Today

This is one of the questions I care about most, because it's easy to get backwards. Let me start with an uncomfortable judgment: most “simplicity” is really sweeping complexity from the designer's desk onto the user's feet. The interface becomes emptier, but people have to remember more—which layer it's hidden in, which gesture, which long press. That isn't simplicity; it's putting the bill on someone else's tab. The real question isn't “what else can we remove?” but: what is the person trying to accomplish right now, and does he or she have enough attention to spare? Take the Ferrari Luce as an example. We didn't start by deciding to “have fewer knobs.” We first mapped the structure of attention while driving—which operations must be done by touch without looking, which can wait for a glance at the screen—and then assigned controls according to that hierarchy. The physical controls on the steering wheel and the central screen carry different tasks. That's not an aesthetic choice; it's dictated by the conditions of attention. Removing one visible part in exchange for one more glance down while driving is a losing trade. So every time I make a subtraction, I ask the same question: after you remove it, where does the burden go? You have to distinguish three outcomes: It disappeared—because that step was never necessary. That's real progress. It got absorbed—the internal structure became more integrated, and the person actually has less to do. It got hidden—moved out of the interface and into memory. That's the most common fake simplicity, and the one I worry about most. The most dangerous thing about the third kind is that you can't see it in a demo. Renders and screenshots look clean. The cost only appears when a real person uses it for the first time, the second time, and again a week later. So here's what I suggest instead of continuing to subtract on screen: Find a specific person, a specific task, and walk through it in a real environment. Not playing around, but really trying to accomplish something. Watch where he pauses, where he hesitates, what he has to recall, and how he recovers after a mistake. Write down those “I'm not sure what's next” moments—they're more honest than any design review. Then take that list back to the design. The things that should really be removed are those that don't bring understanding, only steps. What remains—even a visible handle, a clear click, a part that takes up a little more space—deserves to stay as long as it helps someone confirm “I did it.” Simplicity isn't emptiness. It's digesting complex relationships so the purpose emerges on its own. Emptiness is undigested—it just pushes the relationships back onto the user. I want to ask you one thing, and it will directly change your judgment: do you want to remove something people need to understand, or something people need to remember? The former can often be reorganized more clearly; the latter, if it can only be maintained by memory, shouldn't be removed.

I don’t know. Anyone who gives you a “final price” — including me — is guessing. I’m not dodging. A podcast once asked me where this cycle would top out, and I said “somewhere between $500,000 and $1 million.” That was a guess, not a model, and I don’t use it to make decisions. Why shouldn’t you take numbers like that seriously? Look at analysts’ forecast ranges for 2026: $60,000 to $137,000; for 2027: $55,000 to $200,000; long-term, anywhere from $300,000 to $3 million. A tenfold spread shows this isn’t prediction — it’s sentiment. First, calibrate where we are. BTC is around $86,000 now; the all-time high was $126,073 on October 6, 2025; it’s down about 26% over the past twelve months, with the year’s low near $57,000 in July. In other words, the people shouting six figures last year have already been schooled by the market. “Bitcoin only goes up” isn’t a premise; it’s an illusion. Only one thing really determines that number: adoption. How many people actually treat it as a store of value, not a trading vehicle? How many institutions, sovereign funds, and pension funds can hold it compliantly? Are custody, on/off ramps, tax, and regulatory channels usable? What share of global wealth will ultimately be denominated in Bitcoin? Today that share is so small it’s almost negligible. If it becomes a few percentage points, the price will naturally be six or even seven figures; if it stays where it is today, the price will look like this. This isn’t a price question — it’s an adoption question on a thirty-year scale. My view: long-term up, extremely volatile path, medium conviction — and there will be several more 70% drawdowns along the way. There’s another layer you need to think through yourself: what is that number for you? If you hold it, what you should actually write down isn’t a target price — it’s three things: why you hold it, what share of your total assets it is, and under what conditions you’ll admit your thesis was wrong. Years ago I sold a house to buy crypto, not because I thought it would definitely go up, but because even if it went to zero I could rely on my skills to find work again. That tolerance came from my situation, not from a one-size-fits-all investment script. Don’t treat it as a template. So my answer is: it might end up very high, or far lower than I said — I really don’t know. Don’t ask me for a number; write down those three things. That’s more useful than any target price.

Depends on which singularity you mean. If you mean a specific moment—some day superintelligence appears and everything changes irreversibly—that hasn't happened, and it probably won't happen that way anytime soon. But if singularity means the process I described in The Gentle Singularity, then the answer is we're already in it: 'We have passed the event horizon; takeoff has begun.' It doesn't look like a singularity, because a singularity never looks like a singularity. Back in 2017 I wrote that human-machine integration is gradual, and gradual processes are hard to notice; at that time I thought it had already been going on for a few years. Today we're still on the same curve: amazement becomes the everyday, and the everyday becomes the new baseline—a few years ago you were stunned that a model could write a beautiful paragraph, now you're asking when it can write a whole novel. The real constraint isn't in algorithms, it's in physics. Getting a data center from groundbreaking to full 1GW takes 1 to 3.6 years, and power availability, grid reliability, and community opposition are becoming the main constraints on new projects; the U.S. plans to add nearly 45GW within three years. So the slope of this phase is set by electricity, land, and permitting—not by ideas. That's the most misread thing of the past two years: the capability curve is still climbing, but the bottleneck has moved one layer upstream. To judge whether it's really accelerating, I don't look at labs' release cadence; I look at two things. One, whether AI doing AI research produces visible acceleration in field-level discovery, not just internal productivity reports. Two, whether self-improvement is only looping inside fixed external scorers. Right now bounded self-refinement is industrial practice, but open-ended recursion hasn't been proven; and a 2026 study found that 73.8% of so-called optimizations produced by code agents only improve proxy metrics, with no real gains. That's getting faster, but it's not yet at the point where it can accelerate itself. So: process yes, moment no. If the day comes when AI's contribution to AI research becomes visible in the field's discovery rate, that's the signal that would make me change my mind.

View full answer