Introduction - Srivatsav Nambi
This post is part of our AI6P interview series on “AI, Software, and Wetware”. Our guests share their experiences with using AI, and how they feel about AI using their data and content.
This interview is available as an audio recording (embedded here in the post, and later in our AI6P external podcasts). This post includes the full, human-edited transcript.
Note: In this article series, “AI” means artificial intelligence and spans classical statistical methods, data analytics, machine learning, generative AI, and other non-generative AI. See this Glossary and “AI Fundamentals #01: What is Artificial Intelligence?” for reference.
Interview - Srivatsav Nambi
Karen Smiley: I’m delighted to welcome Srivatsav Nambi from the USA as my guest today on “AI, Software, and Wetware”. Thank you so much for joining me on this interview! Please tell us about yourself, who you are, and what you do.
Srivatsav Nambi: Thank you so much, Karen. Really glad to be here. I'm Srivatsav Nambi, founding AI engineer at Elementary, a company that builds AI-based quality inspection systems for the manufacturing industry. I'm originally from India. Graduated from ASU from Interactive Robotics Lab. And my current work involves both R&D and production side of AI. I have helped build AI systems in real factories for companies like Yamaha, Unilever, Fiberon, et cetera. These systems inspect a thousand products every hour. I love solving hard problems with real impact.
Karen Smiley: Great. So you mentioned ASU. For our audience who don't know what ASU is, can you explain that a bit?
Srivatsav Nambi: Yeah, it's Arizona State University.
Karen Smiley: And the degrees you completed there were?
Srivatsav Nambi: I did my master's in computer science, where I specialized in machine learning applications for robotics.
Karen Smiley: Very good. Can you describe your level of experience with AI and machine learning and analytics, and how you've used it professionally or if you've used it personally? It sounds like you've studied the technology!
Srivatsav Nambi: I've been in the field close to a decade now. Two years in research, at ASU and over six years in industry. In academia, I worked mainly on deep learning and applications in robotics.
Now in industry I'm mainly working on computer vision based AI solutions, for the manufacturing industry. My recent project has been designing and deploying a self-learning AI system that teaches itself how to distinguish defects from live production data. Like, no need to have any user labels or anything.
Karen Smiley: It's great to hear about finding ways that AI can be accurate without requiring human data workers to label the data first. Can you elaborate on this project a bit, to the extent that it's not confidential to your company?
Srivatsav Nambi: So basically we deploy a camera on a production line, like, on a conveyor belt where each product is passing by. We place the camera on it. And the AI system, it's going to read each image and understand about the product. Like as each image passes by, it learns the distribution of the images.
So whenever it sees anything out of the distribution, it's going to flag it, telling that "There is no cap on this bottle." Or "Hey, there is a dent on this part here. This is not the usual thing that I've seen before in the other parts." So basically it has an understanding of the product and whenever something comes out of the distribution, it flags it.
Karen Smiley: So how good does the quality need to be, or how consistent does it need to be, for this to work?
Srivatsav Nambi: It depends on the number of samples. For our system, even if you have 10 to 15 good samples, it can understand what a good product is. It gets up to 99% or 98% of the defects, roughly around there.
But also there is another step, where we can actually train with the label data. But yeah, it understands like 90 to 99% of the defects by itself.
Karen Smiley: What I'm wondering about is, for instance, if you had a production line where 25% of the units had a defect, it would probably be harder for the AI to figure out that's a defect. That's not part of the normal distribution. But if you had a production line that was 90 plus percent good, or 95 or 99%, then it would be much easier for the AI to pick it out without having any labeled data to start with, saying,
"This is good. This is not good.", right?
Srivatsav Nambi: Oh, yeah, definitely. That definitely happens. If it doesn't understand what the good parts are, of course, it's not going to perform well. In our app, we have the ability to reset the model. If a lot of bad data went through, you can always reset it. If bad data gets in, it learns to shape the model based on the bad data.
Karen Smiley: So there's some sort of a feedback process where at the end there's some human inspector saying, "Hey, this shouldn't have passed through. It wasn't really good."?
Srivatsav Nambi: Yes, definitely. Our end system also has a human in the loop. So it basically accelerates the process for them. It's not a one-to-one replacement. Of course the human and the quality inspector would have much more knowledge on the product.
Karen Smiley: Very cool. A lot of people talk about artificial intelligence and how it might replace people, but it seems like it adds more value when it can augment human capabilities and human workers.
Srivatsav Nambi: Exactly. Yeah, that's exactly what I believe too.
