AI Is Becoming a Scientist — And It’s Moving Fast

Welcome to the Sentient Code, where intelligence is engineered, autonomy is emerging, and a line between human and machine grows thinner. Each episode, we decode the algorithms, explore the robotics, and examine the ideas shaping the future of artificial minds.

Welcome back. It is Wednesday, February eighteenth, twenty twenty six. And if you've been looking at the tech headlines this morning, or really any morning for the past month, you're seeing one phrase just repeated over and over.

You really are. It's everywhere.

The industry is trying very, very hard to make the AI coscientist the defining concept of the year.

It's absolutely the buzzword of the quarter. You can't open a journal or a tech blog without seeing it.

But I want to push back on this a little. Right from the start, we've been calling AI an assistant or a copilot since what twenty twenty three at least. Yeah, we've had tools that can summarize PDFs for us and write email a drafts for years now. So my question is is co scientist just a branding pivot? Is this just marketing trying to sell us the same old llms with a new code of paint or is there an actual fundamental architectural difference in the stack that we're seeing today.

That's actually, that is exactly the right question to be asking. And it's fair to be skeptical because the hype cycle is very real. It's always real, always, But if you actually look at the engineering, the shift that we're seeing in early twenty twenty six is not just marketing. It really isn't. If you look back at that, say, twenty twenty three, twenty twenty four era, the architecture was fundamentally something called retrieval augmented generation RIGG.

We've talked about that before.

Exactly ria you'd ask a question, the model would go look up a document, summarize it, and then give you an answer based on what it found. It was a librarian, a very very fast library, very fast, incredibly well read librarian. Sure, but it was still just retrieving existing information, right.

It's fetching, it's not it's not thinking, it's not creating.

Precisely, the shift to coscientists here in twenty twenty six is really about two core things, agency and closed loop systems. We are not just talking about a chatbot that answers a prompt anymore.

Okay, so that sounds like engineering jargon. Let's break that down close. Loop, what does it actually look like in a real lab setting?

So it means the system is self correcting. It observes data, it generates a hypothesis. It then writes the code to test that hypothesis. It executes a simulation or even a physical experiment. Okay, it reads the error logs or the results. And then, and this is the absolute key part, it refines its own hypothesis based on whether it's succeeded or failed. It doesn't stop and ask the human what should I do next?

Ah?

It iterates that self correction loop is what earns it the title of scientist rather than just search engine.

So it's the difference between asking a sus chef to chop an onion for you and asking a chef to invent a completely new soup.

That's a great analogy.

The chef tastes, it realizes that it needs more salt, adds the salt tasted. Again, that loop is internal. It doesn't need external direction for every.

Step precisely, and the capabilities that we're seeing now things like designing experiments from the ground up, simulating molecular interactions at a dynamic level, not static, and even controlling physical robots in a lab. These are things the copilots of twenty twenty four simply could not do.

There was this quote from the research steck we looked at that really stuck with me. AI has moved from summarizing papers to making discoveries.

That's the headline.

It's a bold claim, and today we're going to test that claim. We're going to look at the evolution of this technology, specifically, how we got from you know, the static images of alpha fold to the dynamic simulation engines we have now, and we have to open.

Up the hood and look at the brain of these systems, the whole egenic workflow, this test time compute that everyone's talking about.

Absolutely, and then we have to talk about the physical layer, the robotics, because that's where the digital world actually touches the real one.

That's where code moves.

Yeah, and we'll get into the medical break twos that are happening right now, real world stuff in cancer, in Alzheimer's, and we'll wrap up with how you our listeners can actually get your hands on these tools, because apparently you don't need a PhD and a million dollar grant to play in the sandbox anymore.

No, not at all. The barrier to entry has completely collapsed. It is a very very different world than it was even eighteen months ago.

Okay, so let's rewind a little bit to set the foundation here. To really understand twenty twenty six, we have to understand where we came from. And you can't talk about AI and science without talking about alpha fold.

