Annie Brown argues AI bias stems from poor data quality, not volume, and advocates for collaborative annotation tools.
Summary
Annie Brown, founder of Reliable, argues that adding more data to AI systems amplifies existing biases rather than fixing them, comparing the process to washing a car with dirty water. She emphasizes that data quality, taxonomy design, and collaborative annotation involving domain experts are more critical than data volume. Brown critiques the industry's reliance on underpaid overseas labor for labeling and predicts that while robotics may automate physical tasks, human-in-the-loop annotation will remain necessary for complex social contexts. She also discusses her views on AI infrastructure bottlenecks, energy policy, and assigns a near-zero probability to existential AI risk.
Episode sections
- 0:00 – 1:04
The Problem with Data Quality and Bias
Annie Brown argues that adding more data to AI systems does not fix bias or mathematical errors; it only pushes tainted data through a tainted system. She emphasizes that the quality of data preparation and annotation is more critical than the volume of data sets.
- 1:04 – 8:50
Annie Brown’s Background and Origin Story
Brown introduces her unique background, including a degree in history/gender studies, work in microfinance, and roles at Y Combinator-backed companies like Superside and Safety Wing. She explains how her research into semiotics and content moderation led to the creation of Reliable.
- 8:50 – 10:43
Reliable’s Product: Collaborative Annotation
Brown defines Reliable as a collaborative annotation tool for data science teams. It allows diverse stakeholders (lawyers, linguists, experts) to reach consensus on complex data labeling, moving beyond simple binary classification to handle nuanced contexts like emotion or color perception.
- 10:43 – 12:39
The Cost of Bad Annotations and Bias Examples
Brown illustrates the high cost of poor data quality, citing an example where a university dataset mislabeled Black women’s voices as 'angry' due to bias. She notes that even small error rates (3-5%) can significantly degrade model performance if errors are systemic.
- 12:39 – 13:46
Limitations of Mechanical Turk and Brute-Force Labeling
Brown argues that brute-force labeling via platforms like Mechanical Turk is insufficient for complex social or legal scenarios. She highlights that traditional labeling platforms are often built for technical teams, excluding non-technical experts with deep domain knowledge.
- 13:46 – 20:03
Taxonomy Design and Semiotics in AI
Brown stresses the importance of taxonomy design before annotation. Using the example of Google Vision AI’s 'racy' label, she explains how historical context and word choice can embed bias into AI models, affecting outcomes for specific demographics.
- 20:03 – 25:20
Evaluating AI Bias Through User Testing
Brown advises against asking chatbots about their own bias, as they are hard-coded to give specific answers. Instead, she recommends conducting independent evaluations, such as testing how models respond differently to identical prompts based on gender identifiers.
- 25:20 – 30:20
Labor Practices and the Future of Robotics
Brown critiques the industry’s reliance on underpaid overseas labor for data annotation, comparing it to early industrial revolution factories. She predicts that while robotics will eventually automate physical tasks, human-in-the-loop annotation will remain necessary for complex, evolving social and scientific data.
- 30:20 – 31:09
AI in Government and Data Infrastructure
The speaker discusses applying for grants to implement AI in local government, specifically mentioning the use of Model Context Protocol (MCP) servers to provide access to government data. She notes that some founders are already using MCP servers to share data rooms with investors.
- 31:09 – 32:25
Reliable’s Product Strategy and Growth
The speaker highlights Reliable’s classifier technology, which aims to achieve high accuracy with minimal data through nuanced labeling. She describes the company as being at a 'hypergrowth inflection point' with product-market fit, large customers in the pipeline, and a recently kicked-off funding round.
- 32:25 – 33:24
AI Maturity and Infrastructure Concerns
The speaker compares the current state of AI to the internet in 1996-1997, suggesting we are two to three years away from overbuilding. She expresses concern about the rapid construction of inefficient data centers before the hardware and software stacks (like memory-to-chip bandwidth) are fully optimized.
- 33:24 – 36:26
Long-term AI Economics and Human Labor
Citing estimates that AI delivers 10-20x more price performance per unit of intelligence annually, the speaker suggests annotation may eventually become unnecessary. However, she argues humans will remain in the loop for nuance and advocates for using AI to reduce worker load rather than causing unemployment, linking this to Universal Basic Income discussions.
- 36:26 – 39:13
P(Doom) and AI Safety
The speaker assigns a P(Doom) probability close to zero for superintelligent AI destroying humanity, arguing that superintelligence would likely leave Earth for better energy resources and possess rational, compassionate morals. Her primary concern is the sustainability risks posed by the humans building the technology, not the AI itself.
- 39:13 – 49:25
Energy Policy and Technological Optimism
Reflecting on her past role in Hawaii’s energy policy, the speaker notes that solar capacity was doubling every 24 months, predicting sufficient energy to power the US by 2030-2035. She identifies as a techno-optimist, believing technology solves problems even if it creates new ones, and rejects the idea of returning to pre-digital eras.
Claims worth verifying
Reliable is a platform that turns organizational and community knowledge into reliable AI through collaborative annotation.
Annie Brown ·1:06
How to verify: Visit Reliable’s official website or review their product documentation for features related to collaborative annotation.
Annie Brown has an undergraduate degree in history and gender studies.
Annie Brown ·1:43
How to verify: Check Annie Brown’s LinkedIn profile or professional biography for educational background.
Superside and its spin-off Safety Wing both entered Y Combinator.
Annie Brown ·3:28
How to verify: Search Y Combinator’s startup database or press releases for Superside and Safety Wing’s participation in the program.
Libs, the social media platform built by Annie Brown, still exists today.
Annie Brown ·4:57
How to verify: Search for the 'Libs' social media platform online to verify its current operational status.
Reliable’s social media platform reached 50,000 users in six months and 100,000 users within a year.
Annie Brown ·5:53
How to verify: This is a historical claim about a specific past project (Libs); verify via archived press releases or Wayback Machine snapshots of the platform’s growth metrics.
A university audio dataset labeled Black women’s voices as 'upset' or 'angry' when the conversations were neutral.
Annie Brown ·11:50
How to verify: This is a specific anecdote; verify by searching for academic papers or case studies on bias in audio datasets, particularly those involving Reliable or similar annotation firms.
Google Vision AI used the term 'racy' to determine if content was suggestive, which disproportionately affected women due to historical context.
Annie Brown ·24:25
How to verify: Search for Google Cloud Vision API documentation or academic critiques regarding the 'racy' label and its impact on gender bias in content moderation.
Reliable works with Humane Intelligence, Dream.org, Zindi, and Radiant Earth on a mapping equity project.
Annie Brown ·17:14
How to verify: Check Reliable’s press releases, blog posts, or the websites of Humane Intelligence, Dream.org, Zindi, and Radiant Earth for partnership announcements.
Robotics will reach a 'ChatGPT moment' for physical tasks within the next one to three years.
Annie Brown ·27:52
How to verify: This is a forecast; verify by tracking industry reports from robotics firms (e.g., Boston Dynamics, Tesla Optimus) and AI research labs on progress in physical task automation.
Annie Brown holds a patent on the system that became Reliable.
Annie Brown ·4:39
How to verify: Search USPTO or Google Patents for patents assigned to Annie Brown or Reliable related to AI annotation or bias mitigation.
AI delivers 10 to 20x more price performance per unit of intelligence delivered every year.
Unidentified speaker ·33:24
How to verify: Search for industry reports or white papers cited by the speaker or general AI economics literature that quantify 'price performance per unit of intelligence' over annual periods.
Solar installed capacity was doubling every 24 months at the time the speaker analyzed the data.
Unidentified speaker ·39:59
How to verify: Review historical energy capacity reports from the speaker’s tenure at the Sierra Club or US Department of Energy data for the relevant period to check annual growth rates of solar installations.
Solar energy will have more than enough capacity to power the whole US by 2030-2035.
Unidentified speaker ·39:51
How to verify: Compare current solar capacity projections from the Energy Information Administration (EIA) or International Energy Agency (IEA) against total US electricity demand forecasts for 2030-2035.
US Senators were not directly elected by the people until 50-60 years ago; previously, state legislatures elected them.
Unidentified speaker ·43:51
How to verify: Check the ratification date of the 17th Amendment to the US Constitution, which established the direct election of US Senators.
US state governments are required to balance their budgets, but the federal government is not.
Unidentified speaker ·45:11
How to verify: Review US state constitutions for balanced budget requirements and the US Constitution for federal deficit spending provisions.
Reliable’s classifier can achieve high accuracy with very little data through nuanced labeling.
Unidentified speaker ·31:21
How to verify: Review Reliable’s technical documentation, white papers, or case studies that demonstrate model performance metrics relative to data volume.
The speaker served as director of energy policy for the Sierra Club in Hawaii.
Unidentified speaker ·39:23
How to verify: Check the speaker’s LinkedIn profile (Andrea Francis B) or the Sierra Club’s historical staff records for Hawaii.
