Loading summary
Transcript39 lines
- [00:00]
Paige Desorbo
This is Paige desorbo from Gigli Squad. Boost Mobile is no longer that prepaid wireless company you remember. They've invested billions into building their own 5G towers across America. With Boost Mobile's networks, customers enjoy the speed and service they'd expect from the big three, plus groundbreaking benefits you'd only get from a true challenger of the industry. Boost Mobile will let you try the network risk free for 30 days. So visit your nearest Boost Mobile store or find us online@boostmobile.com today.
- [00:33]
Dr. Samantha Amien
Lab grown meat. Whether it excites you or weirds you out, it's been making headlines lately, and for good reason. Scientists are growing real meat from animal cells without the need to slaughter actual animals. It sounds pretty futuristic, right? But the future might be closer than we think. From climate change concerns to animal welfare and food security, Lab Grown Meat is raising big questions and offering some bold solutions. We'll get into all of that in just a bit, but before we do, I'll speak with Dr. James Tso, who's an associate professor of biomedical data science at Stanford. We're going to discuss how he's created a research lab made up of generative AI scientists. They ask questions, they hypothesize, they debate. It's a wild process. Then we dive in on how the party drug MDMA could be used to treat PTSD in a scientific setting. Obviously not at a rave. I'm Dr. Samantha Amien, and this is Curiosity Weekly from Discovery. Let's dive in. What if you could eat chicken nuggets without the chicken ever clucking? That's the promise of Lab Grown Meat, a futuristic food tech that turns just a few animal cells into a sizzling steak. No slaughterhouses involved. Well, this is a real thing. But expecting it in the supermarket anytime soon is a whole other question. Unlike plant based burgers, Think beyond meat or impossible burgers. Cultivated meat is actual animal tissue, just grown differently. Scientists take a small sample of animal cells, the kind that can turn into muscle fat or connective tissue, and place it in a bioreactor, which is sort of like a giant fermentation tank. They feed the cells nutrients, amino acids, sugars, lipids, and coax them to grow, forming muscle and fat tissue. Kind of like what happens inside an animal's body, but without the animal. As the cells multiply, scientists tweak the conditions, adjusting the temperature, oxygen and nutrient mix to guide them into becoming specific types of tissue. If the goal is to make a burger, for example, they'll focus on growing fat and muscle cells in a loosely structured form. But if they're aiming for something more complex, like a steak, they use edible scaffolding to help organize the cells into layers that resemble the real texture of meat. After a few weeks, they harvest the tissue, shape it into patties, nuggets or sausages, and then you can cook it just like the meat you're used to. The end result is nearly identical to conventional meat, but vegetarian. Maybe if researchers can scale up production using more affordable food grade materials. Think brewing beer. But for meat, it could be one of the biggest food innovations in decades. And in theory, this could be a game changer for the climate too. Some models show that cultivated meat uses up to 90% less land than traditional livestock farming. No need for massive feedlots or water guzzling corn fields. Plus no animals are harmed in the process. It's a win for sustainability and ethics, right? Well, not everyone agrees. Critics of the tech often call it franken meat, and they argue that it's not as green as it sounds because growing meat in a lab requires purified nutrients, sterile environments and lots of electricity to maintain those bioreactors. A 2024 study from UC Davis even suggested that the carbon footprint of lab grown meat might be worse than beef, depending on the growth medium used for the cells. And that's really saying something considering cows are one of the most emission heavy animals on the planet. There's also the issue of scale. While startups like Upside Foods and Good Meat have met the regulatory requirements from the FDA and USDA to legally sell cultivated chicken in the U.S. it's still incredibly expensive to produce.
- [04:24]
Dr. James Tso
Not exactly ready for the drive thru window.
