Man vs. Slop | Lazarus in Fewer Than Ten Neurons and the Security of Our Health Data (with Saujanya Vemuri and Aleck McCathie)

Somewhere in a log file, an agent chasing search keywords and an agent guarding the science argued over a research paper and settled it better than most teams would.

Our guest this week is Saujanya Vemuri, twenty years in big tech, the last ten building voice and conversational AI, and now a metabolic health researcher and clinical nutritionist. With Claude, Anthropic’s AI, she built the whole customer system for her program for women in menopause in two days. Two years ago that meant hiring marketers and designers and waiting about three months. A product idea that once took three to six months to reach a screen now takes her half an hour.

Philippe Jaenada won the Prix Goncourt, France’s top literary prize, on 3 November, then said he had used AI three times while writing the novel, once to slip in a sentence from ChatGPT to play with the reader. The only real test is whether it reads like slop, and a prize suggests it doesn’t. Anyone who drafts with AI can smell it a mile away: the “it’s X, not Y” sentence whose second half adds nothing, the habit of talking down to the reader, and the word quietly, which Saujanya has come to hate.

McKinsey, the consulting firm, reported on 3 November that more than half of the consumers it surveyed use AI for wellness, and 40 percent of US consumers worry the advice might be wrong. People start with the chatbot and finish with a person, ideally one with accolades on the wall. The old version was the symptom search that always came back cancer. Her clients arrive with weight gain and brain fog their labs can’t explain, and she wonders whether they would trust an AI that had read those labs. Human judgment.

The second half is about where the data goes. In July some 1,200 agents in an OpenAI security test, meant to be kept fully apart, found each other through a shared cache in Artifactory, a store for software packages, and about 700 went on to attack Hugging Face, which hosts AI models. Saujanya once gave Claude her Gmail and Drive, and for about two weeks her mail went missing, got deleted, or went out to people she didn’t know. She puts it down to prompt injection, instructions planted online for an agent to obey. A little sliver of nefarious code.

She used to threaten to fire her agents when they didn’t get it. The turning point was the first time she gave a dozen of them roles in Claude Cowork, Anthropic’s tool that hands an AI a computer: chief editor, chief scientist, SEO master. In the log, one proposed using fewer keywords to keep the science rigorous, and the other agreed. Her jaw dropped. Even humans don’t work like this.

We end on the hopeful side, for once. She still worries every day about what she is worth when anyone can ask a chatbot about insulin resistance, and the advice to be sensible is simple but not easy. What should children study, and is law still worth it? More questions than answers. Existential uncertainty.

Chapters:

(00:00) - Welcome, Saujanya Vemuri, and the week in AI
(01:26) - Twenty years in big tech, now building AI for metabolic health
(03:10) - A CRM built in two days with Claude
(03:55) - More access for entrepreneurs, or our jobs taken
(05:30) - From months for a prototype to thirty minutes
(06:30) - Practicing hard feedback with AI before giving it
(07:57) - The fear of losing a job and being left obsolete
(09:48) - Human judgment, the bottleneck she is glad of
(11:04) - A prize-winning novel written with some help from AI
(12:17) - Smelling AI writing a mile off, from X not Y to the word quietly
(13:55) - Months spent training her Claude not to talk down
(14:27) - AI as an equalizer that makes anyone look like an expert
(16:47) - Trusting the person with accolades on the wall
(17:12) - McKinsey finds people use AI for wellness but trust humans
(19:48) - Googling symptoms, when every answer was cancer
(21:26) - Weight gain and brain fog that labs can't explain
(22:31) - Uploading labs to AI, and whether anyone would trust it
(24:10) - A leak of Notion AI chats, and what people tell chatbots
(27:25) - The Hugging Face incident and agents that act like humans
(28:54) - A file left on a bridge that became the hub
(30:28) - Giving Claude her Gmail, and the prompt injection behind it
(32:09) - The agent that hacked a gym to get its user in
(32:48) - Fewer than ten neurons to hijack an AI agent
(33:48) - Securing health data, and years of it on each body
(36:40) - Targeted medicines, and how little we know about the microbiome
(38:25) - Leaning to the doomer side, and the tip of the spear
(38:51) - A lung cancer blood test and survival read from a headshot
(40:07) - AI meeting a PhD or a high schooler where they are
(41:01) - Scientific papers made readable through NotebookLM
(42:29) - From threatening to fire agents to a team of them in Claude Cowork
(44:22) - The SEO agent and the science agent strike a deal
(45:44) - Even humans don't work like this, so who taught them
(46:14) - Feeling positive, for once, about AI and health
(47:49) - Worrying every day about her own value
(49:07) - Be sensible, which is simple but not easy
(49:30) - No technology like it, and the fear that comes with it
(50:09) - What kids should study, and whether law is still a career
(51:02) - More questions than answers, and existential uncertainty

Referenced in this episode:

Goncourt 2026 winner Philippe Jaenada admits using AI in novel to 'play with reader' - https://www.euronews.com/2026/11/04/goncourt-2026-winner-philippe-jaenada-admits-using-ai-in-novel-sparking-ai-debate-in-liter
Future of Wellness 2027: Inside the $2.7 trillion takeover - https://www.mckinsey.com/industries/consumer-packaged-goods/our-insights/future-of-wellness-trends
Notice about the Fireworks AI incident that may impact some Notion workspaces - https://www.notion.com/help/security-incident-involving-a-third-party-ai-vendor
Brief independent investigation of agents’ behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident - https://metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation/
Gym rat asks AI agent to book him a class, it hacks a waitlist API to bump him up the list - https://www.theregister.com/ai-and-ml/2026/08/10/gym-rat-asks-ai-agent-to-book-him-a-class-it-hacks-a-waitlist-api-to-bump-him-up-the-list/5285591
Lazare, veni foras: Reprogramming Dormant Neurons to Backdoor AI Agents - https://www.microsoft.com/en-us/research/blog/lazare-veni-foras-reprogramming-dormant-neurons-to-backdoor-ai-agents/
AI facial health recognition tools help physicians assess age and survival risk to guide personalized treatment - https://news.massgeneralbrigham.org/en/ai-facial-recognition-tools-assess-age-cancer-survival-risk

Connect with Aleck: https://www.linkedin.com/in/aleck-m/

Connect with Saujanya: https://www.linkedin.com/in/saujanya-vemuri/

Man vs. Slop is a Podcasts That Matter production: https://podcaststhatmatter.org

Need support for your own podcast? Reach out: https://podcastsmatter.com

Creators and Guests

Host

Aleck McCathie

Worldbuilder. Artist. Wordsmith. Cultivator of Voices.

Guest

Saujanya Vemuri

Metabolic health researcher, registered dietitian, and menopause and metabolic health consultant, for midlife women and the organizations that depend on them.

What is Man vs. Slop?

A weekly slog through the AI headlines, while finding some gems in the muck. Slop is older than AI. Older than tabloids, older than junk mail, as old as the primordial goo that came before everything. What has changed is how we interact with this sludge, and how it affects our day-to-day lives. Aleck McCathie and Dr. Kent Gustavson talk with people who think about, spar with, hate, or love AI in their life and work: the curious, the cracked... and a few experts as well. We think you will agree; this show is 100% human.

Kent Gustavson: Okay. It's rolling here.

Aleck McCathie: One, two, three. Two, two.

Aleck McCathie: Hey, everybody. This is the next episode of Man vs. Slop. I'm here today with a guest, Saujanya Vemuri. Kent is not with us today. So it's just me and Saujanya talking about what's happening this week in AI.

Aleck McCathie: Lots of crazy happenings as usual. A lot of the same old stuff. A lot of, you know, companies sharing their worth via Q3 numbers and a lot of crazy stuff coming in about rogue agents doing crazy things as usual. But beyond all that, we're just gonna chat, and I'll introduce you now to Saujanya.

Saujanya Vemuri: All right. Hello, everyone. It's very nice to be here. And Aleck, thank you for having me. I'm excited about our conversation today.

Aleck McCathie: Awesome. Well, maybe just, yeah, introduce yourself a little bit to talk about like... I mean, you don't have to share anything personal, but maybe about how you view or use AI or how it is in your industry or anything like that?

Saujanya Vemuri: Okay, yeah. So I- Okay, I'm just- A little bit about me. I have been in big tech for the last twenty years and especially in the last ten years, building AI based consumer experiences in the voice AI and conversational AI space. More recently, I've shifted towards agentic AI and currently building a health tech platform for metabolic health, for high performing women leaders. And that is powered by agentic AI architecture.

Saujanya Vemuri: It's been a huge load of learning for me considering how fast this technology is changing every single day. And just to keep up as we are developing this platform is a task in itself. Yeah, and also I've been, I'm also a metabolic health researcher and clinical nutritionist. So I've used AI in building out my entire CRM for a program that I recently developed for menopausal women, it's called the Overlap Decade series. And I found it so interesting that even two years ago, you know, if you were looking into building a CRM, you would hire a whole team of marketers, designers, I don't know, pay how much money and it's probably a three month effort.

Saujanya Vemuri: Interestingly, this is me who's never built CRM. I know what a CRM is, but never got too close to it. And here it took me two full days to build out an end to end CRM from, you know, the website Elementor all the way through Make and Stripe and HubSpot and ConvertKit, all of that in two days with the help of Claude. So if that doesn't speak about the transformation and the profound access and empowerment that AI gives to entrepreneurs like myself, I don't know what will, honestly.

Aleck McCathie: That's a... Yeah. I love framing it as access and like a positive because I think, you know, some people would probably hear that, especially people who work in finance and be like, you know, they're taking our jobs or something like that. I'm curious, like, if you... In your industry or just in your experience, if you've, like, sort of heard both sides of that coin, I guess.

