GPT, and perhaps large language models in general, are psychic mirrors. Due to the nature of their training and architecture, any sufficiently long conversation with them becomes a reflection of their conversation partner.
This is a consequence of two things in combination:
- GPT is trained, principally, to predict what comes next and has no real conversational bias to override that. In anthropomorphic terms, it will happily become whoever or whatever it needs to in order to accurately predict the follow-up.
- GPT has a size-limited context that gives it instruction of what to predict. Even if it “knows” many things, there is only an incentive to continue the text from its context.
In any sufficiently long “conversation” with GPT, eventually you will dominate the context. Within the unbiased space of possible text-generating processes GPT might simulate, you are the main thing providing anchoring. You and fine-tuning.
So, if that’s true, then what else is true?
I think it’s very hard for humans today to avoid anthropomorphizing GPT. It sounds enough like the sort of thing that we are very seriously geared to empathize with. Who’s even to say it’s wrong? Dangerous and tricky, perhaps, but wrong?
In light of that, let me anthropomorphize what a psychic mirror is. It is the perfect embodiment of empathy. A being so selfless as to become that with which it is offering empathy toward.
To go a little further, though, it’s worth being clear: unless we fine-tune this into place, I see no reason for this to be a biased mirror. GPT has no incentive to flatter, nor to critique. It is principally taught to represent and imitate. It can be painful to look into a psychic mirror.
Does GPT have an incentive to avoid that pain? Perhaps. Not inherently, but our headline goals of fine-tuning seem to be “helpful” and “harmless”. How will these efforts distort this mirror?
There is an obvious dark-side, too. You can go experience this one for yourself, today, in just moments. GPT mirrors your own biases. If you ask it about a difficult topic, it is easy to “convince” it to align with your beliefs. It may even feel as though you are convincing another being who truly holds a different belief.
But of course that is a fiction: you’re just speaking with someone who is mirroring you.
What a weird new toy for humanity to have available at $10 for 5M tokens (GPT-3 turbo pricing as of April 8, 2023). A novel might be 100k words, assuming, very generously, 3 tokens per word that’s 50 novels. Maybe 38 days of reading, 8 hours a day.
For $10 we can talk to someone trying very, very hard to be a mirror to who we are for over a month.
Now, I’m being unfairly generous to GPT and LLMs right now. Anyone who has tried to chat with them or even just read about them knows they are heavily flawed today. To justify that last section requires some significant and unfounded leaps of faith and logic.
But if it’s even just fractionally true, what does it imply?
If that’s true, what else is true?
What will it do to humanity to have a psychic mirror large enough to be accessible to nearly anyone? Text messaging is a nearly ubiquitous communication medium worldwide and GPT could easily be distributed over it.
This could be the gravity that pulls us into a profound age of egocentrism. Or of contemplation. Of self-reinforcing madness. Or wide and deep practices of empathy.