Not too long ago, a wonderous tool called ChatGPT quickly became a hit success overnight. Since then, the media and world have been taken by storm by how AI is going to change everything. Specifically, I think most of them are referring to language models since that seems to be the most popular form of AI discussed in the news (although image generation is a close second).

Having worked in AI for a bit and seeing how these technologies are changing over the last few years, I wanted to give my take on how I think LLMs will actually change over the next few years.

Going forward, I’ll use the word “AI” as synonymous with “foundational models” in this article, since that’s the main type of AI that everyone seems to be talking about.

I have three goals with this post – let me know if you think I’ve actually achieved them:

1. No clickbaity exaggerations. I don’t want to hack clicks onto this article because I promise a rosy future where robots do your laundry and dishes. If you read this article, it should be because you found it interesting enough to send to a friend.

2. Be realistic about what’s imminent. Every future scenario I paint should be something you and hundreds of millions of others experience within the next three years.

3. Easy enough to explain to your grandma. I hate to pick on her, but if it’s not easy enough for your grandma to understand, I’ve made it too complicated.

Note: going by these goals, the PM in me is thinking that if I had a north-star metric, it would be number of readers who share this article with their grandmas. If you share it with your grandma, please let me know!

Share

First things first

So, laid out at the beginning, here are my three realistic predictions of how we’ll use AI in the next few years:

1. AI will excel at analyzing and summarizing large stores of text (codebases, books, etc.), but you’ll still be limited to generating only small pieces of writing with AI.

2. Interacting with AI will be just like like interacting with another person rather than chatting with a chatbot.

3. AI models will know everything about you and naturally interact with the tools you use daily.

For each of these, I’m going to lay out the facts, show some data that supports the facts, and then paint a picture of what that’ll actually look like with examples.

AI for analyzing, but not so much for writing.

The prediction

AI will excel at analyzing and summarizing large stores of text (codebases, books, etc.), but you’ll still be limited to generating only small pieces of writing with AI.

The data

Two things matter for (generative) AI models: the context window and the max output token limit. Said simply: context window = input (what goes into the model) + output tokens (what the model responds with).

The simple fact is that, as they get better, AI models can handle bigger context windows. However, the output tends to become random and useless after a certain length. In other words, the practical input windows of new models are growing much faster than the output windows.1

The latest and greatest models from Google can handle up to 2m tokens (3k pages of text or 250k LOC). Other models in the works in the research stages can handle up to 10m tokens (15k pages of text or 1.25m LOC). But those same models are limited to only ~8000 tokens of output – around 125 pages. That’s a 244x smaller output window than input – and that’s the absolute cutting edge with the best models.

Even OpenAI only just launched an alpha program for long-context GPT-4o, with just a 64k output token window (1k pages of text or around 8k LOC). The trend of output windows being orders of magnitude smaller than input windows will remain true as models get more advanced.

The implications

Because of the discrepancy between input vs output windows, AI will excel at summary and analysis style tasks rather than generation tasks.

For example, here are some realistic scenarios we can expect within the next three years within the software engineering space:

• LLMs will be able to read entire codebases and suggest improvements, optimizations, security vulnerabilities, etc.

• LLMs will be able to write a single file in a codebase while considering the architecture, style, and nuances of the rest of the codebase.

• Large codebases will still need engineers to architect, write, and maintain them.

This can be mapped to other industries:

• Researchers will be able to analyze hundreds or thousands of published papers on a topic and relate them to a study they’re conducting, but they should still expect to write the research papers themselves.

• Lawyers can analyze hundreds of briefs, precedents, and legal proceedings, but still will need to write their own contracts.

• Knowledge workers can search across everything within their company to get a useful answer but still need to write their documents (although hopefully not emails!)

Human or chatbot? We may never know.

Interacting with AI will be just like like interacting with another person rather than chatting with a chatbot.

The data

The latest AI models are multimodal, meaning they can take in text, audio, image, and video input. Furthermore, model output latency for audio is going down from a few seconds to a few hundred milliseconds – on par with normal human conversation.

The latest multimodal LLM, GPT-4o, can process text, audio, video, and images in real-time. The “o” in the name actually stands for “omni” as in “omniIn addition, it can also return audio output back with an average of 320ms latency, down from about 3 seconds in prior models, since it’s built from the ground up to support audio input/output. The demos are truly incredible.

Not to be forgotten, Google’s Gemini Live allows you to interact in real-time using video, audio, and text and get low-latency responses.

These models are also able to make sense of images and videos, giving them the ability to answer requests like “Help me understand this code I’m looking at” or “What does this paragraph mean?” – they’ll automatically see what you see and interpret it as another human would.

The implications

This model of interacting is identical to how humans interact. Because humans assume other humans have the same ability to see, hear, and reason, we use lots of shortcuts in our speech that primitive models can’t handle.

Consider the simple example: “What does this line in the code mean?” There are several problems with this for AI models:

1. They don’t know what “this” line is – they can’t see what you’re seeing!

2. Understanding how to interpret “mean” requires context on what you know – are you a lawyer or a software engineer? If I’m explaining a line of code to an engineer, I’ll use wildly different terminology and give less context than if I’m explaining to a lawyer.

Here are some realistic scenarios I expect will happen in the next three years:

• In your pocket, you’ll have a translator who knows every language, a Mike Ross lawyer who knows everything about law, the best engineer to pair program with, and so much more. And you’ll be able to talk with them on the phone as if they were a real human helping you out.

• Some people will wear glasses that analyze everything they’re looking at and provide useful information about whatever they see. They’ll be able to ask questions like “What is this plant I’m looking at?” or “I don’t remember this person’s name – what is it?” and get correct answers.

Know thyself, wield thine tools.

The prediction

AI models will know everything about you and naturally interact with the tools you use daily.

The data

Many popular techniques exist to enrich LLMs with information they weren’t trained on, resulting in higher quality and more specific output. The most popular method is called Retrieval Augmented Generation (RAG). Over the last few years, cloud companies have started providing RAG-as-a-service for other developers (Google Cloud, AWS, Azure).

And many models today allow for function calling, which lets them interact with your day-to-day tools like calendar and email without additional work. OpenAI and Google Gemini already allow for function calling through their APIs.

But the real magic will be unlocked when the companies that make the technology we use daily, like our phones, start combining AI with all the information they already have about us.

Apple and Google are diving deep into incorporating AI into their phones and combining it with the treasure trove of data they have about the user to allow AI models to perform seemingly magical things for us.

The implications

Apple Intelligence, built into the upcoming iPhone 16 models, will allow you to say things like, “What time is my mom’s flight landing?” From this seemingly simple query, it’ll do a couple of things:

1. Look up which contact is your mom

2. Look at the exchanges of texts between you and your mom

3. See the flight number that she sent to you

4. Pull up the latest flight information from online

All of this happens in seconds, and out comes your response, just as if you had asked another person who knows everything about your life.

And Google’s latest Gemini assistant built into Android and their Pixel line of phones will perform similar tasks. For example, you’ll be able to ask, “What’s my Airbnb address again?” and it’ll do the following:

1. See that you just arrived in Greece

2. Look through your email

3. Find an Airbnb email confirmation for Greece

4. Extract the address from the email

And on the other hand, these models will also be able to interact with your tools using function calling. Here are some example queries that will work seamlessly:

1. Put my mom’s flight tomorrow into my calendar

2. Forward the Airbnb confirmation email to my wife

3. Reminder me to get all the groceries that Joanna texted me about when I get to the grocery store

The net effect is that models can get all relevant information about you (via RAG) and also interact with your day-to-day tools (via function calling).

It’ll feel like having a personalized model at your fingertips that’s aware of all relevant information concerning you and can interact with your calendar, email, etc.

In summary

I hope reading this has built some (justified!) excitement about all the amazing things that will happen to AI models over the next few years.

If you enjoyed reading this article and want more non-clickbaity content in your emails, throw me a subscribe.

And if you found this article easy to understand and want to share it with your grandmother as I mentioned at the beginning, go ahead.

After having carefully considered and weighed the witness testimony and evidence, the court reaches the following conclusion: Google is a monopolist, and it has acted as one to maintain its monopoly. It has violated Section 2 of the Sherman Act.

This seems pretty damning. I mean, we created the Sherman Act back in 1890 to deal with anticompetitive practices by large corporations and give the government a way to combat companies that were being anticompetitive.

So what does this mean? Microsoft lost a similar lawsuit over 20 years ago in which the government initially pushed to have the company split up (this was later overturned on appeals). Does this mean Google may face a similar fate, or will it just be a slap on the wrist and some fines?

Let’s tackle this with a few simple but very reasonable questions.

What is a monopoly?

Well, according to Wikipedia:

Monopoly is a multiplayer economics-themed board game. In the game, players roll two dice to move around the game board, buying and trading–

Shoot. Sorry, it seems like ChatGPT is hallucinating when I told it to write this blog post for me.1 I guess I’ll have to do it the old-fashioned way from here on out – with my mind…

Surprisingly, a monopoly is not really defined in most US legislation nor in charters for any part of the government bodies that promote competition and consumer protection.

The generally agreed-upon definition is something like this: a monopoly is a market in which one player is the only supplier of a good or service. So, first of all, monopoly ≠ company. Google is not a monopoly. Google can have a monopoly on a specific market.

But then this begs the question – what market? Well, that’s actually a really relevant question. One of the first and most important things in these cases is to actually define what market the relevant company has a monopoly over. In this case, the court ruled that Google has a monopoly over the “general search services” and the “general search text ads” markets.

