GPT Rosalind | OpenAI’s Frontier Model for Drug Discovery acceleration

I want you to just uh imagine for a

second that you're walking into a

massive research laboratory.

>> Okay?

>> It's like 3:00 am. The hallways are

completely pitch dark, but inside one of

the rooms, sitting at a lab bench, is a

scientist who is still working,

>> right? Burning the midnight oil.

>> Exactly. But imagine the scientist

literally never sleeps. They never take

a weekend off. and uh they happen to

have memorized every single scientific

paper, every genomic sequence and every

clinical trial outcome that has ever

been published in human history.

>> I mean, that's the dream, right? Right.

They've internalized all that data. And

because they can process it all

simultaneously, they're um they're

constantly spotting these invisible

connections that every other researcher

on the planet has missed,

>> right?

>> Simply because no human can hold that

much information in their head at one

time. And that concept, that exact idea

is the core of our deep dive today.

We're exploring what is honestly a

monumental shift in medicine.

>> Yeah, it really is.

>> We're talking about the launch of GPT

Rosalin on April 16th, 2026. This is the

very first model in OpenAI's brand new

dedicated life sciences series.

>> It's a huge deal.

>> It is. And to really break this down for

you, we've pulled together a pretty

extensive stack of sources. Yeah, we've

got the uh OpenAI official technical

announcement of course,

>> right? Along with in-depth industry

reports from Reuters, Fierce Biotech,

Venturebe, and we're cross-referencing

all of that with some really sobering

academic reviews from Jamon Tus.

>> Yeah, those TUS reports are key for

understanding the uh the stark economic

realities of drug development right now.

>> Totally. So, our mission today, our goal

for this deep dive is to give you a

really clear look under the hood. We are

going to decode the mechanics of how

this new AI actually does science

>> because it's not just generating text

anymore.

>> Exactly. We'll look at why it has the

potential to drastically compress the

honestly agonizing timeline for new

medicines. And uh we'll also look at the

catch, the ethical and operational

hurdles that are keeping it from being

deployed everywhere immediately.

>> Right? Because there's always a catch.

>> Always. To give you a sense of the leap

we are taking here, think about how

traditional scientific research works

right now. It's like um reading the

entire internet, but you're stuck on a

1990s dialup connection.

>> Oh man, the sound of that dialup modem

just played in my head,

>> right? It takes forever to load one

page. GPT Rosland is like suddenly

getting upgraded to gigabit fiber optic.

You have all the information instantly.

>> That's a great way to frame it. But

before we get into the shiny mute tech,

we really have to understand the massive

problem it's trying to solve,

>> right? The bottleneck.

>> Yeah. The gap between gathering the

ingredients and actually cooking the

meal.

>> Yeah.

>> The status quo in drug discovery is um

it's incredibly grim.

>> Grim is a good word for it.

>> Right now, it takes an average of 10 to

15 years just to get a single drug from

the moment a biological target is

identified all the way through to

regulatory approval. 10 to 15 years,

which is just I mean an unimaginably

long time if you are a patient waiting

for a treatment.

>> It's a grueling weight and the timeline

is really only half the problem. The

failure rate is brutal.

>> What are the numbers on that?

>> Well, according to the data from the

tough center for the study of drug

development, only about 10% of the

compounds that even make it into

pre-clinical trials actually survive the

process and reach the market.

>> Wow. 90% failure rate.

>> Yeah. And when you factor in the

financial weight of all those dead ends,

you know, the flawed hypotheses, the

clinical trials that fail in phase two

or three,

>> which are the really expensive ones.

>> Exactly. The capital cost of a single

successful drug is currently sitting

somewhere between 800 million and $2.6

billion.

>> Okay. Wait, 2.6 billion with a B

>> with a B up to $2.6 billion for one

approved drug. So traditional drug

discovery is basically like buying a $2

billion lottery ticket that takes 15

years to scratch off

>> that captures the financial risk

perfectly. Yeah. You are placing these

massive bets on incredibly fragmented,

really noisy biological data.

>> Right.

>> And finding the signal in that noise is

exactly what inspired the namesake of

this new model. GPT Roselyn is named

after the British crystalallographer

Rosalyn Franklin.

>> Oh, right. She was the scientist whose

work with X-ray defraction. Um,

specifically that famous image photo 51,

>> right? Photo 51.

>> That's what actually revealed the double

helix structure of DNA back in what

1953.

>> Exactly. And the thing about her work

was she was looking at what essentially

seems like a blurry, chaotic scattering

of X-rays to anyone else.

>> Just noise.