- [04:27]
Dr. Samantha Amien
That's part of what's fueling the bans. Lawmakers in Florida, Mississippi and South Dakota argue that they're protecting local agriculture and food traditions. Critics of the bans say these laws are premature or that the environmental claims are being overhyped before the industry has even fully matured. Some also worry that cultivated meat could become a tech monopoly controlled by just a few companies. For now, lab grown meat will stay off the shelves in Florida, Mississippi and South Dakota. Alabama also enacted a similar ban in 2024, but the federal government hasn't taken a stance either way, and cultivated meat is still legal in most other states. It's a patchwork legal landscape that mirrors the messy birth of many food innovations, but personally, I'm excited to try it one day.
- [05:16]
Whole Foods Market
Only the best will do for mom so make Whole Foods Market your Mother's Day destination. Shop the floral department for vibrant blooms like tulips, orchids, peonies, and expert crafted bouquets. Then head to the wellness and beauty department and give mom a spa like experience with scented candles and more. And if you're hosting brunch or dinner, order flavorful Whole Foods market catering by May 9. Celebrate Mother's Day with Whole Foods Market in store and online.
- [05:51]
Dr. James Tso
We're very much in the era of what Can't AI Do. It feels impossible to keep up with all the new milestones in AI, in large part thanks to advancements in generative AI, particularly large language models. They're so good at language understanding and generation tasks that we're integrating them into our everyday lives. But I've always been a bit skeptical about gen AIs when it comes to science. Every time I ask an AI chatbot a question about science, it cites the wrong source or says something without a credible source at all. The research focused ones are slightly better, but never at a level that can fool me to think it's being written by an actual scientist. Then I was at a scientific conference and saw a talk by Dr. James Tso claiming he's made an entire research lab of AI scientists who do their own hypothesis generation and come up with experiments for how to test it. Dr. James Tso is an associate professor of Biomedical Data Science at Stanford. Welcome to Curiosity Weekly, James.
- [06:49]
Thank you so much Sam for having me Excited to be here.
- [06:51]
What do you see as some of the main challenges with AI in science and medicine right now?
- [06:57]
I think a couple of the big challenges is one is that AI is quite expensive. It consumes a lot of resources, both like money and energy, so that relatively few groups would have the resources to be able to train these models and also to use these models, which is often makes it harder, especially for groups in different countries or in different universities institutions to be able to actually benefit from the technology and to really get get to the full impact. That's why I think I'm very excited about these more open source models. And we try to everything that we do here, like the virtual lab, we open source, all of that. And we want to make these open models open source and also relatively easy and efficient to use so that people could try to benefit from the technology without having to really spend a lot of electricity or a lot of dollars.
- [07:49]
I was gobsmacked by your talk. I was a skeptic and you had me convinced. So I'm always excited when that happens. Please tell us more about how you're creating AI science agents.
- [08:00]
Yeah, so we've recently built this platform that we call the Virtual Lab the virtual lab is basically a research group consisting of AI scientist agents with different expertise, for example, in chemistry or computational biology or data science. They have their own one on one meetings where one of the agents would meet with the professor agent and then they will progress through those meetings, sort of tackle quite challenging research problems.
- [08:24]
The idea that the AIs are also in meetings kind of made me laugh, like, all right, they got to do it too. Not even AI can escape meetings.
- [08:32]
That's true. Although their meetings, I have to say, are more efficient than our meetings. So one of their meetings maybe would take a few seconds or maybe a minute so that in the time that we had our coffee break, they could run hundreds of meetings and probably make some interesting research discoveries.
- [08:47]
Let's dig into that more. What does a day in the life of an AI scientist look like?
- [08:51]
Yeah, so basically, so we provide those high level problem descriptions to the AI professor and then the professor would actually try to create and then to train the different student agents that have the relevant expertise in tackling that problem. So for example, if the goal is to design vaccines, maybe the professor decides it's useful to have an immunologist agent. Then it's actually the job of the professor to train and to create that immunologist agent. So they have like this virtual lab school try to become a deeper expert into the relevant areas. And then the other half of their time and they start to do research, they have their meetings, they write the code, implement, execute, and then do data analysis.
