1
00:00:00,200 --> 00:00:03,680
Imagine one player entering
an unfamiliar world.
想象一个玩家走进陌生的世界。

2
00:00:04,120 --> 00:00:07,680
How can foundation models
learn from failure, recover,
基础模型如何从失败中学习、恢复，

3
00:00:08,060 --> 00:00:10,680
and carry that experience into the next game?
并将这份经验带入下一个游戏？

4
00:00:11,420 --> 00:00:12,880
It starts by interacting:
一切始于交互：

5
00:00:13,220 --> 00:00:16,220
observing, acting, and
discovering what follows.
观察、行动，再去发现结果。

6
00:00:16,900 --> 00:00:19,630
Success and failure can
reveal where a strategy works,
成功与失败可以揭示策略何时有效，

7
00:00:19,640 --> 00:00:22,380
and what the next attempt should change.
以及下一次尝试该如何调整。

8
00:00:22,775 --> 00:00:25,995
But growth also depends on
the worlds we learn from.
但成长也取决于我们从中学习的世界。

9
00:00:26,635 --> 00:00:29,725
Stable rules, persistent
state, meaningful goals,
稳定的规则、持续的状态、有意义的目标，

10
00:00:29,735 --> 00:00:33,595
and reliable feedback give
each decision consequences.
以及可靠的反馈，让每个决策都产生后果。

11
00:00:34,515 --> 00:00:37,495
Our survey connects four
pillars around the player:
我们的综述以玩家为中心，连接四个支柱：

12
00:00:37,915 --> 00:00:41,195
datasets for experience,
models for capability,
数据集提供经验，模型承载能力，

13
00:00:41,535 --> 00:00:44,875
harnesses for interaction,
and benchmarks for evidence.
运行框架支持交互，基准测试提供证据。

14
00:00:45,640 --> 00:00:47,060
Growth can take two paths.
成长可以沿着两条路径展开。

15
00:00:47,520 --> 00:00:49,240
Training updates model weights;
训练会更新模型权重；

16
00:00:49,520 --> 00:00:54,160
with weights fixed, feedback can still
reshape memory, beliefs, and plans.
权重固定时，反馈仍能调整记忆、判断与计划。

17
00:00:54,640 --> 00:00:56,680
Each path needs different evidence.
每条路径都需要不同的证据。

18
00:00:57,280 --> 00:00:59,600
Now change the layout, controls, or goal.
现在，改变地图、操作方式或目标。

19
00:01:00,080 --> 00:01:01,860
Familiar strategies may break,
熟悉的策略可能失效，

20
00:01:01,900 --> 00:01:05,620
opening new opportunities to
adapt as the challenges expand.
而不断扩展的挑战，也带来了新的适应机会。

21
00:01:06,360 --> 00:01:08,300
So how do we know the player improved?
那么，如何知道玩家真的进步了？

22
00:01:09,025 --> 00:01:12,185
Measure depth, coverage
at a stated skill level,
衡量能力深度，以及既定能力标准下的覆盖范围，

23
00:01:12,225 --> 00:01:14,665
and transfer to unseen games.
以及向未见游戏的迁移。

24
00:01:15,065 --> 00:01:18,425
Verify those gains on
fresh, independent episodes.
再用全新且独立的游戏回合验证这些进步。

25
00:01:19,125 --> 00:01:21,335
Explore Towards Generalist Game Players
阅读《迈向通用游戏玩家》，

26
00:01:21,345 --> 00:01:23,425
and our living research collection.
探索我们持续更新的研究合集。

27
00:01:23,845 --> 00:01:25,755
Discover the players that learn to act,
关注学会行动的玩家，

28
00:01:25,765 --> 00:01:28,045
and the worlds that help them grow.
以及帮助它们成长的世界。