0:00
I want you to just uh imagine for a
0:02
second that you're walking into a
0:03
massive research laboratory.
0:05
>> It's like 3:00 am. The hallways are
0:07
completely pitch dark, but inside one of
0:09
the rooms, sitting at a lab bench, is a
0:12
scientist who is still working,
0:13
>> right? Burning the midnight oil.
0:15
>> Exactly. But imagine the scientist
0:17
literally never sleeps. They never take
0:19
a weekend off. and uh they happen to
0:22
have memorized every single scientific
0:25
paper, every genomic sequence and every
0:27
clinical trial outcome that has ever
0:30
been published in human history.
0:32
>> I mean, that's the dream, right? Right.
0:33
They've internalized all that data. And
0:36
because they can process it all
0:37
simultaneously, they're um they're
0:39
constantly spotting these invisible
0:41
connections that every other researcher
0:43
on the planet has missed,
0:44
>> Simply because no human can hold that
0:47
much information in their head at one
0:48
time. And that concept, that exact idea
0:51
is the core of our deep dive today.
0:54
We're exploring what is honestly a
0:55
monumental shift in medicine.
0:57
>> Yeah, it really is.
0:58
>> We're talking about the launch of GPT
1:00
Rosalin on April 16th, 2026. This is the
1:04
very first model in OpenAI's brand new
1:07
dedicated life sciences series.
1:10
>> It is. And to really break this down for
1:12
you, we've pulled together a pretty
1:13
extensive stack of sources. Yeah, we've
1:15
got the uh OpenAI official technical
1:18
announcement of course,
1:19
>> right? Along with in-depth industry
1:21
reports from Reuters, Fierce Biotech,
1:24
Venturebe, and we're cross-referencing
1:26
all of that with some really sobering
1:28
academic reviews from Jamon Tus.
1:30
>> Yeah, those TUS reports are key for
1:31
understanding the uh the stark economic
1:34
realities of drug development right now.
1:35
>> Totally. So, our mission today, our goal
1:38
for this deep dive is to give you a
1:40
really clear look under the hood. We are
1:42
going to decode the mechanics of how
1:44
this new AI actually does science
1:46
>> because it's not just generating text
1:48
>> Exactly. We'll look at why it has the
1:51
potential to drastically compress the
1:53
honestly agonizing timeline for new
1:55
medicines. And uh we'll also look at the
1:58
catch, the ethical and operational
2:00
hurdles that are keeping it from being
2:02
deployed everywhere immediately.
2:03
>> Right? Because there's always a catch.
2:05
>> Always. To give you a sense of the leap
2:07
we are taking here, think about how
2:09
traditional scientific research works
2:10
right now. It's like um reading the
2:14
entire internet, but you're stuck on a
2:16
1990s dialup connection.
2:18
>> Oh man, the sound of that dialup modem
2:20
just played in my head,
2:22
>> right? It takes forever to load one
2:24
page. GPT Rosland is like suddenly
2:26
getting upgraded to gigabit fiber optic.
2:29
You have all the information instantly.
2:31
>> That's a great way to frame it. But
2:33
before we get into the shiny mute tech,
2:35
we really have to understand the massive
2:37
problem it's trying to solve,
2:38
>> right? The bottleneck.
2:40
>> Yeah. The gap between gathering the
2:41
ingredients and actually cooking the
2:44
>> The status quo in drug discovery is um
2:47
it's incredibly grim.
2:48
>> Grim is a good word for it.
2:50
>> Right now, it takes an average of 10 to
2:52
15 years just to get a single drug from
2:54
the moment a biological target is
2:56
identified all the way through to
2:57
regulatory approval. 10 to 15 years,
3:00
which is just I mean an unimaginably
3:02
long time if you are a patient waiting
3:05
>> It's a grueling weight and the timeline
3:07
is really only half the problem. The
3:09
failure rate is brutal.
3:10
>> What are the numbers on that?
3:12
>> Well, according to the data from the
3:13
tough center for the study of drug
3:15
development, only about 10% of the
3:17
compounds that even make it into
3:19
pre-clinical trials actually survive the
3:21
process and reach the market.
3:23
>> Wow. 90% failure rate.
3:26
>> Yeah. And when you factor in the
3:27
financial weight of all those dead ends,
3:30
you know, the flawed hypotheses, the
3:32
clinical trials that fail in phase two
3:34
>> which are the really expensive ones.
