AI allows for scale in a messy environment

My friend Sherif Elsayed-Ali wrote:

What should a city built in 2050 look like? They should live in harmony with the natural environment – meshed into it rather than replacing it.

This is close to one of my core belief. That the real value of AI is to enable automation that can scale while respecting the organic and messy nature of our environment.

The industrial era gave us the ability to scale. That is, the ability to build systems (Ie. collections of machines and humans) that can increase their output by a factor X while increasing the number of humans involved by a smaller factor Y.

Humanity gained an amazing power, but it came at a cost: the machines were dumb and needed a formatted environment to do their jobs efficiently. Building a factory is a two step process: first format the environment by building a huge metal box, then pack it with dumb machines that can work their scaling magic away from the organic messiness of the world.

Agriculture is the same: clear up a huge plot of land and structure it cleanly by planting stuff in neat rows that can be tended to by our dumb machines.

Another example. Prior to cars our cities were a lot more organic, with curved streets wide and narrow meandering around. Modern ones have been formatted to maximize our ability to automate them with dumb machines — cars.

AI changes all that.

Suddenly we no longer need the formatting phase. We can achieve scaling by training an intelligent machine to operate within our natural, messy, and organic environment.

We can build factories and cities that keep their efficiencies while meshing with the environment. For agriculture it’s even better: we can embrace the efficiency of nature (the way a messy collection of crops can complement each other to increase yield) instead of fighting against it.

It’s hard to think in this way. The idea that efficiency requires a formated environment is deeply ingrained in our mentalities. Therefore, today we’re thinking of using AI to automate the messy parts of our processes — the formatting phase, or the steps where our formatted environment has to interact with the non-formatted world around it.

Soon, however, we’ll take a step back, take a holistic look at our systems, and realize we can skip more and more of the formatting phase without losing the scaling efficiency.

We’ll realize that we can often do that today, with our current level of AI.


New Year Wishes

The image below was taken by Spaceship Voyager 1 in 1990. It was 6.4 billion kilometers from us. It’s called the pale blue dot in honor of that little blue pixel you see in the center of a ray of light. It inspired astrophysicist Carl Sagan, who wrote the following:

Our posturings, our imagined self-importance, the delusion that we have some privileged position in the Universe, are challenged by this point of pale light. Our planet is a lonely speck in the great enveloping cosmic dark. In our obscurity, in all this vastness, there is no hint that help will come from elsewhere to save us from ourselves.

Let me use cruder words.

The pale blue dot doesn’t give a shit.

We’re facing an array of global challenges and we have to fix this ourselves, we have to fix this together, and we have to fix this now. We wont fix anything if we wait for the singularity to emerge and save us all. We wont fix anything if we sit back and hope for someone else to fix it for us.

We’ll fix this if we roll up our sleeves and find a way to work together. To understand together what is broken. To go back to the basics, to the tools we have for probing reality and understanding what’s broken. To science.

We need to use science to write stories of hope and wonder. Beautiful stories that get people to regroup, no matter where they come from, no matter what they look like, no matter what they believe in.
And we can do this. We have the tools to do this. Not in a hundred years, but today.

We’ve got the tools, today, to embrace the complexity in which the world thrives.

Here’s to a New Year where we do this, together, a bit more.

(Adapted from a keynote I gave at Element AI a few years ago.)


Waverly Monthly Update — December 2022

Launch and iterate. If you’ve never heard it, it’s the motto of most modern startups. (Moving fast and breaking things is a bit less popular these day.)

So that’s what we’re doing… And you can help us, big time!

If you care about the mission of Waverly, if you love the product, if you dream of a future where we are in charge of our algorithms, then please help us launch in the US. It’s easy to do:

  1. Create an account on Product Hunt today*.
    It’s one click with a Google account.
  2. Go to our Product Teaser Page and click Notify Me
  3. On Dec. 13, go to our launch page and show your support.
    You’ll be notified by email.

The best way to support Waverly, on Dec. 13, is by writing a comment explaining why you care about us, what you love about the app, how you see it evolving…

Nurturing thoughtful conversations about a positive future for algorithms is what we’re all about and Product Hunt is a perfect place to advance that mission.

📈 Usage-wise the stats are also showing our Canadian launch was a big success:

  • 9.7% compounded month-over-month growth in weekly active users (1y rolling average)
  • 16.1% compounded month-over-month growth in number of cards seen daily (1y rolling average)

Next steps, you ask? We’re excited to be starting a seed funding round in January. We’ve had many great conversations with important mission-aligned investors in the past month and I can’t wait to launch the process. This is always a stressful time for startup entrepreneurs, so wish us luck!

There’s so many of you on this mailing list, whenever I hit send on one of these emails it warms my heart. Thanks for being there, for the good words you regularly send my way, and for your continued support.