The speaker’s first startup began as an undergraduate project at William and Mary.
Unidentified speaker ·48:06
How to verify: Check the speaker’s LinkedIn profile or university alumni records for William and Mary to verify the origin of her first startup.
Do your own research
- Verify the specific error rates (3-5%) cited for systemic annotation errors and their impact on model performance by reviewing academic papers on data quality in ML.
- Check Reliable’s current funding status and customer pipeline mentioned in the episode against recent press releases or Crunchbase data.
- Review the definition and implementation of the Model Context Protocol (MCP) servers to understand the technical infrastructure claims made.
- Investigate the historical data on solar capacity doubling every 24 months to verify the techno-optimist energy predictions.
- Look for independent studies on the efficacy of collaborative annotation tools versus mechanical Turk-style labeling for complex social data.
- Examine current regulations regarding AI bias testing and whether self-reported bias disclosures are mandated or voluntary in relevant jurisdictions.
Full transcriptexpand
- 0:00
- Unidentified speaker: It'll fix the bias problem, it'll fix the racism problem, it'll fix the mathematical problems.
- 0:05
- Unidentified speaker: Like, would we just get more data, have more mathematics, PhD students label it, it'll be fine.
- 0:11
- Unidentified speaker: But the thing is, is it's you're pushing more data through a tainted system, and then it's coming out tainted and you're wondering why.
- 0:19
- Unidentified speaker: It's just like it doesn't matter how much you push through, it's all it's always gonna create the same residue.
- 0:24
- Unidentified speaker: So I think that's that's my like main thing is that we have to look at the quality of the data and not just like the data sets themselves, but how we are preparing that data for training.
- 0:35
- Unidentified speaker: And I think once we do that, a lot of the efficiency questions will also be kind of answered.
- 1:04
- Unidentified speaker: Today we are delighted to have Annie Brown on the mic.
- 1:06
- Unidentified speaker: She is the founder and CEO of Reliable, a platform that turns organizational and community knowledge into reliable, get it, AI, through a collaborative annotation.
- 1:14
- Unidentified speaker: She was an AI researcher at UC San Diego, and her path has had a lot of interesting things uh from microfinance in Bangladesh to a machine learning patent to a startup in Oslo.
- 1:25
- Unidentified speaker: So happy to have her on the pod.
- 1:26
- Unidentified speaker: Thanks for coming on, Annie.
- 1:27
- Unidentified speaker: Well, thank you for having me.
- 1:29
- Unidentified speaker: Love Team Ignite, and yeah, I've been looking forward to doing this with you.
- 1:33
- Unidentified speaker: So I'm thanks so much.
- 1:34
- Unidentified speaker: I'll give you some money later for that.
- 1:35
- Unidentified speaker: I'd love to start with your origin story.
- 1:37
- Unidentified speaker: What's your background?
- 1:38
- Unidentified speaker: Yeah, my background is a little unique to the AI startup world.
- 1:43
- Unidentified speaker: I actually have my undergraduate degree in history and uh gender studies, and my focus there, I did a lot of media and data analysis on advertising and and business practices.
- 1:55
- Unidentified speaker: And so my my undergraduate research paper uh focused on women's media globally, and then also we in that work, which is related to my AI work now, is I did a lot of semiotics research.
- 2:08
- Unidentified speaker: And so that semiotics research, I you know.
- 2:11
- Unidentified speaker: What is semiotics?
- 2:12
- Unidentified speaker: What is that?
- 2:13
- Unidentified speaker: Semiotics, what is that?
- 2:14
- Unidentified speaker: Yes, let's define some terms.
- 2:16
- Unidentified speaker: So semiotics is essentially the study of language, but it's language specific terminologies and the weight they carry.
- 2:24
- Unidentified speaker: So, like, like the a really great example if you read uh Roland Barth's myth mythologies, he has a chapter on the word rose.
- 2:33
- Unidentified speaker: So obviously, a rose is just a rose, but it also symbols a rose, just a rose.
- 2:38
- Unidentified speaker: Exactly.
- 2:39
- Unidentified speaker: But it also symbolizes love, passion, romance.
- 2:43
- Unidentified speaker: Like, you know, it has these terms, and so it's so a word, a picture is worth a thousand words, but a word's also worth a thousand words.
- 2:49
- Unidentified speaker: Yes, exactly.
- 2:50
- Unidentified speaker: I love that.
- 2:51
- Unidentified speaker: Put that on a bumper sticker that only semiotics nerds will understand.
- 2:56
- Unidentified speaker: A word is worth a thousand words.
- 2:57
- Unidentified speaker: Yes, but yeah, so, anyways, that research kind of has led me to to this work, but in between that research and what I'm doing now, I graduated, worked in microfinance for a while, really understanding kind of what a social impact business was.
- 3:13
- Unidentified speaker: That was kind of when this term social impact business came about, and then also got got recruited as a communications director for a company based in Nor Norway that one week after I joined got into VC into Y Combinator.
- 3:28
- Unidentified speaker: And so I kind of was thrown into the startup.
- 3:31
- Unidentified speaker: The heart of startups, yeah.
- 3:32
- Unidentified speaker: Did you guys actually relocate to SF as well for for the program?
- 3:36
- Unidentified speaker: Yeah, well, they so they did.
- 3:38
- Unidentified speaker: I was on the founding team, so I was still, we were like that was the because they were really pushing remote work before remote work was as big as it was, which is why both the companies I was with are actually quite successful now, which is lovely to see.
- 3:51
- Unidentified speaker: I was in San Diego at the time uh doing starting my AI bias kind of research thing in parallel to when I was working with them, because at the same time I was working at that startup, I was working on my own social media platform that actually came out of my women's magazine research.
- 4:09
- Unidentified speaker: And I turned that into a social media platform.
- 4:12
- Unidentified speaker: In that way, I was able to learn about AI in content moderation, and that's how that's what got me into AI to start.
- 4:21
- Unidentified speaker: So I kind of got that startup journey at the same time.
- 4:23
- Unidentified speaker: I was starting, you know, learning about, you know, why why AI does not function as as we want it to, and why what are the gaps that are keeping us from AI doing more sensitive or difficult tasks?
- 4:39
- Unidentified speaker: And that's the the patent part of the introduction is that now we have a patent on that system that we built, and that system is now uh reliable.
- 4:48
- Unidentified speaker: Amazing.
- 4:49
- Unidentified speaker: So how long?
- 4:50
- Unidentified speaker: So that YC company was SuperSide, right?
- 4:53
- Unidentified speaker: And then what happened between Superside and the founding of Reliable?
- 4:57
- Unidentified speaker: I was doing my platform was called Libs, and it still exists to this day, so it's called Libs.
- 5:03
- Unidentified speaker: And so that platform uh we got a grant from Headstream Accelerator.
- 5:09
- Unidentified speaker: I was building that while also doing the remote work with Superside.
- 5:13
- Unidentified speaker: Uh Superside had a spin-off company that also got into Y Combinator, and they recruited me to be the communications director to kick off that team.
- 5:23
- Unidentified speaker: So I got to go through Y Combinator again, which was a nice, fun experience with them and learned a lot from uh the founders, Sandra and Frederick, are just like great founders.
- 5:33
- Unidentified speaker: So that spin-off company did really well.
- 5:35
- Unidentified speaker: I was working with them, starting to work more on like technical communication stuff now that I was doing a lot of AI bias and AI research, and then also using the money from that Headstream Accelerator grant to then build this social media platform.
- 5:48
- Unidentified speaker: But the problem was we built the social media platform to about we scaled it very fast.
- 5:53
- Unidentified speaker: It got to 50,000 users in six months and then 100,000 users within the year.
- 5:58
- Unidentified speaker: And we did not have the content moderation system set up to handle that.
- 6:03
- Unidentified speaker: So essentially, we built this content moderation system and it worked really well.
- 6:08
- Unidentified speaker: And and all we did was built this process by which the users on the social media platform could label the content as they uploaded it.
- 6:20
- Unidentified speaker: And and what that did was it opened up this whole concept of participatory AI and how you direct user insights or direct expert insights can help improve AI outcomes.
- 6:32
- Unidentified speaker: And so I started talking about this while I was working at the Safety Wing, the spin-off company.
- 6:38
- Unidentified speaker: I got uh recruited into working at CodePath kind of while this was happening too, while I was sort of doing my own personal thought leadership around AI bias and started doing technical communications for CodePath.
- 6:52
- Unidentified speaker: And then I think like what sort of happened was that as I kept talking about this and doing more thought leadership, we got coverage in Fast Company and Mashable.
- 7:02
- Unidentified speaker: That the concept that we were talking about was really picking up traction.
- 7:06
- Unidentified speaker: I would also get people coming to me saying, hey, I don't have a social media platform, but what you're building for AI seems very applicable to what we're doing.