So, a monopoly is a market in which one player dominates, and Google was declared a monopolist in the “general search services” and “general search text ads” markets.

Is having a monopoly illegal?

Nope. Surprisingly, it’s not. This is counter to what many people seem to think. With the Sherman Act, two things were made illegal:

1. Getting a monopoly through anticompetitive actions (exclusive contracts, predatory pricing, etc.).

2. Agreements between businesses to limit trade or competition (price-fixing, dividing markets among competitors, etc.)

The act also gives the government the ability to investigate and prosecute when companies are being anticompetitive:

1. The FTC can investigate and fine companies

2. The DOJ can file lawsuits against companies

And we frequently see this happening, with some popular instances of the government keeping an eye on anti-competition:

• DOJ launches two antitrust probes into NVIDIA (Aug 2024)

• DOJ sues Apple over monopolizing smartphone market (Mar 2024)

• FTC launches probe to investigate generative AI data collection of Google, Amazon, Anthropic, OpenAI, Microsoft (Jan 2024)

• FTC sues Amazon for anticompetitive practices to maintain a monopoly (Sept 2023)

In short, having a monopoly is not illegal, but getting a monopoly through anticompetitive actions is illegal.

So what did Google do?

Google has a standing exclusive agreement with Apple to be the default search engine in Safari. Back in 2002, when this agreement started, no money exchanged hands.

But it quickly became an ad revenue sharing agreement, which in 2022 resulted in Google paying Apple over $20 billion. Even on an Apple scale, this is no chump change – it’s over 15% of Apple’s annual operating income coming directly from Google!

It’s this agreement that the DOJ alleged and the court agreed was anti-competitive for two reasons:

1. It helped Google get a monopoly since it foreclosed the rest of the market and gave them >90% market share of search queries on mobile devices.

2. It limited competition from others because it was exclusive and essentially priced out competitors.

But let’s consider some very reasonable rebuttals (some of which Google tried to make in their case)…

It’s not like Google stopped people from changing their search engine on Safari. If people really wanted to, couldn’t they just change it manually?

Yup. But defaults matter, and the fact that Google was willing to pay so much in the first place to be the default meant they realized that too. In addition, many users don’t even realize they can change the search engine, which requires going through multiple settings.

Outside of Safari, most people still choose to search on Google because it’s the best. Doesn’t that mean it’s in consumers’ best interest for Google to be the default on Safari?

Not really. The goal of the court is to stop anticompetitive actions. In this case, setting a default allows Google to maintain a monopoly and stifle competition. In addition, just because Google is the best today doesn’t mean it’ll always be the best, and stifling competition in the long term hurts consumers because it makes it harder for someone to invent an even better search engine.

Okay, but if Google already has a monopoly, why does it even need to pay Apple? Doesn’t the fact that they need to pay so much to Apple mean they don’t have a monopoly?

The court didn’t really get into this, but a relatively simple take is to reframe the payment from Google to Apple as a payoff for Apple not to compete. I mean, look at the smartphone market – the fact that Google and Apple compete so much for the global market is only because there’s no clear dominant player worldwide. And look at Google Maps vs. Apple Maps, which continues seeing improvements for consumers because they’re both competiting.

Apple could probably build a great search engine (and has thought about it before!) but chooses not to because of the guaranteed payment from Google every year. It’s basically a payoff.

So what’s next?

My goal for this post was to simplify what exactly this trial was about and what the court ruled without all the technical or legal jargon. The court basically uncovered that one of the biggest companies in the world was paying off the other biggest in multi-billion dollar payments to reduce competition. And the DOJ and court said, “Nuh uh. Not on my watch.”

But what happens next? Well, both sides (Google and the DOJ) have to work together to agree on the penalties for Google. But Google is also appealing this ruling and could keep appealing all the way to the Supreme Court, which means we may not know what happens for a few years.

My guesses?

1. Google is forced to stop its exclusivity default agreement with Apple.

2. Apple decides to use Siri to better answer search queries in Safari rather than going to a search engine at all.

3. If they do have to go to a search engine, Apple still uses Google (because, apparently, they think Bing sucks).

Having worked in the AI space for the past five years at Grammarly, I’m here to argue the opposite. I’m here to share why I think generative AI, an incredible technological advancement, is not disruptive but rather incremental. I’m claiming the way it will impact society will be through small percentage improvements on a long tail of industries rather than unlocking new ways of living and working.

What makes a technology disruptive in the first place?

At its core, disruptive technology makes people do something in the world radically differently than they do today. Obviously, that’s a retroactive definition, but there are many leading indicators we can use to gauge how disruptive a technology will be. A great leading indicator for what could be a disruptive technology is if it makes it 100x faster or easier to do something foundational people are doing today.

For example, cars were 100x faster than horses, and planes were 100x faster than cars. Both of these were disruptive technologies. However, what ultimately made them disruptive is unique. Cars allowed people to live farther away from city centers and quickly get around to anywhere connected by a road. Infrastructure investment in roads and the federal highway system quickly took advantage of this in the US, allowing cars to become the primary tool for daily transportation across the country. Today, the average American spends an hour a day in a car.

Said differently, cars allow us to live in one place and work somewhere else. It’s a similar story for airplanes and commercial air travel, which allowed us to get from one point to another in less than 24 hours. Both of those capabilities didn’t exist for humanity before those technologies. Some other popular examples of disruptive technologies were the printing press, which allowed us to print books 100x times faster than before, and the internet, which allowed us to communicate 100x times faster.

Another characteristic of disruptive technologies is that they create significant market value for their industries and have several other multi-billion dollar industries adjacent to or dependent on them. For example, the automobile industry is massive, but equally massive are the rental, residential real estate, and restaurant industries in the US, which depend on consumers needing cars to get around and live away from work. Cars allowed US residents to spread out and purchase their own homes in the suburbs, while Michelin, a well-known tire company, even started its Michelin Guide to encourage people to use their cars more to travel to famous restaurants on road trips.

Similarly, for planes, we have the entire tourism industry, including rental cars and hospitality, built on top of planes. And it’s very fresh on everyone’s minds after the COVID pandemic what happens when air travel suddenly stops – the dependent industries suffer too.

So, we have three (non-exhaustive) characteristics of disruptive technologies:

1. They make something people want to do 100x faster/easier.

2. They generate significant market value.

3. They spawn other dependent industries also of significant market value.

What do we call non-disruptive technologies?

I call it incremental. The vast majority of technologies fall into this bucket since human progress by nature is incremental, with once-in-a-century innovations that are truly disruptive.

Think of the yearly iPhone updates that Apple makes, which, rightfully so, seem like small changes. Surely, today’s iPhone is leaps and bounds ahead of the first iPhone released, but they still resemble each other. And we got to today’s through years and years of iteration and small improvements, feedback, technological advancements, experimentation, and so on.

But the first iPhone introduced back in 2007 seemed, at the time, revolutionary and disruptive in that it brought on the true worldwide adoption of touch-based smartphones that could browse the internet similar to a computer.

Most technologies make percentage point improvements to a process, capability, or functionality and allow us to do something slightly better and faster than before. Of course, this can compound over time, and we may find ourselves slowly doing things differently. But that’s quite different from disruptive technology, which creates a new status quo for society seemingly overnight.

What bucket does generative AI fall into?

First, some quick background. When companies have a problem they think can be solved with machine learning (ML), they need to follow a pretty extensive process to develop an ML model specifically for the task, domain, or application. For example, if you want to do sentiment analysis on some type of text, you’d need to collect a lot of sentences, label their sentiment, and then train a specific ML model on that data. If you also wanted to modify the text to make it more positive, you’d usually have to train a new model with new training data that models the task you’re trying to accomplish.

But collecting the data, training a model, and deploying it to be used within your product isn’t trivial and usually requires very specific and expensive expertise. This is why, for most businesses, ML is either outsourced or avoided entirely. At the end of the day, lots of businesses can’t justify the upfront cost of building an ML team just to solve what, for them, is usually an optimization problem.

However, things we frequently call “generative AI” are tools like large language models (LLM) used in ChatGPT or large vision models used in DALL-E or Stable Diffusion. Those fall into a different category called foundational models and work quite differently.

Foundational models are inordinately complex ML models trained on a mind-boggling amount of data and function like general-purpose models that can do almost everything well. For example, the LLM that powers ChatGPT, GPT-4, can do all kinds of complex tasks with text, such as sentiment analysis, grammatical improvements, text generation, text modification, code generation, etc. Similarly, tools like DALL-E can do image generation, image editing, conversion, and more.

So, how does this change how someone can use ML in their business? Whereas previously, to use ML, you needed to train a custom model, which requires significant time and expertise, now you can just use an off-the-shelf foundational model and get pretty good results just by telling the model what you want it to do in human speak (aka “prompt engineering”).

This is an essential point. These foundational models allow businesses that previously couldn’t solve complex problems with machine learning to use machine learning overnight. And what types of problems are we talking about? Most likely, optimization and efficiency problems, such as:

• Optimizing shipping routes for companies operating fleets with international and domestic shipping. (bearing.ai)

• Improving crop yield and reducing waste in agriculture. (taranis.com)

• Helping large automotive industries with EV battery generation. (aionics.io)

• Allowing companies to model and engineer proteins and improve their synthesis. (cradle.bio)

These are all examples of optimization problems – shipping routes, crop yield, battery manufacturing, and protein engineering. The technology of foundational ML models makes it easier for businesses to solve complex optimization problems without in-house ML expertise.