Right. Alpha fold is really patient zero for this entire revolution. If we go back to that twenty twenty one to twenty twenty four window, Google deep Mind effectively solved the protein folding.

Problem, which was a fifty year old grand challenge in biology. This wasn't a minor problem, oh no.

It was the grand challenge for half a century. We knew the chemical formula of a protein, you know the list of ingreen and seemino ass, but we had no reliable way of knowing how it folded up into a complex three D shape.

And in biology, shape is everything everything.

Shape determines function. If you don't know the shape, you don't know if that protein is going to build a muscle or digest your food, or you know, causet disease. It's the whole ballgame.

I remember the analogy everyone was using back then. It was like Google Maps for the molecular world.

It's a pretty decent analogy. Actually, before alpha fold, we were navigating the cellular world by stars and intuition, doing painstaking X ray crystallography.

Which could take months or years for a single protein exactly.

Suddenly, with alpha fold we had a street view for nearly every protein known to science. Then alpha fold three came along and predicted the structures of not just proteins, but their interactions with DNA, RNA other molecules.

It was a massive leap, and the creators won the Nobel Prize in Chemistry for it in twenty twenty four.

Rightfully so it was the moment of validation. It was the proof that AI had truly arrived in serious hard science. It wasn't just for you know, identifying cats and pictures anymore.

But here we are in twenty twenty six and looking at the research we have in front of us for today, alpha fold almost looks I don't want to say basic because that feels disrespectful. No I know em but it looks like just the starting line.

It was the foundation. Absolutely, the key limitation of alpha fold, and this is so important to understand the current moment, is that it gave you a static structure. It gave you a snapshot, a beautiful, high resolution statue of a protein.

But biology is in the statue, not at all.

Biology is wet, it's warm, it's noisy. Proteins don't just sit still. They vibrate, they twist, they dance, They adopt all these different conformations based on what they're touching, or the temperature of the cell or the pH level.

And computing that dance is exponentially harder than just computing the one static shape, right yeah, because you aren't just solving for X, you're solving for eds over time in a changing environment exactly.

You're dealing with what's called an energy landscape. Alpha fold was incredible at finding the lowest energy state, the most stable resting pose.

Of the protein, the bottom of the valley, the.

Very bottom of valley. But the twenty twenty six systems, these successors, like alpha proteo and a SOD are probabilistic. They don't just give you the bottom of the valley. They map out the entire landscape. They tell you the probability of a protein shifting into a different shape that can, say, suddenly, lock onto a drug molecule.

So it's a difference between a photograph and a movie.

It's a movie where you can see all the possible alternate endings, and even better, a movie where you can influence which ending happens. This gets us to the concept of inverse folding.

Okay, inverse folding, I saw that in the isomorphic labs papers. Break that down for us.

Okay, So traditional protein folding is here is a sequence of amino acids. What three D shape does it make?

Input is the sequence? Output is the shape?

Right, inverse folding flips that entire problem on its head. It says, here is the shape I need. For example, I need a shape that perfectly blocks this specific virus receptor. Now you tell me this sequence of amino acids that will reliably build that exact shape.

So it's generative. It's like using an image generator like Dally, but instead of the prompt being an astronaut riding a horse on the moon. Okay, the prompt is a protein that binds to and neutralizes the COVID twenty six spike.

Precisely. It is generative biological engineering, and this one capability to design interactions, not just predict them, has completely utterly collapsed the timeline for drug discovery.

Okay, let's talk about those timelines, because traditionally, discovering a new drug is a marathon. We're talking ten to fifteen years from initial idea to pharmacy shelf.

And billions of dollars, and you have to remember most of that time is just failure. It's high throughput screening where you're just trying a key in a lock and it doesn't fit, so you file it down a little, You try again, it still doesn't fit. You do that

It's a huge shift in efficiency. The sources mentioned that AI designed molecules. Molecules that were invented in silico entered phase three clinical trials back in twenty twenty five.

That is the holy grail. Phase three is the final most expensive step right before FDA approval. Getting a drug to phase three and twenty twenty five means the foundational discovery work probably started in what twenty twenty two or twenty twenty three.