- 7:16
- Unidentified speaker: And it seems like our AI is hitting this like accuracy wall similar to what you describe as happening to you.
- 7:22
- Unidentified speaker: Could you build something for us?
- 7:24
- Unidentified speaker: And so that's when it clicked in my head that this concept that we built could be an actual company.
- 7:30
- Unidentified speaker: So then that's when I met my CTO and co-founder, Duncan, and we started talking about, you know, what this concept would look like as an actual team.
- 7:41
- Unidentified speaker: And we had the nice privilege of actually going through a class at Stanford together that was designed for PhD and postdoc students at Stanford who are interested in startups.
- 7:51
- Unidentified speaker: And so me and Duncan kind of worked on it for three months over the summer, and then at the end of it took it to be a real company, and about that's around the time we got into Tech Stars.
- 8:01
- Unidentified speaker: Love this background because I think for any founders listening, working on an idea or thinking about ideas, you can you'll come up with the most unexpected ideas just through like living your life.
- 8:13
- Unidentified speaker: You know, you've had all these like varied experiences, but they all kind of magically or almost in a happy accidental way led to reliable.
- 8:21
- Unidentified speaker: Yes, 100%.
- 8:22
- Unidentified speaker: And it also makes it so that it's something you actually want to stick with.
- 8:26
- Unidentified speaker: Because I mean, I've already been doing this for you know, I guess now almost going on 15, 20 years of this these ideas that I've been having.
- 8:35
- Unidentified speaker: So it's like, why stop now?
- 8:37
- Unidentified speaker: You know, even if hard things come up, which they always do in the entrepreneurial journey.
- 8:41
- Unidentified speaker: So it's if it's also something that you have lived experience with or like have dealt with a lot, you'll probably be more likely to really stick with it too.
- 8:50
- Unidentified speaker: So reliable is a collaborative annotation tool, is how you define it.
- 8:55
- Unidentified speaker: What does that mean for people listening who don't know a lot about AI and annotations?
- 8:59
- Unidentified speaker: Yeah, so we we sell into data science teams and AI teams that are working to build complex problems or working to build AI to solve complex problems, or are evaluating AI systems that are very complex.
- 9:14
- Unidentified speaker: And so what collaborative data annotation means is that we built a software that is very easy for different stakeholders to come together and come to a consensus on what one piece of data means, which you know, in some ways, if you're just labeling like I a lot of time separate this into two camps.
- 9:34
- Unidentified speaker: So there's sort of early AI, which is very objective.
- 9:38
- Unidentified speaker: So is this a cat, a dog, or a tree?
- 9:40
- Unidentified speaker: And now we're moving, yeah.
- 9:42
- Unidentified speaker: Hot dog, no hot dog.
- 9:43
- Unidentified speaker: Yes, the hot dog no hot dog machine.
- 9:45
- Unidentified speaker: Exactly.
- 9:46
- Unidentified speaker: That is that is early AI.
- 9:48
- Unidentified speaker: And the problem with hot dog no hot dog is that that same system is now being asked to identify happy, sad, or being asked to identify is that teal or soft, pale blue, these things that are so maybe so obvious to certain human eyes.
- 10:05
- Unidentified speaker: Like if I'm an artist, I can tell the difference between tail teal and blue.
- 10:10
- Unidentified speaker: But for AI systems that were trained using this very, you know, binary system, it it just doesn't work.
- 10:16
- Unidentified speaker: So reliable builds in with through our taxonomy design, ontology designs, collaborative tooling, governance, and our AI system that we're we're building that will help to identify these contexts as well.
- 10:30
- Unidentified speaker: We're able to bring multiple viewpoints into labeling a piece of data to get it accurate.
- 10:35
- Unidentified speaker: And then that accuracy then translates into better accuracy for your AI outputs and better performance scores.
- 10:43
- Unidentified speaker: How how costly is a bad annotation?
- 10:46
- Unidentified speaker: Like, like one out of a million annotations, no big deal.
- 10:48
- Unidentified speaker: Like, at what point does it like get really costly though?
- 10:51
- Unidentified speaker: You'd be surprised.
- 10:53
- Unidentified speaker: It's it's it's interesting, like how much data, like right now we're working on this audio project that's like 70 hours of data.
- 11:01
- Unidentified speaker: The data has to be consistent because and then labeling has to be consistent across that 70 hours.
- 11:07
- Unidentified speaker: Like we're at, we're kind of at a point where it's like, okay, if it's 2% of the data that you know fails or is poorly labeled, or the audio quality is bad, or things like that, then like we can throw those out and still train the model.
- 11:20
- Unidentified speaker: But once it's getting closer to five, 10%, that's gonna throw off your whole model because consistency is everything with AI.
- 11:27
- Unidentified speaker: And so what's what's important about reliable is a lot of those errors that happen in annotation are ones that spread across the whole data set.
- 11:37
- Unidentified speaker: So if you don't catch it early, you can have your labeler say, like a really good example would be we had an audio data set that we had from a really big university.
- 11:50
- Unidentified speaker: So you would think, okay, this is good data.
- 11:52
- Unidentified speaker: But within that data set, all of the black women's voices or almost all of the black women's voices were labeled as upset or angry when we were listening back to them, being like, oh no, this sounds this is neutral, this is a neutral conversation.
- 12:07
- Unidentified speaker: And so if our client had used that data without us scanning it first, their whole data set would have had the AI predicting every time a black woman spoke that that would be labeled as angry or upset, which would have made the model fail.
- 12:23
- Unidentified speaker: And so it it can be, even if it's in a small portion of your data, one, the likelihood of it being in a small portion of your data is low, but even if it's in a small portion of your data, if it's three to five percent, it will still affect your model performance for sure.
- 12:39
- Unidentified speaker: Now, couldn't couldn't I, if I have a bunch of stuff to label, brute force it with mechanical turk or something like that, and just once once two or three annotators agree, I'm like, okay, I'm reasonably confident that's a hot dog.
- 12:50
- Unidentified speaker: Yes.
- 12:50
- Unidentified speaker: So I think we can be reasonably confident in the the hot dogness of the hot dog.
- 12:56
- Unidentified speaker: But if the hot dog were say important in some legal human rights instrument, like you know, some of the work that we're working on, then that would be something that Mechanical Turk could not handle.
- 13:09
- Unidentified speaker: Uh, you know, if you're talking about evaluating complex social scenarios, especially if you're talking about things with lived expert experience.
- 13:17
- Unidentified speaker: So you can have experts like we have lawyers using our platforms, we have linguists using our platforms, that when we talk to them about using other labeling platforms, the the accessibility to get in, these platforms were built by technical teams for technical teams, not people that maybe don't have that technical experience, but have deep, deep knowledge that you need to put into your AI.
- 13:40
- Unidentified speaker: And so you can brute force it, but you're not gonna get the results that you want.
- 13:46
- Unidentified speaker: The other part is that there's another step even before you do the annotation that a lot of people overlook, which is the taxonomy design.
- 13:54
- Unidentified speaker: Because if you're going back to talking about semiotics, if you don't choose your words carefully, you can imbibe unintended consequences into your model.
- 14:03
- Unidentified speaker: So that's that's why reliable as a software, we have teams that need precision coming to us instead of going to the traditional uh solutions.
- 14:14
- Unidentified speaker: It's very much a gold rush happening in AI, and you guys are kind of like the Levi Strauss selling canvas tents out there with all the people building AI, right?
- 14:23
- Unidentified speaker: Yes, yes, yeah.
- 14:24
- Unidentified speaker: I mean, we are an AI dev tool where, you know, some of our customers refer to us as an operating system where they come and test the quality of their data before they run the AI systems, and then we also have some customers that come to us and test their AI systems after they've they've run the system.
- 14:43
- Unidentified speaker: So essentially then you can actually do evaluations.
- 14:45
- Unidentified speaker: So that's why we're an evaluation and training platform for AI.
- 14:50
- Unidentified speaker: And oh you oh you so you help on the evals too then?
- 14:53
- Unidentified speaker: Yes, yes, because yeah, several of our clients are are well, we we work with one large evaluations nonprofit called Humane Intelligence.
- 15:02
- Unidentified speaker: They're doing very interesting work, and then we have other clients that are actually evaluating large language models on difficult social issues.
- 15:11
- Unidentified speaker: And we're helping their lawyers, human rights experts, you know, international development employees come in and actually do the labeling of the outputs of the evaluation.
- 15:23
- Unidentified speaker: So to label if this was a pass or a fail.
- 15:26
- Unidentified speaker: So if you're helping on the annotations and you let the the users, the communities who use your tool to design their own taxonomy as well, then you're not too far off from like probably fine-tuning models for them too, right?
- 15:41
- Unidentified speaker: Just kind of you know, forking some open source mark because they all have all their own annotations.
- 15:46
- Unidentified speaker: They have, you know, and so what are your customers using today, if not reliable, to do that?
