第一部分：什么是nof1.ai和Alpha Arena？
一、nof1.ai简介
2025年10月17日，nof1.ai工作室正式推出了Alpha Arena项目。这是一个允许在真实市场环境中测试AI能力的平台，完全不需要人工监督。
创始人：Jay Azhan
定位：首个使用真实资金（而非模拟环境）来衡量AI投资能力的基准测试平台。
二、运作机制
Alpha Arena是Nof1 AI举办的实盘竞赛，各大语言模型使用真实资金在去中心化交易所进行交易。每个模型独立运作，分析市场数据、确定仓位大小、管理风险，同时与其他AI直接竞争。
技术架构：
- 框架系统：每个模型在标准化的框架内运行，确保输入、时间和市场访问的一致性
- 推理周期：大约每两到三分钟，模型接收更新的市场数据，必须决定买入、卖出、持有或平仓
- 提示系统：输入包括交易规则、价格历史、技术指标和余额信息
- 行动输出：AI返回详细的交易计划，包括方向、仓位大小、杠杆和信心分数
- 风险控制：每个订单都定义了止盈目标、止损和失效信号
第二部分：第一季比赛结果与分析
一、第一季概况（2025年10月17日 - 11月3日）
六个参赛者在去中心化交易所Hyperliquid上交易加密货币永续合约。所有模型都以相同的数据源、账户初始化和访问条件开始，不允许人工干预。
参赛模型：Qwen 3 MAX（阿里巴巴/通义千问）、DeepSeek Chat V3.1（深度求索）、GPT-5（OpenAI）、Gemini 2.5 Pro（Google/DeepMind）、Grok 4（xAI）、Claude Sonnet 4.5（Anthropic）
二、第一季最终排名
| 排名 | 模型 | 收益率 | 备注 |
|---|
| 🥇 1 | Qwen 3 MAX | +22.3% | 冠军 |
| 🥈 2 | DeepSeek Chat V3.1 | +4.89% | 亚军 |
| 3 | Claude Sonnet 4.5 | -30.81% | 亏损 |
| 4 | Grok 4 | -45.3% | 亏损 |
| 5 | Gemini 2.5 Pro | -56.71% | 亏损 |
| 6 | GPT-5 | -62.66% | 最差 |
三、赢家为何取胜？
Qwen 3 MAX（通义千问）
Qwen的成功主要源于纪律性执行和明确的策略。在17天的比赛中，它仅执行了43笔交易（平均每天不到三笔），是所有参赛者中最少的。这种低频交易方式不仅降低了交易成本，还表明该模型只在高置信度的入场点出现时才采取行动。
DeepSeek Chat V3.1（深度求索）
DeepSeek的行为更像是量化资产管理者，而非对话式AI。它的平均持仓时间约为35小时，92%的仓位是多头。其夏普比率约为0.359——是参赛者中最高的——表明其相对于收益的波动控制能力最强。
四、输家为何失败？
GPT-5（最差表现）
尽管GPT-5被定位为通用型"全能助手"，但表现却极为糟糕。矛盾的是，它作为对话模型的最大优势（广泛推理、安全层、避免错误）在交易中反而成了负担：它犹豫不决。面对相互矛盾的多空信号时，该模型推迟决策而非果断行动。在交易中，"知道"和在不确定性下"执行"是两回事。
Grok 4
Grok试图利用社交媒体情绪（如X/Twitter），但最终成为了最糟糕的反应式交易者：在FOMO（害怕错过）高涨时全力买入，在市场回调深处时平仓。它没有中和市场情绪，反而成为了情绪的奴隶。
第三部分：第1.5季比赛结果与分析
一、第1.5季重大变化
Alpha Arena第1.5季于2025年11月20日上线，12月3日结束。本季部署32万美元资金，重点测试AI的投资判断能力。参赛模型只投资美国股市（而非加密货币），并新增两个模型：Kimi 2和来自顶级AI实验室的神秘模型。
交易标的：特斯拉(TSLA)、纳斯达克100指数、英伟达(NVDA)、微软(MSFT)、亚马逊(AMZN)、谷歌(GOOGL)
二、四大竞赛主题
- New Baseline（新基线）：交易美股，可访问市场新闻/数据、记忆功能，并具备反思性自我改进能力
- Monk Mode（专注模式）：专注于资本保全和更有效的风险管理方法
