Mind and Iron: The AI company that predicts death dates
Also, are large language models, like a bad NASCAR driver, hitting a wall?
Hi and welcome back to another glistening episode of Mind and Iron. I'm Steven Zeitchik, veteran of The Washington Post and Los Angeles Times and lead shopper at this newsy department store.
Every Thursday we come at you with all the tech, AI and future news you could ever want. At least the stuff that (we think) matters. Please consider getting behind our independent mission.
Hope you all had a good Thanksgiving weekend. I got to spend mine out in some warmer climes while also doubling up a birthday celebration that same weekend. I'm thankful to my parents for having me late enough in the fall that my birthday can't actually fall on Thanksgiving Day. And for having me, period; that was nice of them too.
This week we're getting into the swing of our December to Remember with a double-hose spray of stories. To wit (and speaking of birthdays):
AI is now predicting our estimated day of death. Would you want to know?
And: The last few weeks has seen a battle wage over whether AI models are reaching their natural end point…with some pretty big consequences over how our lives will/won’t be shaped. What it all means.
First, the future-world quote of the week:
“There’s probably not a more important date in your life than the day that you’re going to die.”
— Entrepreneur Brent Franson, telling Bloomberg why we should let AI tip us off to it.
Let's get to the messy business of building the future.
IronSupplement
Everything you do — and don’t — need to know in future-world this week
Knowing your (gulp) expiration date?; The model isn’t alright
1. HAVE WE REACHED THE MACABRE-SATIRE PORTION OF THE AI REVOLUTION? OR JUST ITS PERFECT USE CASE?
That was the question dancing in my head as news emerged this past week of the AI Death Clock, an app that....well you really don't need me to explain it, do you.
The idea of this program, devised by an entrepreneur named Brent Franson, is to take all the health and lifestyle data it can get its hands on about you (a tranche that surely, definitely, unquestionably can never be exploited or raise privacy concerns), compare it to millions of other people's health and lifestyle data and then determine when you might die. Like, literally, it flashes a date.
Think of it as the world's most comprehensive doctor crossed with that hooded old figure in a Peter Jackson movie who cackles witchy predictions. But, you know, data-driven.
Some 53 million health records across 1,200 life-expectancy studies have already been programmed into the Death Clock’s algorithm.
(Btw this is perhaps testament to my grim-minded teenagehood but some friends used to play a game in which they would predict each other's death ages with faux-certainty. This is like that, but, again, with data.)
One big idea with the clock is that it’s designed to change in real time. The app isn’t this granular yet, but it’s ostensibly programmed to respond to our changing health. So if you had a really indulgent meal, or climbed Kilimanjaro, or got on the 405 during rush hour, you'd see the date move up. The one thing worse than knowing your death date is seeing it surge closer.
Not that a postponement seems so great either. Sure it offers temporary relief, like getting the results of some encouraging bloodwork. But doesn't it lower your guard to order that extra serving of French Fries and in turn move the date back up again? You can never truly reprogram a Death Clock.
This isn’t the only such app, either. Last year we learned of an algorithm known as “life2vec” that forecasts life expectancy the way ChatGPT predicts the next word. Call it the mortality-prediction industry: feed in the data, learn how long you’ll be around.
(The 2011 Justin Timberlake movie “In Time” tried a version of this conceit, with digital wrist projections that counted down to people’s deaths, though that was more of a parable for the class disparities in health-care than a fable about knowing too much.)
The accuracy of such apps is not just unknown but unknowable, of course, unless it's predicting deaths in a hospice. So in a way, a brilliant marketing move; by the time we know if it's accurate too many years will have passed. Plus we'll be gone so what are we gonna do about it.
I asked an actuary I know what he thought about the Death Clock program, and he said at least in theory such an app could improve on current models, which really are built to make predictions about large populations, not customize likely outcomes for individuals. Still, an improvement over current models doesn't say much when such models aren't even trying to predict our future.
The big problem is that there's too much data that can't be accounted for. The chance of a cardiac event in the next five years for someone with your health profile may be 1-3%, a seemingly knowable range, but when you combine that with 20 other life-threatening conditions that also run a risk of 1-3%, you're suddenly in a range of 20%-60%, which doesn't seem so useful.
Meanwhile people are struck with deadly diseases all the time with no family history or exposed to environmental factors whose toxicity we don't yet know. And the third-leading cause of death in the U.S. is accidents, which unless you ride a motorcycle around blind turns every day, no computer system can plausibly predict. Even with 53 million of other people’s health records.
(For the really dark underside of using AI to project health outcomes, United Healthcare under Brian Thompson has been deploying an allegedly flawed AI algorithm named nH Predict that aggressively and erroneously turns down claims. Subject for another session.)
The even bigger problem is that a machine predicting a death at 75 for someone who is currently 35 doesn't really mean much since a lot can happen over the next 40 years. An insurance company of course needs to know your life expectancy at that age so they can adjust rates accordingly. But it's not really meaningful for you to know that so many decades out. You could in the interim take up smoking or that motorcycle habit; on the other hand you could drastically overhaul your diet. I get that this is supposed to be more of a motivational tool than a prophetic revelation. (“Friendly reminder that life is slipping away,” one version of the program cheekily notes.) But it functions like the latter, even though it definitionally lacks the information to pull it off.
