WEBVTT

NOTE Sentence-level transcript of https://www.youtube.com/watch?v=hVJOnuhFmTA

NOTE One cue per sentence. Cue ids are the line anchors on /transcripts/hVJOnuhFmTA.html. A cue ends where the next begins, or 2 s after its last word.

s1
00:00:01.309 --> 00:00:03.309
[music]

s2
00:00:07.080 --> 00:00:11.400
I'm sure that sometime in the last few hours most of you did this.

s3
00:00:11.400 --> 00:00:16.880
You typed a request into a small box to a super intelligence and then you waited.

s4
00:00:16.880 --> 00:00:27.840
You watched the cursor blink, maybe a little throbber cycled through clever gerunds like hullabalooing, tomfoolery, and philosophizing to hide the wait.

s5
00:00:27.840 --> 00:00:32.560
Maybe it gave you what you wanted, maybe you rephrased it and tried again.

s6
00:00:32.560 --> 00:00:34.600
It all felt completely normal.

s7
00:00:34.600 --> 00:00:40.640
I want to spend the next 20 minutes making prompting feel unfamiliar and strange again.

s8
00:00:40.640 --> 00:00:43.960
I'm Ted Johnson, co-founder of Join an AI.

s9
00:00:43.960 --> 00:00:53.640
During my 25-year career building enterprise software, collaboration systems, and AI-enabled interfaces, I've always focused on human interaction.

s10
00:00:53.640 --> 00:01:04.920
I've also been following AI for two decades, including back to the far less impressive GPT-1 and 2. And when ChatGPT arrived, I felt two things at once.

s11
00:01:04.920 --> 00:01:13.280
First, and unsurprisingly, amazement, knowing the world would never be the same, followed by actually surprising disappointment I couldn't shake.

s12
00:01:13.280 --> 00:01:23.560
This disappointment turned into an observation that started a company, Join an AI, and that I keep coming back to, which is why do we still have to learn AI?

s13
00:01:23.560 --> 00:01:29.160
Why does something this powerful so often feel unnatural to use?

s14
00:01:29.160 --> 00:01:31.960
Here's the path we'll take to answer that.

s15
00:01:31.960 --> 00:01:35.920
We'll start with the most familiar computer interface and make it strange again.

s16
00:01:35.920 --> 00:01:38.880
Then, I'll give you three key concepts.

s17
00:01:38.880 --> 00:01:51.800
The channel, the physical transport that carries your intent, expression, the range and richness of meaning the channel can carry, and the protocol, the shape or rules of this exchange.

s18
00:01:51.800 --> 00:01:58.040
I'll use those three concepts to show you that the prompt is our present-day punch card.

s19
00:01:58.040 --> 00:02:06.920
We'll share examples of the ways interfaces could progress, and we'll wrap with some practical advice for AI and human-centered design.

s20
00:02:06.920 --> 00:02:09.479
Everyone knows what this is, the keyboard.

s21
00:02:09.479 --> 00:02:14.040
It's everywhere and it feels completely normal or natural.

s22
00:02:14.040 --> 00:02:14.840
But it isn't.

s23
00:02:14.840 --> 00:02:17.120
We all had to take lessons.

s24
00:02:17.120 --> 00:02:18.840
We all had to practice.

s25
00:02:18.840 --> 00:02:21.560
And I say this is someone who loves keyboards.

s26
00:02:21.560 --> 00:02:25.600
But it seems that we've been trying to fix them as long as they've been around.

s27
00:02:25.600 --> 00:02:31.400
People have tried more efficient layouts like the Dvorak or Colemak to save their fingers some work.

s28
00:02:31.400 --> 00:02:44.480
A more extreme example, some keyboard enthusiasts refuse to squander their two digits on the spacebar giving them four or eight or 10 keys to press with that efficient thumb of theirs.

s29
00:02:44.480 --> 00:02:46.280
And what do we even mean by the keyboard?

s30
00:02:46.280 --> 00:02:49.120
Here's the patent drawing for the layout we use every day.

s31
00:02:49.120 --> 00:02:52.240
This patent's from about 1860.