- [09:27]
I can't stop smiling. Like, it's just so funny to think of these, I don't know, AI algorithm, I don't even know how to explain it. These agents, as you say, they're going to school, they're learning, they're becoming experts. And then perhaps someone who themselves is at school, maybe an undergraduate student or a graduate student struggling with a research project, wanting to bounce ideas off, an expert would then go, is that the integration would then go to this AI expert and chat. I bounce ideas off them?
- [09:55]
That's right, yes. This is particularly useful when we want assistance or help on some of these more technical domains. For example, there could be some recent advances in a particular drug or clinical trial. And then we basically would ask the agent to go to school to learn about that technical topic. And then after they've gone to the school, they would do some of their own self assessments to see how well did they learn that topic. If they didn't learn it well enough, they have to go back to school and keep learning it. But after They've learned that topic well enough, then they can sort of assist us and assist other agents in tackling those technical problems.
- [10:29]
Something I think about is so much of science is kind of unwritten, even some nuances in protocols or interpretation and peer review and critique, that often is changing a little bit. But a lot of the peer review happens privately. It's not open sourced. So is that another one that you're thinking about when it comes to limitations for the AI models?
- [10:49]
It's funny that you mentioned peer review because we actually, as a part of the virtual app, have a critic agent, sort of like a professional critic, whose job is to basically be a reviewer that sits in and participates in every meeting by the other agents and try to provide some more skeptical and feedback to try to ground to the other agents. And we do find that having this critic agent is quite important to keep the other agents more grounded and to reduce these hallucinations.
- [11:17]
The proverbial reviewer number three.
- [11:20]
Yes. Yeah, yeah. It's actually kind of funny in that when the reviewer and the different agents have their meetings, sometimes they get into disagreements, similar to how the human researchers. We can have disagreements with human reviewers, but the agents actually are very polite when they have these disagreements. So they have to somehow come to some consensus. But they're very nice about it, which is something that maybe we can even learn from.
- [11:44]
I was going to say there's this misconception that scientists don't debate and we all disagree with one another. In fact, debate that gets quite heated is the core of what it's actually like, I think, to be in research. So it's funny that it'd be kind of cool if people could just see a simulation of that, even just to see the process of science.
- [12:01]
Yeah, I think actually debates and scientific disagreements is where we see some of the most interesting and deepest insights being shared, both by the AI, but also by humans.
- [12:12]
I couldn't agree more. So you have a critic agent who's particularly good at pushing new ideas to come. What other checks are in place to ensure we're not making faulty conclusions based on an AI scientist's work? When does a real human come in or is that even necessary?
- [12:29]
Yeah, that's a great question. And I think where the AI scientists interact with the human researchers with the real world, that's really critical. So in our experiments as a virtual lab, so we have the AI scientists, for example, come up with some new designs for nanobodies and for different proteins, we still need to really do the real world experiments to make those proteins and then to test them in the real world in the lab to see do they actually work as they are expected to. I think that's still really critical because we don't want to entirely just trust on the computational predictions from the AI agents.
- [13:03]
I will say one of my big concerns with AI generally are the inherent biases encoded. We have biases in our society and they're reflected in training data sets, which means they end up in the AI systems. We've seen this in countless examples and I'm excited to talk to you about it because you've done a lot of research in this area looking at the biases. Everything from looking at dermatology, you looked at. I think you created the data set that was the first publicly available expertly curated image dataset with diverse skin tones to help improve dermatology AIs. You did analyses of AI medical devices and you even did a large study showing that large language models which are often used in AI applications, have these really harmful biases associating Muslims and violence. So this is an area where you are an expert in doing really, really incredible work. So can you tell us a little bit more about how you deal with and how broadly we deal with these biases in AI?