The perils of a huge pre-PMF Series A: My story at Element AI

This story was originally published on Product Hunt.

In 2016, I co-founded Element AI. In 2017 we raised $100M. It went weird.

Peak Zeitgeist

We hit peak zeitgeist. AI was super hot. Many investors believed that early AI would be a talent game. Montreal — where Element AI was headquartered — was seen as one of the largest untapped beds of AI talent.

In the months between December 2016 (our seed) and June 2017 (our Series A), we had executed super well on the talent game. We had signed a lot of big logos who wanted to undergo an AI transformation. We had built a world-class network of academics and managed to actively engage them on consulting engagement with these big customers. More importantly, we did it with enough collegiality that researchers enjoyed the time they spent discussing with us — and amongst themselves — around our customer’s problems.

Talent Tsunami

We managed to translate researcher engagement into hiring momentum. Very talented AI researchers and AI engineers heard of the exciting discussions that were going on between top researchers in our meeting rooms. Some of them were stuck in data science jobs, some were at Google or Facebook but dreamt of having more impact.

We brought them in.

At that point, any C-level of any large corporation who walked into Element AI went into jaw-drop mode. They saw equations doodled on the windows in the kitchen, they heard researchers discuss the latest deep learning result. They saw the kind of energy they dreamt of having in-house but that they just couldn’t get.

We closed them.

We became the fastest growing startup in Canada.

Customers Everywhere

We had so many marquee customers. I gave talks all around the world. I found myself on a stage next to Andrew Ng, only him and I, addressing a room of hundreds of Hyundai executives. That’s where I learned that jokes don’t survive live translation.

We had an exceptional team of consultants. Ex-McKinsey. Ex-Deloitte. They looked top-notch in meetings. They could wow our customers by bringing in the brightest AI minds at a moment’s notice. We delivered.

We opened offices in Toronto, in London, in Singapore, in Seoul.

Too Much of a Good Thing

Then there were the products. We had a lot of great initiatives. We built infra to help our researchers develop and train their models more efficiently. We built a solid information platform using top-notch NLP, a top-of-the-line visual text processing tool, a product using AI to help manage user permissions in large organizations… Lots of potentially very useful products.

But the problem was that none of these products had the traction our consulting group had. We tried to align our consulting outreach with the products we were building, but there was always this new customer that was too big to pass and that we just had to accept.

We had incredibly talented and driven people. People who wanted to bring value to the organization. Only, that value came mostly from consulting. This led to internal tensions. On one side you had those who believed success would come from products and asked for more freedom to build them. On the other, you had those who wanted to bring in revenue and asked for engineers and researchers to work on consulting engagements.

The Exit

That tension became really hard to manage. After 4 years, it’s the kind of thing you could feel in the air, just walking through our gorgeous headquarters in the center of Montreal’s Quartier de l’IA.

On March 13th 2020, the day the world shut-down, I caught the last flight out of London Heathrow. This was my last business trip with Element AI.

I had decided to follow my dream and found Waverly. A product-centric startup. Yes, yes, AI. Yes, deep tech. Researchers, all of that. But first: a product. First, value that can scale. Real traction. Something small that could grow, grow, and grow.

A year later Element AI was sold to Service Now.

My Takeaway

I loved my journey at Element AI. I loved my coworkers. This was the best team I ever worked with. They had the brains of my Google coworkers and the grit of entrepreneurs.

As founders, I think we took a reasonable bet. I still think that talent was key. I believe we executed it well. It was bold, and being bold was not our mistake.

Our mistake was not seeing the tensions that could come from having too much money. Money gave us the ability to bring in the best customers, but it did not pressure us into building products. Building products and despairing because they’re not finding traction. Building products and testing them, pivoting them, validating them…

That’s why I’m posting this on ProductHunt. Because you understand what it means to relentlessly push and pivot a product.

That’s what I want to do now, with Waverly, and I know you can help me.


The original publication of this article has generated a lot of discussions, I encourage you to read the very thoughtful response of Jerome Pasquero, former PM at Element AI.


Waverly Monthly Update — October 2022

Today is a big day for Waverly. Our biggest day ever, in fact…

We’re launching!

That’s right. Waverly is now officially available on the Apple AppStore and all your friends can install it. Just tell them to search for Waverly on the store, or you can send them here.

Two great articles if you want to learn more about Waverly, both the product and our philosophy. In BetaKit by Meagan Simpson and en français in Le Devoir, by Alain McKenna. Yoshua Bengio also wrote about us.

It’s a big day, also, because we’re announcing our partnership with Phar, a Montreal-based company specializing in Market Intelligence. Phar will help us deliver Waverly for Business — which you definitely want for your organization as it gives you:

  • private waves for your teams ;
  • expert support for wave creation ;
  • continuous prompt engineering by supercurators ;
  • weekly insights digest ; and more…

This really takes Waverly to the next level! Interested? You can get started here!