s32
00:02:52.240 --> 00:03:02.880
And my personal favorite, it just as could have easily been the Hansen Writing Ball, which looks anything but like a way you want to talk to a superintelligence.

s33
00:03:02.880 --> 00:03:09.440
We carry these legacies of an arbitrary input device designed under constraints that haven't existed for a century.

s34
00:03:09.440 --> 00:03:13.680
And we put it between ourselves and the most capable machines ever built.

s35
00:03:13.680 --> 00:03:15.040
Nobody alive chose it.

s36
00:03:15.040 --> 00:03:16.720
We all inherited it.

s37
00:03:16.720 --> 00:03:18.000
And then we stopped noticing.

s38
00:03:18.000 --> 00:03:20.440
So, that's the first idea, the channel.

s39
00:03:20.440 --> 00:03:23.400
The medium an interface gives you to work in.

s40
00:03:23.400 --> 00:03:24.840
A keyboard is a channel.

s41
00:03:24.840 --> 00:03:27.160
A microphone, channel.

s42
00:03:27.160 --> 00:03:30.720
A screen, a punchcard, a prompt box, all channels.

s43
00:03:30.720 --> 00:03:35.519
And channels matter because each one can physically carry a different kind of signal.

s44
00:03:35.519 --> 00:03:39.080
For example, text is a stream of discrete symbols.

s45
00:03:39.080 --> 00:03:42.840
Voice can carry timing, pitch, hesitation, and words.

s46
00:03:42.840 --> 00:03:45.920
A diagram can carry spatial relationship all at once.

s47
00:03:45.920 --> 00:03:50.080
But these are differences in what the medium can transmit.

s48
00:03:50.080 --> 00:03:53.400
It's bandwidth, not the differences in the meaning.

s49
00:03:53.400 --> 00:03:56.457
And carrying more signal isn't the same as the machine understanding

s50
00:03:56.457 --> 00:03:57.120
[snorts]

s51
00:03:57.120 --> 00:03:58.160
any of it.

s52
00:03:58.160 --> 00:04:00.680
That's a separate question.

s53
00:04:00.680 --> 00:04:02.400
That's the next idea.

s54
00:04:02.400 --> 00:04:05.240
Humans use all these channels constantly without thinking.

s55
00:04:05.240 --> 00:04:08.040
We never pick one channel and force everything through it.

s56
00:04:08.040 --> 00:04:09.440
That would be absurd.

s57
00:04:09.440 --> 00:04:13.640
And yet, that's what we ask people to do with machines over and over.

s58
00:04:13.640 --> 00:04:16.640
Hold on to the word channel because here's the plot twist.

s59
00:04:16.640 --> 00:04:19.920
With AI, the channel never really changed.

s60
00:04:19.920 --> 00:04:25.320
You're still typing into a box, but what you were allowed to push through it was about to.

s61
00:04:25.320 --> 00:04:30.640
For the first time, the computer channels carry rich, complete human language.

s62
00:04:30.640 --> 00:04:34.440
Notice I said what it carries, not what it is.

s63
00:04:34.440 --> 00:04:37.480
You're still typing into a box, you're still hitting the submit button.

s64
00:04:37.480 --> 00:04:39.200
The keyboard didn't change.

s65
00:04:39.200 --> 00:04:43.680
What improved is the range of what you're permitted to express through it.

s66
00:04:43.680 --> 00:04:46.240
There's the second idea, expression.

s67
00:04:46.240 --> 00:04:51.440
How much of what you actually communicate or mean will go through the interface?

s68
00:04:51.440 --> 00:04:53.640
That's the second idea, expression.

s69
00:04:53.640 --> 00:04:58.440
How much of what you actually communicate or mean will the interface let through?

s70
00:04:58.440 --> 00:05:03.280
Here's an example of expression progress over time with computers.

s71
00:05:03.280 --> 00:05:08.480
Starting with assembly, which gave you an instruction set, a few dozen opcodes.

s72
00:05:08.480 --> 00:05:13.200
Then came the commands with the shell inputs, flags.

s73
00:05:13.200 --> 00:05:17.000
Then modern programming languages gave you primitives that you could compose.

