에피소드 137: Why AI Hype Is Always Wrong (And What Actually Happens)

This is the nature of what will go viral and you're probably on the platform because you want to yell at something, but I also think it's what happens in every major market cycle.

Every time there's a market shift, a bunch of people will show up and say, "Oh my god, this is going to change the world tomorrow and our lives are going to be different and that's the best thing ever." Another group of people show up and say, "Oh my god, this is going to change the world tomorrow and our lives are going to be absolutely disastrous." And rarely is it either of those things.

Hi, my name is Demetri Panici and I'm a content creator, agency owner, and AI enthusiast. You're listening to the AI Agents podcast brought to you by Jform and featuring our very own CEO and founder, Aytekin Tank. This is the show where artificial intelligence meets innovation, productivity, and the tools shaping the future of work. Enjoy the show.

Hello and welcome back to another episode of the AI Agents Podcast. Here we have the CEO and founder of Polar Grid, Rad Kvachovich. How are you doing today, Rad?

I'm doing well and way to go on the name.

I did it. I was literally doing that before we pressed record, guys.

So just give me a like, subscribe, and review on the Apple podcast so that they're like, "Wow, this guy can pronounce words." No, but in all seriousness, Rod, really excited to have you on the show. Tell us a little bit about who you are and how you got into AI in the first place.

Yeah, totally. So I've been an entrepreneur and founder my whole life. I likely am also unemployable is why. I started my first company when I was 12 years old way back in the 90s building websites and e-commerce when that was like the brand new thing that no one knew about.

I worked with some buddies in the 2010s in a different space and we took a company from a slide deck and then we built it to over a 20 billion valuation on the NY in under five years and now building out Polar Grid and super pumped to be working in the AI space.

Yeah, absolutely. What are you most excited about with what you're doing right now? Tell us a little bit about Polar Grid in general and then tell us first and foremost what are you stoked about right now?

I think what gets me most excited is when there's massive market shifts. It's where I've always worked across different categories and there is no doubt that AI represents that on a global basis.

It's super fun because you get to see everything in the world get disrupted in different ways. There's really no industry that's safe or able to remain stagnant through this market shift.

You get to see it and what it means for my own kids and their schooling. You get to see it in what your friends work and how it's impacted, and for us we get to build at the space.

I just think when you have massive movement on the technology front and looking back to how the 90s and 2000s played out, it's one of those things where we all think about what the impact will be 5 to 10 years down the road.

But we got to live through those years to find out what we got right, what we got wrong, and what takes less time or more time than we expected.

Absolutely. What do you right now? I know that you kind of are, if I'm not wrong, the goal of your company is real-time AI interface at the edge. Now, I'll be honest, there's a bit of buzzwords there.

So I would love to hear just one other thing that I mean obviously under that it says that you're cutting network latency by 70% per request. What is the kind of issue that's happening right now with latency and what prompted you to think that AI was going to be a good solution for this?

Let me take a step back and maybe tell you how we got into the space. So about a year and a half ago we were reading a bunch of books that were written in the late 90s, really looking for analogies with the dot-com boom, things that rhyme that would apply today.

If I take you back to my first company building websites when I was 12 years old, we were putting images on the internet and people were super pumped because now you don't have to send an image through the mail. You can see it on the internet and share it with others and it takes only 30 seconds to load and that feels like real-time magic.

Then I'm dinging myself here, but you go back to 1997 and we have this thing called Adobe Flash with animated introductions to websites with animated pictures moving. Again, it feels real time. It takes only 30 seconds to load.

Then Napster comes along, if you remember those days, and we're able to download single songs in only 12 minutes on dial-up modems, absolutely amazing because we don't have to go to them all anymore.

In all three of these cases, we were enamored with them at the time, but if you look at them now, any of the service providers take single-digit milliseconds from a latency perspective.

If it took a second, you would bounce from the application or the page in a heartbeat. We looked at that and said that feels like it rhymes with what's happening with AI.

The thesis we came up with is when you have large market shifts, we all get enamored with all these new features and capabilities, and rightly so. It's super cool all the things you can do.

But it often takes the market a couple of years to realize that it needs to focus on user expectations. As user adoption increases, user expectations continue to grow.

As competition comes into play, product managers start worrying about user experience. We got thinking about building an edge network to drive real-time AI by comparing the late 90s and now and the confluence across them.

In 1997, there's a company called Akamai who's still around and doing great things today. They figured out that if they put servers at the edge, like in every city, they can cache static data and make things load faster.

At the same time, people figured out how to optimize image files. When you put those two together, images went from taking 30 seconds to render on the internet to being instantaneous.

The analog to today is we have our ChatGPT moment. There are all sorts of models being used across AI products. A lot of focus has been on improving chipsets and models to make things run faster.

That sped up end-to-end latency but the gap that's remained is on the networking infrastructure. We still can't load stuff quick enough to make it actual real time.

That's where we've zeroed in. I can explain why that latency exists and what's different from all the work done across the original internet, mobile, and streaming over the last 30 years and why that was essentially undone as AI came on board.