This trend will continue across the long-tail of businesses and industries as well, as famous AI researcher and investor Andrew Ng talks about, over the next few years as foundational models become better and easier to use via lightweight UIs such as ChatGPT and other startups that will inevitably crop up.

So, in conclusion, the impact of “generative AI,” or really, foundational machine learning models, is not disruptive. There’s nothing vastly new that these models will necessarily allow us to do that will change how we live our lives. But that’s not to say I’m discounting the significant world impact that AI will have. Giving all businesses the ability to make a 10% improvement to their complex optimization problems is not, by any means, insignificant. It is definitely an incremental and not disruptive change, but a 10% incremental improvement across almost every industry will have an incredible impact on the productivity of our workers, countries, and, eventually, people.

Too often, the incentives are aligned that the people talking about the technology are the most incentivized to hype it up, and this creates a situation where its usefulness and disruptive power are blown way out of proportion and create unnecessary hype, which I believe does a disservice to the actual productive impact this technology actually has on businesses and people. 

I hope this article adds more nuance to your thinking the next time you read about how generative AI will “disrupt” the world and change everything we know and are accustomed to today.

I wanted to use this post to share a few thoughts on how I’ve noticed and experienced burnout. I’m no expert in anything remotely related to burnout, so please treat this as self-reflection and a collection of thoughts rather than expert testimony. I hope it’s useful to read these thoughts and reflect on them to see if they give a little more structure to evaluate your own state of burnout and what you can do to mitigate and reverse it.

I’m going to try to take an ethos, pathos, and logos approach to defining and understanding what burnout is. In other words, I’ll try to define burnout from three different angles:

1. Ethos: How thoughtful people define burnout

2. Pathos: What burnout feels like (to me, but potentially to you as well)

3. Logos: What science says burnout is and how to counter it

How thoughtful people define burnout

First, I wanted to describe how well-known public figures describe burnout. Those people are Sam Altman, Vivek Murthy, and Barack Obama, individuals who work or have worked in prominent, high-stress environments and have vaulted achievements in their respective fields.

Altman and Murthy both talk about an important point related to burnout: burnout isn’t caused by working long hours but rather a sustained misalignment between the day-to-day work and the longer-term goal of the individual. In an interview with his brother, Altman mentions that “burnout is caused by failing and lack of momentum, not working too hard.” This analogy makes sense for those attempting to start a company, a good chunk of Altman’s audience, for whom the goal is close to finding momentum with their business. 

Murthy similarly mentions in a perspective piece for the New England Journal of Medicine that “Burnout is not only about long hours. It’s about the fundamental disconnect between health workers and the mission to serve that motivates them.” Health workers’ primary motivations come from serving others, but a constant daily struggle of paperwork and red tape, as well as failing, at times, to save their patients or make an impact in the desired way, can lead to a misalignment of their longer-term goals with their daily work.

In fact, long hours may not even be a required ingredient that leads to burnout. There are many examples of people working very reasonable hours and still getting burnt out. Tech hours generally aren’t seen as excessively long for extended periods. You’ll frequently find people working 9 to 5 or even fewer, with frequent breaks throughout the day for errands, going to the gym, walking around, getting coffee, etc. But burnout is common in tech, even in such working environments.

What can be done about this? Well, shorter-term fixes can be as simple as taking a walk, which Obama did frequently when he was the president. To be fair, he probably still goes on walks now. Something as simple as walking and disconnecting from work, especially in the middle of the day, can help in the short term. However, in the long term, there needs to be a closer alignment with the work you’re doing every day and your long-term goals. And the more hours you put in without that alignment, the more likely it will lead to burnout.

What burnout feels like (to me)

The second dimension we can think about burnout is how it makes someone feel. My personal insight is that burnout isn’t an acute pain you wake up to one day, but a chronic pain that gradually increases over weeks and months. Every day it gets a bit more prominent, making it easy to slowly get used to over time.

Over the last few years, my own ability to recognize burnout was always a posteriori. I realized what I felt for the past few weeks was burnout, rather than realizing the trajectory I was on was tending towards burnout in the future. I’d spend a few months working without reflecting significantly how how I was doing emotionally and mentally. Eventually, I would stumble into a vacation or a long weekend, find some time to reflect and disconnect, and finally realize over the past few weeks that the buzzing demotivation in the back of my mind over the last few weeks was burnout.

This highlighted the importance of creating the time and space to honestly self-reflect on how I’m doing about work. Let me explain more what I mean by this. 

To properly recognize and diagnose my state of being burnt out, I needed to properly self-reflect on how I was feeling about work, life, and the balance between the two. Only I could decide I was burnt out. Nobody else could truly convince me for or against it otherwise. 

But how do I come to that realization? Well, every person’s strategy for reflection is different. What worked for me was talking with close friends and family, thinking, and writing. And when I say talking with close friends, I don’t mean a 30-minute walk for coffee in the middle of the day. These are heavy conversations and require the mental capacity and space to not be within a working context. This is why my burnout realizations only came after a long weekend or vacation, in which I finally stopped being in work-mode autopilot and truly stepped back. That’s what I mean by creating the time and space. I needed not to be thinking about work and not physically be close to or in the same area where my mind is in work mode. If you normally work from home, this can be as simple as getting out of town or going to the cafe or park for the day with a pen, paper, and no phone.

The last ingredient is honesty. I find it easy to open up and be vulnerable when writing, and it also happens to be the best way I think. So, I force myself to journal and write down my thoughts extensively, even if no one will ever read them but myself. I got into journaling every month a few years ago, and it singlehandedly helped me make sense of my thoughts and be true to how I was feeling.

In short, burnout doesn’t suddenly hit you one day; it creeps up slowly and sneakily. My recommendation for recognizing when you’re burnt out is to deliberately carve out regular time and space to honestly self-reflect on how you’re doing.

What science says burnout is

What can we learn from the actual experts? Burnout has been a topic of increased interest in the scientific community over the past few decades, and there has been a decent amount of research on it.

Let’s start with the WHO, which classifies burnout as an occupational phenomenon (not a medical condition):

“Burn-out is a syndrome conceptualized as resulting from chronic workplace stress that has not been successfully managed. It is characterized by three dimensions:

• feelings of energy depletion or exhaustion;

• increased mental distance from one’s job, or feelings of negativism or cynicism related to one’s job; and

• reduced professional efficacy.

Burn-out refers specifically to phenomena in the occupational context and should not be applied to describe experiences in other areas of life.”

In fact, burnout became recognized by German-American psychologist Herbert Freudenberger back in the 70s. He noticed several physicians at his free clinic in NYC as well as himself experiencing a “physical or mental collapsed caused by overwork or stress.” Sounds awfully familiar. He diagnosed himself and his colleagues with “burnout syndrome.” It took some time for scientists to agree that the key distinction between exhaustion and burnout, however, is a feeling of fulfillment and satisfaction. Painting a room or running a marathon can be exhausting, too, but there’s a sense of completion and fulfillment of one’s goal that motivates us to continue.

And while companies have started becoming aware of the problem, their well-intentioned solutions may exacerbate burnout. Josh Cohen, a professor and psychoanalyst at the University of London, mentions in a piece for The Economist that companies have started offering mindfulness and meditation sessions during the workday, which inadvertently becomes another thing that employees can succeed or fail at, all the while taking away from their working time and not properly allowing them to disconnect from work.

Christina Maslach, a psychology professor at UC Berkeley, classifies six factors that contribute to burnout:

1. Excessive hours. When you’re overstretched and overworking consistently.

2. Lack of autonomy. When your actions don’t give you control over the outcomes.

3. Lack of reward. When you don’t get recognized for your work.

4. Hostile workplace environment. When your workplace culture is competitive and intense rather than collaborative.

5. Politics and bureaucracy. When there are excessive policies that aren’t enforced fairly or reasonably.

6. Lack of meaning. When you don’t find value in the goal of your work.

A great topic to close out on is the recognition that burnout and depression exhibit very overlapping symptoms. In a longitudinal research study by Renzo Bianchi, burnout may be best understood as a combination of “work-related depressive symptoms.” In addition, consistently experiencing burnout-causing factors without an outlet to resolve the problem or overcome or neutralize the stressors may begin experiencing deeper depressive symptoms.

The takeaway

In summary, burnout is a well-understood work-related disorder, but recently, it has gained more attention due to its exacerbation throughout the pandemic, specifically on healthcare and front-line workers. But there are common ways to mitigate burnout before it happens, identify when you are burnt out, and prevent and reduce it to return to a healthier work-life relationship.

I hope this post was a useful overview of my journey experiencing and dealing with burnout and a bird’s-eye view of some of the research being conducted in this space to help us better understand how the mind and body feel and react to burnout.

I wanted to use this post to codify the most important lessons I’ve learned from my last half decade at Grammarly. Some of these lessons may be useful to you, some might be things you already knew, and some might not be relevant at all. That’s okay. My goal in writing them in this blog post is to ensure I have a quick reference guide for myself.

The five lessons I’ve learned are:

1. Invest for success.

2. Boring = successful.

3. People >> work.

4. 10% vs. 10x.

5. Don’t forget about fun.

Invest for success.

I’ve seen many examples of people spreading themselves too thin by simultaneously working on too many things. I frequently made this problem early on in my career, working on my full-time job, this blog, my fitness, one or multiple side projects, and so on. In reality, I was really only making progress on my full-time job and making marginal to negligible progress on everything else. Either that or I was getting only a few hours of sleep every night and would get sick every other week.