- [14:03]
Yeah, well, thank you very much for bringing up this important topic. It is important to be aware of the fact that these models can pick up and also can amplify biases in its training data. So in that sense it's definitely not like a silver bullet. So if there are biases in the training data which could be in the form of certain, if it stereotypes about different groups or about different scientific findings, then the model, as they learn from that training data, would also pick up on those biases and even exaggerate some of those potential negative connotations or incorrect results. And I think especially in scientific discovery, there are a lot of biases in the existing literature that people are aware of. For example, there's definitely a lot of biases in the form of confirmation bias and also a lot of biases in the type that people typically do not publish negative results. If something doesn't work, then you don't publish that. So things that you do see tends to be maybe overly optimistic that are being published. There are also a lot of biases in terms of which groups, which demographic groups ends up being studied in clinical trials, in different biological healthcare studies?
- [15:14]
Yeah, absolutely. I mean, if you're not studying representative populations, then the finding can't be applicable to all those populations. And we see women historically excluded from research Trans and gender non conforming people excluded from research, people of different racial backgrounds being excluded from research or not properly represented or confusingly represented. The dermatology AI, when you looked at, for example, a lot of dermatology textbooks are just based on white skin. And so it makes it harder to identify conditions at the same early onset in people with darker skin tones. And that's hugely horrible for the way that different patients get treated. And so that was really cool to see that you're actually trying to do something about it and create a higher quality data set to then go train the AIs off so we can have better impacts for everyone, actually.
- [16:07]
Thank you. And I think that's really important, especially as the AI models become more powerful and more flexible. We need to have better ways of continuously monitoring and what are its decisions, how it's making these different decisions and what are the potential biases in both its training data and also in the data that's producers of output.
- [16:26]
You also do a lot of work on machine learning approaches for new biotechnologies. So for example, I saw that you're looking at combining new machine learning breakthroughs in genomic technologies to study different human diseases and you used real world data and AI to broaden eligibility criteria for oncology trials. Can you tell us a little bit about that? Because it was so cool to see very real world applications and important problems that you're trying to use AI to solve.
- [16:53]
Yes, happy to. So the clinical trial one is actually closely related to, to this issue of the biases that we were just talking about. A lot of the existing standard clinical trials are often very restrictive in the sense that they're only basically testing the drugs, the medicine people that essentially come from a very narrow set of populations. Often these are males, middle aged and relatively healthy among the patients. And the reason for that is that the companies often want to go for these more sort of healthier looking patients in order to reduce their risks. But what that also means is that how well the drugs work, whether so called the Olympic athletes of the patients, may not reflect how well the drugs work on everyday patients like us, especially people from different backgrounds. So that's been recognized to be a huge challenge and limitation of existing clinical trials. What we showed in a collaboration actually with several pharma companies, is that we could use AI and use real world data, which meaning data from electronic health records to help to identify which are these patients. Maybe they are older, maybe they're more diverse, so they would not have been eligible under the original more restrictive clinical trial. Eligibility, but they could actually safely benefit from the treatments in those clinical trials. So we showed that we could actually more than double the size of the eligible patients, enrolling more women, more diverse patients, more older patients, while still maintaining the safety and the standards across these patients.
- [18:26]
Love to see it. That is so exciting to know that type of work exists. I have to ask you what you're most excited about when it comes to AI in science and medicine.
- [18:35]
I am quite excited about this idea of having AI scientists that can make new fundamental discoveries beyond what human experts could make. And if we can have these AI scientists that can make these new discoveries that could lead to a lot of new treatments or better ways of using existing treatments to help the broader population communities.
- [18:55]
I have to agree, that's pretty exciting too. Thanks so much for chatting with us, James. It was so interesting and it was great to have you.
- [19:01]
It's really a pleasure to be here. Thank you.
