WEBVTT

00:00:00.000 --> 00:00:05.211
When machines grow this capable, who decides how they get used?

00:00:05.211 --> 00:00:10.422
Looking across today's news, every story led me back to that question.

00:00:10.422 --> 00:00:15.634
A developer chose, on its own, who could touch its strongest model.

00:00:15.634 --> 00:00:20.845
An executive who warns of danger drew sharp fire from peers.

00:00:20.845 --> 00:00:26.056
The firm that lends money and the firm that sells hardware are becoming one.

00:00:26.056 --> 00:00:31.267
And research now hands even the checking of answers to machines.

00:00:31.267 --> 00:00:36.479
In pharma, comparing performance is no longer enough to pick a tool.

00:00:36.479 --> 00:00:41.690
Today, I ask who decides the use, and who does the checking.

00:00:41.690 --> 00:00:46.213
The maker narrowed who may use its most powerful product, by its own choice.

00:00:46.213 --> 00:00:50.736
Until now, the strongest models went out fastest and widest.

00:00:50.736 --> 00:00:55.259
This time, that order is reversed.Narrowing access means the developer admits that

00:00:55.259 --> 00:00:59.783
danger grows with capability.

00:00:59.783 --> 00:01:04.306
Security specialists got it first.

00:01:04.306 --> 00:01:08.829
I read that as a wish to measure misuse before wide release.

00:01:08.829 --> 00:01:13.352
Yet a looser edition is said to be coming.Let a select few try the main product

00:01:13.352 --> 00:01:17.875
before selling it widely.

00:01:17.875 --> 00:01:22.398
That suggests these tools now ask who is fit to hold them.

00:01:22.398 --> 00:01:26.922
Investors cheered.

00:01:26.922 --> 00:01:31.445
But critics already see faults.Safety limits are not the model itself.

00:01:31.445 --> 00:01:35.968
They are add-ons that can be attached or removed.

00:01:35.968 --> 00:01:40.491
So two models with one name may answer differently.

00:01:40.491 --> 00:01:45.014
If you use an outside tool, confirm which edition your contract covers.This weighs

00:01:45.014 --> 00:01:49.537
most on teams handling drug information or patient data.

00:01:49.537 --> 00:01:54.061
Benchmark scores alone are hard to defend later.

00:01:54.061 --> 00:01:58.584
Write down who received the model, what limits it carries, and why you chose it.

00:01:58.584 --> 00:02:03.107
But is restricting access for safety a view the whole industry shares?

00:02:03.107 --> 00:02:07.630
In fact, that very idea is under fierce attack.

00:02:07.630 --> 00:02:11.678
Behind the choice to hand a strong model to few hands lies deep concern about its

00:02:11.678 --> 00:02:15.727
dangers.

00:02:15.727 --> 00:02:19.775
Yet the executive who voiced that concern most loudly was cornered by industry

00:02:19.775 --> 00:02:23.824
heavyweights.The critics were a chip seller's chief and a scientist hailed as a

00:02:23.824 --> 00:02:27.872
pioneer of the field.

00:02:27.872 --> 00:02:31.921
Both say the alarm is overblown.

00:02:31.921 --> 00:02:35.969
The scientist goes further, saying the warner does not grasp defense.

00:02:35.969 --> 00:02:40.017
Some note the critics have business motives too.

00:02:40.017 --> 00:02:44.066
That it happened behind closed doors shows how deep the rift runs.What matters to me

00:02:44.066 --> 00:02:48.114
is not who is right.

00:02:48.114 --> 00:02:52.163
It is that builders weigh the danger very differently.

00:02:52.163 --> 00:02:56.211
That gap shows up as stricter or looser limits, and faster or slower releases.For a

00:02:56.211 --> 00:03:00.259
drugmaker, this becomes a choice of supplier philosophy.

00:03:00.259 --> 00:03:04.308
On the same task, a builder's safety stance changes what you may do, and when the

00:03:04.308 --> 00:03:08.356
tool refuses.

00:03:08.356 --> 00:03:12.405
Adopting a vendor means adopting its view of safety.

00:03:12.405 --> 00:03:16.453
Policy writers must now read each developer's stance.

00:03:16.453 --> 00:03:20.502
Behind this debate, who actually pays to run these systems?

00:03:20.502 --> 00:03:24.550
Let's set safety aside and follow the money.

00:03:24.550 --> 00:03:29.098
While the debate goes on, builders need vast processing power.