And I wasn’t the exception. People tend to set too many goals rather than not enough goals. The issue isn’t a lack of ambition. It’s a lack of focus. Jeff Bezos famously said that leaders don’t need to make hundreds of decisions every day. Instead, they need to make several high-quality decisions every day. I strongly resonate with this idea and believe it applies to tasks and projects I’m working on as well. It’s a question of focus.

But over the last few years, I have found it harder to focus. I read a great book on this topic called Stolen Focus. It aims to explain the seemingly universal feeling we all seem to have that we can’t focus as well as we used to. The author tries to explain whether that feeling we all share is, in fact, the reality (spoiler alert: it is) and, if so, what might be causing this inability to focus. However, I don’t think it has great strategies for how to actually solve the problem of reduced focus, so if you’re looking for actual strategies and tactics, I’d recommend a Huberman podcast or some tried-and-true books (Atomic Habits, Flow).

Anyway, the strange thing about this? Ruthlessly prioritizing what to do and ultimately deciding to do fewer but more impactful things was my job at Grammarly. As a product manager, most of the job is figuring out how to get the most impact for the business with limited resources. Many of the decisions I’d make are to say no to things we could do in favor of something more impactful. It wasn’t just a question of the end result but of the return on investment.

So here’s the lesson I’ve learned: My time and energy are assets I can invest. I can choose to spread that investment across a lot of things (i.e., spread myself too thin), or I can choose to invest in a few very impactful things and be successful.

Boring = successful.

Think about the most impactful product or tool you use daily. Perhaps it’s your phone or car. Now, how well do you actually know how those underlying technologies work?

I don’t really know how computers work. And by extension, I don’t really know how phones work. How do silicon chips come together to make the internet that lets me browse memes on Instagram or FaceTime with my friends and family worldwide?

If you take the time to look at the technology powering phones, it’s absolutely insane. Layers upon layers of complexity are all built on top of each other and incorporate decades of research from the brightest minds we’ve ever seen.

As a user, I see none of this complexity. I just use the magical box in my hand that can call, text, browse, navigate, charge, and even fall without breaking. And trust me, I drop my phone often.

At Grammarly, I noticed an interesting phenomenon that I couldn’t quite explain at first. We spent so much attention and care to what seemed like quite simple and trivial things – should we show this underline on top of the other? Should we stop showing this pop-up after three times or four times? How many milliseconds after the user stops writing do we start giving them suggestions?

At first, these questions seemed trivial and potentially even overkill. But the reality is that getting to that level of optimization and detail takes decades of work. Every small decision made can either improve or degrade the product, and a concatenation of thousands or millions of improving decisions leads to an incredible and successful product.

In fact, to the user, the product should seem boring. If a user considers it boring and commonplace that we can analyze their text using AI in real-time and make suggestions right in the text field on any device or application, that’s a win. Because although I consider my phone boring, I use it every day for multiple hours.

My lesson learned here: strive to build something that abstracts away the complexity of the underlying technology and instead create something that seems so simple people find it boring.

People >> work.

The third lesson I learned was about the importance of the people you’re working with over the work you’re doing.

First, a caveat – the work you do does matter. Everyone’s preferences are different, and some people may care more or less about the actual role or the field they’re working in. My point here isn’t that you should do any job you’re offered as long as the people are cool. Rather, I’m trying to capture the relative importance.

So why are the people more important than the work itself? Because what you’re working on can radically change day-to-day or quarter-to-quarter, but the people you work with and the meaningful relationships you make will follow you through as friends, mentors, investors, and advocates. Even if you’re no longer working with them.

As part of a project at Grammarly, I interviewed dozens of People and HR executives at companies ranging from newly IPOed startups to Fortune 500 companies. I asked them two simple questions.

First, I asked, “What is the top reason employees at your company quit?” The #1 response? A bad manager. Then, I asked, “What is the top reason employees choose to stay?” The #1 response? Close friends at work. Both of those responses have nothing to do with the work itself but rather the people you’re working with.

The lesson I learned here: Think of my professional growth as a series of mentorships and chances to work with talented, driven, and experienced people across the company.

10% vs 10x.

Something that seems obvious in retrospect but was difficult to grasp at the time was the difference between a 10% change and a 10x change. Let me explain what I mean.

When building a product, you’re constantly looking for ways to improve it. In software, those improvements are usually realized through features.

Most of the changes will lead to small, incremental improvements in the product, on the order of a few percentage points. Those are the 10% features. They take less time and effort to build and are less risky because you can A/B test them to measure the improvement and undo the change if there’s no positive impact.

Other features are risky, big, and difficult to A/B test, but they also lead to huge improvements in the product if they’re successful. They’re high-risk, high-reward. Those are the 10x features. Some famous examples are Microsoft switching Office from a pay-once to a subscription model or Amazon deciding to launch 2-day shipping. It’s a high cost and hard to test, but if it pays off, it can reshape the company.

But these two different features require very different levels of investment and different levels of conviction, skillsets, and feedback loops. You might need a team that’s very good at execution for the 10% features but a chaotic and creative team for the 10x features.

How does this relate to my life outside of Grammarly and the software world? This metaphor of 10% vs. 10x improvements applies to almost every goal you can have in life.

Say you want to become more healthy and fit. Well, what kind of improvement are you actually looking for? If you want to be slightly healthier and lose a few pounds, that sounds like a 10% improvement. You can probably accomplish it by going to the gym more often and skipping the pastry or bagel with your morning coffee.

But if your definition of becoming more healthy and fit is to run a marathon, having never run one before, that starts to look like a 10x change. And that 10x change requires a mentality shift and lots of daily training. You need to radically alter your day-to-day habits to achieve that goal.

The lesson here is to be honest about what type of return I’m looking for with every goal and ensure I’m properly investing to achieve that return. A 10% goal is very different from a 10x goal.

Don’t forget about fun.

The final lesson is probably the most important. On paper, my journey at Grammarly over the past five years looks intense. I started when there were 100 people across two offices. I worked on various teams and complex projects, and the company today has over 1000 people and is one of the biggest unicorns in the world. When I tell people how long I’ve been there, I often hear, “That must have been an intense few years.”

There were certainly times when it felt intense, but it never felt unsustainably intense because of how much fun I had every day. And I don’t just mean the classic corporate sense of fun. Don’t get me wrong, I did enjoy almost every project I worked on – they were challenging, interesting, impactful, and taught me a ton. As someone who loves challenges and personal growth, that is fun.

But I also had fun in the general sense of the word. I traveled to cities like Vancouver, Berlin, and Kyiv and met incredible people with incredible life stories and experiences. I also took advantage of the remote-first lifestyle and lived in places like Hawaii for several months, and surfed every day with incredible coworkers I consider some of my closest friends.

These are experiences I wouldn’t have if I didn’t consciously try to balance working hard and enjoying my life alongside work. Had I not done this, I probably would’ve found the past five years unsustainably intense.

The lesson learned here, despite its apparent cliché, is to not forget about enjoying life alongside your work.

In the previous post, I talked about how without proper checks and balances, most products become bloated in size and user experience. But it raises an important question – do teams always need to keep building new features? When is a product considered good enough or done? And what should a team do once it’s done?

Definition of done

We’ll start where most discussions should start – by defining some important terms. The biggest word to define is done, since it’s ambiguous and subjective.

All software products and their features are built to solve a user problem. Imagine your product is an actual person. Now, you can think of the problem as a user or a company “hiring” your product to do a task or solve a problem. For example, I hire Airbnb to help me find and book a stay at someone’s place. I hire my laptop to help me write things and check my email and read things. So we can redefine the question to when has my product solved the problem it was hired to fix?

But that’s still quite vague – how do we know when the problem is solved? And for how many users? Okay, let’s make it simpler. Say you could ask every user, after they use your product, whether their problem was solved? I’m making lots of assumptions here, such as:

• Assume the user is the one buying the product (not true for most B2B products!)

• Assume the user hires your product to solve a single problem

• Assume the user can evaluate whether the problem was solved or not

• …

Accounting for all those assumptions in the real world is left as an exercise to the reader. For now we’ll assume that you can magically ask all users of your product whether their problem was solved or not.

That’s a pretty reasonable way to measure how “done” a product is: what percentage of users will say your product solved their problem.

The importance of leading and lagging indicators

For most products, it’s not feasible to ask every user whether the product solved their problem. For starters, users may not care to tell you. But there may just be too many users for you to feasibly ask – if you have a million users, how would you ask every single one of them?

Sure, you can try to send a survey. But if you saw a survey from Google asking “Hey there user, did Google solve your problem,” what would you say? No, the reality is the question requires too much context to ask in a survey, and to provide every user that context is infeasible for widely-used products.

So what do you do instead? Let’s consider a hypothetical (and laughably oversimplified) example.

I have a dog that needs food every day. But I also have to go to work every day. So, I decide to build a machine that automatically dispenses dog food when I press a button on my phone. The user is my dog, and it solves the problem of my dog getting too hungry by the time I’m home.

Now, without installing a camera, how would you know whether the machine works?

Well, I could ask my dog. But I don’t speak dog, so that doesn’t work. Next idea.

What if I filled the bowl when I got home, and if my dog finished the food immediately, then the machine didn’t dispense enough food and it didn't work. But what if my dog is just really hungry all the time, and this will actually make him overeat? Next idea.

What if I monitored the weight of the food at the beginning and end of the day, and also monitored my dog’s weight to ensure he’s not overeating? Not a bad idea.

You can imagine if we continued iterating, we could come up with better and better indicators of success. But in both instances, I utilize leading and lagging indicators to get at the underlying idea of whether my product solved the problem for the user.