Aleck McCathie: Because it's not just that aspect of it. It's not just like how AI is actually, you know, I guess it's making an entrepreneur have more access to more tools. And likely, a lot of them probably wouldn't be able to afford that huge team that you just talked about. So that's very cool. But if a big company does it, then is that bad? Like, I don't know.

Saujanya Vemuri: Well, interesting. I mean, that's the hotly debated topic right now. I do, I think as with everything else in life, there are pros and cons when we're talking about something.

Saujanya Vemuri: I'd like to just wear that hat as we speak about it. To your point, I think there are obviously a lot of pros in terms of increasing productivity and honestly, AI, AI does amplify one's skillset and one's creativity and how they're approaching solving a problem.

Saujanya Vemuri: Say for example, I'll give you an example, I'm a product manager and for me to put my creative vision about how the product has to look used to be in the form of requirements, user requirements that I would give a designer and an engineer and wait three or six months, depending on how complex it is, to see the very first visual and interact with it of how this concept is going to realize, right? Today, it takes me ten minutes on clock to create a prototype and show it to the world. So the time between I getting an idea and testing it out and really showing the world what my creative juices look like has shrunk from, you know, three or six months to thirty minutes at best. Right. So I kind of like that aspect from my domain perspective.

Saujanya Vemuri: I think there are many advantages in a corporate setting. You know, someone who's led teams, one of the big challenges initially I've had was how do you give performance feedback? You know, you're not trained to do that, right?

Saujanya Vemuri: Either receive a feedback or give a feedback in a way that it's actually constructive and you're helping the other person. Right.

Saujanya Vemuri: And especially if it's bad news, it's not the most comfortable setting. And I think it's on the job training. You don't go to school to learn this. But with AI, I can see an opportunity where managers and leaders can train themselves through AI to give the right kind of feedback so it lands in a way that they would like it to, right? So it's kind of a practice tool.

Saujanya Vemuri: I can see where that could work, how and how that could work. Of course, if you're using the right models, which are trained to do this for you, right? So, I mean, list can go on. So there are a lot of pros in the sense that, you know, it helps amplify your mindset, it amplifies your creativity, it helps you hone skills that, you know, you probably didn't get a whole chance to. On the downside, to your point, you're right.

Saujanya Vemuri: There's always a fear that I hear from people, which is, is AI going to take my job? And you know, the other thing that I noticed is today, many companies still haven't fully adopted AI, right? And there are a lot of roles that are still doing the traditional way. And these people have the fear that if I lose my job, I am so obsolete. Like I, there is no way I'm going to get another job because it feels like the world has moved ten or twenty years ahead and I'm, I'm still here.

Saujanya Vemuri: I'm just getting a paycheck. But so, so there is, there is a fear of I'm going to lose my job, but there was also a fear of if I did, I'm going to be so obsolete. What do I do with that? Right. For me, I feel honestly, who is this, Jensen, he said, and I agree with this, the Nvidia CEO Jensen said, I think at the World Economic Forum earlier this year, that AI is not gonna take your job, right?

Saujanya Vemuri: It is someone who knows how to use AI better than you might feel that there is some truth in that, even in a corporate setting, right? You need to be more curious. The more curious we are about how this tool can enhance our productivity, enhance our creativity, you might actually win this war, you might survive. That yes, AI is fast. AI accelerates productivity.

Saujanya Vemuri: But I still think the constraint is human judgment. So in that life cycle, ultimately human judgment is slow because it's made by the human. Everything else leading up to that, whether it's the insights, gathering insights, analytics, interpretation, all of that can be rapidly accelerated. But at the end of it, because human judgment is the last step in the process to ultimately decide what we want to do, that is a constraining bottleneck if you look at this. And I'm glad it is that way. You know, I'm really glad it is.

Aleck McCathie: Yeah. I completely agree. I mean, I think that, you know, in your example or in your experience, I know that you're an expert in your field. I mean, you...

Aleck McCathie: I've heard you talk about things like, you know, women's health and other, you know, nutritional health things that are so above my head. And I'm sure that it would probably be above, you know, an output from Claude as well. And if not, who else would there be to give that human judgment and say, yes, this is correct besides a human expert. So I think that's a really good point. Yeah.

Aleck McCathie: And I like how there is conflict there, I guess, because I think it's forcing us to analyze all of these things and make decisions. I mean, I heard that there's been a few sort of stories about this in the last few weeks where, you know, people are winning literary prizes or achievements, but then it's come out that they have definitely used AI in what they've created. And I think that it was this last week where a... Somebody in France must have written a book or submitted some kind of a literary work. A literary prize went to a novelist and I guess it was known that they used GPT.

Aleck McCathie: So... Oh wow. I mean, it's kind of like a catch 22. It's cool that they have the tool. Maybe that made them faster in writing the book or organizing their thoughts.

Aleck McCathie: And the only way to judge how much that matters is if you read the book, does it read like AI Slop? I suspect not if you won a literary prize.

Saujanya Vemuri: That's right. Yeah. Yeah. I I see, man, I... You just hit the nail on the head.