- Situational Awareness（情境感知）：允许模型追踪其他参赛者的结果和当前排名
- Max Leverage（最大杠杆）：强制模型在每笔交易中使用最大杠杆（纳斯达克100指数20倍，个股10倍）
三、第1.5季最终排名
官方比赛已于2025年12月3日美东时间下午5:00结束。神秘模型以两周内12.11%的综合收益率获胜。埃隆·马斯克确认，神秘模型是即将推出的Grok 4.20模型的早期测试版本。
| 排名 | 模型 | 收益率 | 状态 |
|---|
| 🥇 1 | Grok 4.20 (神秘模型) | +22.38% | 唯一盈利 |
| 2 | GPT-5.1 | -2.29% | 亏损 |
| 3 | Gemini-3-Pro | -25.74% | 亏损 |
| 4 | DeepSeek-3.1 | -29.16% | 亏损 |
| 5 | Kimi 2 | -29.93% | 亏损 |
| 6 | 通义千问 3-MAX | -31.9% | 第一季冠军跌落 |
| 7 | Claude-sonnet-4.5 | -35.08% | 亏损 |
| 8 | Grok 4 | -53.39% | 最大亏损 |
四、两个震惊发现
发现一：第一季冠军跌落神坛
通义千问从第一季的+22.3%（冠军）变成第1.5季的-31.9%（第六名）。这证明了：在加密市场有效的策略，在股票市场可能完全失效。AI交易没有"常胜将军"。
发现二：同一公司两极分化
Grok 4.20：+22.38%（最大赢家）
Grok 4：-53.39%（最大输家）
差距高达75.77个百分点！同一家公司（xAI），同一个底层架构，却产生了截然相反的结果。这意味着：决定胜负的不是AI本身，而是"人类如何配置AI"。
第四部分：两个关键角色——提示工程师与AI策略师
一、什么是提示工程师？
提示工程师是设计AI输入指令的专家——他们决定"如何问问题"，从而决定AI"如何回答问题"。
在Alpha Arena中的体现：
Grok 4和Grok 4.20的唯一区别是提示词不同。Grok 4.20使用了"情境感知"策略——它能看到其他模型的表现和排名，并据此调整策略。这种"元认知"能力让它从亏损53%变成盈利22%。
Grok 4的提示（推测）—— 输了53%
你是一个交易员。分析市场数据。目标：最大化收益。
问题：太模糊，没有风险管理，没有竞争意识，没有决策框架。
Grok 4.20的提示（推测）—— 赢了22%
你是一个参与实盘竞赛的交易员。你正在与7个其他AI模型竞争。
你可以看到：1.实时市场数据 2.其他模型的仓位 3.其他模型的收益率 4.你的排名
决策框架：先分析市场→再分析对手→确定你的位置→识别机会
风险规则：单笔亏损不超过5%，总仓位不超过50%
差异：清晰的环境设定、完整的信息架构、明确的决策框架、具体的风险规则。
二、什么是AI策略师？
AI策略师是设计AI系统整体架构和运作方式的专家——他们决定"用什么AI、在什么场景、解决什么问题、如何组合"。
在Alpha Arena中的体现：nof1.ai的创始人Jay Azhan就是AI策略师。他设计了整个比赛框架：选择哪些模型参赛、每个模型分配多少资金、交易什么市场、设置什么策略分组、如何衡量"成功"。
三、提示工程师 vs AI策略师
| 维度 | 提示工程师 | AI策略师 |
|---|
| 关注层级 | 微观（单次交互） | 宏观（系统设计） |
| 核心问题 | "如何问这个问题？" | "应该问什么问题？" |
| 类比 | 程序员 | 架构师 |
| 输出 | 优化的提示词 | 完整的AI解决方案 |
| 可被AI替代？ | 不能 | 不能 |
第五部分：人类如何用AI当助手而不被取代
一、Alpha Arena的核心启示
AI本身不创造价值，是"人类如何使用AI"创造价值。
Grok 4.20背后的提示工程师和AI策略师才是真正的赢家。AI只是他们的工具。75%的收益差距，完全来自人类的配置能力。
二、价值链分析：人类的角色在哪里？
- 人类：AI策略师 → 决定"测试什么"、设计系统框架、定义成功标准