That’s the gist of it. You can’t ask every user whether their problem was solved, so you need to use metrics (aka heuristics, indicators) to determine whether the problem is solved. And you might use the wrong metrics the first few times and realize you didn’t solve the problem as much as you wanted. This happens all the time, and it’s part of the process.

But seriously, how done is done enough?

That’s up to the team working on the product. Now, you might think this entire post was a bait-and-switch, and you’d be right. I said I’d tell you how to know when something was done, and then at the end I said that there’s no way for me to tell you when something is done.

But here’s why this post is still useful – you now have a framework by which to evaluate done-ness. Don’t discount the importance of this framework.

Here it is defined more succinctly:

1. Identify what problem your product or feature solves. It should be a simple and clear problem. If you asked any user of your product whether that problem was solved, they should be able to give you a yes or no answer.

2. Formulate leading and lagging indicators of that problem being solved. These should include a good mix of metrics that help you measure, at scale, whether you’re solving that user problem or not. For earlier products or B2B products, you might have such few users that you can talk to them directly. But you still want to ask them questions that help you understand whether you’re solving their problem.

3. Create a definition of done for those indicators to help you decide when to move on to another problem. This is critical for most teams, and in the real world, you’d probably be constantly evaluating your indicators and goals, trying out features and optimizations to reach those goals, and re-evaluating. It’s an iterative process that teams undergo every quarter. Don’t feel like your goals or metrics are ever set in stone.

4. Rinse and repeat. This answers the “what you do next” part of the article. The reality is that you can always invest in solving new problems, and this is the challenge of working on software products. True success comes from constantly asking yourself what other opportunities exist and how you can create more value for users.

This framework can certainly be applied to other areas as well. Software development has many overlaps with problem solving in general and how to think about time management and resource allocation. Hopefully applying this framework can be useful for you when you’re working on complex problems that require an investment of your or others’ time, effort, and resources.

But there seems to be some truth to this reaction. I’m sure you’ve noticed how certain apps on your phone, over time, have gotten so complicated and have so many features that you never use. Take Facebook, or Yelp, or even iOS itself with all its latest bells and whistles that seem to make it more complicated.

Why do all these products become more and more bloated? What does bloat even mean? And is it bad, or are users just overly sensitive to change? I cover these questions in this week’s post.

Defining product bloat

It’d probably help to start by defining the word bloat as I use it.

When it comes to software development, there are two broadly agreed-upon definitions of bloat. The first is bloat in the app size (or bundle size, installation size, etc.). Essentially, “how much space does it take up on the user’s device?” For websites, app size bloat can be the size of the website that needs to be downloaded from the web server.

App size is usually measured in megabytes (MB) or gigabytes (GB). It’s easy to see why this matters. People’s time is limited, and if something takes too long to load or download, then you’ll lose people before they even see your product. Sometimes, there are even platform restrictions – until recently, Apple used to restrict cellular app downloads to a maximum of 200 MB. Anything greater than that limit would require WiFi to download. App size bloat impacts new user acquisition – if it takes too long, I won’t wait for it to finish loading.

The second kind of bloat is the user experience (UX) bloat. This is harder to measure, and the topic of debate among product, design, and engineering teams everywhere. UX bloat can be experienced as a cluttered interface or lots of information being presented at once. But it comes in many other flavors. The impact here is also more subtle, but quite sinister. If your product’s UX is bloated, newer users can get overwhelmed and not know how to do what they want in your product. And more experienced users can get confused by the unfamiliarity if you continue adding bloat – finally when you try to fix it, those users will stop using your app because now everything is different than what they remember. UX bloat impacts the retention of users – if I can’t easily use your product, I’ll stop using it.

Why does bloat happen?

Hopefully by now you understand what bloat is and how it negatively impacts users. But how do we end up in a situation where almost every successful product slowly becomes more and more bloated? Surely the most talented, intelligent people are working on these apps – are we all just doing something wrong?

Like most things in life, it comes down to tradeoffs and incentives. Bloat happens because product development incentivizes shipping new features and functionality and doesn’t explicitly disincentivize bloat. Let’s consider some examples.

Example 1: A team is faced with a decision of whether to ship a new feature that adds 10 MB to an app’s download size. Optimizing the app size would take an additional month of work. But shipping the feature quickly reflects positively on their performance. They’re more incentivized to ship it without optimizations, so they do. The end result is the app size goes up by 10 MB, increasing the app size bloat. User installation rates drop by 0.1%.

Example 2: A team wants to ship a new feature that they’ve been working really hard on. They want to ensure lots of users use that feature because it’ll reflect positively on their performance. But putting the feature’s button on the homepage could make it feel cluttered, since most users on the homepage probably don’t care about this feature. They’re more incentivized to put the button on the homepage, so they do. The end result is that the app’s homepage gets slightly more cluttered, causing UX bloat. User retention drops by 0.1%.

In both of those situations, product development teams aren’t sufficiently incentivized to minimize bloat. Tragedy of the commons says bloat will inherently go up.

But why is this bad? Notice that there’s a pretty sinister impact of this bloat. In both of those examples, bloat’s downstream impact on the business metrics (installation rates and user retention) is so subtle that it might not be detected, even if they were monitoring that metric. Unless there are millions of data points per day, a 0.1% drop in any metric might require months of data collection, and most teams are moving too quickly to measure every new feature for that long.

Over time, bloat causes death by a thousand cuts to critical business metrics. But anyone bloat-increasing change may not be noticed on its own.

What can teams do to avoid bloat?

The reality is that some bloat is inevitable. Most of the time, bloat is a negative externality of shipping features. But now that you’ve seen some of the pitfalls of bloat and how it can sneakily impact the business, let’s discuss some strategies to minimize it.

The best way to combat bloat is to think about it as carbon emissions that require a carbon tax. A bloat tax. Again, it all comes down to incentives. Teams and leaders need to disincentivize bloat enough to allow teams to weigh the tradeoffs of shipping features against the increase in bloat.

Here are some steps you can follow.

First, define ways to measure bloat as objectively as possible. With app size, this is easy – just use the actual size increase. With UX this is more challenging, and you’ll have to get creative. One way to do this is with t-shirt sizes: S/M/L/XL, with the design team voting on how much UX bloat a particular feature adds. Another way is to have a point system with voting. Try out different versions to see what works best for your org.

Second, assign a maximum bloat tolerance that you’re willing to incur for a feature before you start working on it. Align on this with the leadership team and all relevant stakeholders. For example, for a very large, important new feature or initiative, you may be willing to tolerate a large increase in the bloat. If Uber is shipping Uber Pool, they’re willing to incur more size and UX bloat since it’s strategic and has high revenue potential. However, Uber may only assign a small bloat tolerance for scheduled rides, as it’s not core to their business or strategy.

Lastly, and most importantly, encourage teams to stay within bloat guidelines, and help them adjust on a case-by-case basis. It’s important to treat a bloat tax as a guideline rather than a hard-and-fast rule, especially in the beginning. Never stifle creativity at the expense of process. Eventually, teams should feel ownership over their bloat tax similar to how they feel ownership over their deadlines. If the tax is being used as a punishment or to evaluate performance, there are bigger underlying problems around expectations and ownership.

Overall, bloat is a systemic and slowly eroding problem that eats away at business metrics and causes users to become frustrated or eventually churn. But hopefully by bringing attention to bloat and starting to create a bloat tax, you can tip the incentives in such a way that teams actively seek out ways to minimize and mitigate bloat.

Turbulence on the rocketship launch of the internet

The vision of the internet was to create a decentralized, global network that facilitated the free and open exchange of information.

But near the turn of the century, China started building its Great Firewall (GFW) project with the aim to moderate the internet and establish a philosophy of “internet sovereignty” – the idea that the internet inside a country should be controlled by the country’s government. This idea seemed contrarian to the very name of the “World Wide Web, which was created upon the premise of openness, decentralization, and freedom.

Yet the GFW project trekked on, completing both major phases of development in 2008, just 20 years after it was started in 1998, and just 24 years after the internet arrived in China. The project was complex even by today’s standards, allowing the CAC, or the Cyberspace Administration of China, to perform a variety of filtering including IP range bans, DNS spoofing, and even execute man-in-the-middle attacks with TLS by impersonating any valid TLS certificate. It’s estimated that over 50,000 police officials were involved in building the GFW.

The internet continues to get splintered

Let's fast forward a bit to the last decade. The internet is no longer the panacea of the open exchange of information it was envisioned to become. In fact, most would characterize the internet as of the last ten years or so as a splinternet – an internet split by sovereign states and entities to serve their national interests and laws.

Other countries such as Russia have followed China's footsteps and created their own set of laws and tools to passively monitor and actively filter the internet, including creating their own domestic DNS, creating a further divide in what was intended to be a universal, open standard. There are, obviously, benefits to restricting the availability of content on the internet. For example, the US and Australia have been in active discussions to create a firewall to block weapon-making instructions and other such unseemly content’s availability and accessibility on the internet.

The aggregation of internet access through phones and apps

Fast-forwarding to today, the modern internet, for most people, is not through their computers, but rather through their mobile phones. Those phones run operating systems developed either by Google or Apple, a duopoly that’s started to get more scrutiny by the US government and its legislative bodies. And over 90% of the time spent accessing the internet through smartphones in the US occurs through apps, not websites. Those apps are governed by restrictions where they’re listed – on the Apple App Store and the Google Play store.