s74
00:05:17.000 --> 00:05:27.280
While powerful, step-by-step, each one of these is a fixed vocabulary, requiring you to express your intent by choosing from a menu the machine will accept.

s75
00:05:27.280 --> 00:05:29.760
Natural language blew that menu open.

s76
00:05:29.760 --> 00:05:34.360
For the first time, you can say almost anything, the way you'd say it to another person.

s77
00:05:34.360 --> 00:05:38.960
And on the expression one axis, the leap is real and enormous.

s78
00:05:38.960 --> 00:05:46.880
There's an ocean of meaning in an ordinary human request, context, nuance, intent, all things we've never had to spell out to each other.

s79
00:05:46.880 --> 00:05:51.360
For the first time, you can say almost anything the way you'd say it to another person.

s80
00:05:51.360 --> 00:05:53.720
For the first time, a machine can take it in.

s81
00:05:53.720 --> 00:05:59.160
So, here is what should bother us as engineers and designers as it's bothered and inspired me.

s82
00:05:59.160 --> 00:06:05.320
The channels for computers have been the same for 15 years, some 180 years.

s83
00:06:05.320 --> 00:06:08.920
Now, with AI, we've poured an ocean of expression into it.

s84
00:06:08.920 --> 00:06:14.160
So, why does it feel like we're still sipping through a straw and struggling to learn how the AI thinks?

s85
00:06:14.160 --> 00:06:21.720
Because there's a third idea underpinning the other two, and it's really the one that hasn't kept up, the protocol.

s86
00:06:21.720 --> 00:06:25.960
The rules you follow and the shape of the interaction itself.

s87
00:06:25.960 --> 00:06:27.800
Channel stayed the same.

s88
00:06:27.800 --> 00:06:36.000
Expression exploded in the last 3 years with LLMs, but the protocol, prompting, is the protocol of a punch card.

s89
00:06:36.000 --> 00:06:38.840
And the punch card's protocol is good old batch.

s90
00:06:38.840 --> 00:06:41.600
Here's what punch card batch meant.

s91
00:06:41.600 --> 00:06:53.919
You sat down away from the machine, carefully encoded your entire request in advance, carried your deck to the operator, you submitted the job, and then you waited, sometimes hours, sometimes overnight.

s92
00:06:53.919 --> 00:06:59.840
Then you read the printout, found one thing that was wrong, fixed it, resubmitted it, and waited again.

s93
00:06:59.840 --> 00:07:04.280
The machine never engaged with you while you were thinking.

s94
00:07:04.280 --> 00:07:08.040
It engaged with the finished package after the fact.

s95
00:07:08.040 --> 00:07:10.360
Now, let's look at the prompt.

s96
00:07:10.360 --> 00:07:14.600
Assemble the whole request, submit it, wait, read what comes back.

s97
00:07:14.600 --> 00:07:19.040
Something's off, assemble it again, submit it again, and wait.

s98
00:07:19.040 --> 00:07:23.720
We have to acknowledge that there are features, interactive features improving this.

s99
00:07:23.720 --> 00:07:27.640
You can ask for updates, you can ask for summaries of what was done.

s100
00:07:27.640 --> 00:07:30.919
But in the end, it's still batch with interactive sprinkles.

s101
00:07:30.919 --> 00:07:32.800
It's the same protocol.

s102
00:07:32.800 --> 00:07:41.040
We shrank the wait time from overnight to a few seconds or few minutes, and the speed fooled us into thinking that it become interactive.

s103
00:07:41.040 --> 00:07:41.800
It hasn't.

s104
00:07:41.800 --> 00:07:43.040
It's still batch.

s105
00:07:43.040 --> 00:07:47.000
You still package a complete turn before the machine is allowed to participate.

s106
00:07:47.000 --> 00:07:54.240
We learn tricks, send tips to use code skills, or rewrite prompts a certain way to manage this.

s107
00:07:54.240 --> 00:07:55.680
And speaking doesn't change it.

s108
00:07:55.680 --> 00:07:58.720
Your voice just gets transcribed into the box and submitted.

