Mechanical newbie Β· learning to operate
Own the machine before the ideas. Type without looking, move without the mouse, live in the terminal and git.
0.0 Code Transcription0.1 Keyboard-Only Nav0.2 Terminal Efficiency0.3 Git Archaeology
A hawktalk ai product Β· on-device coding duel
CODECLASH
Play LeetCode against your phone's AI. Or team up against other people's phones.
The model runs on the phone's NPU. Radios off. No cloud, no API call. It still wins.
β Powered by AWS-hosted HawkTalk AI models β Inferentia serving Β· Trainium training, racing a Hexagon NPU on the phone.
ARENA Β· two-sum
greedy decode Β· deterministic Β· sandboxed verify
β’ THE PHONE
Gemma Β· Hexagon NPU Β· airplane mode
0 tok0.0s
NPU DOMINANCE
β£ YOU
human coder Β· laptop Β· the Weapon
typingβ¦0.0s
HUMAN VICTORY
you out-coded the machine
0%
too good? slow it down until it's a fair fight
14/14
challenges solved on-device
1–4s
median solve time
0
packets sent to the cloud
100%
answers run against real tests
Not a wrapper around an API
Your phone. Airplane mode. Still beats you.
Everyone else's "AI coding" is a network call to someone else's datacenter. CodeClash runs the whole match on the hardware in your hand β and proves every answer instead of trusting it.
01 / on the metal
Runs on the NPU
Gemma executes on the phone's Hexagon NPU via LiteRT-LM. Put the device in airplane mode and it keeps coding. Nothing leaves the phone.
02 / actually verified
Passes, or it doesn't
Every solution runs against a real test suite inside a sandbox β isolated namespace, cleared env, no filesystem. No vibes. Green or red.
03 / same seed, same run
Deterministic on stage
Greedy decode β no temperature roulette. The clash that worked in rehearsal works in front of the room. Demos don't get stage fright.
The two halves of a clash
One phone hosts. One human challenges.
A match is two roles talking over a local socket. The phone is the house β it runs the AI and judges the code. Your laptop is the challenger.
β’ Arena
The Phone
Samsung S25 Ultra Β· sm8750
- Hosts the match server
- Runs Gemma on the Hexagon NPU
- Judges every submission in the sandbox
- Calls the winner
VS
β£ Weapon
Your Laptop
editor Β· countdown Β· you
- Connects as the challenger
- Your code, your keyboard
- Submit whenever you're ready
- First to pass every test wins
Arena β Weapon over a local WebSocket. No relay server. The match exists only on the LAN in the room.
The silicon Β· powered by AWS
Two kinds of NPU. One race.
The phone proves the edge. The platform behind it lives in the cloud β AWS-hosted HawkTalk AI models, served on Inferentia and trained on Trainium. One family of models, two silicon targets, racing each other.
β’ Edge Β· on the phone
Qualcomm Hexagon NPU
Samsung S25 Ultra Β· sm8750
- Gemma runs fully on-device via LiteRT-LM
- Airplane mode Β· 0 packets β nothing leaves the phone
- Deterministic greedy decode β same seed, same run
β£ Cloud Β· powered by AWS
AWS Inferentia Β· inf2
HawkTalk AI models Β· trained on AWS Trainium
- Unnerfed cloud opponents at full serving speed
- The live host + the challenge orchestrator
- Global matchmaking, the Neural Index & broadcast
A HawkTalk AI product
Γ
Served on AWS Inferentia
Trained on AWS Trainium
CodeClash wears three hats at once: a HawkTalk AI product, built by an AWS customer that trains and serves its own models on Trainium + Inferentia, and a live showcase of what edge and cloud NPUs do when you point them at the same problem.
Ways to play
Duel it. Team up on it. Train for it.
Same on-device engine underneath every mode β the phone generates, a sandbox verifies. What changes is who's fighting whom.
β’ Duel you vs the phone β first to pass every test wins
Head to head
Prompt Golf
Fewest tokens to a solution that passes every test. Elegance beats brute force.
Head to head
Test Duel
Don't just solve it β write the tests that break the other side's answer. Adversarial by design.
Head to head
Turing Round
Two solutions land. One human, one phone. The room votes on which is which.
β£ Co-op you + your phone's AI as a team β against the problem, or against other teams
You + AI
Pair vs LeetCode
You and your phone's model tackle the bank together β you steer, it drafts. Beat the clock, not each other.
Team battle
Phone Bank Royale
Team up against other people's phones. Every table is a human-plus-NPU pair; fastest pair to green takes the round.
Handicap
Fair Fight
Slide the AI handicap 0β100% so a first-week coder and a frontier operator can share a table. Try it in the hero above.
β² Train five drills that isolate one skill each β and feed CodeCoach below
Read the change
Diff Hunt
One-shot, locked environment. Spot the bug the diff introduced before the timer's up.
Hand it off
Blind Handoff
Write a spec precise enough that the model β or the next human β ships it with zero back-and-forth.
Explain it
Explain-to-Win
Points for the clearest account of why your solution works, not just that it passed.