Karen Smiley: Can you share a specific story about how you have used a tool that included AI or machine learning features? You've talked about your example of using machine learning for your business, but how about the way that you use the tools yourself? I'm wondering what your thoughts are about how the AI features of those tools worked for you, what went well, and what didn't go so well when you used them?
Srivatsav Nambi: Yeah, definitely. I use LLMs almost daily, I would say. Sometimes to write the boilerplate code or sometimes to break down any complex technical things that I don't understand, like in fields like medicine, immigration, law, et cetera.
Karen Smiley: Yeah. I'm curious about your experiences with using AI on things that you don't already know. Some of my guests have reported horror stories about getting really bad answers when the LLM asked them a clarifying question, and they didn't know enough about the problem domain to be able to answer the question well.
ChatGPT told one of my recent guests, Noemi Apetri, that she had cancer. Thankfully it turned out ChatGPT was wrong! And it's because, when it asked her a clarifying question, she didn't know enough about the medical area to say, "Well, tell me this instead of that." That's one of the reasons she got a wrong answer.
So I'm always really curious to hear about using it in domains that you know well. You obviously know coding very well, and you'd know when it's giving you a bad answer. Do you have any examples of when you asked the LLM about something from a technical area that you don't know as well? And the answer was either really good or really bad?
Srivatsav Nambi: Yeah, no, definitely. I keep experimenting with them, to see what its boundaries are. So I do have a similar example. I had a toothache, so I scheduled my appointment with the dentist. But I know that LLMs can do object detection or object description pretty well.
So I just sent it a snapshot of my tooth and said "Do you know why this hurts? What's going on and all?" And it immediately said "I'm very sure that this is a cavity." And the next day I had the appointment with a dentist and all, and he said "No, your teeth are perfectly fine. Looks like there's a crack on your teeth because you hurt yourself. But it's perfectly all right."
So it definitely does a lot of mistakes, even in the domains that we don't know. I would never replace it for an actual expert or take it as the 100% source of truth.
The main reason I like using it, though, is when it breaks down the complex topics for me. Like when I'm reading about medical stuff, there's a lot of technical jargon. So sometimes I just ask "Hey, can you explain to me like I'm 5, or like I'm 10? Can you summarize this for me?"
I would definitely do my own research. I would not rely 100%. It hallucinates a lot. The thing with current LLMs is that it gives you a strong yes or no. It will never give an answer where it's uncertain. We always have to keep that in mind when we're using the services.
Karen Smiley: Right. I just read a study. They were looking at how often the LLM gave the right or wrong answer, and then how often it was confident when it was wrong, even though it was wrong. And I think it was something like 47% of the time when it was wrong, it was still highly confident that it was correct.
Srivatsav Nambi: Yeah, that happens. I think people need to be aware, even when it's confident, it speaks with, like, 100% certainty.
Karen Smiley: You mentioned writing code. Can you talk a little bit about an example where you've used it to help you write code and how that has worked out?
Srivatsav Nambi: Yeah, definitely. I don't use any of our company's code because of privacy concerns. To give an example, if there's a new hardware or GPU out in the market and I want to test how fast it is, I just ask it to write boiler plate code to load a model and run it for me.
And then I would look at the latency numbers. Instead of me writing it, if it writes it, I can know the answer within 10 minutes. I use it for such purposes. The way you did it in the medical stuff, like, it's the same thing for code too.
It hallucinates a lot. One common issue that I see happening over and over again is it hallucinates the APIs. It tried to make software calls that doesn't even exist. That's the most common issue that I see. It's kind of weird. The entire thing, not just the endpoint, but the payload and the response type and everything. So that's a weird part.
Karen Smiley: I had a similar experience when I was using Python and using the LLM to help me write some code to call this library for processing some medical data in a certain file format. And it was EDF file format. It had a library. But it made up endpoints. Even when it got the endpoint right, the parameter it passed in was not the right thing at all. So the compiler caught one. The other error I had to catch just from knowing what it was supposed to be doing.
Srivatsav Nambi: Yeah, yeah. It's funny.
Karen Smiley: So this is a good overview of how you've used AI-based tools. I'd like to hear if you've avoided using AI-based tools for anything or for some things? And if you can share an example of when, and why you chose not to use AI for that?
Srivatsav Nambi: I would never use AI tools for any of our confidential data sets, company source code, or company's research. If you're working on any patents I would not share the details. Or even writing the proposals or anything. I would not use any of the LLMs, or I wouldn't completely rely on them.
Karen Smiley: The question of confidentiality is a concern. I know there are some cases where they say that if you have a paid account and you turn off some of the data storage and sharing, that it's supposed to be confidential. But I don't know how much that can actually be trusted.
