A do-it-yourself Python script
Google Sheets template: https://docs.google.com/spreadsheets/d/1I951dcIb423RpkqZekGnomzsrznYi1_W5aCvWzsH7GA/Python script: https://github.com/liamtrotzuk/gpt3-dream-diary
As an experiment, I stored a easy dream diary in Google Sheets for about 3 months, recording about 35 desires on this interval. I then used a easy Python script to learn the final 10 desires and provides them to the GPT-3 language machine studying mannequin from OpenAI, asking the mannequin to summarize the desires and search for symbolic developments, as a method to simply hold operating neutral tabs on my basic unconscious psychological state.
I used to be compelled to do that fast challenge by a newfound need to mirror on the unusual eventualities my thoughts conjures when it sleeps, coupled with B. a disinterest in spending time myself making an attempt to summarize the desires (by both qualitative or quantitative means) because of the potential for private bias, wedded with C. an unwillingness to let one other particular person summarize my desires because of the non-public nature of what’s in them.
I’d been taking part in with GPT-3 on the time, and it struck me that an neutral machine that has already confirmed itself very expert at writing good summaries of extremely summary works of fiction (corresponding to ebook studies) could be a superb, discrete, and time-efficient software for summarizing the unusual non-sequiturs and odd symbolism of a human being’s desires.
A random pattern of three out of the 35 desires that I recorded:
A. “You had been watching planes taking off and touchdown from an airport within the Bronx, its runway fringed with timber. Previous to that, you’d been using a subway line that did a loop round a park within the Bronx that felt so much like Van Cortlandt. Whilst you watched the planes, nevertheless, there was an aerial risk. Witches, on broomsticks, who may assault us from above with some sort of flaming pumpkin grenades. It was a continuing, enduring risk, however not sufficient to make us cease plane-watching.”
B. “A ironmongery store opened beneath the condominium, and also you had been very excited.”
C. “You had been taking the subway. You had been above-ground someplace in Brooklyn, making an attempt to get additional into BK. You went the unsuitable method (coming into an elevator for the Manhattan route), however corrected. Then you definately entered an enormous, cavernous station, a chamber that sloped down, with a fancy sequence of openings within the high that allowed within the solar. A pool/fountain cascaded its method down the size of the station. You had been alone, although you knew that the underside led to a much bigger station with extra individuals. You needed so as to add it to your ‘High 3’ favourite subway station record. Subsequent factor you knew, you had been paddle-boarding at a seashore (felt prefer it was in Vancouver, although you weren’t certain), and also you had paddled out to Tom Hanks’s yacht, which was moored off the seashore. Tom was on deck, and he exchanged a couple of phrases with you. He stated that they had been removing quickly. The solar was starting to set. The seashore was crowded, and some different boats inhabited the waters. You fell off the paddleboard a couple of occasions — your expertise felt rusty. Lastly, you started to move in, with the conclusion that you need to fish extra.”
All 3 of those desires are among the many final 10 desires that I’ve remembered, so due to this fact they had been amongst these 10 that had been learn by the script. I fed these 3 desires plus 7 extra into the Da Vinci GPT-3 mannequin, probably the most highly effective of OpenAI’s fashions (for business functions, customers would possibly select a sooner mannequin on the expense of energy). I set the ‘temperature’ — a measure of the ‘danger’ the mannequin is keen take take — to 0.9, fairly excessive, which yielded probably the most attention-grabbing and non-literal outcomes. You may learn extra about how OpenAI categorizes ‘riskiness’ in AI responses right here.
After experimenting with varied prompts, you possibly can shortly achieve a basic sense of the place the machine demonstrates consistency and cohesiveness in its textual evaluation. I finally settled on asking 3 usually helpful questions, from most-specific to least-specific, that may hopefully yield a helpful high-level abstract of how my desires had been thematically trending:
#1openai.Completion.create(mannequin=“text-davinci-002",immediate=“Is the next record of desires predominately optimistic or predominately unfavourable or predominately impartial? What number of are optimistic? What number of are unfavourable? What number of are impartial?” + STR_dreams_last_10,max_tokens=2000,temperature=0.9)#2openai.Completion.create(
mannequin=“text-davinci-002",immediate=“What does the next record of desires say concerning the dreamer’s psychological state?” + STR_dreams_last_10,max_tokens=2000,temperature=0.9)#3openai.Completion.create(mannequin=“text-davinci-002",immediate=“What's the overarching theme of the next record of desires?” + STR_dreams_last_10,max_tokens=2000,temperature=0.9)
The first time I ran these prompts, the mannequin returned the next for every:
- “textual content”: “nnThere are 5 optimistic desires, 2 unfavourable desires, and three impartial desires.”
- “textual content”: “nnThe dreamer could also be experiencing nervousness about an upcoming occasion.”
- “textual content”: “nnThe overarching theme of the next desires is escape.”
All 3 responses are attention-grabbing, and at the least 1 + 2 are largely correct. The first immediate — ‘predominately optimistic or predominately unfavourable?’ — is simply tough sentiment evaluation, and the best for which to gauge accuracy.
After tallying up every dream with my very own judgments of relative positivity or negativity, I concurred with the GPT-3 rely of 5 optimistic (one in all which was the ironmongery store dream), 2 unfavourable (one in all which was the plane-watching, witch-attack dream), and three impartial (one in all which was the subway station/paddleboarding by Tom Hanks’ yacht dream).
That’s spectacular, regardless of the potential for bias in my very own sentiment evaluation of my very own desires (a really neutral evaluation would have requested others to rank my desires to be able to check towards GPT-3’s claims, however I didn’t hassle doing so), and we will usually rely this 1 out of the three prompts as helpful and correct thus far.
The script’s response to the 2nd immediate — nervousness about an upcoming occasion — is extra qualitative, however the machine’s reply feels correct for this immediate as effectively. I’m pretty easy-going and am typically not notably involved about future occasions, however there are a number of looming occasions in my life that I’m positively apprehensive about at current — far more so than typical — so GPT-3 is making a verifiable declare that usually aligns with particular realities in my life, totally different from the established order.
Once more, an experiment hewing to extra empirical strategies would most likely have requested a pattern set of out of doors observers of my life to rank the accuracy of the machine’s assertion given their information of my precise life and psychological state, however a easy self-administered sanity examine ought to suffice for the needs of this little experiment.
The machine’s response to the third immediate is extra akin to a newspaper-column horoscope designed to imply all issues to all individuals — GPT-3 may have possible despatched again virtually any nebulous idea, from the precise ‘escape’ to ‘braveness’ to ‘heat’ or any generalized phrases of that nature, and I’d possible have cherry-picked the required occasions and ideas in my life to help that response. ‘Escape’ does really feel like an vital theme in my life, however when does it not? Doesn’t everybody search an escape? Nonetheless, I take pleasure in a foolish verbal Rorschach check as a lot as the following individual, so I made a decision to maintain that third immediate in there as a enjoyable reminder {that a} machine skilled on human information will be as dumb and imprecise as people are.
And I’ll plan to maintain utilizing this easy-to-run script to get a nifty high-level abstract of my desires sooner or later. Whereas I’m nonetheless a silly novice within the craft of experimenting with AI language fashions, this train was a helpful excuse to arrange programmatic entry to GPT-3 and mess around with it, nevertheless trivially.