AI, accuracy and reliability

ChatGPT hit 1 million users in 5 days. If you haven’t seen the screenshots on Twitter or LinkedIn, ChatGPT is an AI chatbot developed by OpenAI which answers anything you ask it without accessing the internet (for now). It’s the most impressive “front-end” application of AI that we’ve seen. I asked it how to write a piece of code, and I was genuinely blown away by how it broke down the problem.

One of the issues with ChatGPT, which Sam Altman - the CEO of Open AI - admits himself, is that it can sound accurate even if it’s not. Users of a product don’t have the same context as the builders of a product, and products can have unintended consequences (see: social media).

In this week’s deep-dive, I’m going to dig into this problem and how we might solve it. As part of this esssay, I interviewed Sridhar Ramaswamy, CEO of Neeva. Neeva is a privacy-focussed search engine that helps users find exactly what they are looking for. They’ve raised $77.5 million to date and have bold ambitions. I’ll cover why I’m excited by Neeva’s mission and how it relates to the problem we’re discussing.

Solving for accuracy or authenticity

Artificial intelligence is a game of probability. When you train an AI model, you take a dataset and divide into two parts: training and test. We need labelled data to train a model, i.e. if we have 100 observations, we need to know the output for the variable we’re trying to predict for every single observation. For simplicity, let’s assume we’re predicting negative or positive sentiment for these 100 rows.

Let’s assume we divide the sample in a 80 (training): 20 (test). You train 80 rows by giving it the input and the output (negative / positive in the example above). You then apply it to the remaining 20 rows and check how accurate it is. You can do this because you know the answers for the 20 rows. You keep optimising until you get the best predictions.

I’m obviously trivialising the task of machine learning. The intent here is to explain that every prediction by AI is a probabilistic prediction. At its core, ChatGPT is a model that aims to predict the next word given a set of inputs. It can easily convince you that something is accurate. This is because it feels like someone is talking to you and you judge based on things like tone.

This can be dangerous because people will start to rely on it for things they shouldn’t, such as health advice (please don’t). Whilst the accuracy will improve as models get better, it can’t be eliminated because - by definition - it’s a probabilisitic exercise.

StackOverflow recently banned all responses from ChatGPT. It found that users were flooding the site with responses from ChatGPT. The responses were often incorrect but sounded accurate. The ban was required to maintain trust on the website.

There are a few ways to overcome this accuracy problem. The first way to achieve this is by augmenting with human input. Someone says X is true or false. The downside of this is that you need a human in the loop. Also, how does one know whether the human is right or wrong? Another way to alleviate the problem is by giving the user a bunch of options and letting them choose. In effect, this is what Google Search does. Finally, you could give the user a prediction score — this doesn’t solve the problem but gives the user context on accuracy.

It’s a really hard problem to solve. I wanted to dig deep and decided to reach out to a company that’s trying to solve this: Neeva. Here’s why I’m intrigued by them followed by my full interview with Sridhar.

Why I’m intrigued by Neeva

Search is a massive market that is dominated by Google. The graphs below speak for themselves. Taking on Google in their own backyard is a bold mission and I admire it.


The team is impressive. Sridhar spent 15 years at Google working on search. If there’s someone who knows how to execute on this vision, it’s the Neeva team.

Ensuring that information is accurate is front and centre to Neeva’s vision. If Neeva can pull this off, they might become a generational business.

Finally, as users, we’ve been conditioned to free lunches. Can you imagine paying for Google Search? Neeva is trying to challenge this. Changing consumer behaviour is hard, changing consumer behaviour and getting people to pay for something that was free is really hard.

Let’s dive into the interview.

Interview with Sridhar, CEO of Neeva

What is Neeva’s mission?

To re-imagine search. By rejecting the ad-support model, Neeva is building a search experience that puts the user first, does not compromise on quality or privacy, and is leading a new frontier by leveraging in house large language models (LLMs) and refined training models with its full system search stack to bring authentic real time AI search to the masses.


Why is this an important problem to solve?

Search is not only one of the most complicated technological challenges to solve, it’s a daily use function that billions of people around the world rely on for the most important information. This information has an enormous impact on individuals, communities and society. It’s incredibly important to deliver the highest quality and most useful information.

How do you see AI progressing in 2023 and over the next 2-3 years?

