WEBVTT

00:00:00.000 --> 00:00:04.309
On the same day, two moves pointed in opposite directions.

00:00:04.309 --> 00:00:08.618
A US agency began examining the people who build AI.

00:00:08.618 --> 00:00:12.927
At the same time, the government and those same builders agreed to police

00:00:12.927 --> 00:00:17.236
themselves.

00:00:17.236 --> 00:00:21.545
Will rules bind from outside, or will firms be trusted to guard from within?

00:00:21.545 --> 00:00:25.855
Nobody knows yet which path will win.

00:00:25.855 --> 00:00:30.164
Meanwhile, AI work in pharma grows every day.

00:00:30.164 --> 00:00:34.473
Users cannot wait for the rules to settle.

00:00:34.473 --> 00:00:38.782
On what basis should they choose tools, and how much should they hand over?

00:00:38.782 --> 00:00:43.091
I look for answers in shifting rules and money flows.

00:00:43.091 --> 00:00:47.400
I also look at the code under the tools, and at how records are kept.

00:00:47.400 --> 00:00:51.575
This morning's top story: AI builders are now the ones under scrutiny.

00:00:51.575 --> 00:00:55.750
A government body has started asking whether their tools hurt the people who use

00:00:55.750 --> 00:00:59.925
them.

00:00:59.925 --> 00:01:04.100
That matters for users like us, too.This changes who judges the danger.

00:01:04.100 --> 00:01:08.275
Until now, builders described the dangers in their own words.

00:01:08.275 --> 00:01:12.450
From here, an outside body gathers evidence and decides.

00:01:12.450 --> 00:01:16.625
And the worry is not a distant future.

00:01:16.625 --> 00:01:20.800
It is harm to people using these tools today.

00:01:20.800 --> 00:01:24.975
Tools called clever must now pass the safety test any product faces.

00:01:24.975 --> 00:01:29.150
Many outlets ran the story at once, so I expect it to last.This body's original job

00:01:29.150 --> 00:01:33.325
is to protect buyers.

00:01:33.325 --> 00:01:37.500
With it stepping in, AI danger becomes a question about the deal between maker and

00:01:37.500 --> 00:01:41.675
user.

00:01:41.675 --> 00:01:45.850
I think the era of simply trusting the builders' goodwill is nearing its end.For

00:01:45.850 --> 00:01:50.025
pharma, this is not a faraway story.

00:01:50.025 --> 00:01:54.200
If an in-house tool's maker comes under review, your terms could change midstream.

00:01:54.200 --> 00:01:58.375
Which tool, for which task, checked in which way?

00:01:58.375 --> 00:02:02.550
Companies that keep that record stay calm when terms shift.

00:02:02.550 --> 00:02:06.725
Yet on the very day the review began, the builders were elsewhere, promising to keep

00:02:06.725 --> 00:02:10.900
themselves in check.

00:02:10.900 --> 00:02:14.736
So on one day, the government set out two opposite approaches.

00:02:14.736 --> 00:02:18.571
One restrains from outside.

00:02:18.571 --> 00:02:22.407
The other asks firms to hold back from within.

00:02:22.407 --> 00:02:26.243
Which one leads will shape how users prepare.Side by side, the gap is about

00:02:26.243 --> 00:02:30.078
penalties.

00:02:30.078 --> 00:02:33.914
The signed promise does not say what happens if broken, so people question its

00:02:33.914 --> 00:02:37.750
force.

00:02:37.750 --> 00:02:41.585
The review, by contrast, can end in enforcement.

00:02:41.585 --> 00:02:45.421
Running a weak promise and a real review together suggests the government has not

00:02:45.421 --> 00:02:49.257
chosen a course.

00:02:49.257 --> 00:02:53.092
The signed paper even carried a basic error, which undercut it on day one.Those who

00:02:53.092 --> 00:02:56.928
built AI also disagree.

00:02:56.928 --> 00:03:00.763
A veteran researcher warned against taking the danger of our species dying out

00:03:00.763 --> 00:03:04.599
lightly.

00:03:04.599 --> 00:03:08.435
The head of a chipmaker told worried labs they could simply quit.

00:03:08.435 --> 00:03:12.270
If builders cannot agree on the danger, each signer's promise will mean something

00:03:12.270 --> 00:03:16.106
different.Pharma has long relied on both law and industry codes to keep its conduct

00:03:16.106 --> 00:03:19.942
in line.

00:03:19.942 --> 00:03:23.777
People in the field know neither is enough alone.

00:03:23.777 --> 00:03:27.613
So companies set the limits for in-house AI use now, without waiting for outside

00:03:27.613 --> 00:03:31.449
rules.