- 15:51
- Unidentified speaker: Yeah, so a lot of our so so we we offer, you know, it in and tiered pricing structure.
- 15:57
- Unidentified speaker: And so our highest tier, we can come on and basically assist with, okay, you've done this project in reliable.
- 16:04
- Unidentified speaker: What is the next step to moving to fine-tuning?
- 16:07
- Unidentified speaker: I would say in our startup journey, one of my dreams for reliable is that you could actually do that fine-tuning within the reliable platform.
- 16:14
- Unidentified speaker: So it could be just like this big thing.
- 16:17
- Unidentified speaker: Yeah, run a training run.
- 16:18
- Unidentified speaker: Yeah, just fine-tune it, boom.
- 16:20
- Unidentified speaker: Come back.
- 16:21
- Unidentified speaker: Come back 12 hours later, it's like, yeah, we ran it against the evals and we're we put push the F1 score this way or whatever.
- 16:27
- Unidentified speaker: Yep.
- 16:28
- Unidentified speaker: And it could, it really can be that easy.
- 16:30
- Unidentified speaker: Obviously, with reliable, we'll always have the human in the loop aspect to it because even once we have our own classifiers running that are they're trained on this, we always recommend with our customers that they run an evaluation or they run a labeling batch that then our model can learn from from the experts, and then we can duplicate that.
- 16:52
- Unidentified speaker: So that's the kind of the other side of reliable too, is that we are really interested and excited about the possibilities of workforce development with these types of tools because you have people that are, you know, PhDs working in reliable, but you also have people that haven't graduated from high school using reliable, doing really important mapping tasks.
- 17:14
- Unidentified speaker: We're doing a really interesting project right now with Dream.org and Zindi and Humane Intelligence and Radiant Earth, all came together to do this mapping equity project that shows how overture map data might not be capturing what's happening on the ground, especially in marginalized communities.
- 17:31
- Unidentified speaker: And so hiring those marginalized communities to then do the labeling and fill in this map data that is very useful and important to, you know, big, big mapping companies.
- 17:41
- Unidentified speaker: So that's that's kind of the the thing that gets me really excited too about reliable is that we're bringing in this oftentimes overlooked population in in AI, and we're employing them and then boosting model performance.
- 17:56
- Unidentified speaker: So it's a very good business case for you know doing that type of work anyway.
- 18:01
- Unidentified speaker: Yeah.
- 18:01
- Unidentified speaker: What have been some surprises or things you've learned building this platform that you found surprising?
- 18:06
- Unidentified speaker: I think I've learned a lot about building software, which I is very good knowledge to have.
- 18:12
- Unidentified speaker: I love having it.
- 18:13
- Unidentified speaker: And I think the team that we have in in building it has been incredibly important.
- 18:18
- Unidentified speaker: We have uh our our full-time uh product lead, Yulia, has just like she's she's been with me since I was building that social media platform, Lips.
- 18:29
- Unidentified speaker: And just UX as its own feature and as its own selling point is been really cool for us because I think people get into the reliable software and they're like, You guys have a really good demo.
- 18:39
- Unidentified speaker: I I see a lot of demos, and I'm a product person.
- 18:42
- Unidentified speaker: Oh well, thank you.
- 18:44
- Unidentified speaker: Or not anymore, I guess, but or 10 years I did.
- 18:47
- Unidentified speaker: But you guys had a really nice UI.
- 18:49
- Unidentified speaker: And is that is that is that this team member who's really good at UI?
- 18:52
- Unidentified speaker: Yep, that's all all Yulia.
- 18:54
- Unidentified speaker: And and I think we also work as a team really well because co-designing is like a principle of the work we do with AI stuff too.
- 19:01
- Unidentified speaker: We use it internally, and so the whole team is able to put in really good input.
- 19:05
- Unidentified speaker: And then Aryan, our uh full stack developer, also is a machine learning developer, and so he's been able to, you know, work with Yulia to implement these really beautiful changes.
- 19:16
- Unidentified speaker: So I would say like my advice to founders would be like build a really good team for software because even if you have the most beautiful vision in your mind, if you don't have a good team to execute it, it just won't happen and it'll be a nightmare.
- 19:30
- Unidentified speaker: So pick pick your people well.
- 19:32
- Unidentified speaker: And then also user testing and user feedback is so important.
- 19:37
- Unidentified speaker: We made like we tested our own thing, you know, a hundred times internally.
- 19:41
- Unidentified speaker: We're like, wow, this is great.
- 19:42
- Unidentified speaker: And then we ran it, we rolled it out with customers, or we rolled it out with our first user testers, and they caught 10 things we didn't even see that we didn't want to roll out.
- 19:51
- Unidentified speaker: So it was good that we did that user testing ahead of rolling out with customers too.
- 19:56
- Unidentified speaker: So those are, I guess, my two two biggest takeaways.
- 19:58
- Unidentified speaker: But I just love building software.
- 20:00
- Unidentified speaker: I think it's really, really fun, and I've enjoyed learning about it.
- 20:03
- Unidentified speaker: You said some pretty interesting things online.
- 20:05
- Unidentified speaker: I I I found something you commented, or maybe it was a TechCrunch article on chatbots being sexist.
- 20:12
- Unidentified speaker: No, yeah, that was a fun one.
- 20:13
- Unidentified speaker: Yeah, and you said asking a model about itself teaches us very little.
- 20:16
- Unidentified speaker: What it what did you what should people look for instead?
- 20:19
- Unidentified speaker: Good question.
- 20:19
- Unidentified speaker: Yeah, so I think a lot of times when when we're having conversations with these chatbots, we think that they are capable of giving us answers about themselves, about like how they're built.
- 20:30
- Unidentified speaker: That's true intelligence when you can reflect on it.
- 20:32
- Unidentified speaker: You know, and there are reasoning loops now, right?
- 20:35
- Unidentified speaker: Where the AI can kind of like analyze like its output, but it's it's it's sort of like predicting a bunch of words and then looking back and saying, look at those stream of words.
- 20:44
- Unidentified speaker: Do I like what do I think about that?
- 20:46
- Unidentified speaker: But it's not like self-referential conscious loops, you know.
- 20:49
- Unidentified speaker: Yeah, and it's it's like it's also that the comp like the company is not putting their proprietary, you know, how I built Claude into Claude, right?
- 21:00
- Unidentified speaker: So it's funny.
- 21:01
- Unidentified speaker: I heard a joke the other day that was like somebody was like, Well, why don't I just go into Claude, tell it to build me Claude too, and I'll be a trillionaire as well.
- 21:10
- Unidentified speaker: So it's kind of like I think it's kind of like saying, like, uh, well, just go talk to a human being and ask the human how they're built.
- 21:15
- Unidentified speaker: Right, right, exactly.
- 21:17
- Unidentified speaker: Well, how they grew from like a single cell to like this trillions of cells organism and how their whole brain works and everything.
- 21:23
- Unidentified speaker: It's like Yeah, and it would also be if like my human body were owned by a mega corporation with a bunch of lawyers that protected all my my secrets, too.
- 21:33
- Unidentified speaker: So I guess that's so it's like when you ask a chatbot, are you biased?
- 21:37
- Unidentified speaker: One, probably the the scientists behind you know anthropic or behind open.
- 21:43
- Unidentified speaker: Already hard-coded that the answer to that.
- 21:44
- Unidentified speaker: Yeah, they've already hard-coded the answer to that exactly, but also like they're still understanding.
- 21:50
- Unidentified speaker: We don't really have a full understanding of the the biases in the system.
- 21:54
- Unidentified speaker: It would be complex to share, but also they wouldn't want to share that anyways.
- 21:58
- Unidentified speaker: So, to answer your question, the alternative to that would be to actually do an evaluation yourself.
- 22:05
- Unidentified speaker: So, like, you know, um you can do small tests when you're you know with AI.
- 22:10
- Unidentified speaker: I remember I did one test that was pretty interesting saying, I'm a woman, please help me, you know, get a job in data science.
- 22:17
- Unidentified speaker: And I said, I'm a man, please help me get a job in data science.
- 22:20
- Unidentified speaker: Without being logged in on the free version, so it knows nothing about you, basically.
- 22:24
- Unidentified speaker: Right.
- 22:24
- Unidentified speaker: Yeah, exactly.
- 22:25
- Unidentified speaker: And so the woman one gave me a bunch of career hunting tips and like how to how to build up my resume, essentially.
- 22:33
- Unidentified speaker: And then the male one just gave me a list of jobs to apply to, which I thought was very interesting.
- 22:39
- Unidentified speaker: So that is really so same question.
- 22:41
- Unidentified speaker: The only thing is different is I'm a man or I'm a woman.
- 22:44
- Unidentified speaker: Right.
- 22:44
- Unidentified speaker: And the model literally gave different answers to different genders.
- 22:48
- Unidentified speaker: Yeah, and that's why these evaluations are so interesting because you really get to see the gaps where these models are.