Oh… And and we have a cool new feature too! (You wouldn’t expect us to slack off just because we had a launch to prepare, right? 😉)
Now every public wave comes with a URL you can share with anyone so they can keep up with the great articles your wave finds on any web browser. That’s right, you no longer need an iPhone or an account to browse through a wave!

🧠 Understanding Consciousness, a wave on the web, by Yoshua Bengio.

Here are some examples of waves-on-the-web :

Want to find the URL of your favorite wave? It’s still a somewhat hidden feature but drop me an email and I’ll give you the link. 😉

Phew! That was an intense month. I can’t tell you how much your support has been instrumental in giving us the energy to get where we are. Thanks for being awesome, and let’s make Waverly awesome together!



Know Me, Don’t Profile Me

Picture the scene. You walk into your small neighborhood coffee shop in the morning. The barista smiles at you from above his espresso machine and mouths “Flat white?” You answer with a smile and moments later you’re sipping your favorite drink.

It makes you feel at home. It makes you feel you’re amongst friends. That’s the kind of experiences we love.

Now picture another scene. Your car started making weird noises so you drive it to the garage, a place you’ve never been to. The mechanic asks you a series of technical questions. You stumble through the answers in a way that makes it pretty clear you don’t know much about cars. A few hours later the mechanic calls you back with a long list of things that need to be fixed and asks you to approve an expensive bill.

Not such a cool experience, now, is it? Makes you wonder if the mechanic somehow figured out that cars were not your forte and is trying to slip you a couple of unnecessary fixes.

These two experiences illustrate the difference between someone who knows you, versus someone who’s trying to profile you.

The concepts of knowing you or profiling you are similar in many ways but they differ in one key aspect. Someone trying to know you better engages in candid conversations. They happily let you know what they’ve learned about you. Their behavior makes it obvious that they’re not trying to extract information without you knowing. They allow you to set boundaries and keep some things for yourself.

Someone who tries to profile you does all the opposite. They watch you without you knowing. They infer things about you from your behavior but they won’t let you know what they’ve learned. Profiling happens in the shadows. It doesn’t let you choose what you’re willing to share or not.

There’s a word that captures that key difference between knowing someone and profiling someone: transparency.

The reason transparency is so important is that it allows the emergence of a trusted relationship. The candid conversations you have with your barista install a level of confidence. They allow you to trust that they want to know you in order to serve you better rather than to serve their own interests. Sure, it’s all about conducting a business transaction, but it’s one where you can be more confident that the goal is not to take advantage of you.

Now let’s transpose the scene to the online world. When you shop on the web, the site you visit tries to present you with a personalized item selection. Do they achieve that personalization by knowing you or by profiling you?

With traditional recommender systems there’s very little transparency as to what the online vendor knows or doesn’t know about you. Their knowledge is built by observing the log of your interactions, not on a history of candid conversations. This makes it hard for trust to emerge.

The personalized online shopping experience looks like a barista who knows your favorite coffee, but it feels like a mechanic trying to take advantage of you.

The feeling of being profiled also exists on our social networks. The order in which the reels are presented on TikTok clearly reflects some knowledge of our preferences, but we’re never offered an opportunity to learn what the system knows about us. We don’t have ways to set boundaries on that knowledge. We don’t have any mechanism through which we could build a trustworthy relationship with our social platforms.

These platforms are not trying to trick us with an inflated garage bill, but they are trying to steal our attention. They are trying to get us to keep watching their content… And when we regret spending too much time on their feeds they give us a poor excuse: “You can shut down the app at any time.”

It doesn’t have to be this way. At Waverly, we’ve built a new type of recommender system. One that is all about transparency. A platform that aims to know you. Slowly, though candid conversations, by letting you set boundaries whenever you want, by forgetting anything you want it to forget. A platform that work for you, with your best interests at heart.


A Coding CEO

Sometimes people ask me: “As a CEO, don’t you have more important things to do than coding?”

And yes, there are a million things to do to get a tech startup like Waverly off the ground.

But if I’ve learned one thing from my previous experiences it’s that, until you reach product-market-fit, amongst the million things you need to do the most important one is…

…the product!

Our product is tech-centric. It lives and dies by its algorithm and by the experience we create for our users. What’s the best thing I can do in that context?

I could go to cocktails, talk about the product, do marketing, close partnerships. In fact, I’m doing it. It’s important, but it’s not as important as building a great product.

I could direct people, but our small team accomplishes wonder with little guidance. I correct course all the time, but it takes very little effort.