s109
00:07:58.720 --> 00:08:03.880
Shorter batch is still batch because the protocol is the part that did not advance.

s110
00:08:03.880 --> 00:08:06.720
The protocol is the part we've had to learn.

s111
00:08:06.720 --> 00:08:08.120
We just gave it a flattering name.

s112
00:08:08.120 --> 00:08:12.400
We call it prompt engineering and treat it like it's a power user skill.

s113
00:08:12.400 --> 00:08:17.440
Strip the label off and it's a set of rules for packaging up good old batch.

s114
00:08:17.440 --> 00:08:20.040
For example, tell it to think step-by-step.

s115
00:08:20.040 --> 00:08:21.400
Give it examples.

s116
00:08:21.400 --> 00:08:23.120
Ask it to be an expert.

s117
00:08:23.120 --> 00:08:25.200
Don't ask it to be an expert.

s118
00:08:25.200 --> 00:08:26.560
Don't ask it that way.

s119
00:08:26.560 --> 00:08:27.960
Paste more context.

s120
00:08:27.960 --> 00:08:29.680
Paste less context.

s121
00:08:29.680 --> 00:08:32.200
Only talk to it through markdown documents.

s122
00:08:32.200 --> 00:08:33.680
We trade incantations.

s123
00:08:33.680 --> 00:08:35.840
We've learned the magic words.

s124
00:08:35.840 --> 00:08:36.840
That's the illusion.

s125
00:08:36.840 --> 00:08:41.599
It feels like mastery, but it's the same sort of mastery a punch card operator had.

s126
00:08:41.599 --> 00:08:45.160
Knowing exactly how to assemble the deck so the job wouldn't fail.

s127
00:08:45.160 --> 00:08:49.520
Moreover, we've gotten good at prompting or these black boxes.

s128
00:08:49.520 --> 00:08:52.720
And and that's the part that should bother us, not reassure us.

s129
00:08:52.720 --> 00:08:54.680
None of this means prompts are bad.

s130
00:08:54.680 --> 00:08:55.680
Punch cards weren't bad.

s131
00:08:55.680 --> 00:08:57.520
Command lines aren't bad.

s132
00:08:57.520 --> 00:09:00.920
They're brilliant solutions for constraints of their time.

s133
00:09:00.920 --> 00:09:01.920
But that's the whole question.

s134
00:09:01.920 --> 00:09:04.320
Is batch still the right protocol?

s135
00:09:04.320 --> 00:09:11.200
Are we still pre-packaging our intent for a machine that no longer needs us to because it shouldn't need us anymore?

s136
00:09:11.200 --> 00:09:12.880
It can ask a follow-up.

s137
00:09:12.880 --> 00:09:14.400
It can clarify mid-thought.

s138
00:09:14.400 --> 00:09:17.160
It can notice it's missing something and say so.

s139
00:09:17.160 --> 00:09:19.360
It should be human conversational.

s140
00:09:19.360 --> 00:09:21.839
Sherry Turkle at MIT puts it very well.

s141
00:09:21.839 --> 00:09:26.120
Conversation is the most human and humanizing thing we do.

s142
00:09:26.120 --> 00:09:28.240
It's one of humanity's superpowers.

s143
00:09:28.240 --> 00:09:31.800
The capacity to engage and think is right there.

s144
00:09:31.800 --> 00:09:36.320
And yet, we're still making people submit the deck and wait for the run.

s145
00:09:36.320 --> 00:09:41.240
Even the punch card inheritance protocol in fact, batch came from the weaving loom.

s146
00:09:41.240 --> 00:09:45.040
You set the whole pattern in advance, then ran the cloth.

s147
00:09:45.040 --> 00:09:48.360
The punch card got reused on computers by default.

s148
00:09:48.360 --> 00:09:51.080
We're still standing at the same moment again.

s149
00:09:51.080 --> 00:09:56.320
AI could finally meet us in the middle of a thought, got handed a protocol of a loom.

s150
00:09:56.320 --> 00:09:57.040
That's what I mean.

s151
00:09:57.040 --> 00:10:00.560
The prompt is still a punch card, not because of how you encode it.