That's a compression of the timeline that is almost hard to believe.

It's why we're seeing these market forecasts for AI driven drug discovery just exploding. They're hitting eight to ten billion dollars in twenty twenty six alone. This isn't a future trend, it's the economic reality of the industry right now.

So it's speed, yes, but it's also precision. We aren't just throwing spaghetti at the wall faster. We're guiding each trand of spaghetti to the wall with I don't know, laser beams.

That's a colorful image, but yes, that's it. And that precision comes from the fact that we aren't just using a single monolithic algorithm anymore. We've moved into this era of agentic workflows, and this is where we really need to get into the inside the machine part of our discussion today.

Yes, let's unpack this because when I hear AI, my brain still defaults to a text box where a type of question and he gives me an answer a chatbot. But the co scientists system that Google deep Mind has built, which is running on their new Gemini two point zero model, isn't just one thing. It's a team. The outline describes a cast of agents, right.

And you should think of it almost like a corporate structure or a research lab team. In any functional lab, you have different roles, right. You have the creative postdoc who has one hundred wild ideas a day brainstormer exactly. Then you have the critical senior scientist, the principal investigator, who pokes holes in ninety nine of those ideas and

And the aicoscientist replicates that entire structure using specialized, fine tuned software agents. It does, But hang on a second, why is that necessary? Isn't that just the model talking to itself. Why do we need distinct agents? Why can't I just put it all into one big prompt and say, hey, Gemini, be creative, but also be critical and also manage the project.

You can try that, people have, but you run into a problem that engineers call context contamination, or having conflicting objective functions. Okay, if you ask one single model to be both creative and critical at the very same time, the two objectives bleed into each other. You get mediocre creativity because the model is presensoring its own wild ideas.

It's pulling its punches exactly.

And you get soft, ineffective criticism because it subconsciously wants to support the ideas it just generated. It's not a true adversarial process.

So by splitting them up into separate agents, you create a genuine adversarial dynamic. You let them argue, you let them argue, you optimize them against each other. So you have the generation agent. Its only job is quantity and variety. It looks at all the available literature, all the data, and it generates hundreds of potential hypotheses. It's rewarded purely for novelty. It doesn't care if an idea is perfect or even feasible. It just wants to explore the entire solution space.

Okay, so that's the wild idea engine. Then the supervisor passes those ideas to the next stage, the review and ranking agents.

The critics. Yeah, the peer review panel. These agents take those hundreds of hypotheses and they filter them through a brutal gauntlet. They have three main jobs. First, check for novelty. Has this been published before? Is this just rediscovering something from a nineteen eighty two paper, which is a huge problem in science, a huge problem. Second, check for feasibility. Does this violate the known laws of physics? Is this actually chemically possible? And Third, and this is maybe the

It's philosophy. The ranking agents discard it. So out of one hundred ideas, maybe only five or six make it through this filter.

And then those survivors go to the evolution agent.

The improver, the polisher. It takes the best ideas that survive the critics, and it iterates on them. It refines the hypothesis, It tightens the parameters, It suggests modifications. It turns a rough diamond into a cut gem.

And the conductor of this whole orchestra is the supervisor agent.

The manager, the supervisor is the traffic cop. It decides when to send an idea back to the generation agent for more brainstorming, when to push it forward to the evolution agent, and crucially, it manages the compute budget. It manages this thing called test time compute right.

I saw that term in the papers, test time compute scaling. It sounds incredibly technical. But in plain English, does that just mean thinking longer?

That's a perfect way to put it. Yes, But it's how it thinks longer that's revolutionary. Instead of just giving you the first, most probable answer, it allows the model to generate thousands of internal thoughts or reasoning.

Steps, a chain of thought, a whole.

Branching tree of thoughts, and it explores all the different paths, and most of them it discards as dead ends before it ever gives you the final, polished answer. This is a system too thinking that the psychologist Daniel Khannaman talks about. It's slow, it's deliberate, it's logical.