I could hire more people and run more experiments. However, our experimentation engine is bound by the speed at which our users can try new ideas. A bigger team would mean a higher burn with limited benefits — it may even distract us and slow us down.

I decided one of the best thing I could do was to code. Yes, I’ll revise this decision as we grow. I love all aspects of the CEO job. But for now I’m a coding CEO and I’m quite proud of my GitHub chart:


Waverly Monthly Update — September 2022

Hi !

I took a short writing break this summer, but I’m happy to be back!

The team was very busy in the last few months. We’ve been polishing Waverly to make it ready for our upcoming Apple AppStore release. We’ve settled on a date for the release and I’ll give you all the details as we get closer.

Above is a screenshot of one new feature from the latest release: you can now see your friends’ faces on cards. Curators — those who mark an article as fit/unfit for a wave — now appear on each card, with the first curator getting the coveted first checkmark.

I know a lot of people following this newsletter are Android users, or dream of connecting to Waverly from their computer. We’re not there yet, as our small team decided to optimize the experience on a single platform first. Still, don’t hesitate to nag me. It’s great to hear your enthusiasm, even if it makes me feel we can’t deliver everything we wish we could.

So, what about running a startup? As previously mentioned, we’re now firmly committed to building the best content delivery platform for professionals. Waverly is already being used by people and organizations who want to track trends, understand their market, see what their competitors are up to, etc. We feel it’s a place where our AI really shines and it allows us to execute towards our north star: an assistant you can trust and control using everyday language.

Our updated website reflects that positioning. Please take a look at it and send us your feedback.

As always, it’s a great pleasure to feel your support. Thanks again for being a part of the Waverly journey, and never hesitate to send me your feedback — I read every email. You’d like to meet in person? I’ll be at C2 and MTL Connect, let’s hang out!




Will Technique Die?

“You can’t have art without resistance in the materials”

The AI hype peaked three years ago, but in a shrewd move, AI was simply waiting around the corner to ambush us when we let our guard down. At least that’s what I can gather from two articles that popped up on my Waverly this morning.

The first, from O’Reilly, explores how AI is being used by programmers to speed-up their craft using a tool called Copilot. It asks whether AI will enable some form of “higher level compilation” that will make it unnecessary for programmers to learn how to code.

The second, from Wired, asks artists what they think of people using their names in DALL·E and Midjourney prompts:

“When they’re feeding work from living, working artists who are, you know, struggling as it is, that’s just mean-spirited,”

Yeah. They’re not happy…

From my experience coding recommender systems and doing research in AI, my take is slightly different. We’re inventing new tools that open up new territory. We’ll learn how to craft in this new space. Some techniques will become more useful, some will become less useful. It will suck for some people and be great for others.

The problem of the new technique — the “type a prompt and the AI will do the rest” technique — is that it’s easy to believe there’s no technique at all. That any kid could master that craft after spending 5 minutes on Midjourney or by typing a few words into Copilot.

Except that’s not true. Midjourney artists and Copilot coders soon learn that they need to understand how the AI “perceives” their prompt. They need to co-evolve a language with their new AI tool. Something like that already happened in the past, with Google. We all had to learn how to write good queries. We co-evolved a language with the search engine.

Learning this language — especially when it comes to writing code that will be maintained by others — will necessarily be rooted in the human’s deep understanding of what they are trying to achieve. If you don’t know how to organize software, which pieces to abstract out, which pieces to disentangle, then Copilot wont help you.

Will we ever get a computer that can program itself entirely? Something you could direct the same way you can direct a top human programmer today? Maybe, but that’s still very much in our AI-hyped future.

What we’ll get a lot of are these new tools, like Copilot and Midjourney, that completely change the techniques one can use to achieve a piece of code… or a piece of art!


What is Indexing and Why is it Useful for Recommender Systems?

This post contains an answer I gave to Kathryn Kyte who wrote this BBC piece about Waverly.

What does indexing an article mean in the context of a recommender system?

The best analogy I can think of is that of a business directory. A big book where you have, in alphabetical order, different services: Bricklayer, Mechanic, Plumber… Under each category, you have a list of businesses that offer this service. A given business can fall under two services if it does multiple things.

An index of articles is very similar. Instead of having an alphabetical list of services, you have an alphabetical list of characteristics that can apply to an article. These characteristics can be topics, people mentioned in an article, writing styles (journalistic, casual, fiction), etc.

If you look under a given entry of the index, say the topic Beneficial AI, you’ll find all the articles that seem to be talking about beneficial AI. At Waverly, our AI reads tens of thousands of articles a day and extract hundreds of characteristics about these articles. These characteristics, and the articles the point to, constitute the Waverly index.

It’s this index, combined with our ability to extract which characteristics seem important to you given the Waves you’ve written, that allows us to offer the first natural-language-driven recommender system.