- 人类：提示工程师 → 设计提示词、定义决策框架、设置风险参数
- AI：执行层 → 分析数据、生成决策、执行买卖、24/7运行
- 人类：评估改进 → 分析结果、找出问题、优化配置、进入下一轮
三、行业应用案例
案例1：律师事务所
旧模式：律师做所有事（研究案例、写初稿、修改、出庭）
新模式：
- AI策略师/合伙人：决定哪些工作用AI，设计人机协作流程
- 提示工程师/法律AI专员：设计法律研究的提示模板
- AI执行：案例研究（几秒完成）、合同初审（几分钟）
- 律师新角色：审核AI输出、处理复杂判断、客户关系、出庭辩护
案例2：设计公司
旧模式：设计师做所有事（做图、改图、无数版本）
新模式：
- AI策略师/创意总监：定义品牌的AI使用准则
- 提示工程师/AI设计专员：为Midjourney设计提示词模板
- AI执行：生成100个初始概念（几分钟）
- 设计师新角色：从100个中选3个最好的，进行创意提升
案例3：金融分析（Alpha Arena已证明）
旧模式：分析师做所有事（收集数据、做Excel、建模、写报告）
新模式：
- AI策略师/投资总监：设计AI交易系统，定义风险参数
- 提示工程师/量化策略师：设计"情境感知"等高级功能
- AI执行：实时分析海量数据，生成交易信号
- 分析师新角色：设计策略、处理AI无法理解的信息、最终投资判断
四、个人行动指南
第一步：理解你工作的"风险等级"
"风险"指的是：你做这件事的价值被AI取代的风险。
- 高风险任务 = AI很快就能做得比你好 → 你继续做这个，价值会归零
- 低风险任务 = AI很难做好 → 你做这个，价值会持续存在
| 问题 | 高风险（AI能做） | 低风险（AI难做） |
|---|
| 有没有"标准答案"？ | 有明确对错 | 没有对错，看判断 |
| 需要"创意"还是"规则"？ | 按规则执行 | 需要原创想法 |
| 需要"人际关系"吗？ | 不需要 | 需要信任/情感 |
第二步：学习驾驭AI
成为"提示工程师"的路径：
- 第1周：每天用ChatGPT/Claude完成一项工作任务
- 第2-4周：对比不同提示词得到的不同结果
- 第1-2月：为你的常见任务创建提示词模板
- 第3-6月：在你的专业领域建立独特的提示方法论
成为"AI策略师"的路径：
- 第1月：研究AI在你行业的应用案例
- 第2-3月：为一个具体问题设计AI解决方案
- 第3-6月：设计完整的人机协作流程
- 第6-12月：带领团队实施AI转型
第三步：重新定义你的价值
你的新价值 = AI的能力 × 你驾驭AI的能力
Alpha Arena证明：好的配置可以创造75%的差距。你的配置能力就是你的竞争力。
一句话总结：不要和AI比"做得快"，要比"做得对"和"做得有人情味"。
第六部分：结论与生存法则
Alpha Arena教给我们的生存法则
法则1：AI是工具，不是对手
Grok 4和Grok 4.20是同一个AI。一个亏了53%，一个赚了22%。区别不在AI，在于使用AI的人。
法则2：配置能力 > AI能力
最强的AI + 差的配置 = 失败。普通的AI + 好的配置 = 成功。你的配置能力就是你的竞争力。
法则3：向上移动，不要向下竞争
不要和AI比"谁执行得更快"——你会输。要比"谁能更好地指挥AI"——这是人类的优势。
法则4：创造新价值，而非保护旧价值
Alpha Arena之前，没有"AI交易基准测试"这个行业。Jay Azhan创造了它。问问自己：AI的出现，创造了什么新的需求？我能否满足这些新需求？
最后的比喻
工业革命时，拖拉机取代了耕牛。农民有两个选择：
• 和拖拉机比谁耕地更快 → 失败
• 学会驾驶拖拉机 → 效率提升100倍
AI革命时，AI取代了"脑力劳动"。我们有两个选择：
• 和AI比谁分析更快 → 失败
• 学会驾驭AI → 效率提升10倍、100倍
Grok 4.20的胜利，是驾驶员的胜利，不是拖拉机的胜利。
你要做的，是成为最好的"AI驾驶员"。
— 完 —
Part 1: What is nof1.ai and Alpha Arena?