So let’s recap – in the past few years, the main way people are accessing the internet is through their phones. And this mobile internet usage is largely through apps, not websites. And those apps are controlled by the first and third biggest companies in the world (at the time of publishing this article, Dec 2021). What does this mean for the openness of the internet? Surely putting two Silicon Valley companies created in the liberal state of California in a country that eschews freedom of speech and democracy means that the internet’s fate of openness is safe, right?

Wrong. Enter country-specific app bans

In reality, this consolidation of control has given state governments the ability to quickly and efficiently ban access to applications they find threatening to their national security, political ideologies, or any other reason at their discretion. For example, India famously banned TikTok and other Chinese apps from the App Store and Play Store following conflict at the India-China border. Russia forced Apple and Google to ban an app that coordinated protests in favor of the opposition leader, Navalny. How exactly did Russia force these behemoths to remove the app from their stores? By threatening prosecution against their employees based in Russia. And China has been known to ban VPN, religious, and gaming apps on the App Store and Play Store, despite the decades-long investment into their Great Firewall – it’s just easier to compel two companies to ban an app than it is to develop a national, real-time firewall.

The reality – consolidation of distribution makes censorship easier, not harder. The internet was created with the idea of decentralized distribution. Anybody could start a website, and simply removing something from a search engine wouldn’t stop others from accessing it directly. But now that apps are published and distributed by two large entities, censoring something doesn’t require a complex, national firewall – it just requires two phone calls to Apple and Google.

The takeaway

Censorship, by its very nature, isn't bad. The United States censors websites that facilitate drug enterprises, explicit content of minors, and other criminal content. And Apple and Google aren't wrong for following the laws and regulations of the countries in which they operate. However, what is more concerning is that we seem to have found ourselves in a situation where the very precepts of the internet are no longer applicable in our day-to-day usage across borders. The splinternet is real, and it's very much snuck into our daily lives. And we seem to find ourselves in this situation without ever stopping to ask ourselves - are we making the right tradeoffs?

Arguing that you don't care about the right to privacy because you have nothing to hide is no different than saying you don't care about free speech because you have nothing to say.

In this article, I want to cover three very simple reasons why you should care about privacy, even if you think you have nothing to hide.

Reason 1: You’re probably already breaking the law but don’t know it.

The average person commits three felonies a day due to vaguely written laws. Don’t believe me? The legal definition for a conspiracy, which is a felony, is “An agreement between two or more people to commit an illegal act, along with an intent to achieve the agreement's goal. Most U.S. jurisdictions also require an overt act toward furthering the agreement.” That means if you talked about drinking alcohol with friends while you were underage, and then you went ahead and drank alcohol, you just committed a felony, and you could be charged for a felony in addition to the actual crime you committed. In some situations, you don’t even need to do the actual crime – just conspiring to do it is already a felony that you can be sent to jail over!

Reason 2: Even if you aren’t already breaking the law, you can be made to seem you are.

Cardinal Richelieu said it best: “If one would give me six lines written by the hand of the most honest man, I would find something in them to have him hanged.” The justice system isn’t perfect, and an estimated 6% of state imprisonments are wrongful convictions. And this is when we assume all parties in the trial were good, honest, and following the law.

We have strict rules around discovery and search and seizure for a reason, but we don’t have as strict rules for digital privacy yet, and it’s your responsibility to be careful with what information you share online.

Reason 3: Data is forever, but trust isn’t.

Even if you fully and completely trust an entity with your data right now, there’s no reason to suggest you can trust them indefinitely – company leadership and strategy can change, government policy and laws can change, and people’s motivations and values can also change. The problem is – once the data is out there on the internet, you can never take it back. Even if you “delete” something online, some services will mark it as hidden and not show it but continue to store a copy in their databases for their records.

The takeaway

Privacy is a fundamental human right. Privacy shouldn’t be treated as an optional chip we can choose to spend in return for convenience. Instead, we should value our digital privacy the same way we value our physical privacy and advocate our representatives to respect and uphold those rights through legislation and policy.

I want to end this post with a quote from Glenn Greenwald, a journalist who was one of the first to publish the Snowden files:

The other really destructive and, I think, even more insidious lesson that comes from accepting this mindset is there's an implicit bargain that people who accept this mindset have accepted, and that bargain is this: If you're willing to render yourself sufficiently harmless, sufficiently unthreatening to those who wield political power, then and only then can you be free of the dangers of surveillance.

But some people working in tech, especially in high-demand and high-paying fields (e.g., software engineering), are worried that the increase in talent pool means reduced compensation. The thinking behind this seems sound at first – neoclassical economics (the typical supply/demand curves you’ve probably seen before) says that if the supply (talent pool of employees) increases, and demand (number of job openings), then prices (salary) should decrease.

But there are glaring assumptions baked into this model that may change how the wage market behaves for tech jobs.

In this post, I’ll cover three reasons why tech salaries may stay the same despite a remote-first market.

1. Tech employees are highly differentiated and specialized.

The supply-demand curves you see assume perfect competition with an undifferentiated product. To map this to the wage market, this would mean that compensation for a particular role, for example, software engineer, would be fixed at some market-determined price, and there would be no difference between any two engineers. Obviously, this is not the case – tech jobs require highly specialized expertise even within the same job function. People get hired based on their management experience, leadership ability, prior roles, etc.

2. Sometimes, humans are irrational and altruistic.

The supply and demand model assumes that individuals are perfectly rational and always serving our narrow, economic self-interest. There’s no room for employees to take a pay cut to work for a non-profit or a company they’re passionate about, or for a manager to reward an employee for being loyal or caring, or for an employee to boycott working at companies that violate their ethical values. Thankfully, the world we live in has all such instances, and I think we’re much better off for it.

3. In the labor market, excess supply can exist without getting removed.

The neoclassical model assumes that supply and demand will intersect at the market-clearing, or equilibrium price, resulting in no excess supply. However, labor markets are not market-clearing, as evidenced by the existence of unemployment.

This implies that wages are not at their market-clearing price. One theorized reason for why this might occur is the efficiency wage hypothesis, which suggests that employers might pay more than the equilibrium wage to incentivize employees to work harder or reduce turnover when hiring costs are high. These scenarios are pretty common in tech roles, where wages are generally high and employment benefits/flexibility high to encourage productivity and discourage turnover.

Clearly, some invalid assumptions baked into the neoclassical economic model make it not great at describing the labor market in general, especially the tech labor market. So what’s the takeaway?

For starters, don’t worry. But for every argument that tech compensation will decrease due to increased supply, there’s an equally valid counterargument for why they will remain unchanged. Just sit back and enjoy the newfound flexibility and freedom you have at your disposal.

And if remote working isn’t for you, that’s totally fine. More tech companies are becoming remote – of that, there’s no doubt. But there are many companies still functioning in an in-person or a hybrid working model.

There’s really no telling what will happen in the long-term as we continue to participate in this global experiment of remote-first working. We just have to wait and see.

And in both cases, the copy-cat product had an inferior design and user experience than the original. Instagram Stories initially didn’t let you post old content, didn’t have location filters or stickers, and didn’t have selfie face filters. Microsoft Teams had (and continues to have) a variety of user experience issues and stability problems.

So why is it that startups are worried about larger players with inferior products copying them? Shouldn’t the best product win?

Good enough theory

On the contrary, Snapchat and Slack were right to be extremely worried, due to something I call the good enough theory. This theory is, in short: dominant platforms don’t have to build a better product, just a good enough product to discourage users from leaving their platform.

For example, when Snapchat launched Stories in 2013, they had around 30M users, compared to Instagram’s 100M. When Instagram finally launched Stories in 2016, they had over half a billion users, and there was a precipitous decline in Snapchat’s growth rate. Snapchat built Stories to acquire new users, while Instagram built Stories to be good enough to keep their existing users from switching.

The story repeats with Slack and Teams. By 2015, just two years after Slack launched, they had over 300K users, of which 20% were paying, which was unheard of for a productivity workplace product since Dropbox. And yet again, Microsoft launched Teams in late 2016 and in just three years overtook Slack. While Slack needed to sell to new customers, Microsoft just had to make Teams good enough to bundle it into Office 365 for free. The real value came when customers renewed their Office 365 subscription without needing to separately purchase Slack.

So what can startups do?

What can we learn from this? First of all, building a startup is unfair. Not only do you have to build and sell iteratively while staying afloat, but you’re competing against large companies with established customers and hundreds if not thousands of employees that can be deployed against you at a moment’s notice. But if it was easy, it wouldn’t be exciting.

There are still several things startups can do:

1. Focus on building the best product for your users. People who use Slack love Slack, and that’s how they’ve managed to grow so quickly. As long as they continue doing this, they can maintain their existing user base while the rest of the world realizes that Slack is a superior product to Teams.

2. Solve major problems for a small subset of users. Microsoft won’t invest in any solution that can’t reach the majority of their user base, while Slack can afford to build extremely specialized solutions and integrations that only a small subset of their customers are willing to pay a lot for. Once you've solved for a smaller userbase, take what you've built and learned and expand horizontally or vertically. Eventually, you'll have an established relationship with many paying users while also building a huge barrier to entry for larger players.

3. If all else fails, solve a different problem. Slack’s new strategy is to become the social network for enterprise. With their acquisition of Rimeto and launch of Slack Connect, they’re targeting a different market than Microsoft, which is now positioning Teams as a Zoom competitor with increased video and voice capabilities. This allows Slack to answer the critical question they’ll face when trying to get any new user – "We already have Office 365, which comes with Teams for free. Why should we pay for Slack?"