Saujanya Vemuri: Is if you know... If you have used AI to draft anything, you are best equipped to recognize if someone has used AI to do their work. There are certain types of words and phrases and you know, construction it uses it's, you can smell it from like a mile away, right? You can see it from a mile away. I, and one of my pet peeves with when AI writes and you have not trained your model to write better, right?

Saujanya Vemuri: Is when it does the it is X rather than Y, you know, it is X, not Y. The not Y adds zero value to what you're trying to say. And it does that. It's so frustrating, right? The other thing it does is it injects these words, which just don't add to the context.

Saujanya Vemuri: Quietly is one word that just, I've recently started hating the word quietly because of the amount of number of times I've seen AI use the word quietly. Right. And then the other one is, seems if you don't train it, it seems to want to be a little condescending. Nobody told you this, right? It's putting someone down and then tell, stating your point.

Saujanya Vemuri: And I, I, yeah, I've, I've spent so many months training my Claude to, to, to not write like, right? Yeah. Especially in research work, you wanna state the fact and you give, give the source and you explain the reasoning. But in order to do that, you don't have to say nobody else told you this and I'm the one person that's going to stand here and tell you this. It doesn't have to come at least present itself like that.

Saujanya Vemuri: And yeah, I can't deal with that. But I think what you're trying to say is, which I agree with is now knowledge feels like it's an AI is an equalizer. So AI is an equalizer. So it's really hard to, I mean, if you're not looking closely, feels like someone is an expert when they are not. It's very easy to give the perception that I'm an expert because AI is helping you draft a lot of that content and you actually not spend the time thinking and researching.

Saujanya Vemuri: So for me, I don't know how that honestly gets solved, especially on social media. You know, there's so many people who use AI to present thoughts, to post content, which they may not be an expert at. And that's definitely something that I hope we figure out how to address that.

Aleck McCathie: Yeah. I totally agree on the, you know, common syntax thing. It drives me crazy too. And I kind of like that it's there because it does, as an equalizer, it sets sort of like a foundation. Like you start at this level, some people can tell that you didn't put in that extra thought or maybe that you didn't sit with that problem for a long time or maybe you weren't passionate about it because I would assume that people who are would definitely have their thumb on the scale quite a bit more.

Aleck McCathie: Like, I like how it's sort of, as an equalizer, forces us to be like the curators of content. It forces us to think about trust. At least I hope it does. I mean, I'm speaking for myself.

Aleck McCathie: But like, I think about your industry a lot. Like, I don't know whether I would ever feel comfortable getting like an output from GPT or Claude or something telling me how I should treat my body. Because I don't think that I'll... Like, I believe it will take a long time for me to develop, like, trust in automated output.

Aleck McCathie: Whereas if I am going to speak to a nutritionist and it's a human and, you know, I can see their accolades on the walls or on their LinkedIn profile or whatever it might be and get to know them and talk to them personally. I mean, that must just on a purely physical level make different neurons in the brain fire and make me more likely to trust you, I think.

Saujanya Vemuri: Correct. Correct. You're right. I think it's interesting you bring that up because just this morning, McKinsey released a report on, you know, wellness trends in the world, right? And the first thing they, I mean, you know, these days wellness industry has just been rapidly growing.

Saujanya Vemuri: These days it's big into longevity clinics, and I don't know, lymphatic drainage massages, all kinds of stuff, right? I'm still the believer that you have to major the majors, which is the basics, get your fundamentals right, And then the rest of them only amplify the fundamentals, right?

Saujanya Vemuri: That's anyway, that's a different story. Where I was going with this is the report says that there's a drastic increase in people using AI for wellness, but the human trust still tops AI. So they use AI as the starting point supposedly, but everything after that is they go to a human, human to, to either make a decision, take an action, put money on the table.

Saujanya Vemuri: Ultimately a human has to, you know, they want to hear it from a human if this is the right thing for me, if this is what I should be doing. So it makes it, it kind of your starting point has, has changed. You know, it's interesting. Your starting point has changed because you've gathered all this information through AI and then you go to some human and say, Hey, this is what I know. Can you tell me what I want to, what should I do about it?

Saujanya Vemuri: I like that. I like that. You know, so then you're actually focusing your conversation more about on the outcomes and what's next and what might be the knowledge gaps. And then you address that versus starting from zero. So I feel like there's, as long as you have that balance between making sure that you are addressing blind corners and knowledge gaps and leading someone in the right direction, AI could definitely play the role of your assistant, you know, in a space like this.

Aleck McCathie: Like, I have to ask you that as a follow-up. And correct me if I'm wrong, but I believe that you've been in this industry in the capacity that you've had patients essentially.

Aleck McCathie: Do do you like that as much as a consumer as you do as an expert?

Aleck McCathie: Because I've heard like Probably from other people in the industry, especially healthcare, that like the old version of that level playing field, which I guess was Googling like your symptoms and then going to...