00:03:31.449 --> 00:03:35.284
While rules wavered between promise and review, nations were also drawing lines in

00:03:35.284 --> 00:03:39.120
the code that powers AI.

00:03:39.120 --> 00:03:43.482
Now we move from rules to what is inside the tools.

00:03:43.482 --> 00:03:47.843
The rivalry between nations now reaches past visible hardware.

00:03:47.843 --> 00:03:52.205
It extends to the invisible code that makes that hardware run.The target is not the

00:03:52.205 --> 00:03:56.566
hardware itself.

00:03:56.566 --> 00:04:00.928
A processor only performs well when paired with the code that runs on it.

00:04:00.928 --> 00:04:05.289
Reports say one firm's lead lies partly in that code.

00:04:05.289 --> 00:04:09.651
I think that is because so many developers are used to it.

00:04:09.651 --> 00:04:14.013
Rebuild the same code in a form anyone can use, and moving to other hardware gets

00:04:14.013 --> 00:04:18.374
easier.

00:04:18.374 --> 00:04:22.736
The partners want to cut that effort and stop depending on a single supplier.That

00:04:22.736 --> 00:04:27.097
strength did not come from speed alone.

00:04:27.097 --> 00:04:31.459
Over many years, the company drew developers' habits and their stored programs to

00:04:31.459 --> 00:04:35.821
its side.

00:04:35.821 --> 00:04:40.182
If that bond weakens, the reasons for picking one processor over another change

00:04:40.182 --> 00:04:44.544
too.Pharma research and analysis run on the same code and hardware.

00:04:44.544 --> 00:04:48.905
If these split by nation, trade rules may choose your tools instead of performance.

00:04:48.905 --> 00:04:53.267
Firms running the same analyses at overseas sites are the most exposed.

00:04:53.267 --> 00:04:57.628
As nations draw lines in code, large sums aimed at health care also moved, outside

00:04:57.628 --> 00:05:01.990
the drug companies.

00:05:01.990 --> 00:05:06.209
Where money goes shows where people expect growth next.

00:05:06.209 --> 00:05:10.428
Today's largest sum went not into a drug company, but into a new group formed

00:05:10.428 --> 00:05:14.647
outside one.Its founders worked on AI at drug makers and at big technology firms.

00:05:14.647 --> 00:05:18.866
They set up a health-focused research group outside any one company.

00:05:18.866 --> 00:05:23.085
So drug makers are no longer the only drivers of medical AI.

00:05:23.085 --> 00:05:27.304
But who is paying, and for what, is not clear from this report.

00:05:27.304 --> 00:05:31.523
Nor do we know whether it will team up with drug makers.

00:05:31.523 --> 00:05:35.742
The size arrived before the substance.Also, the sum is a pledge to pay, not money

00:05:35.742 --> 00:05:39.961
already paid.

00:05:39.961 --> 00:05:44.179
Pledged amounts can shrink as conditions change.

00:05:44.179 --> 00:05:48.398
I think it is too early to say money is pouring into this field based on a headline

00:05:48.398 --> 00:05:52.617
figure.

00:05:52.617 --> 00:05:56.836
Whether the pledge is real will show only when payment records appear.Pharma

00:05:56.836 --> 00:06:01.055
professionals see daily the gap between announcing a development plan and winning

00:06:01.055 --> 00:06:05.274
approval.

00:06:05.274 --> 00:06:09.493
Funding promises need the same caution.

00:06:09.493 --> 00:06:13.712
Check who is paying, and when it was paid, before treating them as a partner.

00:06:13.712 --> 00:06:17.931
From money, we turn to AI that carries out actions itself.

00:06:17.931 --> 00:06:22.150
There, a danger that is easy to miss was waiting.

00:06:22.150 --> 00:06:25.930
AI that acts on its own is entering the workplace.

00:06:25.930 --> 00:06:29.710
The question is how to tell which action is dangerous, when, and by looking at

00:06:29.710 --> 00:06:33.490
what.The path the study describes is simple.

00:06:33.490 --> 00:06:37.270
The first request widens who can view a document.

00:06:37.270 --> 00:06:41.050
The next makes an ordinary duplicate.

00:06:41.050 --> 00:06:44.830
Each request alone is harmless.

00:06:44.830 --> 00:06:48.610
Together, though, an internal draft becomes readable from outside.

00:06:48.610 --> 00:06:52.390
The danger was never in the words.

00:06:52.390 --> 00:06:56.170
It sat in the condition left behind by the earlier action.A defense that reads only

00:06:56.170 --> 00:06:59.950
each request's words cannot stop this.

00:06:59.950 --> 00:07:03.730
Taken one by one, no request gives a reason to halt.

00:07:03.730 --> 00:07:07.510
Look at the current settings just before acting, and the overlap shows.

