How to Build an AI Startup

Comparing the Old Way vs. New Way

A decade earlier, building an AI startup was incredibly hard.

I founded my first AI startup, Qurious, right out of college. I had zero work experience, no successful exit under my belt, and no PhD. Looking back, I must’ve been crazy.

Think about all the challenges a regular non-AI startup already faces:

• achieving product/market fit

• delivering value to customers

• sales and marketing distribution

• customer support

• hiring and retaining talent

With an AI startup, formidable challenges were stacked on top of all those:

• curate and preprocess large data sets

• achieve high accuracy in real-world production environments

• train multiple ML models for varied tasks

• continuous QA and finetuning to improve accuracy, handle edge cases, and avoid model drift

• manage all the infrastructure and cloud costs associated with running multiple large models

• raise funding to finance this herculean effort

Today, generative AI has simplified building AI startups, making them accessible to all entrepreneurs, without having to raise millions in funding to get started.

Still, AI startups are not easy. Startups remain a challenging, although rewarding, endeavor. But now, you don’t have a whopping list of formidable technical challenges stacked on top of the fundamental startup challenges.

Old Way vs. New Way

Here’s the Old Way of building an AI startup:

• Hire an expensive AI/ML team, including scientists to train models from scratch and engineers to manage the deployment pipeline.

• Curate, clean, and preprocess large datasets, required for training models from scratch and continuous finetuning

• Train and deploy models, managing all the cloud infrastructure while striving to achieve (often unrealistic) customer expectations of high accuracy in real-world environments

• Train multiple models for distinct tasks, for example, sentiment analysis and summarization.

Here’s the New Way, much faster and leaner:

• Integrate with a generative AI API, like OpenAI, minimal coding required

• Write prompts in plain English, no coding required

• Finetune with small datasets to improve accuracy, rather than curating and preprocessing large datasets for model training from scratch

• A single LLM prompt can handle multiple varied tasks, which would’ve previously required training multiple distinct ML models

• Use AI coding tools like Lovable.dev, Bolt.new, v0, CursorAI, WindsurfAI, or Cline to build your MVP 10x faster than ever

Example

Here's an example showcasing the power of ChatGPT to perform 4 complex NLP tasks, analyzing my original LinkedIn post and its comment thread:

1. Text Classification with custom classes: classify comment as either "sarcastic", "agree", "disagree", "funny", "didntgetit", or "dontcare".

2. Text Summarization: summarize thepost's comments

3. Sentiment Analysis: describe the overall sentiment of comments

4. Text Generation: refute the post's summarization


I fed this prompt into ChatGPT along with the LinkedIn post and comments:

Here’s ChatGPT completing the text classification task:

Here’s ChatGPT completing the summarization and sentiment analysis tasks:

Finally, ChatGPT synthesizes this information and analysis to write a rebuttal:

This complex multi-step analysis took me minutes with ChatGPT.

I was impressed by its high-quality balanced summarization and rebuttal.

It would’ve taken months, possibly years, the Old Way to approach the quality of ChatGPT’s output, especially with hard tasks like summarization and rebuttal writing. What you take for granted today with ChatGPT was near-impossible 10 years ago for any non-FAANG company to achieve.

Generative AI has radically simplified what it takes to bring an AI startup to market. It’s now more time-efficient and cost-effective than ever before to build an AI startup. Many applications - including but not limited to transcription, image recognition, and translation - no longer need an army of AI/ML scientists/engineers just to get started.

Competitive Advantage

What is a startup’s competitive advantage in this New Way?

First, competitive advantage has never been limited to technology. Consider the entire customer experience and value-add you’re able to deliver. Startups can have competitive advantages through superior sales, marketing (e.g. ClickFunnels), customer service (e.g. Zappos), partnerships, pricing, onboarding, and more.

Building a defensible technology moat starts with having a deep understanding of customer pain points and building your unique insights into the product. Often, this takes the form of superior UX, features, workflows, and integrations. Also, there’s tremendous value and IP in finetuning your models and optimizing your prompts and agentic pipeline, leveraging domain-specific or customer-specific data sets, in order to solve pain points really well.

Finetuning

I’ve used the term “finetuning” several times. For AI startups, the goal of finetuning is to get a working model as fast as possible, so that it’s good enough for your customers, thereby minimizing the time to product/market fit. Then, your customers are happy, want to keep paying you, and you get to collect domain-specific data, allowing you to finetune more, while building a competitive data advantage.

In the Old Way, before GPT, you’d often start with a crappy model or train from scratch, then spend tons of time, money, and resources finetuning it.

In the New Way, tasks that were previously very hard (e.g. text summarization) are now much easier with minimal finetuning and out-of-the-box LLMs. You can get pretty far by finetuning your prompts, no coding required. You can improve prompts via prompt engineering techniques or metaprompting. You can adjust variables like temperature and context. You can provide concrete examples to your AI API so it can learn faster, aka few-shot learning. You can employ a RAG to improve reliability and accuracy and reduce hallucinations. You can also leverage, for example, OpenAI’s out-of-the-box finetuning API, which can digest more examples than few-shot learning. You can build an agentic flow that makes decisions dynamically in real-time, enabling you to scale intelligence in a way that was impossible 10 years ago.

I anticipate seeing more add-on observability services you can layer into your Gen AI stack to further improve accuracy via continuous monitoring, QA, and finetuning. I’m also excited to see more tools to address, prevent, and detect hallucinations, a painful and recognized problem especially in enterprise.

Summary

This new paradigm for building AI startups is far simplified compared to 10+ years ago. Having built and sold an AI startup the Old Way, I'm excited about the New Way.

AI startups are now accessible to teams that understand painful domain-specific problems but don't have millions in funding to hire scientists or curate massive datasets.

Timing is everything… and the time is now.