Srivatsav Nambi: So there is this SOCI compliance. They say they don't share the data at all. Especially if it's the company core research and all, I wouldn't still trust it.
But I have to still look deeper into the certifications. From what I'm hearing so far, now they're trying to regulate it so that they don't train it at all. But now I'm kind of half and half, but I do want to have regulations, to be pretty strict.
Karen Smiley: One of the concerns that we see, pretty widespread and growing, is looking at where the AI and ML systems get the data and the content that they train on. A lot of times they're using data that people put into online systems or put into tools,
orwe've published online. For instance, I know you've written some papers, and they've scraped papers and books and all kinds of other content.
Srivatsav Nambi: Yeah.
Karen Smiley: And the companies aren't always transparent about the data we put in when we sign up for the systems. So I'm wondering how you feel about companies that are using this content and if you think that ethical AI tool companies should be required to get consent from and give credit and compensate the people whose data they want to use for training — or what we call the "3C's Rule".
Srivatsav Nambi: Yeah, no, definitely. I think that's the right way, right? The basic thing that the AI companies need to take. If you're using someone's data, and it's helping a lot, I think they should definitely compensate them. But at least credit them.
I think the recent issue has been ongoing with the New York Times. ChatGPT has used millions of New York Times articles and trained their models on it. So whenever someone was querying the ChatGPT, it was either slightly paraphrasing the article or giving it exact word-to-word. I think even the New York Times proved that, "Hey, these guys have used our data without permission." And also I think that's still ongoing and we should see what happens.
And another issue that made me feel sad a little bit, too, is the Ghibli trend.So I used to watch anime. And the creator of Ghibli. It's a type of anime style. And ChatGPT launched a filter saying that you can upload any photo and we'll convert it into that anime style. But they never got consent from the creator, and everyone was using it. And the studio, the creator felt really, really bad. He was not happy about it at all. So that was sad to see them using his style of artwork without his consent.
To summarize, I definitely believe in sticking to the 3Cs rule.
Karen Smiley: Yeah, the studio leader had made that comment about AI and how he would never want it to be used, and he used the word that it was an 'abomination' for AI animation. So he definitely was not a fan of AI. And they certainly never asked his permission to train ChatGPT on the body of works, on his anime films. It seems blatant that they basically stole it and didn't get his consent, didn't give him credit, didn't compensate him.
And the other thing I think people don't always realize is that, all these people say, "Oh, this is so cool. Let me upload my family picture that I just took yesterday." And so now, guess what? Now OpenAI has all these new pictures of people that people gave, and they probably didn't even realize, let them use this for further training on the AI tool.
Srivatsav Nambi: That's true. That's crazy. I think in ChatGPT there is a feature called 'temporary mode' where they don't store your data at all. If you upload your picture there, they don't generate the anime style at all, it's only when you allow them to store the data they generated for you. A lot of people weren't aware they were training with their data.
Karen Smiley: So as someone who has used these AI-based tools, do you feel like the tool providers have been transparent about sharing where the data that was used for the AI models came from and whether the original creators did consent to its use? I think we all know ChatGPT did not ask.
Srivatsav Nambi: Oh yeah, definitely.
Karen Smiley: Are there any other tools that you use? And do you have a sense of whether the tool companies have been transparent with you about where their data came from?
Srivatsav Nambi: I think most of the major LLM services that are out there, definitely, they just give vague answers. Like, "Hey, we, we trained it on the data that we scraped on the internet." They definitely give vague answers. But ideally I think it would be good for them to at least cite the data sets or cite the people whose work that they have used. But I don't think they're doing a good job of it. I think a lot of the details are missing on what exact data they're trained on.
Karen Smiley: How is it at your company?
Srivatsav Nambi: Oh yeah. That's a really good question. I'm glad that you brought it up. In our company, I can speak from the infrastructure standpoint because I know what happens behind the scenes. So in our company, we deploy for the customers in the manufacturing industry too, right? So they also have concerns about their data. In our company, whenever we go to a customer, they'll have a hundred percent authority on what to do with their data. We clearly explain how we're going to use this data. Mostly the data is going to live on their storage devices. We wouldn't use it anywhere else, apart for that customer.
In fact, we have an automatic deletion policy. We don't want to store their data more than 15 days. Unless they say "Hey, can you store the data for us so that we can track?" That's what I'm really happy about with our company, that it's doing everything ethical and no secrets.
Karen Smiley: The attitudes of founders about ethics and about the way that they handle customer data, that tends to be a big factor in whether the company starts out ethically and remains ethical.