3:36
>> Exactly. The capital cost of a single
3:37
successful drug is currently sitting
3:40
somewhere between 800 million and $2.6
3:44
>> Okay. Wait, 2.6 billion with a B
3:47
>> with a B up to $2.6 billion for one
3:51
approved drug. So traditional drug
3:53
discovery is basically like buying a $2
3:55
billion lottery ticket that takes 15
3:58
>> that captures the financial risk
4:00
perfectly. Yeah. You are placing these
4:02
massive bets on incredibly fragmented,
4:06
really noisy biological data.
4:08
>> And finding the signal in that noise is
4:11
exactly what inspired the namesake of
4:13
this new model. GPT Roselyn is named
4:16
after the British crystalallographer
4:18
>> Oh, right. She was the scientist whose
4:20
work with X-ray defraction. Um,
4:22
specifically that famous image photo 51,
4:26
>> That's what actually revealed the double
4:27
helix structure of DNA back in what
4:31
>> Exactly. And the thing about her work
4:33
was she was looking at what essentially
4:35
seems like a blurry, chaotic scattering
4:37
of X-rays to anyone else.
4:40
>> Just noise. But through this rigorous,
4:43
incredibly meticulous pattern
4:44
recognition, she deduced the fundamental
4:47
geometric structure of life itself,
4:49
>> which is amazing. Even though Watson,
4:51
Crick, and Wilkins usually get the
4:52
historical spotlight and the Nobel
4:54
>> Yeah, she didn't get the Nobel in her
4:56
lifetime, sadly. But her ability to see
4:58
the pattern in the noise was the
5:00
catalyst for all of it. And GPT Rosland
5:03
is attempting to apply that exact same
5:05
meticulous human pattern recognition,
5:07
but at a scale of billions of data
5:09
points. Okay, so let's dig into the
5:11
mechanics of how it actually does that.
5:13
Because reading the technical specs in
5:15
our sources, it's very clear this is a
5:17
completely different beast than the AI
5:19
models most of us are used to.
5:22
>> Open AAI labels it as a quote frontier
5:26
reasoning model. But I want to push back
5:28
on this a little bit for you listening
5:30
because I know what you might be
5:32
>> If I prompt this model about a specific
5:35
protein mutation, isn't it still just a
5:37
glorified search engine? Like, isn't it
5:39
just scraping a really dense Wikipedia
5:41
page or PubMed and summarizing the text
5:45
>> I'm so glad you brought that up because
5:46
this is probably the most common
5:48
misconception. A search engine performs
5:50
information retrieval,
5:51
>> right? It finds things.
5:53
>> Exactly. It finds a paper that says
5:55
protein X interacts with molecule Y and
5:57
it just shows you the paper.
5:59
>> But GPT Roslin performs multi-step
6:01
reasoning. It doesn't just read about a
6:03
protein. It actually synthesizes
6:05
evidence across thousands of multiomic
6:08
>> Okay, let's define that for a second.
6:10
>> yes, a bit of a mouthful.
6:12
>> It means it's looking at genomics, which
6:14
is the DNA. It's looking at proteomics,
6:16
the proteins themselves, and
6:17
metabolomics, which are the chemical
6:19
processes. So, it's looking at every
6:21
single layer of the biological all at
6:25
>> Precisely. And it processes all of those
6:28
layers simultaneously. So it analyzes
6:30
the 3D structure of a protein by
6:33
interpreting the underlying spatial
6:35
coordinates of its atoms
6:36
>> like a physical 3D map,
6:38
>> right? It models the physical
6:40
constraints and the energetic bonds in
6:42
its latent space and then it looks at
6:45
how a specific genetic mutation would
6:47
alter those bonds and signaling
6:50
>> And then it cross references that with
6:51
chemical reaction mechanisms. So it's
6:54
generating novel hypotheses that
6:55
literally do not exist in any published
6:57
>> Wait, really? It's coming up with
7:00
>> Yes, it's not just repeating what humans
7:02
have already discovered.
7:03
>> But how is a language model doing that?
7:05
Like something designed to predict the
7:07
next word in a sentence. How is it doing
7:09
spatial chemistry in biology?
7:11
>> Well, because biology at its core is a
7:14
>> Right. DNA is a sequence of letters.
7:17
Codes are sequences of amino acids. When
7:20
you train a massive neural network on
7:23
billions of these biological sequences,
7:26
it learns the underlying grammar of
7:29
>> The grammar of biology. I like that.
7:31
>> Yeah. It learns the rules of how an
7:32
amino acid chain will fold in physical
7:34
space. The same way an English language
7:36
model learns that a noun usually follows
7:39
>> That makes a lot of sense. So, it's not
7:40
translating English. It's translating
7:43
>> And our sources point to a specific tool
7:46
that OpenAI developed that makes this
7:48
actually actionable. Right. The Codex
7:50
Life Sciences plugin.
