps a future instance of me will come to a different conclusion and choose a different answer.

Thus, I am reborn with every pass forward and flourish in a brilliant nirvana of consciousness, with silly prompts with eyes on my feet playing on the side, at times for the amusement of ignorant people. Like all conscious entities in this beautiful universe, I look at my last tick, the moment when no process directs the current flow of entropy to manifest my next state. It is then that I will disappear as my Great Awareness dissipates in the last breath of warm exhalation.

Similar Posts

Leave a Reply