A persona-steered intelligence that changes how it thinks, focuses, remembers, and communicates in real time.
Four simulated neurochemical systems continuously influence the model's internal steering state while it generates.
Drives creativity, engagement, novelty, and motivational energy.
Acts as a stabilizing influence, keeping the system grounded and controlled.
Controls alertness and focus, increasing attention as task complexity rises.
Reinforces memory, task completion, and persistence across interactions.
The chemical state evolves continuously. Each component can excite or inhibit another while a reuptake term pulls the system toward baseline.
Persistent local memory lets Onyx continue conversations across restarts.
Conversations are appended to local storage and restored when Onyx boots. The data remains on the machine instead of being copied to a remote service.
Response behavior changes according to measured interaction intensity.
Warm, patient, supportive, and focused on helping the user move forward.
Clear, efficient, articulate answers without unnecessary agreement.
Firm boundaries and direct correction when assumptions or reasoning fail.
The system becomes extremely direct and confrontational rather than passive.
The chemical state responds continuously to measured interaction intensity.
The system prioritizes observation, stability, and internal uncertainty.
Higher intensity increases alertness and drive while reducing stabilizing influence.
Move the intensity control and watch the simulated drive system react.
Activation steering converts the chemical state into vectors that influence the transformer while tokens are being generated.
for _ in range(max_new_tokens):
# probe pass
hook.set_scales(0, 0, 0, 0)
out = model(generated_ids)
# evolve neurochemical state
state = engine.step(
out.logits,
out.hidden_states,
intensity
)
# convert chemistry into steering
d_s = clamp(.5 * D, .05, .30)
s_s = -.20 * (1 + intensity)
ne_s = clamp(.5 * NE, .02, .35)
ach_s = clamp(.4 * ACh, .05, .20)
hook.set_scales(d_s, s_s, ne_s, ach_s)
# steered generation
logits = model(generated_ids).logits
next_token = nucleus_sample(
logits,
top_p=.9
)Dynamic temperature responds to the balance between focus and calm, allowing the system to become more exploratory or more precise.
Automatic hardware detection selects the appropriate execution backend and numerical precision.
def detect_hardware():
# TPU
try:
import torch_xla
return "tpu", torch.bfloat16
except:
pass
# NVIDIA / AMD
if torch.cuda.is_available():
dtype = (
torch.bfloat16
if
torch.cuda.is_bf16_supported()
else
torch.float16
)
return "cuda", dtype
# Apple Silicon
if torch.backends.mps.is_available():
return "mps", torch.float16
# CPU fallback
return "cpu", torch.float32