WEBVTT

1
00:00:00.600 --> 00:00:06.782
Two summers ago, I sat down to build a presentation about AI in education. I called it…

2
00:00:06.782 --> 00:00:11.329
The Next Wave. It's still online — frozen right where I left it.

3
00:00:11.329 --> 00:00:17.013
And on one page, I tried to map the entire landscape of artificial intelligence.

4
00:00:17.013 --> 00:00:20.850
Two columns. What was real… and what was still theory.

5
00:00:26.850 --> 00:00:33.089
The real column held narrow AI — brilliant at one task, helpless one step outside it.

6
00:00:33.089 --> 00:00:38.227
The theory column held the big ideas: artificial general intelligence…

7
00:00:38.227 --> 00:00:43.879
superintelligence… and something called Artificial Capable Intelligence. ACI.

8
00:00:43.879 --> 00:00:49.163
Mustafa Suleyman's name for an AI that could take a goal… and finish it.

9
00:00:49.163 --> 00:00:54.889
With minimal oversight. My slide granted them, quote, "significant progress" —

10
00:00:54.889 --> 00:01:00.100
but filed all three as "theoretical and developmental." As of mid-2024.

11
00:01:02.300 --> 00:01:03.880
Then… the wave arrived.

12
00:01:09.080 --> 00:01:13.811
Today, an agent can take a goal. Not a prompt — a goal. It plans.

13
00:01:13.811 --> 00:01:20.070
It opens a browser and reads. It writes code, runs it, watches it fail — and fixes it.

14
00:01:20.070 --> 00:01:24.655
It works for hours. It hands pieces of the job to other agents…

15
00:01:24.655 --> 00:01:27.930
and checks their work before you ever see it.

16
00:01:28.730 --> 00:01:33.861
They have names now. Claude Code. Codex. OpenClaw. They don't live in a chat window —

17
00:01:33.861 --> 00:01:38.329
they live in the computer. And you don't need an engineering team anymore.

18
00:01:38.329 --> 00:01:41.830
A teacher with a weekend… can build what used to take one.

19
00:01:44.030 --> 00:01:46.770
Which means this card… is in the wrong column.

20
00:01:51.370 --> 00:01:54.047
Now — honestly — the frontier is jagged.

21
00:01:54.047 --> 00:01:59.871
When Stanford measured them this spring, agents still failed one everyday computer task

22
00:01:59.871 --> 00:02:05.560
in three. Today's live leaderboards say one in six — nudging past the human baseline.

23
00:02:05.560 --> 00:02:11.384
And long, multi-day jobs? Still four fails in five. They work inside permissions, too —

24
00:02:11.384 --> 00:02:16.337
drafting an email is not permission to send it. But barely a few years in…

25
00:02:16.337 --> 00:02:22.160
already this good — and getting better. "Theoretical"? … That word doesn't fit anymore.

26
00:02:24.760 --> 00:02:28.958
So what does this mean for your classroom? When AI can produce the essay —

27
00:02:28.958 --> 00:02:32.817
and act on the plan — the artifact tells us less about the learning.

28
00:02:32.817 --> 00:02:37.129
The questions get more human. What did they notice? Where did they struggle?

29
00:02:37.129 --> 00:02:38.320
What did they choose?

30
00:02:39.520 --> 00:02:44.277
And my old map never showed this: your students may be the generation that decides how

31
00:02:44.277 --> 00:02:47.320
the world treats these systems. They'll draw that line.

32
00:02:52.520 --> 00:02:57.588
One last thing. This little robot began as a single still picture —

33
00:02:57.588 --> 00:03:03.110
AI agents taught it to move. These animations? Agents. The score? Agents.

34
00:03:03.110 --> 00:03:05.758
And this voice… is my voice clone —

35
00:03:05.758 --> 00:03:11.885
built with my consent, disclosed right there on the screen. My voice. My choice —

36
00:03:11.885 --> 00:03:17.180
said out loud, because that's the deal. All of it… in about a weekend.

37
00:03:18.780 --> 00:03:24.545
Back in 2024, I filed this under "theoretical." The question this morning isn't whether

38
00:03:24.545 --> 00:03:29.051
capable machines are coming — they're here. It's… what is learning —

39
00:03:29.051 --> 00:03:32.430
when AI can do the work? … Let's find out together.