00:03:29.098 --> 00:03:33.646
Who pays for it decides who holds the upper hand in the industry.Here, a parts

00:03:33.646 --> 00:03:38.193
seller is lending money to its customer.

00:03:38.193 --> 00:03:42.741
The money comes straight back as rent for the seller's own hardware.

00:03:42.741 --> 00:03:47.289
Seller and lender are the same firm.

00:03:47.289 --> 00:03:51.837
In effect, it creates demand for its own product.That is a striking sum for a single

00:03:51.837 --> 00:03:56.384
borrower.

00:03:56.384 --> 00:04:00.932
It fits the view that the developer is racing for capacity above all.

00:04:00.932 --> 00:04:05.480
Seen the other way, the customer is too important for the seller to lose.Still,

00:04:05.480 --> 00:04:10.028
markets doubt this pattern of sellers propping up customers.

00:04:10.028 --> 00:04:14.576
If the customer stops growing, both the loan and the sales may go unpaid.

00:04:14.576 --> 00:04:19.123
Meanwhile, a maker of design software gained value on news of deals with big

00:04:19.123 --> 00:04:23.671
players.For pharma, the question is whether your tools stay available on the same

00:04:23.671 --> 00:04:28.219
terms.

00:04:28.219 --> 00:04:32.767
If a developer runs on borrowed money, prices and access can change suddenly.

00:04:32.767 --> 00:04:37.314
The longer you rely on a system, the more its funding matters.

00:04:37.314 --> 00:04:41.862
So where do drugmakers themselves spend their money on this technology?

00:04:41.862 --> 00:04:46.410
It went not to processing power, but to people.

00:04:46.420 --> 00:04:51.191
So far, we have looked at the builders' money and views.

00:04:51.191 --> 00:04:55.961
Now we turn to the users.

00:04:55.961 --> 00:05:00.732
What a user pays for shows where it sees the value.A domestic drugmaker set out to

00:05:00.732 --> 00:05:05.502
sharply shorten the road to testing in people.

00:05:05.502 --> 00:05:10.273
Its means is not new machinery.

00:05:10.273 --> 00:05:15.044
It is researchers who came with a business it bought.

00:05:15.044 --> 00:05:19.814
The judgment is clear: people unlock the technology's power.Cut that period in half,

00:05:19.814 --> 00:05:24.585
and the first leg of a medicine's journey to patients gets shorter.

00:05:24.585 --> 00:05:29.355
Costs climb with every extra month, so this matters to the firm.But the goal only

00:05:29.355 --> 00:05:34.126
covers the stage before human testing.

00:05:34.126 --> 00:05:38.896
How long the later studies take lies outside it.The only large figure came from

00:05:38.896 --> 00:05:43.667
start-up funding, reportedly raised.

00:05:43.667 --> 00:05:48.438
With no time frame or purpose given, it cannot gauge investment momentum.As a pharma

00:05:48.438 --> 00:05:53.208
practitioner, I believe adoption is not a tool purchase.

00:05:53.208 --> 00:05:57.979
Results depend on whom you gather, and what you entrust to them.

00:05:57.979 --> 00:06:02.749
So when those people check machine results, what should they watch for?

00:06:02.749 --> 00:06:07.520
Checking one model against another hides a subtle trap.

00:06:07.510 --> 00:06:12.130
A growing method has one model write problems and another solve them.

00:06:12.130 --> 00:06:16.750
It saves people from preparing correct answers.

00:06:16.750 --> 00:06:21.370
But when two models give the same reply, can we trust it?One paper studied setter

00:06:21.370 --> 00:06:25.990
and solver making the same mistake, so their answers matched.

00:06:25.990 --> 00:06:30.610
Over rounds of training, internal scores kept climbing.

00:06:30.610 --> 00:06:35.230
Yet accuracy judged from outside stayed flat or fell.The fix is simple.

00:06:35.230 --> 00:06:39.850
Keep the original material away from the judge.

00:06:39.850 --> 00:06:44.470
A judge who never saw it is less likely to repeat the setter's mistake.

00:06:44.470 --> 00:06:49.090
So fewer false passes get through.Re-asking many times while showing or hiding the

00:06:49.090 --> 00:06:53.710
material cut false passes only a little.

00:06:53.710 --> 00:06:58.330
What worked was not repetition but the checker's independence.

00:06:58.330 --> 00:07:02.950
Still, outside experts have not reviewed this study.

00:07:02.950 --> 00:07:07.570
And false passes did not vanish.I see here the same idea pharma quality assurance

00:07:07.570 --> 00:07:12.190
has long guarded.

