Welcome to Supervised
I'm launching a new newsletter to cover AI, machine learning, and analytics.
The tech industry traditionally has thrown around the phrase “AI” to mean anything that remotely touched any kind of machine learning. They’d literally call any kind of logistic regression an AI.
But in the span of six months, we’ve gone from a cheeky experimental web interface for a highly-sophisticated large language model to one of the single-biggest technological arms races in the industry’s history. Google is now facing down one-time annoyance Microsoft in the most important fight of its life. And a vast and complex network of startups are ripping out and rebuilding the way people operate and raising billions of dollars.
The explosion of the use of large language and diffusion models, built on top of a multi-billion dollar scaffolding assembled in the past few years, is going to power a new and completely unknown generation of technology.
It’s why I’m launching this newsletter to cover that technology: Supervised.
In Supervised, I want to answer the kind of question I got recently from a source: “I’m about to sign a $300,000 contract with this company, and I need to know if it’s going to be around in a year.”
My goal with Supervised is to be a trusted resource for those of you who are navigating this fast-changing market full of both immense hype and immense promise.
Why Now
I started covering technology more than thirteen years ago, writing for VentureBeat, The Wall Street Journal, TechCrunch. I was most recently at Business Insider where I covered big data and AI—ranging from training frameworks like Ray, to orchestration tools, the semantic layer, TensorFlow and its successors, and the associated hardware. In between that I spent around four years working as an analyst, where I got a chance to experience analytics and data science firsthand.
In my ten years plus of covering technology, this is the first colossal shift that feels “real” in the same way the launch of the iPhone felt “real.”
Y Combinator’s most recent batch of companies, 57 companies tagged themselves as “Generative AI” out of 272. It may be hard to come by funding for most companies, but if you work in generative AI and seem “real" enough, you’re probably going to close a seed round in short order.
Perhaps the most exciting part of this is the foundational layers for these tools, like data lakes, warehouses, frameworks, and others, are already well established. Databricks and Snowflake have streamlined the process of managing data, while tools for cleaning, loading, and managing that data are already widely available.
And it’s in that same “real” way that we’re going to undergo such a rapid change in the way we live our lives that it’ll feel like a blur. When the iPhone launched there was a fart app or a beer drinking app for every future Goliath like Instagram. The same will happen in AI, and it will take years—not months—for the signal to emerge from the noise.
Already we’re seeing the first signs of it. You can take just a quick look at Matt Turck’s (Firstmark Capital) landscape for machine learning, artificial intelligence and data. It’s so large it won’t even fit on your screen at full size, much less load in a reasonable manner (and may even crash your browser).
For that reason, it’s never been more important to take a critical and investigative eye to one of the most-hyped technologies in the past decade. Covering the industry doesn’t mean picking winners and losers—it’s about ensuring that people making decisions have the best information for those decisions.
Who Supervised is for
There’s a joke in data science that goes around in interview loops that’s something along the lines of, “did that really need to be a neural net?” It refers to the common problem of (often PhDs) overcomplicating what might be a deceptively simple answer.
That answer often lies in a relatively simple supervised learning algorithm using an off-the-shelf tool like Sci-kit Learn, wherein you get a model that works well enough into prod to solve your problem rather than running down a deep learning rabbit hole.
Supervised isn’t about covering the latest hot shit Generative AI startup. I want to help my readers understand the implications of a technology and the team building it—whether that’s the operational impact of AutoGPT or the impact crater it’s created in the cultural zeitgeist of AI.
These questions are going to become more and more important as time goes on. The cost of API calls for language models may drop, but the scale of usage will increase. And there is of course the unknown side effects of this usage: carbon emissions, prompt vulnerabilities, and more.
I wanted to start a newsletter for a practitioner without a PhD. I aim to build one one that exists somewhere between the broad audiences I tapped at Insider while not going so far as to unpack the specific details of a model’s loss function.
What to expect in Supervised
I’ve described what my work here will be like as similar to some of my stories at Insider, but tuned to a slightly more technical audience. In hope of being transparent to future readers of Supervised, and to hold myself accountable, I wanted to lay out the foundation of the newsletter.
For now, Supervised will be free, and will launch publishing three times a week. In it, you will find:
Analysis of trends, backed up by extensive reporting and experience, to unpack what’s changing in the industry.
Data reports on the industry that can be helpful to decision makers, such as analyses of the job market in AI and big data.
Scoops on what’s happening behind the scenes in the industry (which might be published on a more ad-hoc basis if needed).
Direct work with these tools, such as trying out local LLMs or light prompt engineering, to provide some insight as to what you can do.
Later on, Supervised content—particularly scoops—will appear behind a paywall, while readers will still receive one free issue a week.
All of this is, of course, subject to change based on what readers want. I am naturally a data-driven writer (hence working professionally as an analyst) and am here in service of my audience, rather than the other way around.
Building, once more
I tell people that to be a technology journalist, you have to be a local pessimist while also being a global optimist.
The former means putting your readers first and checking your excitement and assumptions at the door and taking a cooler, more steadied hand when covering important topics. Sometimes people will see your stories as positive coverage, some negative, and you’ll usually be pissing someone off.
The latter, though, I would consider much more important.
To enjoy being a tech journalist—and to be good at it—means embracing hope in the unstoppable force that is human innovation. It means you have to assume that, for whatever great things there are now, there is always something better on the horizon. It means having a certain reverence for the technology and the blood, sweat, and tears that went into building it.
We’ve been through so many red herrings and ZIRP-funded euphoric phases that it’s easy to get jaded. But I felt something for the first time in more than a decade when I opened up ChatGPT and cobbled together an AutoGPT instance: a sense of joy and wonder for the future.
I am convinced, having worked in the industry for more than a decade, there has never been a better time to start a new media organization. But like any lasting company, that starts with doing one thing well.
For me, I hope it is building an audience that I can continuously add value to in this exploding field. If one of my stories can save someone from signing a $300,000 contract with a questionable technology riding a hype wave, I will have done my job.
I hope you all will come along for what will be a very exciting—and very bumpy—ride as we enter this new and extremely weird timeline at a time when we thought tech could not get any more weird.
—Matthew