00:07:07.510 --> 00:07:11.290
The defender's focus shifts from words to conditions.

00:07:11.290 --> 00:07:15.070
But experts have not yet reviewed this work.

00:07:15.070 --> 00:07:18.850
Others still need to test it.Pharma holds many documents that must never leave the

00:07:18.850 --> 00:07:22.630
company.

00:07:22.630 --> 00:07:26.410
Think of pre-approval files and patient records.

00:07:26.410 --> 00:07:30.190
If AI acts for you, something must check the viewing settings before each action.

00:07:30.190 --> 00:07:33.970
Reading logs afterward only reveals a leak once it has happened.

00:07:33.970 --> 00:07:37.750
The idea of checking conditions also links to efforts to record where AI output

00:07:37.750 --> 00:07:41.530
comes from.

00:07:41.530 --> 00:07:45.310
Those efforts are happening right beside our own work.

00:07:45.310 --> 00:07:48.980
Things AI produces often cannot be traced later.

00:07:48.980 --> 00:07:52.649
Who made them, and how?

00:07:52.649 --> 00:07:56.319
Where and how to keep that origin is the next question.One effort places a mark

00:07:56.319 --> 00:07:59.988
inside a substance that AI designed.

00:07:59.988 --> 00:08:03.658
If the mark goes in without harming how the substance works, its origin can be

00:08:03.658 --> 00:08:07.327
traced.

00:08:07.327 --> 00:08:10.997
The other is a drug maker tidying its core records before spreading AI across the

00:08:10.997 --> 00:08:14.666
company.

00:08:14.666 --> 00:08:18.336
It chose to fix the data underneath before the models on top.

00:08:18.336 --> 00:08:22.005
Both aim to keep outputs and inputs explainable later.A mark helps only if it leaves

00:08:22.005 --> 00:08:25.675
the contents working as before.

00:08:25.675 --> 00:08:29.344
If it weakened the result, nobody would use it.

00:08:29.344 --> 00:08:33.014
Showing that it need not do so makes origin tracking a more practical choice for

00:08:33.014 --> 00:08:36.683
users.Pharma has always earned trust through records.

00:08:36.683 --> 00:08:40.352
If you cannot show who checked what, and when, a submission will not be accepted.

00:08:40.352 --> 00:08:44.022
That principle holds with AI too.

00:08:44.022 --> 00:08:47.692
How complete your origin records and core data are will set how much work you can

00:08:47.692 --> 00:08:51.361
hand to AI.

00:08:51.361 --> 00:08:55.030
Apart from tidying origins and records, though, some questions still have no answer

00:08:55.030 --> 00:08:58.700
from anyone.

00:08:58.700 --> 00:09:03.717
Outside the big headlines, several things remain unsettled.

00:09:03.717 --> 00:09:08.735
Listing what is unknown also shows what today's big stories are missing.One major

00:09:08.735 --> 00:09:13.752
company head is expected to stay away when lawmakers question risky AI.

00:09:13.752 --> 00:09:18.769
Avoiding that while under review could shake trust in the company.

00:09:18.769 --> 00:09:23.787
A company under scrutiny needs to explain itself in public.

00:09:23.787 --> 00:09:28.804
Another statute, said to apply from today, was also reported.

00:09:28.804 --> 00:09:33.821
Which one, and what it says, is still unclear.

00:09:33.821 --> 00:09:38.839
One developer claimed a rival nation's well-known system can mount strong attacks.

00:09:38.839 --> 00:09:43.856
Yet here, no outside party has confirmed that claim.

00:09:43.856 --> 00:09:48.873
Claims about a competitor are safer left out of decisions until evidence

00:09:48.873 --> 00:09:53.891
appears.Recording the unknown as unknown is also a core skill in practice.

00:09:53.891 --> 00:09:58.908
Keep verified and unverified reports apart, and later judgments go wrong less often.

00:09:58.908 --> 00:10:03.925
Companies that note open questions can update internal rules fast once a rule's

00:10:03.925 --> 00:10:08.943
content is clear.

00:10:08.943 --> 00:10:13.960
So, with the certain and the uncertain kept apart, what can users actually rely on?

00:10:13.970 --> 00:10:17.741
What users can rely on is not a builder's promise.

00:10:17.741 --> 00:10:21.513
It is the records they can check for themselves.

00:10:21.513 --> 00:10:25.284
When outside rules shift, firms that can verify their tools' condition and origin

00:10:25.284 --> 00:10:29.056
stay calm.

00:10:29.056 --> 00:10:32.827
Tomorrow, we watch how the firms under review respond.

00:10:32.827 --> 00:10:36.599
Will they answer with words, or with records?

00:10:36.599 --> 00:10:40.370
That difference is a first clue to how serious the promises are.
