版主： verdelite， TheMatrix
Cursor我觉得就是这么训练的。 但是没用。我觉得目前的训练技术离大局观还很远。AI目前还没法操盘。Caravel 写了： 2025年 2月 13日 14:52 incremental的训练非常重要
应该从一个原始的github开始，人是怎么计划的，写一个prompt，然后加一部分code
然后再进来一个ticket，加一段code
用这种history 记录训练模型
每一个单步难度远不如ioi
AI要加一个外部存储。Caravel 写了： 2025年 2月 13日 14:52 incremental的训练非常重要
应该从一个原始的github开始，人是怎么计划的，写一个prompt，然后加一部分code
然后再进来一个ticket，加一段code
用这种history 记录训练模型
每一个单步难度远不如ioi
推理模型已经涌现了复杂推理的pattern，比如分支尝试，回滚，验证。wdong 写了： 2025年 2月 13日 15:52 Cursor我觉得就是这么训练的。 但是没用。我觉得目前的训练技术离大局观还很远。AI目前还没法操盘。
目前的训练技术里没有跨越不同逻辑层次的内容。以小学生的方式去训练AI，就是数据再大，我觉得也训练不出来大学生。 Reasoning是这方面努力的开始，但还差得远。
我不会说AI这么就没用。怎么发挥AI的作用是我们的责任。咱的盘子目前还很稳。
嗯。对。关键是怎么解耦。AI做40行的程序是可以信赖的。wdong 写了： 2025年 2月 12日 20:47 这就是HCI说的落不了地。 开始一个事情容易，但是能把起了头的事情干完是一种能力。这种能力就是人也不是个个都有的。靠AI没戏，还是得有人掌控全局。人掌控全局就不能让A。我看好的也仅仅是人+AI能落地。
你这个400行，就是我上次说的碎片化。其实400行都多。 如果每个文件不超过40行，或者每次定位要改的部分在40行之内，那改起来就容易了。怎么碎片化我还没搞定。
我下一步还不是用agent写程序。我打算先让agent做literature survey。
这个是AI打游戏，不是AI写游戏。
个人觉得这种一句话prompt的意义不大，大概率就是记住的。应该和人类那样些，先讨论plan，然后一步步实现，code review修改，通过。TheMatrix 写了： 2025年 2月 13日 16:24 这个是AI打游戏，不是AI写游戏。
不过视频最后说Github CEO演示了一个AI写的贪吃蛇游戏。说用了18分钟。看起来是没有人机交互的。
AI能力可能会越来越强，但是我们要的是人机交互，不能黑匣子。
BNF遵照regular expression的格式写。有?,*,+，代表zero or one，zero or more，one or more。TheMatrix 写了： 2025年 3月 2日 12:45 game: game_init, game_going+
game_init: game_static_init, generate_preview_blocks
game_going: null_mouse_event
game_going: click_drag_release
click_drag_release: click_a_preview_block, drag_a_block*, release_a_block
click_a_preview_block: MOUSE_CLICK, mouse_click_effect
drag_a_block: MOUSE_MOVE, mouse_drag_effect
mouse_drag_effect: not_fit_in_board, drag_block
mouse_drag_effect: fit_in_board, show_fit_in, drag_block
release_a_block: MOUSE_RELEASE, mouse_release_effect
mouse_release_effect: not_fit_in_board, go_back_to_preview
mouse_release_effect: fit_in_board, update_board, update_preview
update_preview: update_preview_blocks, update_game_over
update_preview_blocks: no_preview_blocks, generate_preview_blocks
代码： 全选
import tkinter as tk
import random
class BlockPuzzle:
def __init__(self):
pass
game = BlockPuzzle()
def game_init():
game_static_init()
generate_preview_blocks()
def game_static_init():
game_init_layout()
game_init_canvas()
def generate_preview_blocks():
pass
def game_going(event):
game.event = event
if null_mouse_event(): return
if click_drag_release(): return
def null_mouse_event():
#game over
#not click on a preview block
#clicked on a preview block but the block is empty
#clicked on a preview block but the block is dead
#mouse move but not in a dragging state
#mouse release but not in a dragging state
return False
def click_drag_release():
if click_on_a_preview_block():
while drag_a_block(): pass
release_a_block()
def click_on_a_preview_block():
return True
def drag_a_block():
event_type = tk.EventType(game.event.type).name
if event_type == "Motion":
mouse_drag_effect()
return True
return False
def mouse_drag_effect():
pass
def release_a_block():
event_type = tk.EventType(game.event.type).name
if event_type == "ButtonRelease":
mouse_release_effect()
def mouse_release_effect():
if not fit_in_board():
go_back_to_preview()
else:
update_board()
update_preview()
def go_back_to_preview():
pass
def update_board():
pass
def update_preview():
update_preview_blocks()
update_game_over()
def update_preview_blocks():
if no_preview_blocks():
generate_preview_blocks()
if __name__ == "__main__":
root = tk.Tk()
game = BlockPuzzle()
game_init()
root.mainloop()
自顶向下设计完毕，接下来就简单了。一个函数一个函数的填，相当于localized了。我估计叫deepseek也能出来。这个我要回头试一下。TheMatrix 写了： 2025年 3月 2日 12:51 BNF遵照regular expression的格式写。有?,*,+，代表zero or one，zero or more，one or more。
BNF直接翻译成程序。这里因为事件流是被python tk截获的，所以和自己截获事件流写法稍微有点区别。
代码： 全选
def game_init_layout():
game.game_bg_color = "#1A2B3C"
game.cell_color = game.game_bg_color
game.grid_color = "#708090"
game.padx = 100
game.top_section_pady = 20
game.board_pady = 20
game.preview_pady = 20
game.bottom_pady = 20
game.cell_size = 50
game.preview_cell_size = 24
game.grid_border = 1
game.board_n_cell = 8
game.board_size = game.cell_size * game.board_n_cell
game.top_section_x = game.padx
game.top_section_y = game.top_section_pady
game.top_section_height = 100
game.top_section_width = game.board_size
game.board_x = game.padx
game.board_y = game.top_section_y + game.top_section_height + game.board_pady
game.preview_x = game.padx
game.preview_y = game.board_y + game.board_size + game.preview_pady
game.preview_width = game.board_size
game.preview_height = game.board_size // 2
game.canvas_width = 2 * game.padx + game.board_size
game.canvas_height = game.preview_y + game.preview_height + game.bottom_pady
game.preview_one_size = 5* game.preview_cell_size
game.preview_padx = 10
xx = game.preview_width - 3* game.preview_one_size - 2* game.preview_padx
assert(xx % 2 == 0)
game.preview_lr_padx = xx // 2
game.preview_one_x = [
game.preview_x + game.preview_lr_padx +
i* (game.preview_one_size + game.preview_padx)
for i in range(3)]
game.preview_one_y = game.preview_y
def game_init_canvas():
root.title("Block Puzzle")
root.resizable(False, False)
game.canvas = tk.Canvas(
root,
bg=game.game_bg_color,
highlightthickness=0,
width=game.canvas_width,
height=game.canvas_height,
)
game.canvas.pack()
game.canvas.bind("<Button-1>", game_going)
game.canvas.bind("<B1-Motion>", game_going)
game.canvas.bind("<ButtonRelease-1>", game_going)
代码： 全选
def game_static_init():
game_init_layout()
game_init_canvas()
set_game_blocks()
代码： 全选
def variation_blocks():
def rotate(p):
x,y = p
return -y,x
def normalize(block):
min_x = min(dx for dx,dy in block)
min_y = min(dy for dx,dy in block)
nblock = [(dx-min_x, dy-min_y) for dx,dy in block]
return nblock
def variation_block(block):
nblocks = []
nb = block
for i in range(4):
nb = sorted(nb)
nblocks.append(tuple(nb))
nb = [game.rotate(c) for c in nb]
nb = game.normalize(nb)
nblocks = list(set(nblocks))
return nblocks
game.tetris_blocks = []
for block in game.base_tetris_blocks:
nblocks = game.variation_block(block["shape"])
game.tetris_blocks += [{"shape": b, "color": block["color"]} for b in nblocks]
def set_game_blocks():
game.base_tetris_blocks = [
{"shape": [(0, 0), (0, 1), (0, 2), (0, 3)], "color": "#6DB3AC"}, # I (4-square)
{"shape": [(0, 0), (0, 1), (1, 0), (1, 1)], "color": "#F0F000"}, # O