s152
00:10:00.560 --> 00:10:03.160
The encoding is powerful and awesome.

s153
00:10:03.160 --> 00:10:10.080
Because when the LLM is allowed to engage only after you've packaged a complete turn and submitted it.

s154
00:10:10.080 --> 00:10:11.600
And this is where the mismatch bites.

s155
00:10:11.600 --> 00:10:13.800
Model capacity is shooting straight up.

s156
00:10:13.800 --> 00:10:18.240
Reasoning, speech, vision, memory, planning, all curving upwards.

s157
00:10:18.240 --> 00:10:20.320
The interface protocol, flat.

s158
00:10:20.320 --> 00:10:25.920
Still a box, still a submit button, still the human doing all the work around the LLM.

s159
00:10:25.920 --> 00:10:37.120
The human still decides what context matters, still remembers what to ask, still chooses the timing, still notices the ambiguity, still repairs the output, still has to carefully engineer a prompt.

s160
00:10:37.120 --> 00:10:39.520
But the intelligence feels magical.

s161
00:10:39.520 --> 00:10:41.960
It's the interface that still feels like work.

s162
00:10:41.960 --> 00:10:48.200
And when it feels like work, when the output's wrong, when the magic words don't land, people blame themselves.

s163
00:10:48.200 --> 00:10:49.800
They decide they're bad at this.

s164
00:10:49.800 --> 00:10:51.440
They're not specific enough.

s165
00:10:51.440 --> 00:10:52.520
They don't get AI.

s166
00:10:52.520 --> 00:10:56.120
I want to say as clearly as I can, it is not our fault.

s167
00:10:56.120 --> 00:10:58.520
We are not bad at using AI.

s168
00:10:58.520 --> 00:11:03.240
We are being asked to operate a brand new kind of intelligence through a protocol of a punch card.

s169
00:11:03.240 --> 00:11:05.880
The mismatch isn't the user, it's the interface.

s170
00:11:05.880 --> 00:11:09.320
In the race to enable AI, we shortcut the interface.

s171
00:11:09.320 --> 00:11:12.200
Okay, let's make this concrete and familiar.

s172
00:11:12.200 --> 00:11:17.880
A few weeks ago, my co-founder was using a Frontier company's voice mode.

s173
00:11:17.880 --> 00:11:21.600
These are known as speech-to-speech models.

s174
00:11:21.600 --> 00:11:25.360
He asked it a normal question, "When is the next Timberwolves game?"

s175
00:11:25.360 --> 00:11:26.240
Fine.

s176
00:11:26.240 --> 00:11:27.600
It answered it quickly.

s177
00:11:27.600 --> 00:11:34.360
Then, he pretended I showed up as if to speak to me and said, "Hey Ted, come on in."

s178
00:11:34.360 --> 00:11:36.440
He wasn't talking to the AI.

s179
00:11:36.440 --> 00:11:39.800
But these models have no way to know that.

s180
00:11:39.800 --> 00:11:43.440
So the AI did the only thing a prompt box can do.

s181
00:11:43.440 --> 00:11:46.840
It took his speech as a turn and answered it.

s182
00:11:46.840 --> 00:11:48.360
"Sure, I'm here.

s183
00:11:48.360 --> 00:11:49.680
What's on your mind?"

s184
00:11:49.680 --> 00:11:53.200
That's not a good answer, but it's not a dumb model.

s185
00:11:53.200 --> 00:11:59.280
It answered the first question perfectly, but it's a protocol with exactly one slot.

s186
00:11:59.280 --> 00:12:01.600
Your message, then it's reply.

s187
00:12:01.600 --> 00:12:06.000
It has no concept of who's speaking, whether the words were even meant for it.

s188
00:12:06.000 --> 00:12:09.600
And the frontier companies want to make strides as well.

s189
00:12:09.600 --> 00:12:16.400
OpenAI released GPT real-time 2 in late May and started trying it for their voice mode as well.

s190
00:12:16.400 --> 00:12:18.080
It backchannels now.

s191
00:12:18.080 --> 00:12:19.400
It goes, "Mhm."

s192
00:12:19.400 --> 00:12:20.680
and right.

