Sign up

What matters to you matters to us! Customize your newsletter–tell us what you're most interested in and we'll handle the rest.

loader image




July 12, 2023

Navigating the Future of Generative AI: Challenges and Strategies for Entrepreneurs

Generative AI may have come into widespread public awareness only in recent months, but some companies have been developing its potential for years. Norwest assembled a panel of leaders in the application of AI to discuss how entrepreneurs and startups can tap its promise and avoid its pitfalls.

Insider’s VC and Startups reporter, Stephanie Palazzolo, moderated the event. The panelists were:

  • Richard Socher is the founder and CEO of, an AI-powered search engine. Richard has also been an adjunct professor at Stanford University and a Ph.D. student at the university. Their research interests include deep learning, machine learning, natural language processing, and computer vision.
  • William Ballance is the co-founder and CEO of Lavender, a real-time email assistant that combines data science, psychology, and artificial intelligence to help users write emails that get more replies in less time.
  • Weiping Peng has more than two decades of experience building platforms and applications, with a focus on bringing bespoke and highly scaled AI solutions to production. Currently, at Airbnb as a distinguished engineer, she is focused on enhancing the Airbnb platform with the latest advances in AI technology.


Below are excerpts from the webinar:

The Rise and Democratization of Gen AI Holds Significant Promise for Startups

SP: As Generative AI has become more mainstream, how have you seen the demand and interest change?

RS: I’ve been involved in natural language processing and AI research for many years, and to now see this crazy moment where everyone knows about it — this inflection point feels incredibly exciting. It’s been a massive change for us. has been around since late 2021 and has grown reasonably well. But since we launched YouChat and made chat the default way to interact with information on the web, we’ve had rocketship growth. It’s been incredibly fortunate timing.

For the first time, users are accepting the fact that there could be a better and different way to search the internet. Even mid last year, for a lot of users it was still too different of an experience from Google. They’d often say ‘I want it to be a little more similar’ and that has changed. And that may mean there’s a lot of opportunity now to disrupt the search engine space.

SP: There’s no doubt that ChatGPT has been a bit of a rising tide for all AI startups out there. On the other hand, what challenges are you facing selling gen AI products to enterprises?

WB: Especially since ChatGPT went mainstream, there’s been a lot of noise in the market and a lot of promises being made. Many times the buyer doesn’t understand the real capabilities of what’s currently available on the market. They think it’s this silver bullet that can do so much — and there are great things coming from these models — but there are also limitations; a lot of hallucinations; a lot of things the models can get wrong, especially if left untrained.

On the one hand, it’s really exciting to see entrepreneurs launch new products. But I also find that a lot of these are just front ends to ChatGPT. They’re putting a wrapper around this off-the-shelf technology, which introduces noise for us. A lot of really great things are coming from these models but there also are limitations; a lot of hallucinations, a lot of things the models can get wrong. As buyers understand the technology — and we’re starting to see that now — they’re able to see the nuanced differences between tools. What’s interesting is the ability to point the technology at specific use cases and build the application layer around it.

Companies of All Sizes Have Many Options for Incorporating Gen AI

SP: How should later-stage startups and established companies think about incorporating generative AI into their offerings?

WP: (I am here to present my personal views; I don’t represent Airbnb.) We all know AI can be used to solve many challenges, but many cannot be solved by AI alone. So, rushing into production with this technology could make you fall into the age-old pattern of a solution looking for a problem. And it will likely take more time than you think. How do you control the hallucinations? How do you make sure you are recording feedback?

One thing I’d recommend for companies to do is what we call a tech stack readiness check. Ensure your technology stack is ready to adopt the technology and that it aligns with your company’s vision and your own knowledge base to make sure it behaves the way you want. Benchmarking is a very important step. You should know what’s good and what’s bad for your specific use case, then do some benchmarking to validate which model is usable for you. Then you can proceed to production sites where you learn how the model is performing for your expected outcome. Without such a measure, it’s almost like you let a puppy run around the house and don’t know where it is.

One thing I’d recommend for companies to do is what we call a tech stack readiness check. Ensure your technology stack is ready to adopt the technology and that it aligns with your company’s vision and your own knowledge base. – Weiping Peng

SP: A big question for up-and-coming generative AI startups is around what model they should build. Some are building on top of a third-party model provider like OpenAI, others are using open-source models or even building their own. I’m curious to hear from William and Richard, why did you choose either approach and how should startups think about that decision?

WB: The ability to build AI with off-the-shelf models is great for entrepreneurs. The democratization of AI allows anyone to start building apps easily. With Lavender in particular, we’re able to progress so rapidly because we were early in incorporating ChatGPT. And then our space of sales emails was quite novel — there was no one really doing that. We were never replacing the salesperson, but we were able to help them write their emails over 50 percent faster while increasing reply rates. At the time, we were bootstrapped. We didn’t have the backing of Norwest. It would have been impossible for us to build an entire platform and implement gen AI within our resources. We’re fortunate that OpenAI released the API for us to use for GPT3.

The key for founders is that it helps them build faster and cheaper, which allows innovation to spread a lot quicker. Instead of spending our limited resources, time and manpower on building text-generation models, we focused our attention on text classification models for sales emails on top of our customer’s unique data. For startup founders, these new generative AI models allow them to get started immediately with very limited resources.