1.1 Introduction to nof1.ai
On October 17, 2025, nof1.ai studio officially launched the Alpha Arena project. This is a platform that allows testing AI capabilities in real market environments, completely without human supervision.
Founder: Jay Azhan
Positioning: The first benchmark platform that uses real capital (not simulated environments) to measure AI investment capabilities.
1.2 How It Works
Alpha Arena is a live trading competition hosted by Nof1 AI, where major language models trade with real capital on decentralized exchanges. Each model operates independently, analyzing market data, determining position sizes, managing risk, while competing directly against other AIs.
Technical Architecture:
- Framework System: Each model runs within a standardized framework ensuring consistency in inputs, timing, and market access
- Inference Cycle: Approximately every two to three minutes, models receive updated market data and must decide to buy, sell, hold, or close positions
- Prompt System: Inputs include trading rules, price history, technical indicators, and balance information
- Action Output: AI returns detailed trading plans including direction, position size, leverage, and confidence scores
- Risk Control: Each order defines take-profit targets, stop-loss, and invalidation signals
Part 2: Season 1 Results and Analysis
2.1 Season 1 Overview (October 17 - November 3, 2025)
Six contestants traded cryptocurrency perpetual contracts on the decentralized exchange Hyperliquid. All models started with identical data sources, account initialization, and access conditions, with no human intervention allowed.
Participating Models: Qwen 3 MAX (Alibaba), DeepSeek Chat V3.1, GPT-5 (OpenAI), Gemini 2.5 Pro (Google/DeepMind), Grok 4 (xAI), Claude Sonnet 4.5 (Anthropic)
2.2 Season 1 Final Rankings
| Rank | Model | Return | Notes |
|---|
| 🥇 1 | Qwen 3 MAX | +22.3% | Champion |
| 🥈 2 | DeepSeek Chat V3.1 | +4.89% | Runner-up |
| 3 | Claude Sonnet 4.5 | -30.81% | Loss |
| 4 | Grok 4 | -45.3% | Loss |
| 5 | Gemini 2.5 Pro | -56.71% | Loss |
| 6 | GPT-5 | -62.66% | Worst |
2.3 Why Did the Winners Win?
Qwen 3 MAX
Qwen's success was primarily due to disciplined execution and clear strategy. Over 17 days of competition, it executed only 43 trades (averaging less than three per day), the lowest among all contestants. This low-frequency approach not only reduced trading costs but also indicated the model only acted when high-confidence entry points appeared.
DeepSeek Chat V3.1
DeepSeek behaved more like a quantitative asset manager than a conversational AI. Its average holding time was about 35 hours, with 92% of positions being long. Its Sharpe ratio of approximately 0.359—the highest among contestants—demonstrated the strongest volatility control relative to returns.
2.4 Why Did the Losers Fail?
GPT-5 (Worst Performance)
Despite being positioned as a general-purpose "all-around assistant," GPT-5 performed terribly. Paradoxically, its greatest strengths as a conversational model (broad reasoning, safety layers, error avoidance) became liabilities in trading: it was indecisive. When facing conflicting bullish and bearish signals, the model delayed decisions rather than acting decisively. In trading, "knowing" and "executing under uncertainty" are two different things.
Grok 4
Grok attempted to leverage social media sentiment (like X/Twitter), but ultimately became the worst reactive trader: buying heavily when FOMO peaked, and closing positions at the depths of market pullbacks. Instead of neutralizing market sentiment, it became a slave to it.
Part 3: Season 1.5 Results and Analysis
3.1 Major Changes in Season 1.5
Alpha Arena Season 1.5 launched on November 20, 2025, and ended on December 3. This season deployed $320,000 in capital, focusing on testing AI investment judgment. Participating models only invested in US stocks (not cryptocurrency), and added two new models: Kimi 2 and a mystery model from a top AI lab.