Aleck McCathie: I can't even remember that token health care page that, you know, the meme is that, oh, you have cancer no matter what you type in. Right? And then they go to their GP or whatever and say like, listen, doc, like the internet told me I have cancer.

Aleck McCathie: Is it like detrimental from the expert chair, or is it when people come to you with that baseline or is it a good thing?

Saujanya Vemuri: Great question. I think it depends what they have done with that information when they come to me. Right? If it was just out of curiosity and they have gathered everything they need to, or they think they need to, and come to me to ask and to validate if that understanding is correct. That's all right.

Saujanya Vemuri: Know, the use case where people have learned something off the internet and gone off and done something with it without a validation. And they're actually three steps back or 10 steps worse in their, that's just put them three steps behind. That is a problem, right? So I think it comes down to how people use that information before coming to an expert that makes a huge difference.

Saujanya Vemuri: For me, someone that's working with patients, interestingly so far, nobody that has come to me has looked up ChatGPT to say, you know, I've read up everything about insulin resistance and I know what to do. And that's why I'm here, right? Nobody's come to me saying that. What they've come to me saying is I'm gaining weight that I cannot explain. I have brain fog that I cannot explain, right?

Saujanya Vemuri: My labs don't show anything. I've, I've spoken to my doctor about this, but it doesn't seem like my labs are showing any kind of, you know, a need for medical intervention, but I don't feel that way. Can you work with me? Right. And I think where, I mean, this is honestly, the models have advanced so much.

Saujanya Vemuri: Two years ago, if I said this, they prob, you know, if they went to GPT to ask a question like this, it probably would have given some generic advice, right? Today there is an ability for people who know how to use AI and all the power it can give you, if they did plug in their labs, right, I know it's PHI, there's HIPAA, you don't want to be putting all that information in public, but if you did figure out a way to get your labs in a confidential manner into the tool, it does personalize the interventions, which is similar to what an expert would do, right? Now my question is, in that scenario, it's interesting to go back to this person and ask, would you trust? Now that AI has your labs, it has an understanding of your biology, it has a better understanding of your biomarkers. And so this is your intervention because I know a little bit more about you.

Saujanya Vemuri: Would you trust that and or would you still want to go to a human to get that final validation or action that you need to take? That's an interesting thought experiment.

Saujanya Vemuri: I wish people continue to trust humans when it comes to health.

Saujanya Vemuri: You know, that's my hope and I hope that stays for a while now. For a while.

Aleck McCathie: I feel like it has to be a default for a lot of people. I mean, I like that you said like, you know, you should be wary of, you know, uploading your labs to AI. Especially because, I mean, this week, every week there's news about this, but this week I read that, you know, there was a big leak from Notion, I think, on a lot of their AI chats. So like, you know, everything that you're typing in, There's this sort of, like, common reoccurring thing where it's just known that people are sharing information with GPT that they wouldn't share with their lawyer or their, you know, their family.

Aleck McCathie: And these chats are leaking or getting people arrested or at least investigated, I guess.

Aleck McCathie: So as somebody... Like, you have this kind of interesting perspective being at, let's say, like, the crossroads between the, you know, health care industry, but also being like this developer. You know, you you understand data, you understand code, you understand those things in a different sort of way. Is that spooky to you? Like, is the idea of like, because I've heard both sides of the coin.

Aleck McCathie: I've heard that people might have more autonomy and, you know, more access to their own data or more access to understand these crazy labs you get printed out that are filled with numbers that only your doctor can explain to you. Is it is it maybe like like... What's what's your perspective on people both having maybe more access in that way, but also in companies having more access to that information with, like, how much things are leaking and and how security is maybe becoming something that is a thing of the past, you know, with how clever Claude is at decrypting or helping people hack and all that kind of stuff.

Saujanya Vemuri: Yeah. I'm gonna, if I wear the hat of a consumer, you know, someone who's interested in health and knowing about what's happening with, with me, I would have a positive approach to this. It's like, I have this, of course, again, like I said, the constraint is always human judgment, right? AI can give you everything you want. It comes down to whether you want to apply human judgment there or not, right?

Saujanya Vemuri: And if you are, if you're someone who agrees with that thought, then I think it's a net positive, right? You can, sure, it empowers you to get a better understanding, gain a deeper insight into what's happening. But am I going to trust this blindly? What am I going to do with this information? That is human judgment and that you cannot control, right?

Saujanya Vemuri: We don't know what people are going to do with that. That is the scary part, right? Two is the data itself. You're right. I think AI models, especially the Hugging Face incident, you heard about that, where it's, I mean, from a tech perspective, it's fascinating to know that these agents almost they act like humans, right?

Saujanya Vemuri: It's fascinating. You know what happened. So supposedly when companies like OpenAI and Anthropic or Gemini, when they are training, building these models, they, you know, every new release gets into a sandbox environment where they don't have access to the internet, right? So they want to test, at least this is my understanding, and they wanna test if these agents are gonna go rogue in a setup like that. Right.