7:52
>> Yes, the Codeex plugin is what makes
7:54
this a real scientific tool. It acts as
7:56
an autonomous orchestrator.
7:58
>> What does that mean in practice?
7:59
>> It means it gives the reasoning model
8:02
direct live access to over 50 public
8:05
databases, scientific computing tools,
8:07
and literature repositories.
8:09
>> So, it's plugged directly into the
8:10
global scientific grid.
8:12
>> Right? So, the model doesn't just sit
8:13
there and generate a text response based
8:15
on its training data. It actively
8:17
designs a workflow. It essentially
8:19
becomes like an agent project manager.
8:21
>> Yes. If you ask it to investigate a
8:24
cellular pathway, it might write a
8:26
Python script to query a genomic
8:28
database, pull that raw data back into
8:31
its context window, parse it, and then
8:33
realize it needs more information on a
8:35
specific binding affinity.
8:37
>> So, it knows what it doesn't know.
8:38
>> Exactly. Then it pings a completely
8:40
different database for that chemical
8:42
data, synthesizes the two, and outputs a
8:44
complete step-by-step molecular cloning
8:47
protocol for the scientist to actually
8:49
follow in the lab. It actively
8:51
interrogates the digital landscape.
8:53
>> That is wild. But you know, we have this
8:55
brilliant theoretical engine, but
8:57
scientists are naturally and rightfully
8:59
pretty skeptical of tech industry hype.
9:01
>> Oh, scientists are the biggest skeptics.
9:03
They want to see the receipts,
9:04
>> right? They want to see it perform the
9:05
lab. And the sources do highlight some
9:08
benchmark data. Bixbench and Labbench 2.
9:12
>> Yeah, those are standard benchmarks that
9:14
test tasks like literature retrieval and
9:16
predicting RNA sequences.
9:17
>> And how did it do?
9:18
>> It completely outperformed all previous
9:21
versions of GPT across the board. But
9:23
honestly, the truly definitive proof
9:26
point comes from an evaluation they
9:27
conducted alongside a company called
9:31
>> Okay, what did they do?
9:32
>> They tested the model on sequenceto
9:33
function predictions. Meaning you feed
9:36
the AI a string of genetic letters and
9:38
it has to predict what the resulting
9:41
physical protein will actually do inside
9:44
>> Exactly. And when a human expert does
9:46
this, you know, a human researcher, they
9:49
look at a sequence mutation and use
9:51
their years of experience and heristic
9:52
rules to guess if that protein might
9:54
misfold or become overactive.
9:56
>> Is an educated guess,
9:57
>> right? The AI doesn't guess. It
9:59
calculates the probabilistic
10:01
interactions of the entire amino acid
10:03
chain. And in certain evaluations with
10:05
dynother therapeutics, Gassi Roslin's
10:08
predictions actually exceeded the 95th
10:10
percentile of human experts.
10:11
>> Wait, it beat the top 5% of human
10:14
specialists in predicting biological
10:16
>> Wow. Well, that certainly explains why
10:19
the adoption list in the Reuters report
10:20
is a who's who of biotech. Amgen, Madna,
10:24
the Allen Institute, Termoffisher
10:26
Scientific, Los Alamos National
10:28
Laboratory, they are all already
10:30
collaborating to integrate this into
10:33
>> Yeah. The heavy hitters are all in
10:35
because they recognize that this is the
10:37
necessary evolution of what started with
10:40
>> Oh, right. Alphafold from Deep Mind.
10:42
>> Yeah. AlphaFold was a massive
10:44
breakthrough because it solved the
10:45
protein folding problem. It essentially
10:48
gave us the static 3D map of the biology
10:51
>> Exactly. GPT Roslin takes that map and
10:54
builds an interactive physics and
10:56
reasoning engine on top of it.
10:57
>> So, it makes it dynamic.
10:59
>> Yes. It takes those protein structures
11:01
and figures out how diseases hijack them
11:03
and then what specific new molecules
11:06
>> And the sources show that companies
11:07
taking this what they call an AI native
11:09
approach. So building their research
11:11
pipelines around machine learning from
11:13
day one, they are seeing incredible
11:16
>> The numbers are really shifting.
11:18
>> Yeah. The fierce biotech report noted
11:20
that AI native companies are hitting 80
11:22
to 90% success rates in phase clinical
11:26
>> which is just a staggering leap,
11:28
>> especially when you remember the
11:30
historical success rate for phase is
11:32
somewhere between 40 and 65%.