The introduction of accessible AI such as the recent explosion of Chat GPT,  represents one of the most exciting technological advances since the iPhone. The potential impacts to our daily lives are impossible to quantify, but we are likely to look back at and mark this moment as BAI (before AI) and AAI (after AI).

Artificial General Intelligence is likely to make search one of the first and most clearly disrupted industries. Soon enough the traditional 10 blue links on a search engine results page will seem as archaic as dial-up internet and rotary phones.

Neeva is leading this new frontier by leveraging in house LLMs and refined training models with its full system search stack to bring authentic real time AI search to the masses.

Neeva currently has an app in beta where the entire blue link experience fades away. A query returns an immersive experience with ML summaries from all the top results, rich photos and an Instagram/TikTok vibe where swipe-able cards highlight authoritative information about that topic. It’s authoritative because it is careful with citations and benefits from the power of authority of the web! (Mayoclinic is a lot more believable than an unknown health blog). The cards also suggest other important follow up questions. Users have the option to find out more on almost everything in the card. This is a search engine engaging in friendly and authentic conversation about a user query backed up with primary references.

This is only the start, and the next 2-3 years will see incredible leaps in the power of AI, the application, and the creation of transformational uses when it comes to search and other industries.

Building a search engine given Google’s dominance feels like a big technical challenge. How have you found competing with Google search?

We are fortunate to have an exceptionally talented team some of whom helped build the core systems at Google. We are now able to leverage advances in cloud computing and deep learning to run one of the largest search stacks outside of Google and Bing. But it’s not just about running a large search stack, by flipping the business model Neeva has the unique freedom to create new features and experiences that deliver faster and more useful results in a visually richer way than traditional search engines.

How do you compare with other options like DuckDuckGo?

For the most part until now, alternatives to Google have come with some kind of compromise. While they may have delivered more privacy, it often came at the expense of quality while still being subjected to ads. Neeva takes a different approach. Instead of relying on a search stack from an API that can’t be personalised or adapted to create exciting user-first experiences, Neeva built its own independent search stack that crawls hundreds of millions of pages a day, an index with billions and serves at scale. So it can offer the best of privacy and quality without any compromise.

Providing search for customers without tracking or ads is an ambitious goal. Are customers willing to pay for a search engine?

What we have learned over the past 10 years is that there is no such thing as a “free” product. If a product is free, then you are the product. We pay for free products with our time and attention.

And when advertisers pay for advertising to get our attention, they turn around and charge us. That $150b made in search advertising, that is coming from consumers (in terms of higher prices for the products they advertise).

Our own research and experience as well as that of other subscription models like Netflix and Spotify show that if you provide value to consumers they will pay for it. Whether it’s the clean, worry free no ad model, only seeing real results, a completely reimagined search experience that leverages LLMs/AI with our real time authentic search results, the ability to connect personal accounts to search, or customising for preferences in news or retailers, as users experience Neeva they see the value and are willing to pay.

How have you found acquiring users given Google’s dominance in the space?

We increasingly see users are getting fed up with a worsening search experience that prioritises advertisers at the expense of user privacy and quality results. As a result, there is a growing demand for alternatives that put the focus back on the user and deliver a search experience that does not compromise on either quality or privacy. Neeva is fortunate to see a positive response in the form of growing users from 600K monthly users as of September to more than 1 million today. The the most recent launch in the United Kingdom, France and Germany further confirms the appetite among users and the opportunity to challenge the traditional search model.

One of the hardest challenges with AI is knowing whether information is right or wrong. How does Neeva try to solve this?

Neeva integrates search and AI deep in its core. Thanks to Neeva’s independent search stack, as we look at a page, we understand its contents, incoming links and other authority signals that tell us whether the page is important and useful. We use AI to process the information on that page to better understand it. When a user types a query, we not only find the best pages but use AI to distill the content of the page into a bite sized nugget. Users can then easily discern what the page is trying to say. Of course, they visit the page to learn more and/or take actions (like buying a product) on the page. The magic of Neeva is in doing all of this in real time as the web changes.

To close

Information reliability, accuracy and authenticity is going to be a problem to solve in the age of AI. The easier content is to produce, the harder it is to know if something is right or wrong. I’m going to follow Neeva because their mission is bold and exciting. One of our most valuable traits as human being is knowing how to adapt and making decisions with the right amount of information. AI is an incredible piece of technology if we use it to gather information and make a decision. AI is dangerous piece of technology if we delegate all of our thinking to it.