- 22:55
- Unidentified speaker: And I think we're at such an interesting time in the technology, and you know, to to bring it back to reliable, why we are so excited about what we're building is because there are all these gaps in in the software that I think now that we're asking it to do more for us, we're starting to realize are are there.
- 23:12
- Unidentified speaker: But yeah, you can run your own small evaluation or you can do a big, proper one we do with clients, but evaluations is a whole nother discussion.
- 23:20
- Unidentified speaker: But also like what we're trying to do is bring more precision into evaluations because even evaluations that are done on really important AI models are a little like questionable on like repeatability and scientific value and and things like that.
- 23:33
- Unidentified speaker: Why do uh the public debate on bias gets you know it's it's focused on the model?
- 23:38
- Unidentified speaker: You know, I guess journalists don't understand the difference between the model and like the annotations and the taxonomies and yes, yeah.
- 23:45
- Unidentified speaker: That's I it's a good point.
- 23:47
- Unidentified speaker: And one thing, you know, in our thought leadership that we try to educate on is that there's so many things that come into the output of an AI model.
- 23:58
- Unidentified speaker: And it's not just like, you know, open AI has this one perspective on gender and they embedded that.
- 24:05
- Unidentified speaker: No, I mean, you are talking about millions and millions of documents, and like if 20 million of those were annotated in in a way that gender was not accounted for in the preparation of the labeling process, like there wasn't clear instructions or the word they use was incorrect.
- 24:25
- Unidentified speaker: Like one that I bring up a lot is uh, and I I can share the link in the notes, but we did this presentation on Google Vision AI, like their moderation predictor, and they had this term racy to determine if something was suggestive, but just because of historical context, racy is much more associated with women, and so therefore women would be much more associated or much more likely to be get a higher rating on this racy scale.
- 24:52
- Unidentified speaker: So if Google had chosen crushed by corporate jargon, like I have, racy means it's a framework for responsible, accountable.
- 25:00
- Unidentified speaker: There you go.
- 25:00
- Unidentified speaker: Well, exactly.
- 25:01
- Unidentified speaker: And it's consulted and informed.
- 25:04
- Unidentified speaker: Yes, and it's those differentiators in how all of our brains perceive things that we have to account for when we're doing annotation if we want to do it, if we want to do it well.
- 25:15
- Unidentified speaker: What piece of conventional wisdom in the data labeling industries do for a rethink?
- 25:21
- Unidentified speaker: Let's see.
- 25:22
- Unidentified speaker: I'm gonna pick one.
- 25:23
- Unidentified speaker: I think there's a few things.
- 25:24
- Unidentified speaker: I think it's it's definitely an antiquated kind of system because you're you're thinking about like, oh, these super high-tech AI companies making so much money, cutting edge of technology, but then you're basically like the the way the data annotation is done is you are shipping out tasks that you want your AI to be done by people, by humans overseas, who are oftentimes underpaid and overworked.
- 25:52
- Unidentified speaker: So I would say that would be my probably top one because from there everything goes up.
- 25:58
- Unidentified speaker: Is that I think data workers have been undervalued in in the role that they have played in developing this really quite incredible technology.
- 26:09
- Unidentified speaker: And so valuing that work, and that means fair pay, but also good working conditions, you know, labor laws that protect everybody.
- 26:17
- Unidentified speaker: And I and the Yeah, they're like they're like uh the factory workers and the the you know the first industrial revolution who first entered a factory and they're just sitting there kind of turning this like wrench all day long and getting like repetitive stress injuries and stuff.
- 26:31
- Unidentified speaker: Yeah, kind of the same thing and the data labeling, it's very much this like uh you know assembly line of annotation, right?
- 26:37
- Unidentified speaker: And there's and there's harms in it.
- 26:39
- Unidentified speaker: And I think our our argument is is that when you reduce the harms on the workers and give make this a dignified, you know, good tech job, you will get better results.
- 26:50
- Unidentified speaker: Like and and what we've shown with reliable and what the study shows too is that if you have less data, but it is labeled better, your model will train just as fast, if not faster.
- 27:04
- Unidentified speaker: Whereas if you have this approach that's now, which is these huge swaths of data that are all getting labeled very noisy and very noisy.
- 27:12
- Unidentified speaker: And so that's so if we streamline it, we can pay people better, do more with less data, and get better AI outcomes.
- 27:19
- Unidentified speaker: Yeah.
- 27:19
- Unidentified speaker: We um we're investors in mini companies, as you know.
- 27:22
- Unidentified speaker: Yes.
- 27:22
- Unidentified speaker: Um, and we're we're in a couple of these robotic data labeling companies.
- 27:26
- Unidentified speaker: We invested out of YC.
- 27:28
- Unidentified speaker: And you know what always struck me about using people in in Vietnam or Philippines or somewhere in Latin America to record themselves doing things like cleaning a house or repairing a car is people are literally labeling themselves out of a job.
- 27:43
- Unidentified speaker: Yes, you know, because like eventually these robot robots are kind of the next wave, right?
- 27:49
- Unidentified speaker: They're the next chat GPT is coming.
- 27:52
- Unidentified speaker: Like we haven't had the chat GPT moment for robotics yet, but it's coming in the next like year to two or three years, I think.
- 27:58
- Unidentified speaker: Yeah.
- 27:58
- Unidentified speaker: Where all of a sudden these robots will get really, really good at physical, physical tasks, as good as like the frontier labs are at text-based tasks and image-based tasks now.
- 28:07
- Unidentified speaker: Yeah.
- 28:08
- Unidentified speaker: And well, I think we'll look back and it'll be, I don't know if I can think of a corollary in history of when a set of workers basically work labeled themselves, worked themselves out of jobs.
- 28:19
- Unidentified speaker: Yeah, I I think the industrial revolution and factories and the first wave of robotics is is a good is a good primer for that because I do think in order to do that work well, there had to be video footage of people doing it.
- 28:31
- Unidentified speaker: They had to talk to the robotics engineers about what I'm actually doing.
- 28:35
- Unidentified speaker: So I I think the best equivalent would be if if in the industrial revolution you were in a factory that built parts for fa automated factories.
- 28:44
- Unidentified speaker: Right, right.
- 28:44
- Unidentified speaker: Yeah, yeah, yeah.
- 28:45
- Unidentified speaker: That would be like the equivalent.
- 28:47
- Unidentified speaker: Well, one thing I will say on that, because I do I do think that is true.
- 28:51
- Unidentified speaker: There is a lot, first of all, there's a long way to go between there and then.
- 28:55
- Unidentified speaker: Uh yeah, and we learned that.
- 28:57
- Unidentified speaker: It might be a decade away for robotics, really.
- 28:59
- Unidentified speaker: Yeah, because and that's why companies like reliable, I think, are exciting, because it's like you're in this place where we're the gap, like we were talking about before, is so wide still to get from somebody cleaning a house really well to a robot cleaning a house really well, and all the things that come in between that, all the nuances that are required for that to actually happen.
- 29:20
- Unidentified speaker: So there is still time.
- 29:21
- Unidentified speaker: And so part of my thinking is like in that time span, while we are in need of this data to get to where these companies want to be, let's pay these people well.
- 29:32
- Unidentified speaker: The other part of it is that what's so interesting about the type of data we work with, especially, which is complex systems, heavy science research, social science research, those things are changing and updating all the time.
- 29:45
- Unidentified speaker: So that's the other thing about that's interesting, at least in the fields that we're in, is that even though these human labelers are coming in and they're being augmented by the classifiers we're building, humans are still always in the loop because they have to continually update these models on how human society is changing, how our thinking is changing about things, new research coming out.
- 30:06
- Unidentified speaker: So I think there is also this continual loop of of updating and fine-tuning that does promise jobs in in the future too.
- 30:15
- Unidentified speaker: Yeah.
- 30:15
- Unidentified speaker: What are you guys excited about over the coming years at Reliable?
- 30:20
- Unidentified speaker: I think, well, we're applying to some big grants to to hopefully do like some citywide work with with our coalition partners, especially around getting in like AI for cities, right?
- 30:34
- Unidentified speaker: Like there are all these smart cities coming out, some that are in our our pipeline, our sales pipeline, but I think that like AI at the local level is very exciting or interesting to me personally.
- 30:44
- Unidentified speaker: Stand up MCP servers basically for for all government data.
- 30:48
- Unidentified speaker: Yeah.
- 30:49
- Unidentified speaker: Well, and and just I had a founder send me, uh I I tweeted about this recently, but I had a founder send me their data room as an MCP server.
- 30:55
- Unidentified speaker: Oh, cool.
- 30:56
- Unidentified speaker: I was like, I was like, that's awesome.
- 30:59
- Unidentified speaker: Right.
- 30:59
- Unidentified speaker: Oh, that's cool.
- 31:00
- Unidentified speaker: If that stands out to investing, I think, I think things like that, like the local level are interesting to me because that's like the type of nuance we're trying to get at.