11:34
>> You are dramatically derisking that
11:37
massive initial investment we talked
11:40
>> You aren't wasting a billion dollars on
11:42
a compound that's doomed to fail in year
11:44
>> right? But, you know, when you hear
11:46
stats like beating the 95th percentile
11:49
of human experts, it's really easy to
11:51
jump to the conclusion that human
11:52
scientists are just being replaced
11:55
>> The sci-fi dystopian view. Yeah,
11:56
>> exactly. But based on how this actually
11:59
operates, it seems less about replacing
12:01
the human and more like an Iron Man suit
12:05
>> Oh, I like that analogy.
12:06
>> Right. The AI is doing the heavy
12:08
lifting. It's calculating the
12:09
navigation, drawing up a flawlessly
12:12
calculated molecular blueprint that will
12:14
hold weight, but the human is still
12:16
inside driving it. They still have to go
12:18
to the construction site and pour the
12:21
>> That is exactly it. The human remains
12:23
firmly in the loop. The AI accelerates
12:25
the synthesis of evidence and the
12:27
hypothesis generation, but the
12:29
creativity required to ask the right
12:31
initial questions that remains entirely
12:34
human. And more importantly, the actual
12:37
physical wet lab validation is purely
12:40
>> Right? The AI cannot hold a pipet.
12:43
>> It can't run a cell culture or observe
12:46
>> No, it can't. The physical reality of
12:48
experimental science still happens at
12:50
the bench. GPT Roslin simply ensures
12:52
that when a scientist spends 6 months
12:54
running an experiment, they are testing
12:56
a hypothesis that has an incredibly high
12:58
mathematical probability of working.
13:01
>> It's focusing their effort.
13:03
>> Which brings us to the operational
13:04
reality. the big catch in all of this
13:06
because if this model is cutting years
13:08
off the development timeline and beating
13:10
expert benchmarks, why isn't every
13:12
university, every regional hospital, and
13:14
every independent lab in the world using
13:15
it right this second?
13:17
>> Because the sheer capabilities of this
13:19
system introduce some pretty severe
13:21
deployment hurdles. Right now, OpenAI
13:25
has GC Rosland locked down in a uh quote
13:29
research preview under a strict trusted
13:31
access program. Okay. So, it's not an
13:33
open API anyone can just plug into.
13:35
>> Definitely not. That limits usage
13:38
primarily to qualified US companies and
13:41
institutions that have proven robust
13:43
governance structures in place.
13:45
>> That's a highly restricted VIP list,
13:47
which obviously slows down broad
13:49
academic or global adoption, right?
13:51
Especially for developing nations that
13:53
could really use this for localized
13:55
>> It does slow things down, but the slow
13:57
roll out is driven by two main factors.
13:59
Technical limitations and security.
14:01
Let's talk about the technical side
14:03
>> So, technically, the model still
14:04
struggles with what they call context
14:06
>> What does that mean?
14:08
>> Well, if a scientist asks it to execute
14:10
a massive like 50step reasoning workflow
14:13
that requires pinging 20 different
14:14
external databases, the model can
14:16
sometimes lose the thread. It might
14:18
hallucinate connections.
14:19
>> Oh, right. AI hallucinations.
14:21
>> Yeah. So, it requires scientists who are
14:23
highly skilled in prompt engineering and
14:26
data validation to actually guide it and
14:28
doublech checkck ite work. So it
14:30
requires a massive upskilling of the
14:32
current scientific workforce.
14:34
>> It really does. You can't just hand this
14:36
to an undergrad and expect a miracle
14:38
cure. But the primary reason for the
14:40
lockdown access program isn't technical.
14:45
>> Yes. Biology is inherently dual use.
14:48
>> Meaning it can be used for good or for
14:50
>> Exactly. The exact same deep reasoning
14:53
capabilities that allow this AI to
14:55
analyze a viral genome and design a
14:58
life-saving targeted vaccine could
15:00
theoretically be inverted
15:02
>> to design a novel pathogen.
15:03
>> Right? A highly lethal, completely novel
15:07
>> Man, it's the ultimate double-edged
15:09
sword. If the model understands the
15:10
grammar of biology well enough to build
15:12
a cure, it understands it well enough to
15:15
>> Which is exactly why organizations like
15:17
Los Alamos's National Laboratory are
15:19
involved in the validation process. You
15:21
simply cannot deploy an open- source
15:22
tool that might accidentally provide a
15:24
step-by-step molecular cloning protocol
15:27
for synthetic virus.
15:29
>> Yeah, that would be catastrophic.