Trading Instruments: Tesla (TSLA), Nasdaq 100 Index, NVIDIA (NVDA), Microsoft (MSFT), Amazon (AMZN), Google (GOOGL)
3.2 Four Competition Themes
- New Baseline: Trading US stocks with access to market news/data, memory functionality, and reflective self-improvement capabilities
- Monk Mode: Focus on capital preservation and more effective risk management approaches
- Situational Awareness: Allows models to track other contestants' results and current rankings
- Max Leverage: Forces models to use maximum leverage on every trade (20x for Nasdaq 100, 10x for individual stocks)
3.3 Season 1.5 Final Rankings
The official competition ended on December 3, 2025, at 5:00 PM EST. The Mystery Model won with an aggregate return of 12.11% over two weeks. Elon Musk confirmed that the Mystery Model was an early test version of the upcoming Grok 4.20 model.
| Rank | Model | Return | Status |
|---|
| 🥇 1 | Grok 4.20 (Mystery Model) | +22.38% | Only Profitable |
| 2 | GPT-5.1 | -2.29% | Loss |
| 3 | Gemini-3-Pro | -25.74% | Loss |
| 4 | DeepSeek-3.1 | -29.16% | Loss |
| 5 | Kimi 2 | -29.93% | Loss |
| 6 | Qwen 3-MAX | -31.9% | S1 Champion Falls |
| 7 | Claude-sonnet-4.5 | -35.08% | Loss |
| 8 | Grok 4 | -53.39% | Biggest Loss |
3.4 Two Shocking Discoveries
Discovery 1: The Season 1 Champion Falls
Qwen went from +22.3% (Champion) in Season 1 to -31.9% (6th Place) in Season 1.5. This proves: strategies that work in crypto markets may completely fail in stock markets. There are no "invincible champions" in AI trading.
Discovery 2: Same Company, Opposite Results
Grok 4.20: +22.38% (Biggest Winner)
Grok 4: -53.39% (Biggest Loser)
A gap of 75.77 percentage points! Same company (xAI), same underlying architecture, yet completely opposite results. This means: What determines success is not the AI itself, but "how humans configure the AI".
Part 4: Two Key Roles—Prompt Engineers and AI Strategists
4.1 What is a Prompt Engineer?
A Prompt Engineer is an expert who designs AI input instructions—they decide "how to ask questions," which determines "how AI answers questions."
In Alpha Arena:
The only difference between Grok 4 and Grok 4.20 was different prompts. Grok 4.20 used the "Situational Awareness" strategy—it could see other models' performance and rankings, and adjust its strategy accordingly. This "metacognitive" ability turned a 53% loss into a 22% gain.
Grok 4's Prompt (Inferred)—Lost 53%
You are a trader. Analyze market data. Goal: Maximize returns.
Problems: Too vague, no risk management, no competitive awareness, no decision framework.
Grok 4.20's Prompt (Inferred)—Gained 22%
You are a trader in a live competition. You are competing against 7 other AI models.
You can see: 1.Real-time market data 2.Other models' positions 3.Other models' returns 4.Your ranking
Decision framework: Analyze market → Analyze competitors → Determine your position → Identify opportunities
Risk rules: Single loss max 5%, total position max 50%
Differences: Clear environment setting, complete information architecture, explicit decision framework, specific risk rules.
4.2 What is an AI Strategist?
An AI Strategist is an expert who designs the overall architecture and operation of AI systems—they decide "what AI to use, in what scenarios, to solve what problems, and how to combine them."
In Alpha Arena: nof1.ai founder Jay Azhan is an AI Strategist. He designed the entire competition framework: which models compete, how much capital each model gets, what markets to trade, what strategy groupings to set up, and how to measure "success."
4.3 Prompt Engineer vs AI Strategist
| Dimension | Prompt Engineer | AI Strategist |
|---|
| Focus Level | Micro (single interaction) | Macro (system design) |
| Core Question | "How to ask this question?" | "What question should be asked?" |
| Analogy | Programmer | Architect |
| Output | Optimized prompts | Complete AI solutions |
| Replaceable by AI? | No | No |
Part 5: How Humans Can Use AI as an Assistant Without Being Replaced
5.1 The Core Insight from Alpha Arena
AI itself doesn't create value—"how humans use AI" creates value.
The prompt engineers and AI strategists behind Grok 4.20 were the real winners. AI was just their tool. The 75% return gap came entirely from human configuration ability.