Srivatsav Nambi: Yeah. Definitely. I wasn't the actual founder, but I'm the founding ML engineer, one of the early employees you could say. I think we have been ethical and I definitely don't see any point, why would we have to break it too? I think we are doing really well and I think we can continue to do so by sticking to that.
Karen Smiley: Yeah, that's great. So as a member of the public, or as a consumer, our personal data and content has probably been used by AI- based tools or systems. In fact, since we're both on LinkedIn, we know that it has been.
Srivatsav Nambi: Yeah, yeah, yeah.
Karen Smiley: Do you know of any other cases where your personal data has been used by any AI-based tool system?
Srivatsav Nambi: I personally, whenever I open my phone and see the app store, there are a ton of apps telling that they would build personalized AI features by using our data. Whether it's my pictures or my website activity, I definitely see it a lot. I'm pretty sure they might have used my data somewhere too.
I try to avoid it as much as I can. Basically turning off all kinds of tracking. The thing is a lot of people are not aware that it's actually tracking by default. It's always enabled. So you have to explicitly make an effort. For them not to track.
Karen Smiley: I was just reading an article yesterday about Apple and iOS and how they have this location privacy. And then Apple has this new feature they call ‘significant locations’. Apparently you have to go in and specifically turn it off, so it doesn't keep track of what they deduced is your home and where you go to church when, where you get your groceries and things like that. But you have to explicitly turn that off. I checked mine, and I already had it turned off, so I must have seen this before. It came up again and a lot of people were like, "Oh, I didn't know about this."
Srivatsav Nambi: Definitely that happens to me too, especially when I'm using social media apps. I learned it the hard way too. I talked to my friends with something and suddenly I started seeing ads about it. Like, I talk something about in India and all, but I'm in US. I'm supposed to be shown US products. It kind of proves that it definitely heard the conversations to recommend such a product with a hundred percent precision and all.
So, it does happen. And I used to think that I just disable the tracking on the app and it's done. But, yeah, when I went to iOS and saw how to disable all kinds of tracking, there are four to five different places that you need to disable tracking. And I was like, "Yeah, this is not the right way.” It feels very backward. That should be the first thing. There should be a single switch to control if someone doesn't want to, allow tracking.
Karen Smiley: Do you know of any company that you gave your data or content to that made you aware upfront that they were going to use your data for training AI or machine learning? Or did you get surprised by finding out? I guess LinkedIn would be an example there.
Srivatsav Nambi: Lately I've been seeing prompts on the apps like,"Hey, if you're using our free version, we would like to use your data to make our models better.” And I see some of the apps asking permissions. But I don't know if everyone is following that, strictly.
Karen Smiley: Yeah, I know it's very hard sometimes to find the opt-outs. They're often buried in the terms and conditions. I think there's a very small percentage of people that actually ever read those, and you can't blame them because they're so dense.
Srivatsav Nambi: No, I agree. Hundred percent.
Karen Smiley: People have talked about trying to train an LLM to analyze terms and conditions and say, "Okay, what's really in here?"
Srivatsav Nambi: Yeah.
Karen Smiley: So has a company's use of your personal data and content ever created any specific issues for you, like privacy or phishing or loss of income?
Srivatsav Nambi: Yeah, phishing. I just get at least one phishing message, like one in two days, either on my email, on my phone with specific manager or company details and all., Definitely. As I said, people scrape the data on LinkedIn or something. and they try personalized text that they want you to believe that it's true. And it's kind of unfortunate, how hard it is.
Now there are services that charge you to remove data from all the brokers. I don't remember the name exactly, but you pay every month to ensure your data is not stolen or not anywhere, it definitely happens with me. So far I've been ignoring all the phishing emails. But it's very backward that you have to pay to protect yourself. That's definitely sad.
Karen Smiley: Yeah. Phishing is pretty common here. In fact, one thing that really annoys me is that my mom's cell phone was on my plan for a while, and somebody sold that data. And now my poor mom gets phishing texts that are aimed at me.
Srivatsav Nambi: Oh God.
Karen Smiley: But I just keep saying, "Mom, just delete them. Just ignore them. If it's not from one of us, just ignore the message." But that really annoys me. That's not right.
Srivatsav Nambi: I'm glad that you are there for your mom, to explain that's phishing. And for so many elderly people, they don't understand. I think the latest one that's going on in Arizona is that they call you saying they're from MVD [motor vehicle department] and you have some tickets. Some elderly people give into that. They pay the tickets and all. If there is a younger person who understands tech, they would definitely explain it's phishing. But yeah, I'm not happy about that too. Feels ugly.