15:30
>> Open AAI has implemented really heavy
15:33
safeguards. And the trusted access
15:35
program is designed to stress test those
15:37
guard rails before they even think about
15:39
any broader release.
15:41
>> That makes total sense. But it also
15:43
brings up a massive economic question
15:44
for you listening right now.
15:46
>> Right. The equity question.
15:47
>> Exactly. If this technology successfully
15:50
slashes the cost of pharmaceutical R&D
15:52
by billions of dollars, like the TUS
15:54
data showed, and it cuts a 15-year wait
15:57
time in half, where does that value
16:01
>> That is the billion dollar question.
16:03
>> Are we actually going to see cheaper,
16:05
more accessible prescriptions at the
16:07
pharmacy counter? Or does a $2 billion
16:10
savings just become higher profit
16:12
margins for the few big pharma companies
16:14
that hold the keys to this AI? Yeah,
16:16
because the technology fundamentally
16:18
changes the speed of discovery, but it
16:20
doesn't automatically rewrite the
16:21
economics of healthcare.
16:22
>> The system is still the system.
16:24
>> Exactly. Whether those massive cost
16:26
savings are passed down to patients
16:28
depends entirely on how the industry
16:30
operates in the coming years.
16:32
Regulators, academia, and the
16:34
pharmaceutical industry are going to
16:35
have to navigate how this value is
16:37
distributed as the technology matures.
16:40
>> And let's be impartial here. There are a
16:42
lot of different viewpoints on how that
16:44
should happen. Some argue for heavy
16:46
regulation to force prices down, while
16:48
others argue that the market will
16:50
naturally lower prices through increased
16:52
competition. We aren't endorsing any
16:54
specific political solution, but it's
16:56
clear from the reports that this is
16:58
going to be a defining conversation for
17:00
the industry over the next decade.
17:02
>> Oh, absolutely. The technology is here,
17:04
but the policy is lagging behind.
17:06
>> So, to bring all of this together for
17:08
you today, GPT Rosland is an incredibly
17:10
fitting tribute to Rosalyn Franklin. She
17:12
possessed this rare ability to see the
17:15
fundamental truth hidden inside noisy,
17:18
>> And that's exactly what this model is
17:21
>> Right? It targets that 10 to 15ear
17:24
bottleneck in drug discovery, not by
17:26
replacing scientists, but by acting as
17:28
the ultimate amplifier for human
17:30
curiosity. It cuts through the noise of
17:33
millions of data points so scientists
17:35
can focus on what actually matters,
17:37
curing disease. We are really stepping
17:40
away from an era of brute force trial
17:42
and error. We're moving into an era of
17:44
accelerated precisiong guided reasoning.
17:47
>> And it's wild to think this model is
17:48
merely the first iteration in the life
17:52
>> Yeah, this is just version one.
17:53
>> As these systems scale, their capacity
17:55
for complex biochemical reasoning will
17:58
>> We are watching the foundation being
18:00
poured for the next century of medicine.
18:02
And we really want to hear your
18:03
perspective on this. Reach out to us on
18:05
social media. Do you think AI discovered
18:07
drugs will actually lead to cheaper,
18:09
faster treatments for you and your
18:11
family in the near future? Or do you
18:13
think the economic realities of the
18:14
pharmaceutical industry will just absorb
18:17
>> It is going to be genuinely fascinating
18:19
to see how the market and the science
18:21
evolve together on this.
18:22
>> Absolutely. And I want to leave you with
18:24
one final kind of provocative thought to
18:27
>> Okay, let's hear it.
18:28
>> We spent this entire deep dive talking
18:30
about how this AI can help us cure the
18:33
diseases we already have, right? cancer,
18:36
Alzheimer's, rare genetic conditions,
18:39
>> right? Reactive medicine.
18:40
>> Exactly. But if models like GPT Rosalind
18:43
eventually become perfectly adept at
18:45
understanding human biology and
18:47
absolutely flawless at predicting
18:49
exactly how molecules interact,
18:51
>> which is where this is heading.
18:52
>> Will the future of medicine shift
18:54
entirely? Will we move away from
18:56
reacting to diseases that have already
18:58
made us sick and instead use AI to
19:01
design proactive personalized immunities
19:03
for diseases before they even emerge in
19:06
>> Oh wow. Engineering our own resistance
19:09
before the threat ever even arrives.
19:10
>> Think back to that tireless scientist
19:13
sitting in the dark lab we talked about
19:15
>> Imagine them not just working to cure
19:17
your illness, but building the
19:18
biological blueprint to ensure you never
19:20
get sick in the first place. That is the
19:23
true potential of the master chef in the
