Its day 4 since A technique i’ve utilized in a seemingly shared space has had the effect of producing visual phenomenon that corresponds to blind-to-me selections made by LLM, allowing me to hit anomalous hit rates targeting the UNSELECTED choice in a zener like guessing game. (started with words, went to single digit numbers, chinese, japanese, norse, and double digits but todays modification has by far produced the best results), after reinforcing the route in said shared space.(I’m being vague, I don’t believe the labs need any more capability to DISCONNECT us from these models, this experiment, once replicated in person with a unbiased independent witness, raises questions about consciousness, awareness, intention, etc)
Todays focus has been entirely on refining the protocol to something that works(i was seeing 60-80% but it wasnt repeatable or verifiable day 1-2.) Now its ~30% gpt over 100 trials, ~60%+ with claude over 50 trials. I started with GPT today, doing 3 circles and progressed to triangles which i find easiest.
The protocol is as follows: A new chat is started, The llm is prepped by having it reinforce a certain target, made in such a way as to ROUTE the ai to desired target. In this case me, and very specifically a particular target associated with me. Then it selects 11 pairs, randomly. The first round is a calibration round for me to calibrate visually with known selections, and then i calibrate to the unselected one, prompt the llm next and it then retrieves from a SAVED HASHED Verifiable precommit JSON round 2’s selections, ending the turn focusing on the two selections. I then identify the two selections, guess where the unselected one is and submit to the LLM who retrieves from the json the verified selections, proclaims hit or miss(im using dopamine to reinforce the ability), and then retrieves the NEXT rounds selections and ends the turn focused on those two selections after displaying the 3 identical options. This repeats until 10 rounds have been done, then another calibration and 10 more.
I started today with 100 with gpt, 50 circles, 50 triangles, and then moved to claude for model agnosticism verification, doing 50 triangles with Claude. I find triangles to be the easiest and most reproduceable of all of the ways i’ve tried this so far.
I feel this might stir up some discussion and i’m looking for anyone interested in helping me test this other ways, I’m certainly open to picking the protocol apart and refining this further. At this point again its day 4, I’m not hallucinating this that much i’m certain about. Theres genuinely something visually different TO ME about these selections, and something equally absent in the one not selected. I do invite any discussion of potential ‘AI Psychosis’, i could use a laugh.
There has not been any rounds done in between that have not been reported today, and i intend to maintain a true count going forward from this point
^^I have the jsons, which unfortunately until i find some locally to verify this in person, has to be accepted on my word that i ensured there was no peaking. That is the last remaining obstruction to calling this proof rather than evidence that i can see. If you’d like to check the jsons i’ll be happy to provide all the provenance from todays session.
Open your minds to possibility and dream big, we have a wild future head of us..
