Free live lesson · 30 minutes

Replace Jev with Your Own LLM, Then Beat It at Its Own Game

Superlinear Academy
Yan Wang
Your instructor

Yan Wang, PhD

Co-founder of Superlinear Academy · AI at Samsara, Microsoft, Pinterest
  • ~40 papers at venues including CVPR, NeurIPS and KDD; 3,200+ citations
  • Laid the technical foundation for Samsara’s AI platform
  • Key technical contributor to Microsoft Bing Image Search and Pinterest Search
superlinear.academy
13%

Price: $0.042 per million input tokens (Vercel AI Gateway)
Guillermo Rauch (Vercel), launch day: share of active Vercel AI Gateway teams that turned Jev on
Rec. 1

Today

  1. 01
  2. 02
  3. 03

GameIn the browser ·fixed 30 Hz clock Adaptation layerTurns the state intotext; written withClaude Code ~380 tokensPer direction: ghost?how far? pellets?power pellet? ModelPicks one directionat each junction

178

Jev 1.13.0 hosted by TypeSafe · 10 games, seeds 100–109 · real-time clock · p50 latency

050100150200250 tev1:4b · decision model148Jev 1.13.0 · decision model178qwen3.5:4b · chat model197gemma4:e4b · chat model230phi4-mini · chat model232 pellets 148–232

Average pellets, 10 games, seeds 100–109, real-time clock · every model reads the same ~380-token text · a level has 244 pellets

phi4-mini
232pellets
p50 187 ms per decision
Jev 1.13.0
178pellets
p50 115 ms per decision

Average pellets, 10 games, seeds 100–109, real-time clock · paired by seed against Jev

On Pac-Man

qwen3.5:4b, gemma4:e4b and phi4-mini, real-time games
Beyond Pac-Man
102 official tasksJev88.3%SemIf84.5%220 hard questionsJev74.1%SemIf59.5%
External results: SemIf comparison; 220 hard JevBench questions

100200300 333 sRecompute everything72 sReuse the prefix38.8 s+ Batch the suffixes 0 s
Output side

Four open reproductions within 72 hours of launch
SemIf, open-source reimplementation · 777 decisions, total seconds · output side: reading logits vs generating JSON

Real time pellets, 10 games 200 69 148 178 0 tev1:0.8b tev1:4b Jev Lockstep: the game waits separate clock, not comparable with real time 200400600 0 27B · no thinking 27B · thinking 301 531
The price

Average pellets · real time: 10 games, seeds 100–109 · Qwen3.8-27B: lockstep, seeds 100–102, about 2.4 s per decision with thinking · the two clocks are not directly comparable

425pellets
55 ms per decision · real time
Jev, same clock: 178

Distilled 0.8B, trained on labels from Qwen3.8-27B · average of 3 real-time runs (lockstep 417) · reads only the current state

0.000.250.500.751.00 Jev · published by JevBench0.89tev1:0.8b · same size0.66Distilled 0.8B · ours0.53 random 0.32

194 public JevBench judgment questions (intent, routing, policy, severity, extraction; no Pac-Man) · Jev’s 0.89 is JevBench’s published result, not measured by us · the student started from the Qwen3.5-0.8B base

Hosted · JevYour own model
Setup
Latency
Intelligence
Generality
Cost
Jev: hosted by TypeSafe, ~115 ms p50 over the public internet · our 0.8B: 55 ms per decision, real time · cost: Vercel AI Gateway price

  1. 01
  2. 02
  3. 03

C

Next cohort
Price
Access
Taught by
QR code to maven.com/superlinear/aibuilders?promoCode=beyond-jev
Code· 

Questions

QR code to maven.com/superlinear/aibuilders?promoCode=beyond-jev
superlinear.academy

What our engine gave usWhat your product already has

Replace Jev with your own LLM · Free live lesson
Yan Wang, PhD · Superlinear Academy
Make what lasts.