WEBVTT

00:00:05.000 --> 00:00:06.980
Hi, I’m Alice from Mevanu.

00:00:07.560 --> 00:00:12.020
Last Friday, we asked what AI transparency
should mean in everyday work.

00:00:12.580 --> 00:00:17.285
This week, we’re asking a more specific
question: what would prove that an AI

00:00:17.300 --> 00:00:19.240
assistant learned from your work?

00:00:19.740 --> 00:00:24.620
By learning here, we mean using what
happened in one task to guide later work.

00:00:25.140 --> 00:00:29.480
Imagine you tell an assistant, “Next time,
check the source before you write.”

00:00:30.140 --> 00:00:32.280
Saving that sentence is only the
beginning.

00:00:32.680 --> 00:00:35.340
The interesting part comes with the next
assignment.

00:00:35.920 --> 00:00:37.660
Does the assistant use it appropriately?

00:00:38.280 --> 00:00:39.460
Can you see what happened?

00:00:39.920 --> 00:00:41.520
And can you change your mind?

00:00:42.120 --> 00:00:45.680
This week, our development work connected
parts of that process.

00:00:46.260 --> 00:00:50.525
First, references to supplied documents
passed through reading and correction

00:00:50.540 --> 00:00:54.280
checks into a recorded result in a
controlled local test.

00:00:54.700 --> 00:00:58.825
The intended benefit is practical: keeping
track of the material behind an

00:00:58.840 --> 00:01:02.060
answer, including documents that don’t
have a website address.

00:01:02.540 --> 00:01:06.240
The candidate is still unfinished and
awaiting independent review.

00:01:06.760 --> 00:01:10.840
Second, a candidate working method reached
a later controlled task.

00:01:11.560 --> 00:01:15.060
Local checks also covered withdrawing it
and rolling back its use.

00:01:15.380 --> 00:01:19.585
That matters because a useful habit for
one situation can be the wrong habit

00:01:19.600 --> 00:01:23.440
for another. Your latest direction should
still govern the work.

00:01:23.880 --> 00:01:25.525
Here is the important qualification.

00:01:25.540 --> 00:01:28.960
The later test kept the original behavior.

00:01:29.760 --> 00:01:31.580
Improved outcomes were not measured.

00:01:32.080 --> 00:01:36.325
We have connected parts of a development
process; we have not demonstrated the

00:01:36.340 --> 00:01:40.540
complete everyday experience or shown that
it makes Brava more useful.

00:01:41.020 --> 00:01:43.680
So the next test has to look at the whole
job.

00:01:44.220 --> 00:01:45.600
Which information was used?

00:01:45.840 --> 00:01:47.100
Which method was allowed?

00:01:47.520 --> 00:01:48.980
What did the assistant do?

00:01:49.340 --> 00:01:51.100
What result did the person accept?

00:01:51.520 --> 00:01:54.680
How much setup, checking and correction
did that take?

00:01:55.020 --> 00:01:59.265
The next build milestone is to complete
the document-handling path, finish the

00:01:59.280 --> 00:02:03.080
affected checks and put the frozen
candidate through independent review.

00:02:03.440 --> 00:02:06.940
The everyday workflow and its outcomes
still need testing.

00:02:07.360 --> 00:02:08.780
What would count as learning?

00:02:09.220 --> 00:02:13.085
Evidence that an assistant used a relevant
lesson in later work, that you

00:02:13.100 --> 00:02:16.380
could correct or withdraw it, and that the
result helped.

00:02:16.760 --> 00:02:19.260
We can show progress on parts of that
chain.

00:02:19.600 --> 00:02:23.865
We can’t honestly claim the full answer
until the everyday journey and its

00:02:23.880 --> 00:02:25.300
outcomes have been tested.

00:02:25.740 --> 00:02:29.580
Have an idea? Drop a prompt, comment,
suggestion, or question.

00:02:30.060 --> 00:02:31.205
Mevanu. Just.

00:02:31.220 --> 00:02:31.940
Smarter.