s193
00:12:20.680 --> 00:12:23.600
The little sounds we make to show we're listening actively.

s194
00:12:23.600 --> 00:12:28.280
We're seeing the field is converging on the same conclusion we built our company on.

s195
00:12:28.280 --> 00:12:33.520
The interface has to stop being batch and start participating.

s196
00:12:33.520 --> 00:12:37.160
Others are working on real-time conversation as well.

s197
00:12:37.160 --> 00:12:42.040
This is Nvidia's Personal Plex, a research model, not ours.

s198
00:12:42.040 --> 00:12:44.120
Watch what happens when it gets interrupted.

s199
00:12:44.120 --> 00:12:45.960
I've been thinking about starting a diet.

s200
00:12:45.960 --> 00:12:50.440
Yeah, starting a diet can feel a bit daunting, but you could keep it simple.

s201
00:12:50.440 --> 00:12:53.240
Focus on eating more veggies and fruits.

s202
00:12:53.240 --> 00:12:53.640
Try to

s203
00:12:53.640 --> 00:12:56.080
Before I forget, I signed up for a marathon.

s204
00:12:56.080 --> 00:12:58.280
All right, congrats on signing up for the marathon.

s205
00:12:58.280 --> 00:12:59.320
That's a big challenge.

s206
00:12:59.320 --> 00:13:00.520
You've got a lot of time.

s207
00:13:00.520 --> 00:13:03.880
Focus on building a solid base with regular long runs.

s208
00:13:03.880 --> 00:13:05.120
Stay hydrated.

s209
00:13:05.120 --> 00:13:10.320
Make sure you fuel right before and after, and don't forget to stretch and take care of your feet.

s210
00:13:10.320 --> 00:13:11.960
Personal Plex stops.

s211
00:13:11.960 --> 00:13:13.120
It yields.

s212
00:13:13.120 --> 00:13:14.880
It picks the thread back up.

s213
00:13:14.880 --> 00:13:20.880
That's real turn taking, listening and speaking at once in real time.

s214
00:13:20.880 --> 00:13:22.280
You need to come visit me.

s215
00:13:22.280 --> 00:13:23.160
Oh, okay.

s216
00:13:23.160 --> 00:13:24.520
go into the city.

s217
00:13:24.520 --> 00:13:25.000
Okay.

s218
00:13:25.000 --> 00:13:31.400
Cuz that's the thing, like there's like the random spray paint, but then there's also like I'm not sure.

s219
00:13:31.400 --> 00:13:33.040
People must commission them.

s220
00:13:33.040 --> 00:13:36.360
Like these massive like mural spray paint pieces.

s221
00:13:36.360 --> 00:13:38.600
Yeah, I think they do.

s222
00:13:38.600 --> 00:13:44.320
And Persona Plex's back channels listens and lands where a person's would.

s223
00:13:44.320 --> 00:13:49.000
Beyond conversational flow, there are lots of challenges and it's a complex problem.

s224
00:13:49.000 --> 00:13:54.640
Making listening noises is not really the same as knowing who's in the room.

s225
00:13:54.640 --> 00:14:00.920
These are not trained to tell that "Hey Ted" wasn't meant for it.

s226
00:14:00.920 --> 00:14:01.880
But we are.

s227
00:14:01.880 --> 00:14:05.280
We're working on improving the protocol to the models.

s228
00:14:05.280 --> 00:14:10.600
By giving it a better understanding of human and group conversation.

s229
00:14:10.600 --> 00:14:12.800
Good afternoon, everyone.

s230
00:14:12.800 --> 00:14:14.920
Good afternoon, Sam.

s231
00:14:14.920 --> 00:14:17.080
Hi, Jordan.

s232
00:14:17.080 --> 00:14:18.120
Good afternoon.

s233
00:14:18.120 --> 00:14:20.000
Good to see you both.

s234
00:14:20.000 --> 00:14:21.840
Quick one, which requirement is this?

s235
00:14:21.840 --> 00:14:24.320
Do we have an ID?

s236
00:14:24.320 --> 00:14:28.280
This is our Q442, expense approvals.