>> Just noise. But through this rigorous,

incredibly meticulous pattern

recognition, she deduced the fundamental

geometric structure of life itself,

>> which is amazing. Even though Watson,

Crick, and Wilkins usually get the

historical spotlight and the Nobel

Prize.

>> Yeah, she didn't get the Nobel in her

lifetime, sadly. But her ability to see

the pattern in the noise was the

catalyst for all of it. And GPT Rosland

is attempting to apply that exact same

meticulous human pattern recognition,

but at a scale of billions of data

points. Okay, so let's dig into the

mechanics of how it actually does that.

Because reading the technical specs in

our sources, it's very clear this is a

completely different beast than the AI

models most of us are used to.

>> Oh, absolutely.

>> Open AAI labels it as a quote frontier

reasoning model. But I want to push back

on this a little bit for you listening

because I know what you might be

thinking.

>> Sure.

>> If I prompt this model about a specific

protein mutation, isn't it still just a

glorified search engine? Like, isn't it

just scraping a really dense Wikipedia

page or PubMed and summarizing the text

for me?

>> I'm so glad you brought that up because

this is probably the most common

misconception. A search engine performs

information retrieval,

>> right? It finds things.

>> Exactly. It finds a paper that says

protein X interacts with molecule Y and

it just shows you the paper.

>> But GPT Roslin performs multi-step

reasoning. It doesn't just read about a

protein. It actually synthesizes

evidence across thousands of multiomic

databases.

>> Okay, let's define that for a second.

Multiomic,

>> yes, a bit of a mouthful.

>> It means it's looking at genomics, which

is the DNA. It's looking at proteomics,

the proteins themselves, and

metabolomics, which are the chemical

processes. So, it's looking at every

single layer of the biological all at

the same time.

>> Precisely. And it processes all of those

layers simultaneously. So it analyzes

the 3D structure of a protein by

interpreting the underlying spatial

coordinates of its atoms

>> like a physical 3D map,

>> right? It models the physical

constraints and the energetic bonds in

its latent space and then it looks at

how a specific genetic mutation would

alter those bonds and signaling

pathways.

>> Okay.

>> And then it cross references that with

chemical reaction mechanisms. So it's

generating novel hypotheses that

literally do not exist in any published

paper.

>> Wait, really? It's coming up with

entirely new ideas.

>> Yes, it's not just repeating what humans

have already discovered.

>> But how is a language model doing that?

Like something designed to predict the

next word in a sentence. How is it doing

spatial chemistry in biology?

>> Well, because biology at its core is a

language.

>> Oh, interesting.

>> Right. DNA is a sequence of letters.

Codes are sequences of amino acids. When

you train a massive neural network on

billions of these biological sequences,

it learns the underlying grammar of

biology.

>> The grammar of biology. I like that.

>> Yeah. It learns the rules of how an

amino acid chain will fold in physical

space. The same way an English language

model learns that a noun usually follows

an adjective.

>> That makes a lot of sense. So, it's not

translating English. It's translating

chemistry.

>> Exactly.

>> And our sources point to a specific tool

that OpenAI developed that makes this

actually actionable. Right. The Codex

Life Sciences plugin.

>> Yes, the Codeex plugin is what makes

this a real scientific tool. It acts as

an autonomous orchestrator.

>> What does that mean in practice?

>> It means it gives the reasoning model

direct live access to over 50 public

databases, scientific computing tools,

and literature repositories.

>> So, it's plugged directly into the

global scientific grid.

>> Right? So, the model doesn't just sit

there and generate a text response based

on its training data. It actively

designs a workflow. It essentially

becomes like an agent project manager.

>> Yes. If you ask it to investigate a

cellular pathway, it might write a

Python script to query a genomic

database, pull that raw data back into

its context window, parse it, and then

realize it needs more information on a

specific binding affinity.

>> So, it knows what it doesn't know.

>> Exactly. Then it pings a completely

different database for that chemical

data, synthesizes the two, and outputs a

complete step-by-step molecular cloning

protocol for the scientist to actually

follow in the lab. It actively

interrogates the digital landscape.

>> That is wild. But you know, we have this

brilliant theoretical engine, but

scientists are naturally and rightfully

pretty skeptical of tech industry hype.

>> Oh, scientists are the biggest skeptics.

They want to see the receipts,

>> right? They want to see it perform the

lab. And the sources do highlight some

benchmark data. Bixbench and Labbench 2.

>> Yeah, those are standard benchmarks that

test tasks like literature retrieval and

predicting RNA sequences.

>> And how did it do?

>> It completely outperformed all previous

versions of GPT across the board. But

honestly, the truly definitive proof

point comes from an evaluation they

conducted alongside a company called

Dino Therapeutics.

>> Okay, what did they do?

