Life 3.0, read from someone who ships AI for a living
I finally read Max Tegmark's Life 3.0. It came out in 2017, and most of the AI conversation since has caught up to it rather than moved past it, which is a good sign for a book. The frame at the center of it is worth carrying around, so let me set it down plainly, and then tell you where it looks different from the trenches.
Tegmark sorts life into three stages by how much of itself it gets to design. Life 1.0 is biological: a bacterium cannot redesign its body or its behaviour, both are handed to it by evolution. Life 2.0 is us: our hardware is still evolved, but our software is largely designed. We learn languages, pick up skills, change our minds, and pass all of it on without waiting for our bodies to change. Life 3.0 would design both its hardware and its software. It does not exist yet. Whether it arrives this century is the argument the rest of the book is about.
It is a genuinely clarifying way to see things. Once you notice that the story of progress is the story of moving more of yourself from evolved to designed, a lot of the AI noise organizes itself. And Tegmark's central worry, that if we ever build something far smarter than us we had better have made sure it wants what we want, is the correct worry to have in principle. I am not going to argue with the physics or the philosophy. He is better at both than I am.
Where the book and my Tuesday diverge
Here is the honest gap. The book is largely about aligning a superintelligence to human values. My Tuesday is about a company that cannot align a chatbot to a single workflow.
I spend my days close to enterprise AI as it actually gets built, and the failure I see is never that the machine is too powerful to control. It is that the machine was pointed at the wrong problem, or at no clearly defined problem, and quietly returned nothing. Most enterprise AI pilots deliver no measurable result, and it is almost never the model's fault. So while the world debates how to keep a future godlike intelligence in its box, the working reality is that we cannot yet reliably get a very ordinary intelligence to do one useful thing on purpose. Both problems are real. Only one of them is going to decide the next ten years.
Alignment is not only a superintelligence problem
What surprised me is how far Tegmark's big idea scales down. Alignment, at its core, is just the problem of making a system reliably pursue the goal you actually intended rather than a nearby goal that is easier to optimize. That is not a far-future concern. That is the exact thing that goes wrong in a three-month AI project. The pilot optimizes for looking impressive in a demo instead of moving a number the business already cares about. It is misaligned, in the small, boring, expensive way, long before anything is dangerous.
So the discipline the book asks us to develop for a hypothetical superintelligence is the same discipline that separates the AI projects that survive from the ninety-five percent that die. Be ruthlessly clear about what you actually want. Build the system to serve that and not the flattering proxy. Know what not to build. If a company cannot do that with a chatbot, it has no business imagining it will do it with something a thousand times more capable.
Living in Life 2.5
I think we are not building Life 3.0 yet. We are living in Life 2.5. We are still Life 2.0, humans with evolved bodies and designed minds, but our designed software now includes tools that can draft, reason, and build alongside us at a speed that changes what one person can do in an afternoon. That is an enormous shift, and it does not require any superintelligence to matter. The whole game, for the next decade, is whether the people wielding these tools have the judgment to aim them at things worth doing.
That is the quiet lesson under all the drama about 3.0. The future will be shaped less by whether we manage to design a new kind of mind, and more by whether we, the ordinary 2.0 kind, get better at deciding what is worth building. It is the same thing I believe about every piece of software my teams ship, only now the stakes have a longer tail. The machines are getting more capable on their own. Judgment is the part that still has to be earned, and it is the part I would bet on.
My children will grow up somewhere on the road toward 3.0, wherever it actually leads. What I want them to inherit is not a smarter machine. It is a habit: the discipline to ask, before building the thing you can, whether it is the thing you should. That question is old, human, and 2.0 to the core. I suspect it stays the important one no matter how far the hardware goes.
More of how I think about building AI that actually holds up is in this piece on why most enterprise AI fails, or in the rest of my writing.
Book: Life 3.0, Being Human in the Age of Artificial Intelligence, Max Tegmark, 2017.