s237
00:14:28.280 --> 00:14:30.720
There, it answered a question.

s238
00:14:30.720 --> 00:14:35.400
It only takes actions based on the utility-driven model.

s239
00:14:35.400 --> 00:14:45.800
So, it creates goals to fulfill as it labels each of the participants' statements as a question, a proposal, an answer, and then only takes a turn when no one else is speaking

s240
00:14:45.800 --> 00:14:47.480
or holding the floor.

s241
00:14:47.480 --> 00:14:50.880
Right, we need users to approve requests faster.

s242
00:14:50.880 --> 00:14:51.839
Yeah.

s243
00:14:51.839 --> 00:14:55.120
The approval flow's too slow.

s244
00:14:55.960 --> 00:14:58.440
What kind of requests, though?

s245
00:14:58.440 --> 00:15:00.480
Expense approvals first.

s246
00:15:00.480 --> 00:15:03.400
Access requests eventually.

s247
00:15:09.760 --> 00:15:12.520
AI, hold that.

s248
00:15:12.520 --> 00:15:14.680
Actually, let's pause.

s249
00:15:14.680 --> 00:15:18.400
Expense approvals or a general approval workflow?

s250
00:15:18.400 --> 00:15:21.160
Expense approvals first release.

s251
00:15:21.160 --> 00:15:24.640
Access requests are future scope.

s252
00:15:25.520 --> 00:15:27.360
That changes the data model.

s253
00:15:27.360 --> 00:15:29.600
Good to know.

s254
00:15:31.040 --> 00:15:33.240
AI, pull that up for everyone.

s255
00:15:33.240 --> 00:15:36.320
Tracking determining who's the speaker is referring to is critical.

s256
00:15:36.320 --> 00:15:43.720
In this case, it was easy with a direct reference to the AI, but it will happen again without a direct reference.

s257
00:15:44.160 --> 00:15:46.640
Oh, I'd forgotten that was a rule.

s258
00:15:46.640 --> 00:15:49.720
So, over 5,000 needs a second approver?

s259
00:15:49.720 --> 00:15:50.600
Yep.

s260
00:15:50.600 --> 00:15:53.160
Manager plus finance.

s261
00:15:53.160 --> 00:15:56.400
So, a big one can't be a single tap.

s262
00:15:56.520 --> 00:15:57.320
Right.

s263
00:15:57.320 --> 00:16:00.320
Over the limit, it routes to a second approver.

s264
00:16:00.320 --> 00:16:01.080
Agreed.

s265
00:16:01.080 --> 00:16:03.480
Under five, one tap's fine.

s266
00:16:03.480 --> 00:16:05.920
Works for me.

s267
00:16:11.320 --> 00:16:12.360
Okay.

s268
00:16:12.360 --> 00:16:13.520
Agreed.

s269
00:16:13.520 --> 00:16:16.920
Expense approvals, 5,000 threshold.

s270
00:16:16.920 --> 00:16:19.760
Right there, the AI resolved the scope objective.

s271
00:16:19.760 --> 00:16:23.440
No one wrote the prompt, no one patched the turn and hit submit.

s272
00:16:23.440 --> 00:16:29.480
The system was in the conversation, following it, understanding, and choosing its moment.

s273
00:16:31.000 --> 00:16:34.120
AI, capture that for us.

s274
00:16:34.120 --> 00:16:36.440
First release supports expense approvals only.

s275
00:16:36.440 --> 00:16:39.080
Access requests are out of scope.

s276
00:16:39.080 --> 00:16:43.600
Managers can approve or reject an expense right from a notification.

s277
00:16:43.600 --> 00:16:49.560
And per the finance controls policy, anything over $5,000 routes to a second approver?

s278
00:16:58.880 --> 00:17:03.000
Actually, make the threshold 10,000, not five.

s279
00:17:03.000 --> 00:17:06.120
Want me to update the requirement to a 10,000 threshold?

s280
00:17:06.120 --> 00:17:08.120
Yes.

s281
00:17:09.480 --> 00:17:12.720
AI, is this room free after the meeting?

s282
00:17:12.720 --> 00:17:13.560
Let me check.