So it's not just retrieving an answer from its training data. It's actively deliberating. It's reasoning through the problem.

It's simulating multiple paths of reasoning. It's fact checking itself against its own internal knowledge and external tools. It's the difference between a gut reflex and a deep considered contemplation, and the expert analysis is unambiguous. The more time you give the model to think at test time, the better and more robust the reasoning becomes.

We have a perfect real world example of this in the sources the bacterial mystery, and I want to go deep on this one because the timeline is just wild. This is the story about cfpicis.

Right, capsid forming phage, inducible chromosomal.

Islands, the zood height.

Thank you. So these are basically little mobile genetic elements in bacteria. The mystery for science was how do they spread between different bacterial species. They're not viruses, but they act a bit like them.

And humans had been working on this problem for how long, over.

A decade, really smart people publishing papers, making incremental progress, but nobody could figure out the exact mechanism for how they moved around.

So enter the aiicoscientist walk us through exactly what happened here. It didn't just guess the answer.

Right, no, not at all. It followed the exact agentic workflow we just described. The system started by ingesting a massive data set of bacterial genomes and all the existing scientific literature on the topic.

Okay, so read everything.

It read everything. Then the generation agent went to work. It proposed a bunch of hypotheses. One of the most promising was this, maybe the cfpicis steal proteins from actual viruses that happened to be in the same cell. They hijack the viral proteins to build their own transport capsules.

Okay, that's a clever, plausible theory, like a car thief stealing an engine to put in their own chassis.

It's a great analogy. But a theory isn't enough. The review agent immediately kicked in. It said, Okay, if that's true, then we should be able to see a specific cluster of genes activating at a specific time when a bacterium is co infected. It made a testable prediction.

It defined the evidence it needed to define exactly.

So the supervisor agent then directed the system to run a massive search across forty thousand bacterial genomes to look for that specific genetic signature, that coactivation pattern, and it found it. It found a candidate, strong one. But it didn't stop there. The evolution agent took over and refined the hypothesis. It predicted exactly which proteins from the virus

And the total time for this whole process.

The system went from here is a decade old mystery to here is a highly probable data supported mechanism in two days.

Two days versus ten years of human strung that is just staggering.

And it was validated. The human scientist took the AI's output, ran the physical lab experiments and confirmed it. The paper was published in the journal Cell in September of twenty twenty five, and importantly, it wasn't a computer science paper about how cool the AI is. It was a biology paper about a fundamental discovery made by the AI.

That distinction is so critical. The AI isn't the story anymore. The discovery that it enables is the story.

That's when you know the technology has matured. But you know it's not just biology. This deep think capability, this ability to reason through these long, complex chains of logic, is showing up in the most abstract fields like pure mathematics and physics.

Which brings us to open ais. GBT five point two.

Right GPT five point two, which came out late last year. It introduced these thinking modes or deep think modes as a default feature, and the benchmark that just blew everyone away was its performance on the International Mathematics Olympiad the IMO.

This is a competition for the most brilliant high school math kids on the planet. These are problems that can stump professional mathematicians with PhDs.

They're designed to require not just knowledge, but true ingenuity, and GPT five point two achieved a gold metal level performance.

But wait, math has always been the achilles heel of large language models, hasn't it? Because lms are probabilistic. They're just predicting the next most likely word that's right. And math is pure logic. It's precise. If you are ninety nine percent right in a mathematical proof, you are one hundred percent wrong. So how did they finally crack that problem?

They cracked it by combining that test time compute we talked about with sophisticated tool use. When GPT five point two faces a hard math problem, now it doesn't just try to guess the answer in one shot. It breaks the problem down. It says, okay, first I need to solve this integral. Then it actually writes a small Python script to call a symbolic math library like simpi.

Ah. So it's using the computer as a calculator the same way human mathematician would use Wolfram alpha or a physical calculator.

Exactly. It runs the script, it gets a definitive grounded truth as the outpuint, and then it incorporates that result into the next step of its logical proof. It's orchestrating a whole suite of computational tools.