In summary, Slack has a much better outlook than Snapchat, which is latching onto its younger demographic as a last-ditch effort of differentiation. All social networks fade, but enterprise products only die with the company. That’s why companies like Oracle still exist despite subpar products but nobody knows what Myspace looks like today.

One-time Purchase

Back in 1995, a company called Microsoft was dominating the operating system market with Microsoft Windows. Even today, over two decades later, Microsoft Windows holds almost 40% market share.

So how do you pay for Windows? It’s simple – you pay a flat, one-time fee. That’s it. I use the term “value” throughout this post to refer to the benefit the user receives. In this case, the value is an advanced, user-friendly (at the time) operating system and all the security and usability benefits that come with it.

This idea of a one-time purchase was popular in the early dot-com bubble, as it was guaranteed up-front revenue at a time when convincing users to pay for a digital product was rare. Today, this model is rare for software products, for several reasons:

1. Software requires upkeep due to bugs, security vulnerabilities, and platform changes. With a one-time fee, there’s no guaranteed recurring revenue making it difficult to pay for that upkeep.

2. It’s extremely difficult to predict the right price – if you charge too much you’ll get no users but if you charge too little you’ll start losing money quickly when user growth slows down.

Advertising

Then, around 2000, a company called Google was rapidly rising in popularity for their famous search product. And how do you pay for Google Search? Even simpler – you don’t.

Google Search pioneered the advertising pricing model. Google’s advertising model consists of users using their search engine to see search results with ads mixed in. Businesses can purchase those ad slots based on different search keywords, tags, or demographics. The more coveted the ad slot, the more costly the ad slot. For example, ads on searches for “t-shirt” would be more expensive than “black dress shirt with polka dots” since the former is searched more often resulting in more users seeing that ad slot. And also because nobody buys a dress shirt with polka dots. Seriously.

But the users seeing those ads neither earn money nor pay money for using Google Search. Of course, users receive the value of access to a free, high-quality online search engine.

Marketplace

So what does a marketplace model look like? Think Airbnb or Uber – the product is just providing a platform to simplify the connection between supply and demand. For example, Uber is not providing actual rides, but rather connecting drivers (supply) with riders (demand). Similarly, Airbnb isn’t providing actual rooms but is rather connecting homeowners (supply) with people looking for a place to stay (demand).

Marketplace models have become extremely more popular in the past few years and have grown increasingly complex. For example, Opendoor connects homeowners looking to sell with prospective buyers, and they do this by actually purchasing and holding on to the home. Instacart connects grocery stores with people who want to do grocery while also employing contractors to do the shopping and deliver the groceries.

So how do these companies make money? For each transaction (ride given, night booked, house sold, grocery order delivered, etc.)they take a small commission, which is usually reinvested to provide an even better service to both the supply- and demand-side users.

Usage-based

Next we’ll explore the usage-based model, where users pay a fee directly proportional to how much they use the product. This pricing model is primarily used by B2B products, such as Amazon Web Services (AWS).

By using a usage-based pricing model B2B business can offer their service practically for free to small customers. This reduces the barrier for new customers to start using their service. Then, as the customer scales in size the business makes proportionally more money.

In addition, over time customers become more locked in to the business’ service. Switching from AWS to Google Cloud might be cheaper on a per-month basis, but if customers also have to factor in the cost of removing AWS from their software and integrating Google Cloud, which can take weeks, it might not actually be worth it.

Subscription – the next dominating pricing model?

Most companies seem to be adopting a subscription model, in which **users pay a recurring flat fee to continue using the product or receiving the service.** The most popular example is Netflix – having a subscription gives you unlimited access to their software platform and the content available through it.

However, many legacy companies are trying to switch to this pricing model:

• Microsoft has been moving from one-time purchase (Office 2019) towards Office365, their subscription model, cloud service with almost identical functionality.

• Apple is moving from a one-time purchase (iPhone) to a services model, with recurring revenue via iCloud, Apple TV, and Apple News (rumored to be announced in just a few days).

• YouTube is trying to move from an advertisement to a subscription model via YouTube Premium, which gives access to additional content and removes all ads.

• Amazon is testing moving from a marketplace (buyers and sellers of goods) to a subscription model via their Subscribe & Save campaign.

The subscription model is the most superior model for most software companies today for several reasons:

1. The software requires a recurring investment to support its upkeep. A subscription model provides this recurring revenue.

2. Users paying for a subscription aren’t monetized through ethically questionable ways like targeted ads or selling their location or personal data. Rather, they’re paying upfront for something they clearly understand. Remember, if you’re not paying for the product, you are the product!

3. Lastly, companies can focus solely on gaining more users through growth best-practices and through building new features and functionality to increase the amount users are willing to pay.

The subscription bug is spreading like a fire throughout technology companies, and it makes sense. With decreasing infrastructure and scaling costs and increasing regulatory pressure and government involvement you can expect more business to continue making the transition to subscription pricing models.

As we go forward, expect the subscription model to become more nuanced and differentiated, with additional aspects such as freemium models (Grammarly), free trials (Asana), and special pricing tiers for different user segments (Slack), and many more flavors of subscription emerging.

Some people try and summarize their products with a single, all-encompassing metric usually called the “north star metric.” For example, Facebook is famous for using “monthly active users” and Pinterest for using “weekly active repinners and clickers.”

There are situations in which a north star metric is extremely useful – having a single metric allows for a clear aspirational goal to unify the entire company and celebrate success. For example, Facebook set a clear mission to reach 2 billion monthly active users – this number was advertised as their next milestone and was widely celebrated when achieved. Similarly, Airbnb celebrated 1 million nights booked, and Pinterest celebrated 235 million boards created.

But don’t be fooled – these metrics are only used to externally share the company's success. Internally, many different metrics are used to measure growth, and relying on just one key metric to inform growth can lead to myopic decision making and poor tradeoffs when deciding what to build.

To see problems that can emerge from using a single north star metric, let’s take a look at our friend John.

John is working on his podcast app. Using his app, users can subscribe to their favorite podcasts, share them with friends, and get notifications when there are new episodes available. John wants a way to measure how quickly his app is growing, and he decides to pick a north star metric.

John goes on Google, searches “good north star metrics”, and picks daily active users (DAU).

But this is bad. Why? Well, for starters, it’s horribly exposed to seasonality – by looking at daily active users, John is assuming users will use his app in the same way every single day. But when building a podcast app, this might not be the best assumption since users listen to podcasts most frequently on their commutes. On weekends, most people don’t have commutes, so the metric would be useless on Saturday and Sunday.

Okay, well what if John picked monthly active users? Well, in that case he has to wait until the end of each month to measure whether his product grew!

So John settles on weekly active users.

But that’s only half the problem, because John hasn’t even defined what “active” actually means! If I open the app, am I an active user? What if I’m not signed into an account? What if I accidentally tapped on the app, opened it, then immediately closed it? What if your app is running in the background every time I unlock my phone?

My point is this – every product is different, and using one metric just because another successful product used it doesn’t actually work, since how your users actually use your product might be completely different.

So John does a little more thinking and decides to pick a metric tied to the main action users perform on his app, and picks Weekly Episodes Viewed.

You probably know what I’m going to say – this is bad. For starters, it’s ambiguous – if it goes up a lot, does that mean users are viewing more episodes or are you just getting a lot more users? If you’re getting a lot more users, are those high-quality or low-quality users (because it matters if you’re spending a lot of money to acquire them!).

In addition, it’s too narrow – what if people start subscribing to more podcasts? What if people are starting more episodes but not finishing them?

The point of this blog post isn’t to just berate John’s decision making. To be fair, he is picking a frequency – weekly – that makes sense, since users generally listen to podcasts on their commutes. And the user action he’s measuring – episodes viewed – is probably the best indicator of engagement with the product. If users are watching more podcasts, they’re not only discovering new content that will keep them engaged but also getting more value out of the product as a whole.

So what’s the problem? This north star metric is still too narrow, and doesn’t give John or anyone else a good sense of how the product is growing.

## Of course, John could continue creating other, more granular metrics.

He could combine both Weekly Active Users (WAU) and Weekly Episodes Viewed into Episodes Viewed/WAU, which normalizes the metric and incorporates both user growth and engagement growth. But then what if WAU goes down, making Episodes Viewed/WAU go up – clearly this is a bad outcome, but he wouldn’t immediately detect it.

Or he could look at Revenue/Weekly Active User, which captures the value generated for the business – but this doesn’t represent customer engagement. If you suddenly doubled the number of ads shown to users, you’d increase Revenue/DAU significantly, but users would quickly stop using your product as much and in a month or two you might be left with half as many users watching half as many podcasts. Not a good outcome.

Okay, okay, we get it. So what *should* John (and everyone else) do?

Stop treating “north star metrics” like the end-all be-all metric and accept the fact that there is no one metric that summarizes your entire product. Products and users are complex and can’t be summed up by a single “north star” metric.

What would happen if tomorrow, Apple or Facebook decided to stop reporting anything in their quarterly earnings report except one number? Well, aside from potential SEC lawsuits, the stock price would tumble because nobody would have a clue how they were growing!

In the same way, your products growth is not a linear function that goes up and to the right. Understanding your product’s growth is like reading a story, not a graph. What you can do, however, is come up with a list of “guidance metrics” that are informative and actionable.

Let’s help John with this.

1. Create a list of core user actions – what are the three main actions users can do when using your product? For John, this list is:

   1. Viewing a new episode.

   2. Subscribing to a new podcast.

   3. Uploading a new episode for a user’s own podcast.

2. Find a good metric to measure how many users you have. Say Weekly Active Users (WAU), where “active” is defined by someone opening his app for at least 5 seconds.