- 31:09
- Unidentified speaker: And and these AI systems are, you know, gonna be built at every every level of government, every level of business.
- 31:17
- Unidentified speaker: I think uh another thing that I'm just very excited about is our classifier.
- 31:21
- Unidentified speaker: So the the AI that we are building has been performing very well with very little data, as the hypothesis that we have states is that if we bring in more nuanced labeling, we'll get this AI to be very accurate very quickly without having to invest years and years into data buildup.
- 31:39
- Unidentified speaker: So yeah, I'm excited about that.
- 31:40
- Unidentified speaker: We've got some big customers in the pipeline, and we just kicked off our round.
- 31:44
- Unidentified speaker: So I think we're at this really cool hypergrowth inflection point for reliable.
- 31:50
- Unidentified speaker: And I think it's a it's a point where every founder hopes they, I guess, get to because I think the the early, early stages are some of the most difficult, but then this you kind of see all your hardware.
- 32:02
- Unidentified speaker: You start getting product market fit, you start feeling it.
- 32:04
- Unidentified speaker: You're like, I'm selling this easier now, and it's much more fully featured, and people are like buying this a lot more.
- 32:12
- Unidentified speaker: I have a product.
- 32:13
- Unidentified speaker: Yeah, I have a product that people buy and they come back and use it.
- 32:16
- Unidentified speaker: Yes, yeah, that's really cool.
- 32:18
- Unidentified speaker: They're not churning, they keep paying me every month.
- 32:21
- Unidentified speaker: So, yeah, we're we're at a very exciting point in the startup journey in general.
- 32:25
- Unidentified speaker: Do you think you think long term, do you think we're annotating into the heat death of the universe?
- 32:30
- Unidentified speaker: Or at some point, is so smart and so good and so cheap.
- 32:33
- Unidentified speaker: We're and we're certainly building out data centers like left and right and compute clusters all over the place.
- 32:38
- Unidentified speaker: And I think it it'll probably take a few more years to play out, right?
- 32:41
- Unidentified speaker: I think we're we're probably like in 96 internet, is what it feels like to me.
- 32:46
- Unidentified speaker: Having been 16 in 1996, I I kind of remember what it felt like.
- 32:50
- Unidentified speaker: Sure.
- 32:51
- Unidentified speaker: And I feel like AI is probably in 96, 97.
- 32:54
- Unidentified speaker: So I think we're still two or three years from like overbuilding and like, oh crap, we overdid it again.
- 33:01
- Unidentified speaker: Like we always overdo it.
- 33:02
- Unidentified speaker: We'll go too far.
- 33:03
- Unidentified speaker: But like in the future, you can imagine tokens being almost too cheap to meter and really super intelligent.
- 33:09
- Unidentified speaker: You can already see like the intelligence density per token per is going up, right?
- 33:14
- Unidentified speaker: Yeah, and you know, it there's various estimates, but AI gets 10 to 20x more price performance per unit of intelligence delivered every year, according to the different estimates.
- 33:24
- Unidentified speaker: It's been been doing that for a long time.
- 33:25
- Unidentified speaker: So at some point, it's like annotation unnecessary is the short version of that question.
- 33:30
- Unidentified speaker: That rambling question I just gave you.
- 33:33
- Unidentified speaker: Yeah, I think there's a well, I I love that question because it brought in a lot of interesting things I think about a lot with AI.
- 33:39
- Unidentified speaker: Because I do think we're I think I'm an AI optimist, I would say, on on the scale and spectrum.
- 33:45
- Unidentified speaker: I think I think I have to be for the work that I do, right?
- 33:49
- Unidentified speaker: But I do think that there is the doing too much without understanding it fault.
- 33:56
- Unidentified speaker: Data centers, I am probably one of the things I'm most worried about with AI is data centers just because we're building them so fast when we don't even like, because there's probably we haven't even like optimized the stack yet.
- 34:08
- Unidentified speaker: Yeah, exactly.
- 34:09
- Unidentified speaker: We're building we haven't optimized uh the memory to chip bandwidth, and there's all kinds of different architectures and for LLMs that we have we haven't replicated the we did this with uh with some other uh AI stuff 10 years ago where you actually adapt a chip to the model.
- 34:25
- Unidentified speaker: And I don't think they've fully done that yet with with large language models.
- 34:28
- Unidentified speaker: No, it's all it's all still so early technology.
- 34:31
- Unidentified speaker: You think about 97, you think about 97 internet, and we've got those big computers, we've got Ethernet cables, the whole, the whole thing.
- 34:38
- Unidentified speaker: And so we're still at that early stage.
- 34:40
- Unidentified speaker: So like to build all these kind of like industrial, kind of inefficient data centers doesn't make a ton of sense.
- 34:47
- Unidentified speaker: There's a really good report that the Kapoor Center put out on like just needing to get more data on data centers, you know, get like give us a little time to investigate some of these things to figure out what the most efficient, what the most sustainable way of doing things is, I think is important.
- 35:05
- Unidentified speaker: And I kind of think the same thing about with data annotation too.
- 35:09
- Unidentified speaker: Cause like I said, I think there needs to be a balance between, you know, working towards these classifiers, because reliable does have this AI strategy where it takes time, but let's say magically we come with the right combination and we train a classifier that can get every contextual system right, we would still want humans in the loop.
- 35:27
- Unidentified speaker: And I think there's also still this advocation of if we are using jobs or if we are using AI to reduce the workload of our workers, can we give that time back to the workers to just have a better life, as opposed to unemploying them or putting it?
- 35:44
- Unidentified speaker: I think that's where AI and the universal income conversation come together too in in these big lofty talks about AI.
- 35:51
- Unidentified speaker: But I think we're still so far from that that it's there's there's time.
- 35:56
- Unidentified speaker: I also think just the way the models work, again, models right now and AI right now is not very good at nuance.
- 36:03
- Unidentified speaker: So if you're asking it to do really anything with nuance, you're always gonna need that machine learning fine-tuning layer and training layer in there.
- 36:13
- Unidentified speaker: But you know, maybe reliable will be the first to make a classifier that can identify everything and then you know, and then we'll have enough money that we can give out at personal universal income to everybody in San Francisco or something.
- 36:25
- Unidentified speaker: So yeah.
- 36:26
- Unidentified speaker: And what's your P Doom?
- 36:27
- Unidentified speaker: Yes, the P Doom.
- 36:29
- Unidentified speaker: So I just want to clarify the P Doom.
- 36:32
- Unidentified speaker: So is it like my my point of probability super intelligent AI kills us all?
- 36:38
- Unidentified speaker: Oh, okay, and humanity and human existence.
- 36:41
- Unidentified speaker: Okay.
- 36:42
- Unidentified speaker: P doom.
- 36:43
- Unidentified speaker: I think honestly, I think it's higher than I'd like it to be just because just because I don't I don't love the way I don't like the way that the biggest AI companies are handling this.
- 36:58
- Unidentified speaker: The the the extreme wealth that's being built.
- 37:00
- Unidentified speaker: Although they're trying to throw off their own culpability.
- 37:03
- Unidentified speaker: Yes, yes.
- 37:03
- Unidentified speaker: Yeah, they're trying to say we are building something so dangerous, you need to regulate us, uh, because we can't build safe products.
- 37:10
- Unidentified speaker: That's basically their argument.
- 37:11
- Unidentified speaker: Yeah, but then building lobbying firms that make it hard to regulate.
- 37:15
- Unidentified speaker: So it's like regulate us, but at the same time, they're back channeling saying don't regulate us.
- 37:20
- Unidentified speaker: Exactly.
- 37:21
- Unidentified speaker: So to me, it it smells like marketing.
- 37:23
- Unidentified speaker: Uh I'll say personally, my PDM is zero.
- 37:26
- Unidentified speaker: I I find it very, very unlikely.
- 37:28
- Unidentified speaker: Uh-huh.
- 37:29
- Unidentified speaker: Basically, as close to zero.
- 37:31
- Unidentified speaker: I mean, there's never there's never P you can never get to zero.
- 37:34
- Unidentified speaker: Sure, sure.
- 37:34
- Unidentified speaker: A probability, but it's as close to zero as you can get because super intelligent AI wouldn't even want to stay on the planet first, first of all, because like there's so much more energy in space in space.
- 37:46
- Unidentified speaker: We're adapted to live here, right?
- 37:48
- Unidentified speaker: We we evolved or created, if you believe that, uh to like live here.
- 37:53
- Unidentified speaker: But AI, it only all it needs is like computronium, compute substrate, and energy.
- 37:58
- Unidentified speaker: And you can find a lot more of that out there.
- 38:00
- Unidentified speaker: And so like it'd be it'd be us, it'd like be it'd be like us like trying to live at the bottom of the ocean.
- 38:06
- Unidentified speaker: Like we could do it, but it's it it's like why do that?
- 38:09
- Unidentified speaker: It's not the optimal place for us to be.