It's not just a language model anymore. It's a reasoning engine that uses language.

That's the perfect description, and this capability has allowed it and to solve multiple open problems from the famous mathematician Paul Erders.

The Erdor's problems. I remember hearing about these. He was this prolific mathematician who left behind hundreds of unsolved conjectures with cash prizes.

Attached, and GPT five point two managed to crack several of them. But here's the real seal of approval. The proofs that generated were validated by Terrence Taw.

Wow, Terrence Taw For anyone who doesn't know, he's basically considered the Mozart of modern mathematics. It feels metalist.

He's the gold standard. So having Terrence twe review and validate an AI's mathematical proof is like I don't know, having Einstein check your physics homework. It signals that the AI is engaging in genuine, novel mathematical rus not just regurgitating the things that saw in its training data.

And the sources mentioned this new benjmark called frontier math. It sounds like the old tests just got too easy.

They did the AIS were acing them, so they had to build a new harder test. Frontier math consists of these ultra hard problems from the frontiers of research that usually required days or weeks of work by teams of expert mathematicians. GPT five point two is scoring over forty percent on the benchmark.

Forty percent might not sound that high to a layperson, right.

But for this level of difficulty, it's absolutely superhuman. It means it's making significant headway on nearly half of the hardest problems humanity has to offer.

Let's shift gears from the abstract world of math to physics, because we aren't just solving equations on a chalkboard anymore. We are and I can't believe I'm saying this, controlling plasma infusion reactors.

Yeah, this is where it gets really tangible.

The Genesis mission, which was launched by the Trump administration back in November twenty twenty five, specifically mentions a collaboration between OPENING and the Department of Energy on tokeomagmagnets.

Fusion energy the dream of unlimited clean power. The core problem with fusion has always been containment. You have to contain plasma that is literally hotter than the surface of the Sun, and.

You do that with incredibly powerful magnetic.

Fields, right, but the plasma is violently unstable. It wants to burst out. The analogy I've heard is that it's like trying to hold a giant, angry ball of jello together using only rubber bands while the jello is also exploding, and if.

You fail, even for a millisecond, the reaction just.

Fizzles out or you seriously damage the multi billion dollar reactor. The new AI models, specifically the ones developed in this Open AI and DOE collaboration, are now in direct control of these magnetic coils. They can adjust the magnetic field in microseconds, reacting to the plasma's turbulence, faster than any human operator or any traditional control algorithm ever could.

So the AI is essentially surfing the plasma wave in real time.

It is it's predicting the install ability a fraction of a second before it happens, and then preemptively countering it. It's a perfect example of why this high speed inference reasoning is so critical in the physical world.

Speaking of the physical world, this is the perfect transition. Let's move to section four, the physical loop, because this is the part that to me really feels like science fiction becoming everyday reality. We are moving from just code to actual beakers.

This is the concept of closed loop autonomy. It's the final piece of the puzzle. It's the bridge between the digital world of simulation and the physical world of experimentation.

So we have the cycle. The AI generates a hypothesis.

It runs a simulation to see if it's plausible.

And then, and this is the new part, it designs and executes a robotic experiment to test it in the real world.

It analyzes the results from that physical experiment and uses that data to generate a new better hypothesis. The loop is closed.

I get the chemistry side of it, but the robotics and how does a large language model, which fundamentally just deals in tokens and text know how to grip a glass test tube without crushing it. We're not just setting a spreadsheet to a machine anymore.

That is the big breakthrough of what are called VLA models. That's vision language action.

Okay, vision language action. How did that work in practice?

So traditionally lab robots were hard coded. A programmer had to tell them specific coordinates. Move your arm on the X axis ten point five millimeters, rotate your gripper on the Z axis ninety degrees, close scripper. It was incredibly brittle. If the beaker was one inch to the left of where it was supposed to be, a robot would just grab empty air.

Right. They were powerful, but they were dumb robots.