>> They tested the model on sequenceto

function predictions. Meaning you feed

the AI a string of genetic letters and

it has to predict what the resulting

physical protein will actually do inside

a living cell.

>> Exactly. And when a human expert does

this, you know, a human researcher, they

look at a sequence mutation and use

their years of experience and heristic

rules to guess if that protein might

misfold or become overactive.

>> Is an educated guess,

>> right? The AI doesn't guess. It

calculates the probabilistic

interactions of the entire amino acid

chain. And in certain evaluations with

dynother therapeutics, Gassi Roslin's

predictions actually exceeded the 95th

percentile of human experts.

>> Wait, it beat the top 5% of human

specialists in predicting biological

behavior.

>> Wow. Well, that certainly explains why

the adoption list in the Reuters report

is a who's who of biotech. Amgen, Madna,

the Allen Institute, Termoffisher

Scientific, Los Alamos National

Laboratory, they are all already

collaborating to integrate this into

real workflows.

>> Yeah. The heavy hitters are all in

because they recognize that this is the

necessary evolution of what started with

Alphafold.

>> Oh, right. Alphafold from Deep Mind.

>> Yeah. AlphaFold was a massive

breakthrough because it solved the

protein folding problem. It essentially

gave us the static 3D map of the biology

>> like a snapshot.

>> Exactly. GPT Roslin takes that map and

builds an interactive physics and

reasoning engine on top of it.

>> So, it makes it dynamic.

>> Yes. It takes those protein structures

and figures out how diseases hijack them

and then what specific new molecules

could repair them.

>> And the sources show that companies

taking this what they call an AI native

approach. So building their research

pipelines around machine learning from

day one, they are seeing incredible

clinical outcomes.

>> The numbers are really shifting.

>> Yeah. The fierce biotech report noted

that AI native companies are hitting 80

to 90% success rates in phase clinical

trials,

>> which is just a staggering leap,

>> especially when you remember the

historical success rate for phase is

somewhere between 40 and 65%.

>> You are dramatically derisking that

massive initial investment we talked

about earlier.

>> Exactly. Yeah.

>> You aren't wasting a billion dollars on

a compound that's doomed to fail in year

seven,

>> right? But, you know, when you hear

stats like beating the 95th percentile

of human experts, it's really easy to

jump to the conclusion that human

scientists are just being replaced

entirely.

>> The sci-fi dystopian view. Yeah,

>> exactly. But based on how this actually

operates, it seems less about replacing

the human and more like an Iron Man suit

for a scientist.

>> Oh, I like that analogy.

>> Right. The AI is doing the heavy

lifting. It's calculating the

navigation, drawing up a flawlessly

calculated molecular blueprint that will

hold weight, but the human is still

inside driving it. They still have to go

to the construction site and pour the

concrete.

>> That is exactly it. The human remains

firmly in the loop. The AI accelerates

the synthesis of evidence and the

hypothesis generation, but the

creativity required to ask the right

initial questions that remains entirely

human. And more importantly, the actual

physical wet lab validation is purely

human.

>> Right? The AI cannot hold a pipet.

>> It can't run a cell culture or observe

an animal model.

>> No, it can't. The physical reality of

experimental science still happens at

the bench. GPT Roslin simply ensures

that when a scientist spends 6 months

running an experiment, they are testing

a hypothesis that has an incredibly high

mathematical probability of working.

>> It's focusing their effort.

>> Exactly.

>> Which brings us to the operational

reality. the big catch in all of this

because if this model is cutting years

off the development timeline and beating

expert benchmarks, why isn't every

university, every regional hospital, and

every independent lab in the world using

it right this second?

>> Because the sheer capabilities of this

system introduce some pretty severe

deployment hurdles. Right now, OpenAI

has GC Rosland locked down in a uh quote

research preview under a strict trusted

access program. Okay. So, it's not an

open API anyone can just plug into.

>> Definitely not. That limits usage

primarily to qualified US companies and

institutions that have proven robust

governance structures in place.

>> That's a highly restricted VIP list,

which obviously slows down broad

academic or global adoption, right?

Especially for developing nations that

could really use this for localized

disease research.

>> It does slow things down, but the slow

roll out is driven by two main factors.

Technical limitations and security.

Let's talk about the technical side

first.

>> So, technically, the model still

struggles with what they call context

degradation.

>> What does that mean?

>> Well, if a scientist asks it to execute

a massive like 50step reasoning workflow

that requires pinging 20 different

external databases, the model can

sometimes lose the thread. It might

hallucinate connections.

>> Oh, right. AI hallucinations.

>> Yeah. So, it requires scientists who are

highly skilled in prompt engineering and

data validation to actually guide it and

doublech checkck ite work. So it

requires a massive upskilling of the

current scientific workforce.