The ability to build AI with off-the-shelf models is great for entrepreneurs. For startup founders, these new generative AI models allow them to get started immediately with very limited resources. – William Ballance

RS: William brings up a great point. The free training of neural nets both in computer vision and natural language processing enables startups to quickly get to an 80 percent solution, and then collect training data. That gets a flywheel going of using that training data to fine-tune and improve the models. Over time we will see large language models (LLMs) and pre-trained neural nets become more like what databases have evolved into: it doesn’t really matter what database you use, it’s what you do on top of it. We’ll have to get back to the basics: how you go to market, what your distribution strategy is. Which AI you’re using might be less relevant. It’s about how you create a good user experience.

SP: That theme has popped up in a lot of conversations I’ve had around a future where your model isn’t the main differentiator, but it’s everything else that’s traditionally been differentiators for startups whether it’s distribution, go-to-market strategy or product. Weiping, any thoughts here from a technical perspective?

WB: Which AI you’re using will become less relevant. People will probably combine AIs and move their data from one platform to another. Data, not the AI platform, will provide a strong competitive moat for a lot of companies.

How Have You Approached the Ethical Quandaries Around Gen AI?

SP: How do companies use Gen AI in an ethical and safe way?

RS: There potentially are a lot of pitfalls, but in some cases, people are a little overly worried. For example, millions of people like “Game of Thrones” and Stephen King novels – where you have all kinds of horrible things – but we don’t say they are unethical to write them. I think sometimes we measure AI with a different yardstick than in many other areas.

Misinformation is an interesting issue. LLMs can write a lot of stuff, so it’s a matter of how you distribute information and how you teach people to trust it. It feels like we need to have another campaign like when we told our grandparents and parents not to trust everything they read on the internet; check their sources, etc. I think another re-education effort will be needed.

WP: We all put personal information in places we trust, like banks. So, the first thing we should think about is building trust. Trust is hard to gain but really easy to break, so I’d prioritize that over any new features.

RS: The U.S. is learning that as speech becomes easier to create and distribute, it’s pushing the very definition of speech. Different countries and cultures will have different answers.

WB: Free speech will be a hotly contested topic. Does AI itself have free speech? If I use AI to generate tweets or whatever, is that an extension of my free speech?

RS: We need regulation, not for AI in the abstract, not for models that are being researched, but for applications of those models.

WB: Over time, players will come in to detect AI and then reduce the amount of misinformation.

When we started, we intentionally had a very diverse group, engineers on our team and people around the world helping train our models. We enlisted very early on a DEI consultant to make sure we were building models that were ethical, fair, and weren’t going to discriminate.

SP: What about the impact of Gen AI on jobs?

RS: There are many repetitive jobs that feel very boring to most people, and these will be increasingly automated. I don’t think 100 percent of the job category will go to AI, but probably 100 people will be replaced with 10 people that use AI. Some service workers will say ‘I’ve already answered this question 20 times and I shouldn’t have to respond anymore to that kind of query. It should have been automated.’ Many boring jobs going away is exciting in the long term but stressful short-term.

Two centuries ago, over 90 percent of people worked in agriculture, and they would have been shocked if you told them that most will not work in agriculture anymore. You could understand how people would get scared of that future initially. But if you look back now and say ‘who wants to work with their hands every day in sun and the cold?’ no one would say yes.

So long-term I’m very optimistic; the short term, however, will put pressure on people to continuously pursue education, learn new skills, learn how to use AI and be one of those 10 percent that are 10x more productive.

Long-term I’m very optimistic about AI; the short term, however, will put pressure on people to continuously pursue education, learn new skills, learn how to use AI and be one of those 10 percent that are 10x more productive. – Richard Socher

WB: We think about how we build AI in a symbiotic way that keeps the human in the loop, because in our space (sales), it’s all about relationships. There are going to be a lot of jobs and parts of jobs that are quite redundant and monotonous that will be replaced by AI.

Our goal has never been to replace salespeople with AI, but to help them write their emails over 50 percent faster, while also increasing reply rates.

WP: AI won’t replace a human, but a human who doesn’t know about AI at all will be replaced.

Gen AI Holds Both Promise and Pitfalls

SP: When we look down the road five years, what are you most scared or nervous about, and what are the things you’re most excited about?

WB: There are going to be some growing pains in this transition, but in the end, we’ll be harnessing the power of AI to add value in brand new ways. Entrepreneurs are going to apply AI in ways that we can only dream of, and the pace of entrepreneurial innovation is what really excites me. It’s going to create a lot of innovation, fueled by the relative ease of creating new technologies.

RS: Gen AI is going to make things very exciting in the creation of art. If you want to see more art in the world, you’ll be very excited about generative AI for videos, images, music. As artistic creation becomes cheaper and faster, human judgment and fast iteration become more important. Things that will make our lives more interesting.

WP: We’re smart beings. Given a good tool like Gen AI, we will adapt and leverage it and move higher. I look forward to a lot more creativity.

Related Stories


/ Product

April 19, 2023