Extremely dumb. The news systems like the one being used in the cost Scientist project at Carnegie Mellon use these VLA models. They essentially tokenize physical actions in the same way they tokenize words to the AI. The command pick up the vile from the rack is just a sequence of vectors, just like the sentence write a poem about a robot is a sequence of vectors.

So it's predicting the x most likely physical movement in the sequence, the same way chat GPT predicts the next most likely word in a sentence.

Precisely. It learns the grammar and syntax of chemistry experiments by watching videos of humans doing them. It looks at the live camera feed that's the vision part. It understands the high level instruction from the scientist that's the language part, and it outputs the precise motor controls that's the action part.

That is just wild. And because it's a language model, it can read the instruction manuals for the lab equipment.

Yes, it literally reads the PDF manuals for the liquid handler or the spectrometer, so it knows the API calls for the hardware. It writes the Python script to drive the equipment on the fly, tailored to the specific experiment it just designed.

She could literally give it a plain English command like synthesized iboprofen, and it just figures out all the steps. Yes.

In the co Scientists demo from late last year, they did exactly that it looked up the chemical reaction in the literature, It calculated the precise amounts of the reagent, It translated all of that into a sequence of robotic protocols, and it executed the synthesis from start to finish autonomously.

And then there's this other project from Berkeley Lab, the DTCs the Digital Twin for Chemical Science.

That one is so cool. It takes it a step further. It creates a real time, physics based digital replica of the experiment as it's running. So while the robot is mixing chemicals, you have a digital twin of that beaker on your computer screen.

And you can interact with the digital twin.

You can you can adjust parameters on the screen, say you turn up the virtual temperature dial, and the AI uses that to optimize the reaction in the real world in real time. It can compress months of painstaking trial and error optimization into a few minutes.

We're seeing this in industry too. Google's robotic labs are using this approach for materials science.

Yes, specifically for discovering new inorganic materials. In twenty twenty five alone, there are autonomous labs discovered dozens of previously un known stable inorganic compounds. These are materials that could be the basis for the next generation of battery electrolytes, or a new type of superconductor, or a more efficient catalyst for carbon capture.

And they're designing them in silico in the computer and then immediately synthesizing and verifying them autonomously.

The discovery pipeline is becoming fully automated.

This brings us to what is for me, the most important and personal part of this whole deep dive, saving lives Section five, The medical realities in twenty twenty six. Because all this technology is fascinating, but if it doesn't actually help people, it's just a very expensive toy.

And in twenty twenty six it is definitely helping people. The impact is becoming undeniable.

Let's start with Alzheimer's. This is a disease that has baffled researchers for decades. We've had so many false starts, so many promising drugs that looked great in mice and then failed spectacularly in human trials we have.

It's been a graveyard for pharmaceutical R and D. But the AI cooscientist approach has given us a completely new angle on the disease. There's a gene called PHGDH. For a long time, scientists thought it was just a biomarker.

Meaning it was just a signpost a correlation.

Exactly like smoke indicating there might be a fire nearby. It was elevated in patients with Alzheimer's, but we didn't think it was actually causing the fire, just a symptom, right, But the AI systems modeled the incredibly complex three D protein structures and the entire gene regulation network around this gene and found something completely surprising. It's not just smoke, it's part of the fire. How so, the AI discovered that the protein made by the PHGDH gene is an

So it's not a bystander. It's one of the villains.

It is a key villain in the story. And because we now know it's a villain, and thanks to tools like alpha fold and alpha proteo, we know its exact three D shape, we can design a drug to specifically target it and inhibit its function.

New therapeutic candidates have been derived directly from this AI generated insight exactly.

It has opened up a whole new avenue for treatment that we were completely blind to before because the chain of interactions was simply too complex for the human mind to piece together from the data.

That's incredible. And then there's cancer, specifically the field of immunotherapy.

This is a major collaboration we're seeing between the pharmaceutical giant Astrosenica and the AI company Tempeses. AI immunotherapy can be miraculous when it works.

It can literally cure stage four cancer.