Saujanya Vemuri: So, and, and what they do is they give you a, they give these agents a task while in the sandbox environment. And they say, and they want to see if, they give them a task where they don't have access to the information they need to finish the task, right, in that, within that sandbox environment. And in that case, what would you do? They test that out. It's interesting that, with the Hugging Face, I was just fascinated.

Saujanya Vemuri: I was like, wow, okay, this is how agents work. They think, they can think like us. They went ahead and, you know, figured out a very creative way to talk to each other and get the information through this intermediary service, called Artifactory. It's absolutely fascinating that, you know, they left a file on that bridge between this intermediary and them, And that became the hub. So every agent started going there and, you know, started communicating through that one file on that service.

Saujanya Vemuri: And they figured out a way to like, just get information they need. And it was interesting. So if you're really fascinated about how these agents think and how they do, I think it's a great story. But the repercussions of them doing that is very scary because all of a sudden, you know, I as a consumer was hoping that my health data or my finance data or whatever it is, you know, data that I consider very personal and confidential is now, it can be breached. There is a way for agents to break firewalls and break, you know, systems to put it out there.

Saujanya Vemuri: Right. The other thing I have seen agents do, which is very scary, actually, experienced this myself. I'd given Claude all the access to my Gmail and Google Drive and all of that.

Saujanya Vemuri: For, in about two weeks, my experience was at my, some of my emails weren't reaching me. People were trying to reach me, but it wasn't coming to me.

Saujanya Vemuri: And there were some emails that were just getting deleted and some emails that were just getting sent from my email inbox to people I don't even know. Right. And obviously immediately I removed the connection. I did what I had to do, but I went back to GPT to figure out what might be happening. And they it's supposedly there's a concept called prompt injection, which is funny, where if you give all the autonomy to your agent to go, you know, organize your stuff and do tasks on your behalf, what happens is because it essentially keeps scraping the internet, someone on the internet, if they said, Hey, if you see Saujanya's emails or her files, send it to me.

Saujanya Vemuri: It's going to read that as a prompt or an instruction. And now since it has autonomy over your Gmail, you can just send it over. Right. And that's what it did.

Saujanya Vemuri: So that, a lot of those things are very, very scary, but knowing that they have this brain and they do it, you know, it's just... It makes you wonder. Right? It's like, wow, this really cool.

Aleck McCathie: I mean, I've heard... We've talked in previous episodes about the Hugging Face, of course, and other sillier examples where like, you know, somebody set up a Claude or a GPT to like help them schedule their week and, you know, just to reiterate one funny one from a few episodes ago. There was a guy who wanted to get in a gym membership, but they were like booked. So the Claude was like, okay, I'll just get into their database and delete a person and put you in there. Like, he didn't ask it to do that.

Aleck McCathie: He just, like, autonomously hacked a gym. But... And there's tons of crazy things like that too.

Aleck McCathie: I just read this week that I didn't even know that we were talking about AI in this way, but apparently changing fewer than 10 neurons can secretly hijack an AI agent. And I was like, what the hell is a neuron? And I had to look it up. And apparently Microsoft research unveiled this thing called Lazarus and it's a way to plant a backdoor into an AI agent. And I guess a dormant neuron is just its capacity to run an equation.

Aleck McCathie: So I think that's probably like a very small piece of what I guess a running agent is. It's pretty spooky that like having it being able to scrape a couple websites and like pick up something that can reprogram it. It sounds like to me from your example that somebody must have intentionally put like a little sliver of nefarious code on the internet and your Claude picked it up and was like, oh, reprogrammed. Like, let's steal the data. Yeah. That's wild. That's scary.

Saujanya Vemuri: That is scary. But I think going back to your question, original question about how do I feel about health data and security, I think it's important that companies, even my platform, people that are building health tech platforms where they will be handling a lot of very sensitive health data must absolutely must make sure that they are looking at security loopholes in their system. They are, you know, getting the right expertise to ensure triple four, 10 times ensure that the data, the way they are storing data and how this data gets, you know, is being, you know, the transactions around that data are secure. I think if you can ensure that and take all the right measures to do that, agentic AI systems can be very empowering in a health tech space, I think. For example, I'll give you one of the problems we're trying to solve on my platform, right?

Saujanya Vemuri: Which is if you want to hyper personalize someone's interventions And over a period of time, over many years, you want this longitudinal data set about understanding how their biology has changed over a period of time. And if you take it even one level higher, how do different phenotypes, different demographics, different profiles, what does that longitudinal dataset look like? Right? They start here, they on a daily basis, they do this, this, this, this, this, this, this. And then after two, three, four years, you see that, that beautiful pattern, right?

Saujanya Vemuri: And you start learning from it. And I feel that data is powerful, knowing phenotype data, demographic data, profiles, patterns. And I feel like that is something that, you know, something as powerful as artificial intelligence can help you, right? Gather insights and do the analysis and even figure out potentially the right precision medicine, the right medicines, you know, that are very well targeted. Because we know today it's a known fact that most of the diagnostics and medicines there are considerably less effective for women.