>> It really does. You can't just hand this

to an undergrad and expect a miracle

cure. But the primary reason for the

lockdown access program isn't technical.

It's biocurity.

>> Biocurity.

>> Yes. Biology is inherently dual use.

>> Meaning it can be used for good or for

bad.

>> Exactly. The exact same deep reasoning

capabilities that allow this AI to

analyze a viral genome and design a

life-saving targeted vaccine could

theoretically be inverted

>> to design a novel pathogen.

>> Right? A highly lethal, completely novel

biological weapon.

>> Man, it's the ultimate double-edged

sword. If the model understands the

grammar of biology well enough to build

a cure, it understands it well enough to

engineer a weapon.

>> Which is exactly why organizations like

Los Alamos's National Laboratory are

involved in the validation process. You

simply cannot deploy an open- source

tool that might accidentally provide a

step-by-step molecular cloning protocol

for synthetic virus.

>> Yeah, that would be catastrophic.

>> Open AAI has implemented really heavy

safeguards. And the trusted access

program is designed to stress test those

guard rails before they even think about

any broader release.

>> That makes total sense. But it also

brings up a massive economic question

for you listening right now.

>> Right. The equity question.

>> Exactly. If this technology successfully

slashes the cost of pharmaceutical R&D

by billions of dollars, like the TUS

data showed, and it cuts a 15-year wait

time in half, where does that value

actually go?

>> That is the billion dollar question.

>> Are we actually going to see cheaper,

more accessible prescriptions at the

pharmacy counter? Or does a $2 billion

savings just become higher profit

margins for the few big pharma companies

that hold the keys to this AI? Yeah,

because the technology fundamentally

changes the speed of discovery, but it

doesn't automatically rewrite the

economics of healthcare.

>> The system is still the system.

>> Exactly. Whether those massive cost

savings are passed down to patients

depends entirely on how the industry

operates in the coming years.

Regulators, academia, and the

pharmaceutical industry are going to

have to navigate how this value is

distributed as the technology matures.

>> And let's be impartial here. There are a

lot of different viewpoints on how that

should happen. Some argue for heavy

regulation to force prices down, while

others argue that the market will

naturally lower prices through increased

competition. We aren't endorsing any

specific political solution, but it's

clear from the reports that this is

going to be a defining conversation for

the industry over the next decade.

>> Oh, absolutely. The technology is here,

but the policy is lagging behind.

>> So, to bring all of this together for

you today, GPT Rosland is an incredibly

fitting tribute to Rosalyn Franklin. She

possessed this rare ability to see the

fundamental truth hidden inside noisy,

chaotic data.

>> And that's exactly what this model is

doing at scale.

>> Right? It targets that 10 to 15ear

bottleneck in drug discovery, not by

replacing scientists, but by acting as

the ultimate amplifier for human

curiosity. It cuts through the noise of

millions of data points so scientists

can focus on what actually matters,

curing disease. We are really stepping

away from an era of brute force trial

and error. We're moving into an era of

accelerated precisiong guided reasoning.

>> And it's wild to think this model is

merely the first iteration in the life

sciences series.

>> Yeah, this is just version one.

>> As these systems scale, their capacity

for complex biochemical reasoning will

only deepen.

>> We are watching the foundation being

poured for the next century of medicine.

And we really want to hear your

perspective on this. Reach out to us on

social media. Do you think AI discovered

drugs will actually lead to cheaper,

faster treatments for you and your

family in the near future? Or do you

think the economic realities of the

pharmaceutical industry will just absorb

all the benefits?

>> It is going to be genuinely fascinating

to see how the market and the science

evolve together on this.

>> Absolutely. And I want to leave you with

one final kind of provocative thought to

maul over today.

>> Okay, let's hear it.

>> We spent this entire deep dive talking

about how this AI can help us cure the

diseases we already have, right? cancer,

Alzheimer's, rare genetic conditions,

>> right? Reactive medicine.

>> Exactly. But if models like GPT Rosalind

eventually become perfectly adept at

understanding human biology and

absolutely flawless at predicting

exactly how molecules interact,

>> which is where this is heading.

>> Will the future of medicine shift

entirely? Will we move away from

reacting to diseases that have already

made us sick and instead use AI to

design proactive personalized immunities

for diseases before they even emerge in

the population?

>> Oh wow. Engineering our own resistance

before the threat ever even arrives.

>> Think back to that tireless scientist

sitting in the dark lab we talked about

at the start. Yeah.

>> Imagine them not just working to cure

your illness, but building the

biological blueprint to ensure you never

get sick in the first place. That is the

true potential of the master chef in the

kitchen.