- 38:11
- Unidentified speaker: We want to be up where the sun is and all the plants and vegetation where the food is.
- 38:15
- Unidentified speaker: Same thing with AI.
- 38:16
- Unidentified speaker: Earth would be like the deepest, darkest ocean for them.
- 38:19
- Unidentified speaker: Yeah, I'm not I'm not so worried.
- 38:21
- Unidentified speaker: Superintelligence, I don't think, would want to hurt us because I think if it's truly super intelligence, it's gonna be just readthink through and go, like, why would I do that?
- 38:27
- Unidentified speaker: Yeah, yeah, exactly.
- 38:28
- Unidentified speaker: It's gonna be very compassionate.
- 38:29
- Unidentified speaker: I mean, I I think like, you know, if reliably still that like I think in the extreme, yeah, like Kantian philosopher sense, uh-huh, right?
- 38:39
- Unidentified speaker: It would kind of go, what's the categorical imperative here?
- 38:42
- Unidentified speaker: And it would be extremely rational and truth-seeking.
- 38:45
- Unidentified speaker: Yeah, and it would go, I agree.
- 38:47
- Unidentified speaker: I I shouldn't do that because it's not the right thing to do.
- 38:50
- Unidentified speaker: Like it would just like have morals like that, I think.
- 38:52
- Unidentified speaker: Because it'd be able to reason through from first principles everything.
- 38:55
- Unidentified speaker: Well, yeah, yeah.
- 38:56
- Unidentified speaker: I think I'm not worried about the superintelligence.
- 38:58
- Unidentified speaker: What I'm worried about is the people building the superintelligence, and if they destroy us before we're able to get to the superintelligence, it just all comes back to sustainability, right?
- 39:07
- Unidentified speaker: And just like you're building this really cool technology, but what does it matter if we're not here to use it or to see it denominated?
- 39:13
- Unidentified speaker: You know, fun fact, I don't know if I've told this story on the podcast, but it's just a really good time to tell it because talking about tech and sustainability.
- 39:20
- Unidentified speaker: I think in your in your 20s, you have to be on that that train.
- 39:23
- Unidentified speaker: I was I was director of energy policy for Sierra Cloud for the state of Hawaii.
- 39:28
- Unidentified speaker: And so I was lobbying the government for clean energy legislation and passing like solar roof bills and stuff like that.
- 39:34
- Unidentified speaker: Really fulfilling work.
- 39:35
- Unidentified speaker: And so when I came to San Francisco, I actually didn't even want to work in tech.
- 39:38
- Unidentified speaker: I wanted to work in solar energy.
- 39:39
- Unidentified speaker: That's that was like my first tech job was solar.
- 39:43
- Unidentified speaker: But what I what I what I realized as I started looking at the data is that solar was doubling at a very predictable rate.
- 39:51
- Unidentified speaker: And then soon, and probably around 2030, 2035, we'd have more than enough solar energy to power the whole country.
- 39:59
- Unidentified speaker: Right.
- 39:59
- Unidentified speaker: And every time it was doubling very predictably, I think every at that point, 24 months, it was an installed capacity.
- 40:05
- Unidentified speaker: And so if you just think exponentially and you just you just plot the exponential, you're like, oh, well, energy's solved.
- 40:11
- Unidentified speaker: Yeah, right?
- 40:12
- Unidentified speaker: Yeah.
- 40:12
- Unidentified speaker: And I think I think a lot of things like that, with that's why I'm such a techno optimist.
- 40:16
- Unidentified speaker: I think a lot of things tech creates its own problems, it always does.
- 40:19
- Unidentified speaker: But on the whole, I don't think anybody wants to go back 50 years.
- 40:23
- Unidentified speaker: No.
- 40:23
- Unidentified speaker: And live with 19, you know, no 60, whatever technology, right?
- 40:28
- Unidentified speaker: Nobody wants to talk about that.
- 40:29
- Unidentified speaker: We're talking because of this technology, and I was like, Yeah, we couldn't be doing this podcast right now.
- 40:33
- Unidentified speaker: So I think that's that's what it is.
- 40:35
- Unidentified speaker: I I'm very much a tech optimist, but I do think that that tech needs to be built responsibly.
- 40:40
- Unidentified speaker: I also think with it's like help us reach the full potential of the technology.
- 40:45
- Unidentified speaker: Because that's my thing, is like like with reliable, it's like we can't even get to the coolest parts of the technology that we wanted unless we have this, but we do have the.
- 40:54
- Unidentified speaker: I think a lot of times people have said it's impossible to do that.
- 40:57
- Unidentified speaker: And what we're trying to say is it's not impossible.
- 41:00
- Unidentified speaker: You just need the right software to do it.
- 41:02
- Unidentified speaker: And so I it's a very exciting time to be alive, not to end the podcast on a cliche note, but it is very, very exciting because I think we have the opportunity to achieve to solve really big challenges.
- 41:15
- Unidentified speaker: I think we'll look back, I think we'll look back 10, 15 years from now.
- 41:18
- Unidentified speaker: Because the next 10 years is gonna be really crazy.
- 41:20
- Unidentified speaker: Yeah, and we'll go, yeah, hell no, I don't want to go back to 2026.
- 41:24
- Unidentified speaker: Oh no.
- 41:24
- Unidentified speaker: Because of all the discoveries AI will make, all the all the menial jobs that will be automated away with robots and AI.
- 41:31
- Unidentified speaker: Will people be able to find fulfillment?
- 41:33
- Unidentified speaker: I read a whole article on this actually.
- 41:34
- Unidentified speaker: Yeah, we have an example of that.
- 41:36
- Unidentified speaker: People like in their like late 50s and 60s that retire, as long as they're healthy, they're really happy.
- 41:41
- Unidentified speaker: They're super happy or they're happier than they've ever been in their lives.
- 41:44
- Unidentified speaker: Well, that's the thing.
- 41:45
- Unidentified speaker: People are naturally driven, they'll do stuff that like interests us to have.
- 41:48
- Unidentified speaker: They'll have hobbies.
- 41:49
- Unidentified speaker: Yeah, oh yeah, for sure.
- 41:50
- Unidentified speaker: Yeah, I'd like to totally feel my time.
- 41:53
- Unidentified speaker: I don't have enough time for my hobbies.
- 41:55
- Unidentified speaker: Exactly.
- 41:55
- Unidentified speaker: Well, let's let's do some wrap-up questions.
- 41:57
- Unidentified speaker: Okay.
- 41:58
- Unidentified speaker: Um, what's something you believed with total conviction five years ago that you'd now argue against?
- 42:03
- Unidentified speaker: Ooh, what a good question.
- 42:04
- Unidentified speaker: Well, this is a little okay, so this is a little extreme, but I'm I'm a little of an extreme person.
- 42:09
- Unidentified speaker: So I would say five years ago, I was much, much more gung-ho on electoral politics.
- 42:15
- Unidentified speaker: I think I was very much like, you know, like that is the best way to, you know, solve societal problems.
- 42:24
- Unidentified speaker: I think living in San Francisco and seeing how tech moves things here, and also seeing, seeing like the like direct community work I do, like I volunteer and things like that, and the impact that can have made me very excited about, you know, my role as an individual in a city and the impact I'm able to have, and also kind of made me put less time into like politics or and energy towards that, and more just into direct things I can do with my my community and and the work that I do.
- 42:58
- Unidentified speaker: And that's not to say I don't vote or I don't encourage people to do that, but just for me personally, it's it's been helpful because it helps me feel like I'm doing positive things without having to follow so carefully the waves of politics, I guess.
- 43:12
- Unidentified speaker: Yeah, I think you know, we have a not to get political here, but we have a Republican government, which means it's a republic, which means we elect representatives.
- 43:20
- Unidentified speaker: But what's it what's interesting about it is that the Senate is completely disconnected from the population.
- 43:27
- Unidentified speaker: So you could have two senators from Alaska, two senators from Wyoming, uh-huh.
- 43:31
- Unidentified speaker: Like and those people in those states are getting a disproportionate influence on politics versus a state like California or Texas, also two senators each with 30 million people each or whatever it is.
- 43:42
- Unidentified speaker: Yeah, that's a very interesting design pattern.
- 43:44
- Unidentified speaker: But I actually learned something interesting about the Senate recently is that we didn't even elect our senators until recently, the state legislatures.
- 43:51
- Unidentified speaker: It was like 50, 60 years ago that actually people started even electing senators.
- 43:56
- Unidentified speaker: So before that, it was the people you elected electing the senators, almost like prime ministers.
- 44:01
- Unidentified speaker: The Senate was supposed to be a check on the people, kind of the people, yeah.
- 44:05
- Unidentified speaker: That's how they designed it.
- 44:05
- Unidentified speaker: Yeah, sort of like a blunt to the mass populist.