Please do that.

If you go back to the ChatGPT moment, you suddenly have mass public interest in technology created over the decade or 15 years prior and you suddenly have end consumer and business demand coming for inference.

As that demand comes on, hyperscalers rightly say, "Oh my god, we need to supply this demand." They start putting more GPU servers where they already have them, in centralized hyperscale facilities, and rapidly scale those to respond to demand.

That demand has not stopped and they've continued to build out centralized infrastructure to further scale the supply.

The problem is end-user requests go directly from the user to the centralized service provider. We've undone basically 30 years of networking infrastructure.

As a result, network latency is three to 10 times longer than the traditional internet. We're living back in 1995 from an architecture perspective.

We need to recreate the networking foundational layer to enable faster networking roundtrip time which enables real-time AI inference calls.

Where do you think the reason is that this hasn't been sold at mass scale yet and why you think you were one of the people that figured this out?

It goes back to looking at the internet in '95 versus '97 and '99. It makes sense as you have outsized demand that you continue to build up supply to meet that.

Many companies, whether NeoClouds or hyperscalers, are doing well financially and have made the decision to respond to batch processing.

At the same time, user expectation needs to continue to increase. We're seeing a hyper speed of change with AI. It's really only been about three and a half years since ChatGPT's launch.

User adoption has lagged from when it was first known in the tech community. User expectations continue to grow as newer models come out with higher quality.

We're finally at the stage where user adoption is mature enough to expect faster response times, putting pressure on real-time enablement.

Real-time use cases where latency is one second and user interaction needs to be synchronous are where network latency really matters.

Think of voice AI. When working with a voice agent, you expect it to be conversational with free flow of dialogue and quick responses.

Right now, voice takes around 700 milliseconds as best practice to get a response. If we jump on a call and have a laggy connection, we interrupt each other, making conversation no fun.

Voice agents have great reasoning and cloning but the interruption loop breaks the conversation.

Research from companies like Zoom and Teams shows you have to get from 700 milliseconds down to 300 milliseconds for user satisfaction with conversational flow.

Network latency becomes the bottleneck and has to be solved for.

The difference between 300 and 700 milliseconds is about half a second, which is a big deal in conversation.

When you have to think while talking, it creates awkward conversations where it feels like the other person isn't engaged or is distracted.

That breaks trust in voice agents because it doesn't feel like a real interaction.

I remember when my girlfriend's dad was introduced to our agents. He said it's good for phone calls but the wait time is weird and it doesn't feel like talking to somebody.

What applications do you feel this will apply to the quickest?

Where there is a lot of competition and value to be lost if you have dropped calls. In voice, two obvious places are customer support and talent recruitment.

In customer support, voice agents provide no wait time, better quality answers, and responses in customers' languages.

If the voice agent doesn't respond effectively to build trust and have productive conversations, you lose customers, which impacts lifetime value.

No product manager would accept a poor user experience. You want customers to have fantastic responses and love interacting with your brand.

On the voice agent side, interactions shouldn't be equal to dissatisfied call center experiences but absolutely fantastic.

In talent recruitment, voice agents are used for round one interviews to create equal outputs and move more people through the funnel.

If a great potential employee drops because of frustration with voice agents, that materially impacts your business.

Ensuring the best quality interaction on voice, reasoning, and latency is essential for these products.

I think customer support for industries with elderly callers and ordering food could be weird if latency causes irritation.

HR agents and sales reps are probably miserable to talk to now compared to where it could be.

More broadly, voice agents will lead to video agents in two years, where avatars can reach out to customers with trust and resonance.

Models and chipsets will improve but network latency must be addressed for real-time experiences.

Once you have voice and video, it naturally goes into gaming, where latency is a known problem.

My 12-year-old son loves Fortnite and wishes bots were good enough to not know they were bots.

Latency is critical for high-quality gaming experiences, and we're seeing the 90s and 2000s replay where markets move to real time.

Other models will improve latency to be almost good enough for real-time use cases, unlocking new business opportunities and consumer benefits.

What is your favorite part about what you do right now?

How much everything is changing so quickly. I'm not a person that likes stable footing in work.

This market is going at a rapid pace and constantly changing, creating new opportunities all the time.

This is my fourth venture and that's what makes it fun. I wouldn't be working in this category or a startup if I wanted a boring nine-to-five job.

I could be working in a box factory, but that's not the life for me.

What do you think is most interesting recently in AI in general?

The coding stuff is super cool but feels like productivity machinery for coding, producing more output but not revolutionary.

Open Clawbot is super neat, driving excitement and security issues that the tech community quickly tries to solve.

The world gets focused on one idea for a short period, then moves on to the next thing quickly.

It's a very interesting whirlwind we're living in trying to understand where value will be provided two years from now versus the thing of the moment.

Only time will tell who's right.

You called it a hurricane just now. Interesting choice of words.

I feel hurricane has a destructive element but I don't think it's destructive. Maybe whirlwind is better.