{"shape": [(0, 0), (1, 0), (1, 1), (2, 0)], "color": "#BD4DF5"}, # T
{"shape": [(0, 1), (0, 2), (1, 0), (1, 1)], "color": "#00F000"}, # S
{"shape": [(0, 0), (0, 1), (1, 1), (1, 2)], "color": "#F00000"}, # Z
{"shape": [(0, 0), (1, 0), (2, 0), (2, 1)], "color": "#699EEB"}, # J
{"shape": [(0, 1), (1, 1), (2, 0), (2, 1)], "color": "#F0A000"}, # L
# Custom blocks
{"shape": [(0, 0)], "color": "#F09050"}, # 1-square
{"shape": [(0, 0), (0, 1)], "color": "#6FEA85"}, # 2-square
{"shape": [(0, 0), (0, 1), (0, 2)], "color": "#B34DB3"}, # 3-line
{"shape": [(0, 0), (0, 1), (1, 1)], "color": "#A5B350"}, # L-3
# 5-square blocks
{"shape": [(0, 0), (0, 1), (0, 2), (0, 3), (0, 4)], "color": "#D0603C"}, # 5-line (Dark Turquoise)
{"shape": [(0, 0), (0, 1), (0, 2), (1, 2), (2, 2)], "color": "#C8BA4A"}, # Right-angle (Red)
# 9-square block
{"shape": [
(0, 0), (0, 1), (0, 2),
(1, 0), (1, 1), (1, 2),
(2, 0), (2, 1), (2, 2)
], "color": "#FFD700"} # 3x3 (Gold)
]
variation_blocks()
代码： 全选
def game_static_init():
game_init_layout()
game_init_canvas()
set_game_blocks()
game_init_board()
代码： 全选
def game_init_board():
#board grids
for row in range(8):
for col in range(8):
x1 = col * game.cell_size + game.board_x
y1 = row * game.cell_size + game.board_y
x2 = x1+ game.cell_size -1
y2 = y1+ game.cell_size -1
game.canvas.create_rectangle(
x1,y1,x2,y2,
fill=game.cell_color,
outline=game.grid_color,
width=game.grid_border,
)
#board outer border
game.canvas.create_rectangle(
game.board_x-1,
game.board_y-1,
game.board_x+game.board_size,
game.board_y+game.board_size,
#fill=game.cell_color,
outline=game.grid_color,
width=game.grid_border,
)
代码： 全选
def generate_preview_blocks():
nblock = len(game.tetris_blocks)
game.preview_blocks = [random.choice(range(nblock)) for i in range(3)]
game.preview_blocks_present = [True for i in range(3)]
game.preview_blocks_dead = [False for i in range(3)]
game.preview_blocks_rect = [[] for i in range(3)]
for i in range(3):
put_preview_block(i)
代码： 全选
def put_preview_block(i, color=None):
block = game.tetris_blocks[game.preview_blocks[i]]
if color is None:
color = block["color"]
nx = max(dx for dx,dy in block["shape"])+1
ny = max(dy for dx,dy in block["shape"])+1
xx = game.preview_one_size - nx*game.preview_cell_size
assert(xx % 2 == 0)
yy = game.preview_one_size - ny*game.preview_cell_size
assert(yy % 2 == 0)
for dx,dy in block["shape"]:
x1 = dx* game.preview_cell_size + game.preview_one_x[i] + xx//2
y1 = dy* game.preview_cell_size + game.preview_one_y + yy//2
x2 = x1+ game.preview_cell_size -1
y2 = y1+ game.preview_cell_size -1
game.preview_blocks_rect[i].append(
game.canvas.create_rectangle(
x1,y1,x2,y2,
fill=color,
outline=game.game_bg_color,
width=game.grid_border,
))
现在是这样的。已经能看了！TheMatrix 写了： 2025年 3月 2日 14:37 preview的部分填充一下。这样game_init的部分就完成了。接下来是game_going，也就是mouse event的处理。
game: game_init, game_going+
game_init: game_static_init, generate_preview_blocks
AI网页版的是不是每次要把以前所有对话内容全部发给AI？用git history可以省略一大部分对话记录。Caravel 写了： 2025年 2月 13日 14:52 incremental的训练非常重要
应该从一个原始的github开始，人是怎么计划的，写一个prompt，然后加一部分code
然后再进来一个ticket，加一段code
用这种history 记录训练模型
每一个单步难度远不如ioi