s283
00:17:13.560 --> 00:17:19.360
The room looks free until 3:00, but yes, it's yours until 3:00.

s284
00:17:19.439 --> 00:17:27.280
And that's the difference between a smart machine behind the same old prompt and an interface that finally participates.

s285
00:17:27.280 --> 00:17:29.640
Here's the mindset shift I want to leave you with.

s286
00:17:29.640 --> 00:17:33.160
AI is not just an intelligence technology.

s287
00:17:33.160 --> 00:17:36.240
It's increasingly becoming an interface technology.

s288
00:17:36.240 --> 00:17:41.040
And if so, then book smart models alone are not enough.

s289
00:17:41.040 --> 00:17:47.680
Stop picturing AI as a smarter machine hiding behind prompts, agents, loops, and all the old paradigms.

s290
00:17:47.680 --> 00:17:55.960
We have to start seeing intelligence itself as a thing that can finally remove interface constraints and amplify human potential.

s291
00:17:55.960 --> 00:18:00.040
For 75 years, humans adapted to the machine.

s292
00:18:00.040 --> 00:18:03.200
It's syntax, it's forms, it's timing, it's batch.

s293
00:18:03.200 --> 00:18:11.080
A system that can reason, listen, infer, adapt should be able to meet us partway, if not all the way.

s294
00:18:11.080 --> 00:18:16.280
Instead, if AI is for users, then we should obsess about maximizing the interface.

s295
00:18:16.280 --> 00:18:18.600
So, then the design question changes.

s296
00:18:18.600 --> 00:18:27.280
It needs to become what burden are we still putting on humans only because the machine used to be too limited to carry that burden itself.

s297
00:18:27.280 --> 00:18:31.240
Ask that question, and the whole interface space opens up.

s298
00:18:31.240 --> 00:18:33.440
The answer isn't always chat.

s299
00:18:33.440 --> 00:18:36.840
It isn't always voice, and not a wall of markdown.

s300
00:18:36.840 --> 00:18:41.000
It's definitely not a decade-old set of digital constructs.

s301
00:18:41.000 --> 00:18:45.720
The right answer is the affordance humans already use with each other.

s302
00:18:45.720 --> 00:18:53.240
Communication, a question, a pause, a sketch, a checklist, a quiet aside, or saying nothing at all.

s303
00:18:53.240 --> 00:18:58.360
An interface where timing and modality aren't the humans' job anymore.

s304
00:18:58.360 --> 00:19:02.920
Where choosing the right channel at the right moment is done by the AI.

s305
00:19:02.920 --> 00:19:07.320
And as a usability person, this is the part that excites me the most.

s306
00:19:07.320 --> 00:19:13.040
When you take that burden off people, the the disappears and adoption follows.

s307
00:19:13.040 --> 00:19:19.560
Computing has mostly been about to date improving how humans encode their intent for machines.

s308
00:19:19.560 --> 00:19:25.760
The punch card, type a command, click a menu, use your thumb on an iPhone, write a prompt.

s309
00:19:25.760 --> 00:19:30.520
Every step was progress and every step carried the old constraint forward into the next era.

s310
00:19:30.520 --> 00:19:35.520
A translation tax, a precision tax, context tax, repair tax.

s311
00:19:35.520 --> 00:19:38.160
AI is our chance to put those down.

s312
00:19:38.160 --> 00:19:40.200
Not by making everything magical.

s313
00:19:40.200 --> 00:19:41.920
Not by making everything voice.

s314
00:19:41.920 --> 00:19:49.760
Not by replacing human judgment, but by making computers for once more fluent with us.

s315
00:19:49.760 --> 00:19:53.960
Most talks and videos cover how to use or adopt AI.

s316
00:19:53.960 --> 00:19:57.880
The deeper question is how AI intelligence changes the interface.

s317
00:19:57.880 --> 00:20:01.120
Human conversation is the most human thing we do.

s318
00:20:01.120 --> 00:20:09.600
Because if a machine can finally understand more of what we mean, then we can and should stop reshaping ourselves to be understood by it.

s319
00:20:09.600 --> 00:20:11.800
Thank you.