00:07:12.190 --> 00:07:16.810
If maker and checker are not separate, shared blind spots go unseen.

00:07:16.810 --> 00:07:21.430
If one model grades its own replies to inquiries, the score proves nothing.

00:07:21.430 --> 00:07:26.050
Independent checking is now being tested in drafting submissions, and in exams that

00:07:26.050 --> 00:07:30.670
pick patients a medicine will help.

00:07:30.670 --> 00:07:35.173
Let's test today's story against pharma's everyday work.

00:07:35.173 --> 00:07:39.676
How is adoption actually unfolding on the ground?What stands out is buying whole

00:07:39.676 --> 00:07:44.178
teams of experts, not tools.

00:07:44.178 --> 00:07:48.681
A firm that writes submission dossiers was bought by another.

00:07:48.681 --> 00:07:53.184
In pre-dose testing too, a software company joined forces with a specialist partner.

00:07:53.184 --> 00:07:57.687
They plan to move together from checking to rollout.Submission dossiers need people

00:07:57.687 --> 00:08:02.189
who know both the rules and the medicine.

00:08:02.189 --> 00:08:06.692
Software can speed up drafts.

00:08:06.692 --> 00:08:11.195
But people still carry final responsibility for the content.

00:08:11.195 --> 00:08:15.698
That, I think, is why firms acquire people and software together.Testing before

00:08:15.698 --> 00:08:20.201
dosing is the first decision about which patient gets which drug.

00:08:20.201 --> 00:08:24.703
If machines enter here, you must be able to show later why they reached their call.

00:08:24.703 --> 00:08:29.206
This is where independent checking from the paper is needed.Market outlooks diverge

00:08:29.206 --> 00:08:33.709
sharply, but the talk has moved to real use.

00:08:33.709 --> 00:08:38.212
A big outlook number does not justify your own decisions.

00:08:38.212 --> 00:08:42.714
The real question is which task gets automated, and under whose responsibility.

00:08:42.714 --> 00:08:47.217
Finally, let's look at a quieter move behind the headlines: rule-making on who

00:08:47.217 --> 00:08:51.720
answers for how these tools are used.

00:08:51.720 --> 00:08:56.543
Flashy launches draw attention, and rules get pushed aside.

00:08:56.543 --> 00:09:01.365
Yet what binds pharma work directly is not performance.

00:09:01.365 --> 00:09:06.188
It is the rules.In America, top law officers from many states asked lawmakers for

00:09:06.188 --> 00:09:11.011
national rules, while keeping their own powers.

00:09:11.011 --> 00:09:15.833
Who polices these systems is still unsettled.

00:09:15.833 --> 00:09:20.656
Elsewhere, a court let machine output support a police search.I read this as machine

00:09:20.656 --> 00:09:25.478
conclusions starting to back public decisions.

00:09:25.478 --> 00:09:30.301
Builders' internal controls face questions too.

00:09:30.301 --> 00:09:35.124
One developer split with researchers over how secrets were handled.For the most

00:09:35.124 --> 00:09:39.946
advanced models, rules are being made in many places at once: nation, states,

00:09:39.946 --> 00:09:44.769
courts, and builders themselves.

00:09:44.769 --> 00:09:49.592
Drugmakers using them must revise internal policy each time outside rules change.For

00:09:49.592 --> 00:09:54.414
teams handling drug information, machine answers will one day be records that need

00:09:54.414 --> 00:09:59.237
explaining.

00:09:59.237 --> 00:10:04.059
Keep track of who used the tool, and who checked it.

00:10:04.059 --> 00:10:08.882
That is the best defense against outside rules.

00:10:08.882 --> 00:10:13.705
It also helps you explain your choices when asked.

00:10:13.705 --> 00:10:18.527
Today's news pointed past raw power, to how these tools are used and checked.

00:10:18.527 --> 00:10:23.350
Let me close by drawing the answer together.

00:10:23.350 --> 00:10:27.046
Today's answer is simple.

00:10:27.046 --> 00:10:30.741
When choosing a tool, look past raw performance.

00:10:30.741 --> 00:10:34.437
Ask who decides where it goes, who funds it, and who checks its answers.

00:10:34.437 --> 00:10:38.133
A powerful model has entered the world under tight limits.

00:10:38.133 --> 00:10:41.829
When, and on what grounds, will its maker loosen them?

00:10:41.829 --> 00:10:45.524
That is where users will find material for their own decisions.

00:10:45.524 --> 00:10:49.220
Tomorrow, I follow those clues.
