Click, → or scroll to step through. Esc to leave.
UCSC AI HACKATHON
System 1
Playground
Small local models vs Jev vs a text LLM, racing in Flappy Bird and Chrome Dino.
Jev on X
WHAT JEV DOES
Ask a question. Get a decision.
Request · POST api.typesafe.ai/v1/systemone
{
"model": "jev-latest",
"state": "bird_height=212 bird_speed=4 moving_down
next_pipe_distance=96 gap_bottom=180
gap_top=280 gap_center=230
bird_is 18 px BELOW the gap center",
"questions": {
"action": {
"type": "choice",
"instructions": "Flappy Bird. Should the bird flap now?",
"criteria": {
"flap": "Flap: jump up.",
"wait": "Wait: let the bird fall."
}
}
}
}→
Response · 0.25 s round trip
{
"model": "jev-1.13.0",
"answers": {
"action": {
"type": "choice",
"choice": "flap",
"confidence": 0.28,
"probabilities": { "flap": 0.64, "wait": 0.36 }
}
},
"usage": { "input_tokens": 370, "output_tokens": 33 }
}TWO WAYS TO THINK
Slow and careful, or fast and sure
System 2
LLMs think token by token
Thebirdis18pxbelowthegap,so…flap
Many words out, one at a time. Slow.
System 1
Jev just decides
flap0.64
wait0.36
One look, one answer. Fast.
Daniel Kahneman,
Thinking, Fast and Slow
Thinking, Fast and Slow
SIDE-BY-SIDE RACE · SIMULATED
Same question, one answer each
Text LLM
streams tokens, one at a time
0.00 s
thinking…▍
0 / 45 tokens
Jev
one look, one answer
0.00 s
thinking…
WHAT PEOPLE BUILD WITH JEV
Use cases
Tool calling
Robotics
Doom
JEV IS CLOSED SOURCE!!
console.typesafe.ai/usage · every decision is a paid API call
BUT THERE ARE OPEN ONES
Open-source System 1 models
github.com/NandhaKishorM/laya
Kev
github.com/jaredpalmer/kev
OpenJev
github.com/TheoLeeCJ/SemIf-OpenJev
NanoJev
github.com/TianyuCodings/NanoJev
Lev
Turn any open-weight model into a Jev-style System 1 model
HOW WE BUILT LEV
Keep the big model. Swap its last layer.
HOW WE BUILT LEV
Actions run once. Only the state runs live.
PLAYGROUND · JEV VS LEV VS A TEXT LLM
Live race: Flappy Bird
PLAYGROUND · JEV VS LEV VS A TEXT LLM