Anthropic 公开宣称 AI 自我改进进入临界区间——内部代码自动化率突破 80%

"""
最小递归自改进循环演示
让 AI 生成代码 → 自审 → 改进 → 重新评测，循环 N 轮
"""

import anthropic
import subprocess
import sys
import time

client = anthropic.Anthropic(api_key="your-api-key-here")  # ← 替换为你的 Key

TASK_DESCRIPTION = """
写一个 Python 函数 merge_sorted_lists(a, b)，合并两个已排序列表并返回新的排序列表。
要求：时间复杂度 O(n)，不要用 sort()，不要引入第三方库。
"""

EVALUATION_SCRIPT = """
import time, random

def evaluate():
    # 正确性测试
    from solution import merge_sorted_lists
    assert merge_sorted_lists([1,3,5], [2,4,6]) == [1,2,3,4,5,6]
    assert merge_sorted_lists([], [1,2]) == [1,2]
    assert merge_sorted_lists([1,2], []) == [1,2]
    assert merge_sorted_lists([1,1,2], [1,3]) == [1,1,1,2,3]

    # 性能测试
    a = sorted(random.randint(0, 1000000) for _ in range(100000))
    b = sorted(random.randint(0, 1000000) for _ in range(100000))
    start = time.perf_counter()
    result = merge_sorted_lists(a, b)
    elapsed = time.perf_counter() - start
    assert len(result) == 200000
    assert result == sorted(a + b)
    print(f"PASS | 耗时 {elapsed:.4f}s")
    return elapsed

evaluate()
"""

MAX_ROUNDS = 3

def generate_code(task: str, previous_code: str = "", previous_feedback: str = "") -> str:
    """调用 Claude 生成或改进代码"""
    if previous_code and previous_feedback:
        prompt = (
            f"你之前为以下任务写了代码，但评测发现问题：\n"
            f"任务：{task}\n"
            f"之前的代码：\n```python\n{previous_code}\n```\n"
            f"评测反馈：{previous_feedback}\n\n"
            f"请根据反馈改进代码，只输出最终函数代码，不要解释。"
        )
    else:
        prompt = (
            f"为以下任务写 Python 代码，只输出最终函数代码，不要解释：\n{task}"
        )

    resp = client.messages.create(
        model="claude-sonnet-4-20250514",
        max_tokens=1024,
        messages=[{"role": "user", "content": prompt}],
    )
    # 提取代码块
    text = resp.content[0].text
    if "```python" in text:
        code = text.split("```python")[1].split("```")[0].strip()
    elif "```" in text:
        code = text.split("```")[1].split("```")[0].strip()
    else:
        code = text.strip()
    return code

def evaluate_code(code: str) -> tuple[bool, str]:
    """将代码写入文件并运行评测脚本"""
    with open("solution.py", "w") as f:
        f.write(code)
    result = subprocess.run(
        [sys.executable, "-c", EVALUATION_SCRIPT],
        capture_output=True, text=True, timeout=30,
    )
    output = result.stdout + result.stderr
    success = "PASS" in result.stdout
    return success, output

def self_review(code: str, eval_output: str) -> str:
    """让 Claude 自己审阅代码和评测结果，给出改进建议"""
    prompt = (
        f"以下是你的代码和评测结果：\n"
        f"代码：\n```python\n{code}\n```\n"
        f"评测输出：\n{eval_output}\n\n"
        f"请分析问题原因，给出具体的改进建议（不要直接写代码，只给建议）。"
    )
    resp = client.messages.create(
        model="claude-sonnet-4-20250514",
        max_tokens=512,
        messages=[{"role": "user", "content": prompt}],
    )
    return resp.content[0].text

# ---- 主循环 ----
code = ""
feedback = ""
best_time = float("inf")
best_code = ""

for round_num in range(1, MAX_ROUNDS + 1):
    print(f"\n===== 第 {round_num} 轮 =====")
    code = generate_code(TASK_DESCRIPTION, code, feedback)
    print(f"生成代码（前 200 字符）：\n{code[:200]}...")

    success, eval_output = evaluate_code(code)
    print(f"评测结果：{'通过' if success else '失败'}")
    print(eval_output.strip())

    if success:
        # 提取耗时
        for line in eval_output.splitlines():
            if "耗时" in line:
                try:
                    t = float(line.split("耗时")[1].split("s")[0].strip())
                    if t < best_time:
                        best_time = t
                        best_code = code
                except ValueError:
                    pass

    # 自审并生成下一轮反馈
    feedback = self_review(code, eval_output)
    print(f"自审建议（前 150 字符）：{feedback[:150]}...")

    time.sleep(1)  # 避免 rate limit

print(f"\n===== 最终结果 =====")
print(f"最佳耗时：{best_time:.4f}s")
print(f"最佳代码：\n{best_code}")

pip install anthropic
# 设置环境变量
export ANTHROPIC_API_KEY="sk-ant-xxxxx"
python recursive_self_improve.py