It can, but it only works for a fraction of patients. The grand challenge is knowing ahead of time which patient is going to respond. Giving a multi hundred thousand dollars drug to someone it won't help is a terrible outcome.

So they use what they call a predictive biomarker modeling framework. The notes mention contrastive learning. We hear that term a lot these days. Can we demystify it? What is contrastive learning actually doing here?

Sure, standard machine learning often looks for patterns in one big bucket of data. Contrastive learning is more clever, more aggressive. It takes two distinct data sets, in this case, pathology images and genetic data from patients who survived, and the same data from patients who didn't, and it forces the model to find the maximum possible difference between those two groups.

So it's not just looking for a pattern, it's looking for the one thing that most clearly separates winners from losers.

It's forcing the two groups as far apart as possible in the high dimensional data space to find the signal in all the noise. It's like looking for the needle in the haystack by systematically burning away all the hay and what do you find? It identified a complex multimodal signal, a subtle pattern combining features from the pathology slides with specific genetic sequences that no human pathologist would ever be

Therapy and the resulting clinical trials.

A fifteen percent survival benefit.

Fifteen percent that is not a rounding error in oncology. That is a assive victory. That is thousands of lives.

It is. It's the promise of precision medicine finally being delivered. It's about getting the right drug to the right person at the right time.

And finally, on the medical front, antimicrobial resistance, the superbugs. This is the one that really keeps me up at night, the slow motion pandemic.

It should keep everyone up at night. We are running out of effective antibiotics. Bacteria are evolving resistance faster than we can invent new drugs to fight them. But the aiicoscientist is helping us catch up.

How it's helping us anticipate the enemy's moves. There was a paper where the system predicted these highly complex gene transfer mechanisms, the ways that different bacteria swap armor plating and weapons to protect themselves from our drugs, before those mechanisms were ever observed and published experimentally.

It saw the future. It predicted their evolution.

It did, and by understanding these defense mechanisms before they become widespread, we can start designing new antibiotics that bypass them, or even turn their own to against them. It gives us a fighting chance in a race we were starting to lose.

It's just it's breathtaking stuff. But I want to pause here for a second. We've been talking for a while about the AI doing everything generating the ideas, running the robots, predicting the genes. So where do the humans fit in? Are we just becoming the janitors for the server form?

That is the big anxiety, isn't it the existential question? But the expert consensus in twenty twenty six from everyone we've read is remarkably clear. Humans are not only still necessary, they are more important than ever.

Why if it's so smart, why does it still need us? Well, for one, hallucinations, AI models, even the very best ones like GPT five point two, can still confidently make things up. They can sound incredibly convincing while being totally fundamentally wrong about a physical constant or a biological fact.

They don't know what they don't know exactly.

Novelty requires rigorous verification. You cannot just blindly trust the output of the black box, especially when human lives are on the line.

So the human is the safety valve, the final fact checker.

The human is the safety valve, absolutely, but they're also the conductor of the orchestra. The ideal setup we're seeing emerge is the human plus AI team. Think about that workflow we discussed earlier. The AI proposes one hundred different hypotheses. It's not the AI that decides which three of those are the most scientifically important or the most ethically sound to pursue. It's the human. The human brings high level context, intuition, and most importantly, values to the table.

So we picked the destination. The AI figures out the fastest way to build a road.

That's a perfect way to put it. And then you have the massive looming ethical and legal questions. For example, who owns an AI discovered drug? Oh?

The ownership question this is a big one. If I, as a researcher, read a clever prompt for the AI, and the AI invents a new life saving molecule, do I own the patent? Does my university? Does Google? Who built the AI? Or does nobody?

It is a massive legal and philosophical battleground right now. If an AI that was trained on decades of publicly funded research data discovers a cure for cancer, should a private pharmaceutical company be able to hold an exclusive twenty year patent on it?

These are not questions that an ADI can answer for us.

Oh. These are uniquely human problems that require human judgment and societal debate.

So we aren't being replaced. We are being augmented.