Sometimes it feels like low-grade hysteria with freakouts about stuff.

It grinds my gears when people state technically inaccurate things on LinkedIn to get impressions.

What are your thoughts on that? It's one of my biggest pet peeves.

I get less agitated about it. We all know how clickbait works. This is the nature of what goes viral on platforms because people want to yell at something.

Every major market cycle has people saying AI will change the world tomorrow for better or worse, rarely either.

Look at early industrialization where people were concerned about machines replacing labor but it took hundreds of years to get to today.

Look at the internet where people prophesized no retail stores would exist but many still do 30 years later.

AI is exciting but unlikely to cause everyone's doom or replace all jobs in the next few years.

It's the boring middle ground but humans like controversy so it makes sense.

I totally agree. There's a lot of overindexing on freaking out about AI.

Recently there were freakouts about people leaving Claude company and buying Mac Studios for employees forever.

It's whiplashy and I get an aneurysm from the mixed messages.

Where do you stand on the job market and how it's going to be affected in the next three to five years?

I put myself back to the mid 90s as the best analogy. Word processing was new and scary but changed the workforce.

AI will change the workforce and roles but historically creates new, generally higher-earning jobs.

Digital production jobs have been rewarding and well-paid across many companies.

If people resist new tools or don't learn them, productivity will go down and roles will be harder to find.

If you engage with tools and use them effectively, it can be a superpower and increase demand for you.

It's about how we use tools rather than tools having power over us.

Everything will continue to evolve quickly, so don't rest on laurels about where to lean in today.

How do you go about testing every new opportunity with so many tools out there without wasting your day?

Human community comes into value. We encourage everyone in the company to use various tools and share learnings across staff and peer groups.

There's no answer today that gets you to the answer six months from now because tools will be drastically different.

It's about creating a learning pipeline to get signals when new, more effective tools are available.

Do you think the CS degree will go away or change?

I don't think it will go away but it changes. When I started my first company, I thought CS degrees weren't needed but that's wrong.

CS degrees still matter 30 years later. Some learn best asynchronously, others want school structure.

What drives you as a founder to be self-learning and self-earning?

I don't do well with rigid structure. I skipped most classes but got good marks and eventually left university because it wasn't helping me start a business.

I like finding problems and solving them rather than being employed to work on a small part within a structure.

Companies need both personalities: those who like structure and those who like autonomy.

College for me was mostly for running track. I often felt I didn't learn much in college and wouldn't recommend an MBA universally.

It depends on your path. Some degrees are necessary for specific careers like doctors.

Professors said you'd remember two things from a four-month course and forget the rest.

Founders need to get companies moving and momentum doesn't require a university degree.

How are you approaching hiring with AI improvements?

We noticed applicants using AI tools during interviews and wondered if that's good or bad.

We want people using AI tools to augment their process if applicable to the job.

We ask applicants what AI tools they use in case studies and how they'd use them on the job.

If someone isn't fluent with AI tools, they're probably the wrong fit because they won't keep up with productivity.

Some business owners struggle with people using AI tools and see it as cheating, but I see it as knowing how to get stuff done.

Misrepresenting experience is different. We solve this by asking applicants to disclose AI tools used and how they helped.

Like early Microsoft Word spell check, now it's expected to use tools like Grammarly before submitting work.

Using AI tools in interviews should be about tool selection and human-in-the-loop processes, not cheating.

One team member built a management portal in under two weeks that took 20 people six months a decade ago, thanks to AI tools.

You need human oversight to ensure quality in AI-generated work.

Where do AI agents fit in the workforce?

Agents automate workflows more efficiently than historical tools, integrating with applications and databases.

I don't see agents fully replacing human roles yet, more augmenting productivity.

For example, finance teams may shrink from 10 to one augmented by agents but still need human leadership.

Work removed by agents doesn't necessarily mean roles removed; humans may become superhumans doing more.

The goal is for agents to make humans more powerful, not replace them.

I struggle with doomsday AI scenarios; I see AI as a tool that can be used for good or bad.

AI removes tedious tasks but is a massive shift in productivity and creation.

In 20 years, today's AI tools will seem minuscule compared to future technology.

What is your personal favorite AI tool you use to get more work done?

Simple tools like Renol and fellow.ai act as assistants or chiefs of staff, making me incredibly more productive.

Less layers in companies drive productivity, and these tools simplify company operations.

In the long term, who wins: OpenAI or Polar Grid?

I think open source wins. Outside North America, open source scales rapidly and cheaply off proprietary layers.

Startups have competitive advantages with lower resources and prioritization, benefiting open source.

We'll find out in 20 years what was right or wrong, and there's likely a fifth option we haven't heard about yet.

Thanks for the fun conversation. Please check out everything Rod is doing at Polar Grid at polarrid.ai and follow on Axe and LinkedIn.

Thanks for being here and thanks to everyone who listened or watched. Please leave a review on Apple Podcast. We'll see you in the next one. Bye-bye.