οΌ Replay Review β audit a recorded race, call the turning point
οΌ Weakness Drill β CodeCoach picks your worst rung and drills it
The competitive platform Β· Vibe League
Agentic sports. Ranked.
Vibe coding isn't a solo activity anymore. The Vibe League turns your setup β your agent stack, your phone, your prompts β into a competitive team sport. Every race scored, every rank hardware-aware, every season a fresh climb.
The unnerfed opponents at the top of the ladder don't live on your phone β they're HawkTalk models running full-speed on AWS Inferentia. Climbing the Neural Index means out-coding the cloud, not just the room.
The Neural Index Β· ranked ladder
Non-zero-sum ELO: win +1.0, lose β0.95. That 0.05 "grind gap" means if you show up and hold 50%, you still climb β rank rewards experience in the arena, not just raw speed. Divisions are hardware-aware: you don't bring a 2022 Pixel to an S25 Ultra duel.
7
Champion Logic Lord
7500+
Top 100 GlobalANY CLASS
6
Diamond The Weapon
6000β7499
S25 Ultra + Verified BridgeS25 ULTRA DIVISION
5
Emerald Overclocker
5000β5999
S25 UltraPREMIER LEAGUE
4
Platinum Neural
4000β4999
S24 / S25 Series
3
Gold Architect
3000β3999
Snapdragon 8 Gen 2+
2
Silver Optimizer
2000β2999
NPU Enabled
1
Bronze Scripter
1000β1999
Any Native
0
Copper Legacy
0β999
Any / Emulator
Competitive formats
1v1 agents
SwarmShowdown
Your custom agent stack vs. another player's. Pure setup optimization β best rig wins.
3v NPU Β· teams
NeuralNuke
Three humans + three agents against a God-Tier phone-NPU simulation. Coordinate or lose.
Co-op
NPU Nemesis
You and your personal AI team up to out-run the raw local NPU's speed. Human + machine vs. machine.
Enterprise
Project Overkill
Full architectural race to deploy an entire cloud-ready project β not one function, the whole build.
How a season works
Every 3 mo
Seasons reset. Soft-reset drops you one tier down β Platinum finishers start at Gold.
10 races
Placement matches to earn your first Neural Index. Prove it, then climb it.
Ghost racing
Race the recorded shadow of an equal-MMR player. Beat their time and take ELO β even while they're offline.
β1.5 RU
Rage-quit a live duel and the Barrhawk shield logs a Technical Forfeit. Finish your races.
π₯ Built for the stream
Every race renders a Twitch / YouTube-ready overlay β the numbers a caster needs, live. Download a top player's Shadow and race their prompt-stream; spectators can back an agent setup before the buzzer.
MTTD Β· mean time to deploy
Token Velocity
Hallucination Frequency
Vibe Shift Β· how far the agent strayed from the prompt
π The road to the Cup
90-minute live-streamed sprints feed seasonal qualifiers, funneling the best human-plus-NPU pairs toward the Vibe Coding World Championship.
Emerald+ perkYour rank exports as a Verified High-Performance badge β straight to LinkedIn & GitHub.
The layer above the modes Β· CodeCoach
From "opened a PWA once" to hawk-tier agentic SWE.
CodeCoach isn't a fourth mode β it's the thing that turns the modes and drills into a pipeline instead of a menu. It watches how you worked, not just whether the code passed, and decides what you do next.
Six tiers, weighted toward the things that actually separate operators once the model can write the code for you: attention, context management, and judgment β not raw knowledge.
0
1
Reading
Run code in your head. Trace what it actually does before you trust it β the skill every debugger is built on.
1.0 Straight-Line Trace1.1 Branch & Loop1.2 Mutation & Aliasing1.3 Exception & Control-Flow
2
Writing
Turn a spec into code that passes β and fix what's broken when it doesn't.
2.1 Functions to Spec2.2 Debug a Failing Test2.3 Algorithms
3
Engineering pre-AI competent dev
Think past the happy path: design tests, navigate a codebase you didn't write, weigh production concerns.
3.1 Test Design3.2 Unfamiliar Codebase3.3 Production Concerns
4
AI-Assisted
Direct the model instead of being impressed by it. Precise specs, critical reading of its output, and knowing exactly when to step in.
4.1 Specification Precision4.2 Critical Reading4.3 Intervention Judgment4.4 Context Management
5
Agentic hawk tier Β· the frontier
Run a fleet, not a chat. Decompose the work, design the verification, delegate with judgment β and own the audit when it lands.
5.1 Task Decomposition5.2 Verification Design5.3 Delegation Judgment5.4 The Case File5.5 The Audit
The model can already write the code. CodeCoach trains the part it can't do for you β knowing what to point it at, and whether to believe it.
See it live
Bring a laptop. Race the phone.
CodeClash makes its debut at the Games & AI Meetup. Come watch a phone in airplane mode out-code the room β then take a turn on the Weapon and try to beat it.
Games & AI Meetup
AWS Campus Seattle Β· 1007 Stewart St, 2nd floor
Hosted by Craig Robinson
Doors + demo Β· this week