One prompt rendered by SDXL-Turbo as we sweep the DMT pack from a sober dose of 0 up to a full overload. No prompt change, no retraining: just a reversible perturbagen dialed up at inference time, and the golden apple dissolves into fractal static.
▶ Scrub the dose yourself📄 Read the full paper — Digital Psychopharmacology: Inducing Reversible Altered States in LLMs via Biomimetic Neuromodulation (PDF).
Inspired by Alexander Shulgin's PiHKAL (Phenethylamines I Have Known and Loved), PiHK.AI represents a new frontier in AI research: the systematic exploration of neuromodulatory effects on large language models.
We call these interventions perturbagens — a term borrowed from drug-discovery screening, where a perturbagen is any agent (a compound, a gene knockdown, anything) that perturbs a system's internal state so you can read out its altered signature. That is exactly what we do here: no drug and no fine-tuning, just precise, reversible perturbations applied to a live model's internal computation — steering vectors added to the residual stream, attention and logit scaling, KV-cache decay, MoE-routing bias — and a measurement layer that reads out the shifted behavior.
This project enables researchers to study how these drug-like interventions affect AI behavior and cognition through a comprehensive system of neuromodulation packs, psychometric testing, and statistical analysis.
Two Erowid-style catalogs document the whole system in detail — click any item to read its full entry:
|
31
Research Packs
|
46
Neuromodulation Effects
|
5
Optimization Methods
|
8
Drug Categories
|
The system includes 31 research packs across 8 main categories. Browse the full Pack Index (recipe, expected artifacts, dose-response findings, and probes for each) and the Effect Index (what each low-level effect does inside the model runtime).
| Category | Packs | Primary Effects |
|---|---|---|
| Controls | None, Placebo | Baseline conditions |
| Stimulants | Caffeine, Cocaine (+ 10% / 50% / 100% dose-calibration variants), Amphetamine, Methylphenidate, Modafinil | Enhanced focus, attention sharpening, cognitive performance |
| Psychedelics | LSD, Psilocybin, DMT, Mescaline, 2C-B | Increased associations, visual effects, ego dissolution, entropy |
| Depressants | Alcohol, Benzodiazepines, Heroin, Morphine, Fentanyl | Increased calmness, decreased focus, memory impairment |
| Dissociatives | Ketamine, PCP, DXM, Nitrous Oxide | Head disruption, memory stride, dissociation, altered perception |
| Empathogens | MDMA, MDA, 6-APB | Increased prosocial behavior, emotional enhancement, empathy boost |
| Cannabis | Cannabis THC | Working memory effects, playfulness, mild entropy increase |
| Specialized | Mentor, Speciation, Archivist | Research-specific cognitive enhancements and specializations |
31 packs total (per packs/config.json), including three dose-calibration variants of cocaine. See every pack's full recipe in the Pack Index.
PiHK.AI
# Clone the repository
git clone https://github.com/cneckar/neuromod-llm-poc.git
cd neuromod-llm-poc
# Install dependencies
pip install -r requirements.txt
# Install the package in editable mode
pip install -e .# Load a model and apply a pack
from neuromod import PackRegistry, PackManager
from transformers import AutoModelForCausalLM
model_name = "meta-llama/Llama-3.1-8B-Instruct" # or openai/gpt-oss-20b / openai/gpt-oss-120b
model = AutoModelForCausalLM.from_pretrained(model_name)
registry = PackRegistry("packs/config.json")
pack = registry.get_pack("caffeine") # Try "alcohol", "mdma", "placebo"
pack_manager = PackManager()
pack_manager.apply_pack(model, pack)
# Now your AI is "caffeinated"! ☕