- [19:08]
Dr. Samantha Amien
If you've been following the headlines, you've probably heard something about MDMA assisted therapy for ptsd. Mdma, better known as Molly or a big part of what's in Ecstasy, is being studied as a possible treatment for post traumatic stress disorder. It's one of those stories that sparks a lot of curiosity, skepticism, hope, and recently some setbacks. In 2024, the FDA decided not to approve MDMA assisted therapy, at least not yet. See, ptsd, it's really difficult to treat. It can show up after someone experiences something deeply traumatic like war, sexual assault, or a serious car accident. And it often brings flashbacks, nightmares, anxiety, and emotional numbness. Current treatments include talk therapy or medications. But what about for people where that's just not enough? That is where MDMA can come in, maybe in carefully controlled clinical trials, not raves or festivals. To be clear, MDMA seems to be able to help people open up in therapy in a way that traditional treatments don't. MDMA leads to an excess of chemicals like serotonin, dopamine, and norepinephrine in the brain. Those are the feel good social bonding and mood regulating chemical messengers in the brain. What that does is create a kind of therapeutic sweet spot. People feel safe, less fearful, more connected. And the idea here is that makes it easier for them to face and process traumatic memories without being completely overwhelmed in therapy. The rationale is therapy can then go deeper and be more effective. One of the biggest moments for MDMA assisted therapy came from a phase 3 clinical trial. In that study, about 71% of participants who got the MDMA assisted therapy no longer met the criteria for PTSD by the end of the trial. Compare that to just 48% in the placebo group. We're talking about people with moderate to severe PTSD here, which often comes along with other challenges like depression or substance use. Despite the results, the FDA said not yet. In August of 2024, they raised a few concerns. One was safety, things like cardiovascular risks or the potential for some people's mental health symptoms to actually get worse. They also flagged issues with how some of the earlier studies were conducted. Specifically, they pointed out that MDMA's noticeable effects made it nearly impossible to blind participants, which basically means people could tell whether they were getting the real drug or a placebo, potentially skewing the results because their expectations might influence how they respond. There were even some ethical questions about how data was reported. So the fda, they want more evidence.
- [21:48]
Dr. James Tso
The company behind the treatment and many.
- [21:50]
Dr. Samantha Amien
Researchers in the field are already working on next steps. They're planning more studies and more back and forth with the FDA to try and get this therapy across the finish line. And it's not just mdma. There are some groups looking at hallucinogens like lsd. And then there's a whole wave of interest in using psychoactive compounds to treat other mental health issues. Psilocybin, the active ingredient in shrooms, is being studied for depression and anxiety, CBD and thc. They're already widely used by the public for help with sleep, stress and anxiety, but their evidence base, beyond pain relief, is a lot less robust. What we're seeing isn't just a shift in medicine. It's a broader rethinking of how we view substances that were once illegal and heavily stigmatized. These compounds are being reexamined not as threats, but as possible tools for healing, especially when used with intention and under professional guidance. Thanks for listening. Before we head out, I have one question. What are you curious about? If there's a field of study you're into or a subject we haven't gotten to yet. Please just leave us a review on Apple Podcasts and let us know. Tell us how much you love the show and leave us a five star review. We'd love to see that for Warner Bros. Discovery Curiosity Weekly is produced by the team at Wheelhouse DNA. The senior producer and editorial correspondent is Theresa Carey. Our producer is Chiara Noni. Our audio engineer is Nick Karisimi and head of production for Wheelhouse DNA is Cassie berman. And I'm Dr. Samantha Yamin. Thanks for listening.
- [23:32]
Christy Lee
If you like detailed and immersive audio storytelling, you might like Canadian True Crime. Praised by listeners as thoughtful, well researched and empathetic, Canadian True Crime began as a passion project by Aussie Canadian host Christy Lee. With curiosity and a keen eye for detail, Christy carefully unravels the full stories of some of Canada's most compelling cases, going beyond the superficial to uncover the real story behind the crime. Find Canadian True Crime wherever you listen to podcasts.