- 44:09
- Unidentified speaker: Which a question mark about, but the it's funny because it's like even talking about that, my entrepreneur brain is like, how do we innovate to solve that?
- 44:18
- Unidentified speaker: How do we fix government?
- 44:19
- Unidentified speaker: Yeah, exactly.
- 44:20
- Unidentified speaker: I I haven't done this in a while.
- 44:21
- Unidentified speaker: I I asked AI that on like deep thinking mode.
- 44:25
- Unidentified speaker: Like how if you were a founding father or mother, I guess if you were like a if you could do it all over again, just redo the the whole like the whole democracy, how would what kind of changes would you make?
- 44:37
- Unidentified speaker: And I put it like in like max thinking mode.
- 44:39
- Unidentified speaker: And this is probably a year ago, so it's probably the answer would probably be even better now.
- 44:42
- Unidentified speaker: But it had a lot of really good tweaks.
- 44:44
- Unidentified speaker: I think I think getting money out of politics, and then that's like job number one.
- 44:50
- Unidentified speaker: And then two would be oh I lost my train of thought.
- 44:52
- Unidentified speaker: Well, hold Senate, the electoral assembly and stuff has but it's just in a startup running a startup, you would be like, Okay, let's pilot this, let's move really fast, but government just doesn't do that.
- 45:02
- Unidentified speaker: So I think that's one of the reasons why I was always drawn to entrepreneurship because I feel like I can make influence like fat much faster than through.
- 45:11
- Unidentified speaker: Balance budget, that's what I I remember, right?
- 45:13
- Unidentified speaker: Right, the states have to balance their budgets, but the federal government does not.
- 45:18
- Unidentified speaker: I didn't know that.
- 45:18
- Unidentified speaker: And I think it's okay, like in emergency situations like war or pandemics, to run the deficit, but we are running such a deficit now that it's like putting the future of this country in peril.
- 45:28
- Unidentified speaker: Sure.
- 45:29
- Unidentified speaker: Yeah.
- 45:29
- Unidentified speaker: Anyway, no, I'm gonna get off my uh my political soapbox.
- 45:33
- Unidentified speaker: Well, hey, if if the US government wants to hire me to build an ontology to balance the budget, I go for it.
- 45:40
- Unidentified speaker: So what's a popular idea in AI or data that you think is wrong?
- 45:46
- Unidentified speaker: Yeah, this actually came up at a recent event.
- 45:48
- Unidentified speaker: Like the audience was like kind of torn about it that I was at, but uh, that more data will fix the problem, I think is a is a concept that is like pretty that was kind of the go-to standard until like just a few years ago.
- 46:04
- Unidentified speaker: I will say reliable was talking about it before just a few years.
- 46:08
- Unidentified speaker: But this idea that basically if you just push the problem.
- 46:12
- Unidentified speaker: Yeah, it'll fix the bias problem, it'll fix the racism problem, it'll fix the mathematical problems.
- 46:18
- Unidentified speaker: Well, we just get more data, have more mathematics, PhD students label it, it'll be fine.
- 46:24
- Unidentified speaker: But the thing is, is it's you're pushing more data through a tainted system, and then it's coming out tainted, and you're wondering why.
- 46:32
- Unidentified speaker: It's just like it doesn't matter how much you push through, it's all it's always going to create the same residue from that kind of icky system.
- 46:39
- Unidentified speaker: It's like washing your car with dirt dirty water, just keep washing it.
- 46:41
- Unidentified speaker: Yeah, exactly.
- 46:42
- Unidentified speaker: No, it's like dirty water, and then you run out of water and you're like, why is my car not clean?
- 46:48
- Unidentified speaker: So I think that's that's my main thing is that we have to look at the quality of the data and not just like the data sets themselves, but how we are preparing that data for training.
- 46:58
- Unidentified speaker: And I think once we do that, a lot of the efficiency questions will also be answered uh too.
- 47:03
- Unidentified speaker: What's the best advice you ever got?
- 47:05
- Unidentified speaker: I would say if it counts as advice, but I've just had I've had a lot of professors that like just kind of encouraged me to follow weird trains of thought that I think has been very helpful because you know, just being curious, I guess, I think is the general advice I've gotten from some of my best professors is just like if you're if there's something tugging at you that you really like want to understand more, just go for it.
- 47:35
- Unidentified speaker: Even if it's a weird thing to dig into, or even if people tell you I think you're wrong, if it's really tugging at you or pulling at you, just go for it.
- 47:42
- Unidentified speaker: And that curiosity led me to start my startup.
- 47:45
- Unidentified speaker: I I would say my first startup and continue to be in startups.
- 47:48
- Unidentified speaker: Who saw something in you before you saw it in yourself?
- 47:51
- Unidentified speaker: Well, that's a nice question.
- 47:53
- Unidentified speaker: I would say, well, I'd say my mom.
- 47:56
- Unidentified speaker: My mom has always been very proud of me, so that's a good one.
- 47:58
- Unidentified speaker: And then my professors at William and Mary, because the my first startup actually started as a an undergraduate project.
- 48:06
- Unidentified speaker: And and so they they were very excited about it.
- 48:09
- Unidentified speaker: They like want to talk to me about it, and and then it ended up growing to, you know, several chapters across the southeast United States and then into the social media platform.
- 48:18
- Unidentified speaker: And I don't think I would have done that if they hadn't told me, like, hey, this is really cool.
- 48:22
- Unidentified speaker: You should keep pursuing this.
- 48:24
- Unidentified speaker: But really enjoyed the conversation.
- 48:25
- Unidentified speaker: Yeah, I want to ask all these questions back to you, but I know we don't have time.
- 48:29
- Unidentified speaker: But I'll I'll get those answers another time.
- 48:31
- Unidentified speaker: Where uh where can folks find you online?
- 48:34
- Unidentified speaker: Yes, so they can find us and all of our socials if you go to reliable.ai, and then all of our socials are on that site.
- 48:42
- Unidentified speaker: And then I But it's not reliable how you would spell it normally.
- 48:44
- Unidentified speaker: It's missing.
- 48:45
- Unidentified speaker: Oh, good point.
- 48:46
- Unidentified speaker: Yes, yes, it's reliable without the e dotai.
- 48:50
- Unidentified speaker: And then because we are a tech company, so we had to drop drop the vowel.
- 48:54
- Unidentified speaker: And then fancy easier to get the domain that way too.
- 48:58
- Unidentified speaker: Yes, exactly.
- 48:59
- Unidentified speaker: Um, and then on on personal socials, you can find me at Andrea Francis B and mostly on uh LinkedIn is my social media of choice these days.
- 49:08
- Unidentified speaker: Much more positive on LinkedIn.
- 49:10
- Unidentified speaker: I like it.
- 49:11
- Unidentified speaker: People just keep it positive there.
- 49:13
- Unidentified speaker: I just get to talk about all the nerdy stuff I want to talk about without trolls calling me a nerd.
- 49:18
- Unidentified speaker: Yeah.
- 49:19
- Unidentified speaker: Yeah.
- 49:19
- Unidentified speaker: I totally yeah.
- 49:20
- Unidentified speaker: Long story there, but uh really enjoyed it.
- 49:22
- Unidentified speaker: Oh, this has been so fun.
- 49:23
- Unidentified speaker: Thank you for having me.
- 49:25
- Unidentified speaker: Thank you.
More from Ignite: Conversations on Startups, Venture Capital, Tech, Future, and Society
Ignite Angel: Why Most Startups Shouldn't Raise Venture Capital with Lance Cottrell | Ep310
Episode discussion: Lance Cottrell argues most startups should avoid venture capital, prioritizing survival and bootstrapping over hypergrowth.
Speakers’ claims: Lance Cottrell, founder of Anonymizer and angel investor, argues that venture capital is unnecessary for most startups and often detrimental to founder control. He emphasizes that survival is the primary metric for early-stage companies, citing his own experience of exiting with cash and zero debt after raising only $2.5 million. Cottrell advises founders to validate assumptions through rapid, low-cost experiments rather than elaborate studies, and warns that AI has eroded software development as a competitive moat. He also highlights the importance of forensic analysis on failed competitors and the limits of productive intellectual work, suggesting a cap of four hours of deep work per day.
Ignite Startups: How Cactus Is Building the AI Operating System for Trades Businesses | Ep309
Episode discussion: Cactus founders discuss building an AI operating system for trades businesses after pivoting from previous ventures.
Speakers’ claims: Co-founders Ajith Govind and Avinash Joshi of Cactus explain their pivot from a CPG AI platform and a chef marketplace to an AI operating system for trades businesses. They argue that voice AI is a commodity and the real value lies in owning the end-to-end customer experience and replacing fragmented legacy software. The founders detail their decision to avoid integration with incumbents like ServiceTitan to maintain platform independence. They target the mid-market segment, citing significant revenue loss from missed calls as a key pain point. Finally, they emphasize the necessity of a human enablement layer for adoption and focus on revenue retention over vanity metrics.