3. Now, put (1) and (2) together! John’s guidance metrics are:

   1. Episodes Viewed/WAU

   2. Subscribes/WAU

   3. Uploads/WAU.

Now, John can either create a growth model using these metrics to forecast his product’s growth, create dashboards of these metrics over time so everybody can understand how the product is growing in realtime, measure the success of new features based on what user behavior they should drive – the possibilities are endless!

By using the three steps outlined above to come up with guidance metrics, you can ensure you’re constantly measuring your products growth without being overly focused on one north star metric.

But what is growth? What does it mean to have a person or team focused on the growth of a company? How should you start thinking about growth in your products?

The goal of this post is to explain what exactly growth means, and how many successful companies approach growth in a systematic, scientific way. For those familiar with growth, I’ll make things more interesting by not using any popular buzzwords such as retention, paid acquisition, funnels, virality, SEO, north star metric, and DAU/MAU. Let’s get started.

Growth has just three steps

Step 1: Users give you something in exchange for something else.

Step 2: You invest what they give you to tell more people about your product.

Step 3: Some of those people become users!

I call this the growth engine. So what does this actually look like within a product? Let’s check out some examples.

How Candy Crush grew by showing ads.

Ads are by far one the most popular ways to grow, and games such as Candy Crush use ads extensively.

Step 1: Users give Candy Crush money in exchange for in-app purchases. In-app purchases allowed Candy Crush and other games to make a large portion of money from a very small amount of “power” users – for most games, the top 1% of users generate over 50% of the revenue.

Step 2: Candy Crush invests that money to show more ads. The percentage of users that start playing a game via an ad is extremely small, so you’d serve ads to many people every day.

Step 3: Someone sees a Candy Crush ad and starts playing. Voila, a new user is born! Doing some basic math can tell you on average how much you’re spending to get each new user and how much money you make from each new user. As long as you make more money per new user than you spend to acquire them, you’re making money. Easy as pie.

How LinkedIn grew by getting people to create profiles.

LinkedIn and other online forums and directories (Yelp, Quora, Stack Overflow) grew using a completely different growth engine.

Step 1: Users give LinkedIn personal information in exchange for seeing others’ profiles. LinkedIn goes so far as to gamify this process by showing a progress bar to incentivize you to fill out every single field.

Step 2: LinkedIn allows their personal information (profiles) to be crawled by search engines. Companies like Google and Bing have bots that crawl profiles from sites such as LinkedIn so that when you search someone’s name or company, you can see their LinkedIn profile in the search results.

Step 3: Someone clicks a LinkedIn profile from a search results page and signs up. And once again, a new user is born. Sneaky sneaky companies that rely heavily on this process to grow (LinkedIn, Quora, Medium) will block you from accessing the page until you actually create an account. Don’t believe me? Try visiting my LinkedIn profile in Incognito mode and hard refresh the page once (Shift+Refresh).

How Dropbox grew by getting existing users to invite new users.

There’s no story like Dropbox, which created one of the best referral programs and pushed Dropbox up from 100,000 users in 2008 to 4,000,000 users in 2010, just 15 months later. But, once again, they followed the same three steps of any growth engine.

Step 1: Users give Dropbox their friends’ emails in exchange for free Dropbox space. The beauty of giving users free Dropbox space is that it encourages them to use the product even more!

Step 2: Dropbox sends their friends an email inviting them to register. Not only will my friends receive an email, but they’ll receive a text, Facebook message, and voicemail from me asking them to sign up using my referral code. Persistence pays off.

Step 3: Friends click the link in the email and sign up for Dropbox. Voila, a new user is born. And now, once this new user signs up using a referral link, they already know how useful referring someone else can be – they’ll be even more inclined to refer *their* friends and the cycle repeats itself.

How can you make your product grow?

Remember, the growth engine just has three steps:

Step 1: Users give you something (money, personal info, or friends’ email addresses) in exchange for something else.

Step 2: You invest what they give you to tell more people about your product.

Step 3: Some of those people become users!

So how do you get started? Start by picturing your ideal user. Then, ask yourself three questions:

1. What is the problem in their life that your product solves? If your product doesn’t solve a problem, then there’s no reason for any user to continue using it.

2. Now, what can you ask them for in exchange for solving that problem? Don’t default to saying money, since I’ve given you two examples of successful products (LinkedIn and Dropbox) that grew quickly without making users pay them.

3. Finally, how can you leverage what they give you to acquire more users? This requires experimentation and iteration, and you probably won’t succeed at first.

We’ve only covered three examples of growth engines, but there are many more, and successful products will use a combination of many different growth engines to power their growth-mobile. In order for your product to continue growing, you’ll have to continue experimenting and learning from mistakes. By thinking about growth using this framework, you can keep building out more growth engines in your product and continue helping more users every day.

The evolution of News Feed

The News Feed algorithm has changed over time to reflect feedback from users. Since its conception, the focus has been on allowing users to find high quality, relevant content from pages while always prioritizing posts from friends and family.

However, users complained they were feeling disconnected from their News Feed since public content was muddling interactions from friends and family. So, in 2015, 2016, and 2017 Facebook continually made a variety of changes to their ranking algorithm to prioritize content you’d find relevant, based on who posted it, how often you interacted with them, and what topics were trending across all users. Despite these modifications, the focus remained on showing relevant content, where relevant was defined as something you’d engage with.

Then, at the beginning of this year, Mark Zuckerberg claims:

Recently we've gotten feedback from our community that public content -- posts from businesses, brands and media -- is crowding out the personal moments that lead us to connect more with each other.

Duh. Users had been complaining about this since 2013. He goes on further to explain a shift in strategy:

I'm changing the goal I give our product teams from focusing on helping you find relevant content to helping you have more wmeaningful social interactions.

I think this is a positive shift. The problem from the beginning was always this: what should Facebook optimize for? It's clear that users come to the News Feed to get updates from people they already know. On the other hand, businesses come to the News Feed to reach users they don’t know. In the middle is Facebook, struggling to balance the see-saw by prioritizing relevant content. Unfortunately, optimizing for relevant content, in this case, is equivalent to minimizing the user’s pain of seeing posts they don’t care about on their News Feed. By prioritizing meaningful social interactions, users will probably spend less time on the News Feed, but the time they spend will be more engaged and focused. I’m guessing Facebook believes this will increase ad conversion and avoid user churn.

Google Search in the spotlight

Sundar Pichai recently testified in front of the House Judiciary Committee. Several members of the Committee suggested that Google’s search results were biased against conservatives. Some went so far as to suggest the results were intentionally manipulated by Google or Google employees to support the company’s political agenda. A notable quote from Lamar Smith (R-TX):

Google has long-faced criticism for manipulating results to censor conservatives… PJ Media found that 96% of search results for Trump were from liberal media outlets. Not a single right-leaning site appeared on the first page of search results.

The methodology for the study that he’s referring to was as follows:

1. Type “Trump” into news.google.com.

2. Analyze the sources of results against a media bias chart.

Calling this a study is like calling a tweet a dissertation. Besides, the 96% statistic has been completely debunked by reputed researchers, and PJ Media’s supervising editor even agreed the study is “not scientific".

On the other hand, Darrell Issa (R-CA) made an interesting statement:

If you measure the outcome, what you find is there is an appearance of bias.

Put differently – if ranking algorithms are adapted based on user feedback, and users are biased, the algorithms become biased. Extreme cases of this cognitive bias in machine learning have occurred before, such as when Microsoft’s AI chatbot started spewing racist and inflammatory tweets it learned from other people on Twitter, or when Northpointe’s risk assessment score falsely labeled black defendants as future criminals twice as often as white defendants, or even when Google’s sentiment analyzer labeled statements such as “I am Jewish” as negative.

It’s likely that Google’s user base is already biased across a variety of dimensions – the digital marketing consultancy Further found in a 2015 research study that Google users are younger, more likely to be male, and more interested in gaming than Bing or Yahoo users. Google is heavily incentivized to optimize search results to be relevant to its current users since this results in higher engagement and more revenue. But doing so may create a feedback loop resulting in even more biased search results.

So what should they optimize for?

Facebook and Google have both optimized for showing relevant posts or search results. However, the News Feed still has to show ads, which are sometimes irrelevant. And Search users will find articles that validate their biases more relevant.

So what is the right thing to optimize for? Nobody really knows – this question has spawned a new field of research called machine bias.

One possible solution is to design recommendation algorithms to account for bias. This can be done proactively by using unbiased data to improve algorithms. It can also be done reactively by vetting results via third parties and modifying algorithms to disincentivize biased results. New research around “counterfactual fairness” attempts exactly this by negatively weighting algorithms that fail to account for biases in the data.

Another solution is to improve social awareness of how these ranking algorithms work. It’s clear from Zuckerberg and Pichai’s testimonies that we can’t yet rely on Capitol Hill to properly regulate these algorithms, seeing as they still don’t understand Facebook’s business model or Google’s search algorithm. While the onus should be on White House representatives to understand what they’re regulating, maybe Google and Facebook should take the time to explain and educate how these algorithms work. The current status quo of Capital Hill pointing fingers at Silicon Valley, pointing fingers back saying they’re ignorant won’t get us anywhere.

This is not a simple problem that we’re going to solve overnight. Accounting for bias in algorithms is extremely challenging – any algorithm that learns from human decisions will inherently be biased. However, there are steps we can take today to begin the journey towards reducing machine bias, and Google and Facebook, as thought leaders in the machine learning and consumer technology space, are the perfect role models to take this important first step.