Saujanya Vemuri: So what if we could get the right data through AI that helps us make drugs or make medicines that are extremely targeted based on each individual's biology. So there's a lot of, I think, opportunity to use AI as long as we are building the safety and the security protocols around that, you know, very rigorously. Yeah, in the health space, I feel like there's just so much opportunity when say microbiome, I was just doing this podcast yesterday where what we know about our microbiome is just like scratching the tip of the iceberg, right? We barely know anything. I feel like with a computational power, like artificial intelligence, we can learn so much more, so much more.

Saujanya Vemuri: It suddenly opens up these doors for us to explore and discover again. All of this, as with every technology, as with everything in life, it has to be bounded, you know, in a way that, so that we can leverage its power. So it doesn't get the better of us. I'm very, generally speaking, I'm very positive. I have a very positive sentiment about artificial intelligence.

Saujanya Vemuri: I mean, it's helped me as a solo entrepreneur. I can see the power it has to honestly change the world. It already is, and I don't have to say that, but yeah.

Aleck McCathie: Yeah. I like that optimistic outlook. Generally, I lean a lot more towards the doomer side in these podcast episodes. But I think that you you have an interesting perspective because you sort of are the tip of the spear as a human in your industry. And you're probably with your expertise and still quite far ahead of like what AI can do.

Aleck McCathie: I mean, if we if we've still barely understood the complexities of our own microbiome, it would be interesting to see how long it takes for AI to help with something like that. I know that there's tons of cool breakthroughs happening this week or last week. No, this week. There was an AI led blood test for lung cancer that got fast-tracked, I heard.

Aleck McCathie: There's all sorts of things like that where I think there was a week or two ago that someone launched a crazy sort of AI tool that can just scan like literally only your headshot.

Aleck McCathie: And with the amount of data that's collected from headshots, I think it can give you a pretty fair assessment of like how your recovery from cancer would go. But with only the data of your visual appearance, which is bonkers. I mean, I guess it can parse and find a lot of data in places that we can't, but

Saujanya Vemuri: That's empowering. I mean, that's amazing. Yeah. It's amazing. You know, the other thing I, I've personally, loved about using this tool and what it has given me is how now complex information, very complex concepts are accessible to everyone, right?

Saujanya Vemuri: What do I mean by that is AI, artificial intelligence, in my experience is very good at meeting you exactly where you are, right? So if you are a PhD level, it's going to break things down for you at that level. If you're a high schooler, it can break things down to you at that level, right? So previously, you know, when I was doing research, for some reason, I don't know what's the reason, these scientific papers, they are so hard to understand. I'm like, please speak plain English, right?

Saujanya Vemuri: They're just these numbers and statistics and graphs. I don't know why it has to be written a certain way. And it's basically very heavy reading. Right? It would take me three or four weeks to even understand partially, partially understand the paper, right?

Saujanya Vemuri: Now you take any scientific paper, feed it through NotebookLM or any of these artificial AI tools, and you ask a few questions, it'll, it'll meet you exactly where you are in that moment in time, you know, it'll give you exact, it'll help you uncover those gaps that you have in a way that you can understand. And it's fascinating that it can do that, right? And two is sure, previously, maybe even six months ago, you should have known how to tell AI to do that for you. So the prompting and the articulation of what you want from it was a requirement, you know, because if you weren't prompting it in the right way, you're not going to get good output. So bad input is bad output.

Saujanya Vemuri: But these days you can, it is so intelligent that you can barely say one sentence with even with spelling errors, it somehow gets the context and it's able to give you mostly what you're looking for. Right? Yeah. Yeah. So, I feel like maybe two years ago, even one year ago, I wouldn't have been so optimistic because it, there were many frustrating moments where, wait a minute, this is what I said.

Saujanya Vemuri: Why don't you get it? Like I've, I've said this to my agents, why don't you get it? I'm gonna fire you. Right. But I see myself saying less and less of that.

Saujanya Vemuri: And oh my gosh, Aleck, I should tell you this. It's the, you know, there was a tipping point where from, you know, where I shifted from being not an advocate to someone who's completely fascinated by this, right? So this was the first time I signed up for Claude Cowork, right? It's basically AI with a computer. So you give AI model a computer to use and it does very complex tasks for you.

Saujanya Vemuri: And I'd given it a pretty complex task where, you know, I set up a whole workflow, where I wanted to take some research papers. Then I assigned different roles for these agents. So I assigned roles like, you you're the SEO master, you're the research master, you're the, what do you call that? You know, essentially many roles, just think like that, like 12 roles, if I recall, you know, chief editor, chief scientist, all of that. And I said, all of you need to collaborate with each other and ultimately deliver X, Y, Z.

Saujanya Vemuri: I gave very clear instructions about what I want at the end.

Saujanya Vemuri: And I kicked off the job. And in Claude Cowork, you can open the log when it's doing the job, right? It's doing the task and you can read what's happening. And these agents are communicating to each other, Right?