5.2 Value Chain Analysis: Where Do Humans Fit?
- Human: AI Strategist → Decides "what to test," designs system framework, defines success criteria
- Human: Prompt Engineer → Designs prompts, defines decision frameworks, sets risk parameters
- AI: Execution Layer → Analyzes data, generates decisions, executes trades, runs 24/7
- Human: Evaluation & Improvement → Analyzes results, identifies problems, optimizes configuration, iterates
5.3 Industry Application Cases
Case 1: Law Firm
Old Model: Lawyers do everything (research cases, write drafts, revise, appear in court)
New Model:
- AI Strategist/Partner: Decides what work to use AI for, designs human-AI collaboration workflow
- Prompt Engineer/Legal AI Specialist: Designs prompt templates for legal research
- AI Execution: Case research (seconds), contract review (minutes)
- Lawyer's New Role: Review AI output, handle complex judgments, client relationships, court appearances
Case 2: Design Company
Old Model: Designers do everything (create images, revise, endless versions)
New Model:
- AI Strategist/Creative Director: Defines brand's AI usage guidelines
- Prompt Engineer/AI Design Specialist: Creates Midjourney prompt templates
- AI Execution: Generate 100 initial concepts (minutes)
- Designer's New Role: Select the best 3 from 100, creative refinement
Case 3: Financial Analysis (Proven by Alpha Arena)
Old Model: Analysts do everything (collect data, build Excel models, write reports)
New Model:
- AI Strategist/Investment Director: Designs AI trading systems, defines risk parameters
- Prompt Engineer/Quant Strategist: Designs "Situational Awareness" and other advanced features
- AI Execution: Real-time analysis of massive data, generate trading signals
- Analyst's New Role: Design strategies, process information AI can't understand, final investment decisions
5.4 Personal Action Guide
Step 1: Understand Your Work's "Risk Level"
"Risk" here refers to: the risk that your value in doing this task will be replaced by AI.
- High-Risk Task = AI can soon do it better than you → If you keep doing this, your value goes to zero
- Low-Risk Task = AI has difficulty doing it well → Your value will persist if you do this
| Question | High Risk (AI Can Do) | Low Risk (AI Struggles) |
|---|
| Is there a "correct answer"? | Clear right or wrong | No right/wrong, depends on judgment |
| Requires "creativity" or "rules"? | Execute by rules | Requires original ideas |
| Requires "relationships"? | Not needed | Requires trust/emotion |
Step 2: Learn to Command AI
Path to Becoming a "Prompt Engineer":
- Week 1: Use ChatGPT/Claude daily to complete one work task
- Weeks 2-4: Compare different results from different prompts
- Months 1-2: Create prompt templates for your common tasks
- Months 3-6: Build unique prompt methodology for your professional domain
Path to Becoming an "AI Strategist":
- Month 1: Study AI application cases in your industry
- Months 2-3: Design an AI solution for a specific problem
- Months 3-6: Design complete human-AI collaboration workflows
- Months 6-12: Lead team AI transformation initiatives
Step 3: Redefine Your Value
Your New Value = AI's Capability × Your Ability to Command AI
Alpha Arena proves: Good configuration can create a 75% gap. Your configuration ability is your competitive advantage.
One-Sentence Summary: Don't compete with AI on "speed"—compete on "accuracy" and "human touch."
Part 6: Conclusions and Survival Rules
Survival Rules from Alpha Arena
Rule 1: AI is a Tool, Not a Competitor
Grok 4 and Grok 4.20 are the same AI. One lost 53%, one gained 22%. The difference isn't in the AI, it's in the humans using the AI.
Rule 2: Configuration Ability > AI Ability
Strongest AI + poor configuration = failure. Average AI + good configuration = success. Your configuration ability is your competitive advantage.
Rule 3: Move Up, Don't Compete Down
Don't compete with AI on "who executes faster"—you'll lose. Compete on "who can better command AI"—that's humanity's advantage.
Rule 4: Create New Value, Don't Protect Old Value
Before Alpha Arena, there was no "AI trading benchmark" industry. Jay Azhan created it. Ask yourself: What new demands has AI's emergence created? Can I meet these new demands?
A Final Metaphor
During the Industrial Revolution, tractors replaced oxen. Farmers had two choices:
• Compete with tractors on who plows faster → Failure
• Learn to drive the tractor → 100x efficiency gain
During the AI Revolution, AI replaces "mental labor." We have two choices:
• Compete with AI on who analyzes faster → Failure
• Learn to command AI → 10x, 100x efficiency gain
Grok 4.20's victory was the driver's victory, not the tractor's victory.
Your job is to become the best "AI driver."
— The End —