Promoted augmented is the right word. Our jobs are changing. We're moving from being the people physically pipetting liquid into tubes to being the people who are orchestrating fleets of robotic scientists to solve humanity's biggest problems.

This brings us to what I think is the most exciting and empowering part for our listeners, because you might be listening to all of this and thinking, well, this is great for the scientists at Google or after Zeneca with their billion dollar budgets. But I'm just a regular person with a laptop, and that is.

Where you would be completely wrong. The barrier to entry for high level scientific discovery has never ever been lower than it is today. We are seeing the exciting ride of the citizen coscientists.

I love that term. So what can I do today? February eighteenth, twenty twenty six, sitting at home with just my computer.

You have free access to tools that would have cost millions of dollars and required a supercomputer just a decade ago. The Alpha fold server is public. You can go right now and submit a protein sequence and get a state of the art three D structure back free for free.

So let's give our listeners some actual homework. What are some specific actionable things they can try this week?

Okay, sign it number one. Design a new molecule. You can use an open source LM like LAMA three or mistrawl. Combine with a free software library called rd kit, which is a standard toolkit for computational chemistry.

And what would it prompt be?

You could prompt it act as an expert medicinal chemist. My goal is to design a novel small molecule to inhibit the PHGDH protein implicated in Alzheimer's Generate five potential molecular structures. Then you can take those designs and actually simulate how well they might bind to the target protein. Using the public.

Alpha fold server, you can literally do the first steps of drug discovery on your couch.

You can generate the initial hypothesis. Now you can't synthesize it in your kitchen. Please please do not try that disclaimer, right, but you can do the foundational computational work that used to be the exclusive domain of major pharmaceutical companies.

Okay, that's incredible.

Assignment number two, use the advanced chatbots Gemini, Claude, GPT five point two, but prompt them like a professional. Don't just ask a simple question, give them a roll. Say you are my AI coscientist. My grand challenge is to find a new sustainable battery electrolyte that doesn't rely on lithium. Please generate five novel research hypotheses, rank them based on scientific feasibility and potential impact, and for the top one, suggest a concrete experimental plan to test it.

So it's all about the prompt treating it like a brilliant, tireless colleague, not like a simple search engine exactly.

And finally, Assignment number three, join a citizen science platform. There are platforms right now that are feeding training data into the Genesis mission models. By helping to classify images or fold proteins in a game, or even donating your idle GPU time you are actively contributing to the National Scientific effort.

You can be a part of the Genesis mission. That sounds pretty epic.

It is. This is a collective societal effort. Now science is being democratized.

So as we wrap up this deep dive, let's just zoom out one last time. We've covered the history, the tech, the robots, the medicine, the citizen science. What does this all mean for the big picture of human progress?

It means we are in the early days of a new scientific renaissance. I don't think that's an exaggeration. By the end of this year, late twenty twenty six, many experts predicting that we will see the first fully AI led discovery to publication pipeline.

Meaning from start to finish.

From start to finish, the AI has the initial idea, it does all the work, it writes the paper, it submits it to a journal, and it gets published, a completely autonomous discovery.

And this changes the scale of the problems we can even dare to tackle completely.

The grand challenge is curing cancer, solving climate change, achieving fusion energy. They have always been limited by human bandwidth. We only have so many brilliant scientists in the world, and they only have so many hours in the day.

They have to sleep, but silicon doesn't sleep.

Silicon doesn't sleep. The limit on the pace of discovery is no longer a collective bandwidth. It's becoming our collector imagination. It's our willingness to collaborate with these new silicon colleagues.

It's a genuinely hopeful message. I think the AI cooscientist isn't just accelerating the science we already do. It seems like it's redefining what discovery even means.

I think so it allows us to see patterns and ask questions that we're previously too complex for our brains to even formulate. It's a new lens for viewing reality.

That is a powerful thought to leave everyone with. So here is my final question to you listening right now, what hypothesis will you explore with your aicoscientists today? You have the tools, you now have the knowledge. The rest is up to you.

Go discover something new.