Catching the Fourth Wave

Why I left a job I loved — to build the future of Enterprise AI

This is the scariest thing I’ve ever done — I’ve left an amazing job and great colleagues to … do what?

Image generated using reve.art

For a successful career in technology, you need to catch the wave

In my career, there have been four waves of technology disruption. I caught two of them, missed one, and don’t want to sit the fourth on the sidelines.

Wave 1: The Internet (1998–2006)

The first wave of my career was the internet. In 1998, I was working on severe weather prediction at NOAA. A couple of things came together. The US government was nearing the end of deploying a nationwide radar network, but bandwidth limitations meant that each radar data had to be processed onsite at the local NWS office. Our lab was part of Abilene, an “Internet2” network, and so we were early to the promise of the internet and its ability to connect computers across vast distances. I helped build the ability to connect weather radars across the country into a single grid in real-time — this involved quite a bit of innovation in moving data around on T1 lines, real-time data processing, pattern recognition (for data quality control), and agentic data fusion. [Yes, by 2006, I was using agents]. Deep, important, world-leading work that fundamentally changed how people experience weather warnings, aviation, etc. It was extremely rewarding to work on a hard and meaningful problem at a time of massive change. Of course, working in the government was not lucrative — catching the internet wave by working in internet search or e-commerce would have made me more money, but it’s not about the money. It’s about whether you are working on challenging problems during a technology wave.

Wave 2: Mobile & Web 2.0 (2007–2013)

The second technology wave in my career was mobile and web 2.0. This is the technology wave that led to Uber, Instagram, etc. This wave completely passed me by. I did some half-hearted work on getting a better ground truth for my machine learning algorithms by helping build crowd-sourced weather reports that took advantage of GPS on phones. But there was no depth to it. I just didn’t understand mobile, let alone social. It was a huge miss for me, career-wise.

Wave 3: Cloud and Deep Learning (2014–2023)

The third wave in my career started in 2014 or so — public cloud and deep learning. I well-and-truly caught that one. I was fortunate to join Google Cloud and get a seat on that rocket ship. I helped build the learning organization, hire ML engineers for professional services, set up the Advanced Solutions Lab that launched ML initiatives at Google Cloud’s biggest and best customers, and build the solutions engineering team for analytics and AI. My work affected multiple industries and helped thousands of data engineers and data scientists build new careers. Moreover, being in the center of things, where the world was being reinvented, was heady — in 2018, I worked on applying early Transformer technology with Lukasz, one of the authors of the “Attention is all you need” paper. The model could barely generate one line of poetry and I had really low hopes for it …

Wave 4: Gen AI & AI Agents (Now)

I was wrong, of course. LLMs have improved rapidly and they are incredible at all sorts of things — data extraction, natural language understanding, code generation, tool calling, content reformatting, summarization, question-answering, … the list goes on. [Ironically, they are not good at being factual databases, which is what a lot of people try to use them for]. This is now the fourth wave in my career.

How best to catch this wave?

AI agents can bring amazing ROI, but enterprises need help

With cloud and ML plateauing (and having underestimated how fast the LLMs I was playing with at Google would improve), I started to get more and more interested in the business side of things. How could a business use technology (specifically data and AI) to create value? This took me to private equity, where I had the chance to work on transforming multiple businesses and set in motion incredible value creation opportunities. The best way to learn is to do something multiple times, and if you want to learn how to transform companies and industries, PE is a good place to learn how to do it.

A recent paper from MIT’s NANDA initiative confirmed a few things that I have observed in the past couple of years working with management teams on GenAI initiatives:

1. Tools that don’t learn fail. The primary factor keeping organizations on the wrong side of the GenAI Divide is that enterprises build tools that don’t learn. It is quite common for the pilot version of a GenAI tool to only do the happy path, fail to capture nuances, or succeed only a small part of the time. You need to have the ability to diagnose the errors, improve prompts & context, and build in the ability of AI agents to reflect, and improve.
2. Back office > Front office. So many enterprise initiatives are focused on the “front office” (customer-facing functions like customer support, sales, and marketing), yet the ROI is often in the “back office” because you get double the benefit — when your core workflows are more efficient, you can increase your margin and expand the market (because you can address use cases and customers that used to not be profitable).
3. Internal builds fail twice as often. This is often because internal teams don’t design for what small teams of repeat GenAI builders already know and take into account — the importance of intermediate evals for continuous learning, for example. This results in tools that don’t get adopted or are stuck in the pilot stage.
4. User expectations have shifted. Much like consumer web and mobile applications taught enterprise users to expect better user experiences in SaaS, consumer applications of GenAI are training employees to expect to be able to iterate and improve on software outputs and recommendations using natural language prompts. Also, users now expect that they don’t need to issue the same prompt each time — the tools need to remember and personalize.
5. Demand deep customization. Enterprise buyers need to demand deep customization of any GenAI software that they purchase. What you don’t want to do is to perform a bakeoff and choose the tool that happens to perform best on your evaluation dataset (which will happen by random choice; or worse, because the tool has memorized the public dataset that you are giving to the vendors because you don’t want to jump through the hoops of giving them confidential data). Instead, choose the startups that offer software and engineers who will work with you to learn your workflow and improve over time.
6. Golden age for startups. The successful startups of this era are those that solve for continuous learning (see point #1), memory (see point #4), domain-specific workflows (see point #2), and build a forward-deployed team (point #5).

I loved my job — I loved working with management teams on big and meaty strategic initiatives, the impact was real, and the pay was great. But deep down, I knew that staying meant sitting on the sidelines of the fourth wave because my strategic role meant that I wasn’t doing the actual work. (Realistically, at this stage in my career, I lead teams that do the work, so what I was missing was the granular accountability and daily grind.)

I am hankering to build. However, it’s hard to leave a great job, great pay, and great impact. But it has to be done. So, I did something that every atom of my risk-averse-self rebelled against. I forced myself to resign from a job I loved. It’s exhilarating and it’s scary.

So, what next?

Building the future

The future of enterprise technology, I believe, is Vertical AI Agents. Companies and entire industries will be reimagined with the help of this powerful new technology. I want to be actively building this future either directly or indirectly (through the people I lead/mentor/teach/etc.).

What are some ways to play this?

1. I could join one of the established unicorns and build a team of Vertical Agentic AI experts who build vertical agents across multiple industries.
2. I could partner with a domain expert to co-found a startup. We could build AI agents that greatly expand the market in an industry that is critical to the modern world.
3. Something other option? Drop a comment below or DM me.

What should I do? What will I do? Stay tuned!

I don’t have all the answers yet, but I do know this — I want to be build alongside people who are just as obsessed. If you’re exploring this space, building something, or want to collaborate — let’s talk. Drop a comment or DM me.