Karen Smiley: What's interesting is that this started a long time before AI really came on. This is a problem made worse by AI because your data gets pulled in. It's one thing to be able to correct it in a source system. If you find out that someone's got your data and it's wrong, once it's pulled into an LLM, you're never going to get it back out, and never going to get it corrected, and that makes it that much harder.
Srivatsav Nambi: I agree.
Karen Smiley: So one thing that we're seeing with all of the misuse of our data is that the public distrust of AI and tech companies has been growing. What do you think is the most important thing that AI companies need to do to earn and keep your trust? And do you have specific ideas on how they can do that?
Srivatsav Nambi: I think AI is a powerful tool that helps everyone in their day-to-day lives. But the thing is, a lot of the AI companies are actually not doing a good job of telling the user how they have collected the data, and giving them control.
So what I believe is that if the AI companies want to gain the trust of people, I think they need to clearly tell them how they have collected the data, does the user have the control on how their data wants to be shared? And I think for the AI companies they need to have independent auditors to review their AI companies. They can review the entire company — how they're dealing with the data, how they're training their AI and stuff. So have independent auditor certifications or something. I think if you have that, companies could gain their trust back.
Karen Smiley: There's one certification called Fairly Trained. It started up a year ago January, by Ed Newton-Rex. The idea is that a company will invite them in and they will look at where they got their data from. Did they source it fairly, and did they have consent? Or did they commission it from people that were then paid for the work? I think it's still under 20 companies, and it's all the small ones — it's not any of the big ones.
The newer thing is an ISO 42001 certification. But that doesn't really require them to share where they got their data or prove that they've sourced it ethically. Like, Anthropic and Claude have got that certification, but they're also subject to one of those 40+ active lawsuits about where they got their data and whether it was stolen. So that isn't foolproof either. But at least it's something. You have to give people who are trying to do the right things some credit for trying.
Srivatsav Nambi: I think the regulation needs to get much better, for sure. I think there are good AI companies too. I can speak as an engineer behind the scenes. for our company. You can definitely build companies ethically without having to steal data or misuse the control of the user. Basically, there are companies that are doing good too. I believe the regulation needs to get better and stop the bad things from happening.
Karen Smiley: Great. Well, that was my last question, so thank you so much for joining me on this interview today! Is there anything else that you would like to share with our audience?
Srivatsav Nambi: I think this is a good conversation. And, yeah, I would say: everyone using AI to stay curious and stay responsible. Question the company's products that they're using.
And, yeah, personally if anyone is interested in how AI is being used in manufacturing they could follow my LinkedIn. I'm planning to break down those things.
Karen Smiley: I know we're connected, so I'll be looking forward to your posts on manufacturing! We had a long experience with that in the previous company where we worked together, so it's an ongoing interest.
Srivatsav Nambi: Of course.
Karen Smiley: Thank you so much!
Srivatsav Nambi: Alright. Thank you so much.
Interview References and Links
Nambi Srivatsav on LinkedIn
About this interview series and newsletter
This post is part of our AI6P interview series on “AI, Software, and Wetware”. It showcases how real people around the world are using their wetware (brains and human intelligence) with AI-based software tools, or are being affected by AI.
And we’re all being affected by AI nowadays in our daily lives, perhaps more than we realize. For some examples, see post “But I Don’t Use AI”:
We want to hear from a diverse pool of people worldwide in a variety of roles. (No technical experience with AI is required.) If you’re interested in being a featured interview guest, anonymous or with credit, please check out our guest FAQ and get in touch!
6 'P's in AI Pods (AI6P) is a 100% reader-supported publication. (No ads, no affiliate links, no paywalls on new posts). All new posts are FREE to read and listen to. To automatically receive new AI6P posts and support our work, consider becoming a subscriber (it’s free)!
Series Credits and References
Disclaimer: This content is for informational purposes only and does not and should not be considered professional advice. Information is believed to be current at the time of publication but may become outdated. Please verify details before relying on it.
All content, downloads, and services provided through 6 'P's in AI Pods (AI6P) publication are subject to the Publisher Terms available here. By using this content you agree to the Publisher Terms.
Audio Sound Effect from Pixabay
Microphone photo by Michal Czyz on Unsplash (contact Michal Czyz on LinkedIn)
Credit to CIPRI (Cultural Intellectual Property Rights Initiative®) for their “3Cs' Rule: Consent. Credit. Compensation©.”
Credit to Beth Spencer for the “Created With Human Intelligence” badge we use to reflect our commitment that content in these interviews will be human-created:
If you enjoyed this interview, my guest and I would love to have your support via a heart, share, restack, or Note! (One-time tips or voluntary donations via paid subscription are always welcome and appreciated, too 😊)