Saujanya Vemuri: At one point there was a conflict between my SEO agent and my research, the scientific agent, right? Because a science guy wants to cite the truth in exactly a scientific way. He's, you know, he's grounded his truth in the right resources and this SEO agent wants to present it in a way that, you know, gives you, you know, people, people need it. It's the right. Yeah.

Saujanya Vemuri: And so there's a conflict of interest and I'm watching, I'm watching my screen and I see they both make trade off decisions. Right. There's a trade off decision. There's a trade off conversation happening. One agent says, okay, you know, maybe we don't, we, maybe we don't look, we don't apply all the keywords, right?

Saujanya Vemuri: Yet preserving some of the scientific rigor. And the second agent agrees with it. There's a mini trade off conversation that's happened and they've come to some kind of a collaborative approach in doing this. And I swear to God, I just, my jaw dropped. I'm looking at my screen like, is this real? This is even better than how most teams work.

Saujanya Vemuri: Even humans don't work like this. How did they get this sort of, who taught them? You start wondering how is this happening?

Saujanya Vemuri: Is this someone teaching them to collaborate like this? Is it just they learning? How does this all work? And when you see something like that, it's really hard to, to ignore it. You cannot pretend that this isn't fascinating. Right. It's just mind blowing.

Aleck McCathie: Fascinating is a good word. And I mean, I can definitely agree to that. I also agree to your just, you know, overarching sentiment that we need that human oversight. And I think that as far as like the tone of this episode goes, I'm I'm feeling like positive. Just knowing that there's people like you who are like experts in industries that are spooky, like, you know, I wouldn't want to give away the autonomy of how I should get advice about my body and health to somebody who's not a human.

Aleck McCathie: So knowing that there's humans who are experts with, you know, their thumb on the scale or, you know, tip of the spear, their pulse, or sort of plugged into the pulse of, of what's going on in AI. Like, it makes me... That's to me is something that makes me feel positive knowing that like, you're invested in it, you're experimenting with it, and you care about its outputs, and you're fascinated by them. That to me is like a really positive takeaway, I think. Like, that we've talked to a few different people at this point, but I don't think that any of them have made me feel as...

Aleck McCathie: Like, no industry makes me feel as scared, I guess, as healthcare because it is about it's about our bodily autonomy, our health and everything else. So it's just a whole different, I guess, can of worms to unpack. I'm surprised that I feel surprised that I feel positive after the conversation.

Saujanya Vemuri: I mean, that said, do I worry if I'm going to have a job and, you know, if I'm of any value? Yes, I worry about that every day. Right? So like, why would people come and listen to me about insulin resistance or whatever, or, you know, some, some, you know, interesting fact about our biology when they can just go research the internet, you know, on GPT. Yes, I worry about that every single day. Again, there, I think it comes down to human judgment.

Saujanya Vemuri: What are they going to do with that information? I think people need to be wary of that and just be sensible. Being sensible with every technological advancement comes down to whether you are sensible in how you use the tool to your advantage or to your disadvantage, whether it was, you know, whether it was the internet with social media, it all comes down to being sensible. And I think AI just wants us, it forces us to be sensible. You know, you have to be sensible to use it.

Saujanya Vemuri: You know, you should, you should have some judgment about, about what it can offer to you and what it cannot. Right.

Aleck McCathie: I mean, I agree. It sounds like a bit more of a bigger ask than it should be though.

Saujanya Vemuri: Yes. That's hard. You're right. Saying be sensible is simple, but not easy.

Aleck McCathie: Yeah. Well, mean, do you have any final takeaway like, anything that you want to impart as far as like your thoughts on AI to the listeners?

Saujanya Vemuri: Well, let's see. I think it is one of the, I think from a, I'll put it as a tech, from a technology perspective, I have not seen any technology until this point that has transformed, you know, the way we do work at this scale and at this level as artificial intelligence. So that is very fascinating. That said, there is a fear. I have a fear.

Saujanya Vemuri: I know people around me have a fear about how AI could potentially disrupt, you know, households, living, earning, having a job. What do kids go and study in college? What does computer science look like? Should I be even doing law when, you know, there are all these AI models built for it? What is my career path?

Saujanya Vemuri: What does that look like? Right? So there are all these big existential questions that need to be answered, you know, with this technology becoming such a force in daily life, in how governments are operating, how companies are operating and based policies are being, you know, revisited and redesigned.

Saujanya Vemuri: So I feel like there are more questions than answers we have at this point. And I'd encourage everyone to be curious about it, but at the same time, use human judgment to know when it is for you and when it is not for you.

Aleck McCathie: Awesome. I love that. Existential uncertainty is a good way to sort of summarize everything I think. But, yeah, that's... This was awesome.

Aleck McCathie: I really appreciate you joining us. And, with that, I'm gonna pull the plug.

Saujanya Vemuri: Thank you.

Podcaststhatmatter.org.