What the death apps are practicing is something that might be called pseudo-data — the idea that a fact seems knowable when it isn't. Nothing is more certain than death, but unless you've taken up BASE jumping or are in especially poor health, nothing is more uncertain than the exact date it will happen. And thinking that it is certain can lead to all kinds of questionable choices. The mind spins at all the ways we can go astray assuming we’ll die on a day that we won’t. Heck, the mind spins at all the ways we can go astray assuming we’ll die on a day that we do.
Yet there's something unremittingly intriguing about the Death Clock, and on a few occasions this week I found my mind drifting toward it, at once excited and a little scared to try it. (I haven't yet.) Would you want to know the projected date of your death, even if speculated upon by a machine? It's at least as accurate as a fortune-teller, and those people can scare the bejesus out of us.
The models will get better (for how much better, see the next item). But the more important change I suspect will be in our attitude. Hard as it is to conceive from our present perch, we've been on a steadily upward arc when it comes to medical information getting more revealing and predictive. For centuries we had vague ideas of our health from superficial physical diagnostics. Then in the 1950’s blood tests became common. The idea of relying on something so intimately accurate was a little unsettling — know your health future from this liquid inside you? But it felt ho-hum by the time MRI's came along a quarter-century later.
Those were themselves a big deal — magnetically peer into your deepest organs? — until predictive genetic testing arrived a number of years ago. And on it goes. Tests that seem shocking become commonplace, and predictions that seemed impossible become more refined.
The idea that a machine model can, by looking into the data, peer deeper into our future feels unnerving, prompting disbelief and discomfort. And yet sooner or later I suspect such practices will become standard, and we’ll think back on this time as a period as dark and stumble-y as the days before blood tests. I have no idea whether the Death Clock will catch on in this form function or achieve a new level of accuracy. But the history of this stuff suggests that technology predicts our fate with ever-increasing specificity. There's no reason to think that the future will look any different.
2. WHEN SOMEONE SAYS SCALING CONTROVERSY, WHAT COMES TO MIND — A RIGGED WEIGHT SYSTEM, OR MAYBE SOMETHING TO DO WITH FISH?
The AI world has a different idea about what this means. And it’s been percolating a lot in the past few weeks.
First, a little background. There are all kinds of ways for machines to appear to “think” like humans. The one most common in the past two years — more or less since OpenAI released ChatGPT — is Large Language Models. This is the idea that a computer is not really knowing facts or making inferences the way humans do; instead it’s processing large amount of words that have been used for and applied to humans for years and synthesizing them in new ways based on all the ways they’ve been combined before.
This is how GPT-generated language can sound natural, or even like the lyrics of Taylor Swift. It’s also how you can get answers to questions that humans would need to work to answer or maybe not even be able to answer at all, like medical diagnoses or electoral predictions.
LLMs are not knowing or thinking the way we do. But they are, if you will, extracting information by seeing patterns in what we already know.
A core tenet of those who work on LLMs for a while now has been that the more of these words you feed it — the more you stuff it to the gills with information — the more meaningful the results that will spill out. Keep training, it will keep getting more fit. And eventually even possibly solve major challenges like the climate crisis.
How much more can be stuffed in for us to still see corresponding improvements? How much more can this approach — “scaling” — give us? AI evangelists think that the sky’s the limit. That’s articulated most by — who else — Sam Altman. As he put it in September, "To a shocking degree of precision, the more compute and data available, the better it gets at helping people solve hard problems." Scaling to a T.
In a fittingly Shakespearean turn, one of the people arguing against it is Altman’s former partner-turned-rival-and-ideological-opponent Ilya Sutskever.
Sutskever, you may recall, was pushed out of OpenAI in the spring after first trying to push Altman out of OpenAI a year ago. He’s since gone on to start an AI safety company. And a few weeks ago he said that simply adding all this new data, contrary to Altman’s evangelism, wasn’t going to do much of anything.
That the LLM models (they also include Google’s Gemini, Meta’s Llama and Anthropic’s Claude iterations) have basically reached the peak of what they can do. You can stuff it with all the knowledge you want until kingdom come, and it won’t do much more than it does now to solve problems.
“The 2010s were the age of scaling, now we're back in the age of wonder and discovery once again. Everyone is looking for the next thing,” Sutskever told Reuters recently.
(He also said that "scaling the right thing matters more now than ever,” suggesting he hasn’t given up entirely on the current approach. But he’s also duly skeptical that feeding AI whatever you can get your hands on will do much.)
We’ve told you about some other people who’ve been arguing this for a while. Folks like AI skeptic Gary Marcus, who believes that “from the boneheaded errors, hallucinations, and utter inability to fact-check their own work…the cognitive capacities of LLMs are severely limited.” He has said that “there are serious holes in the scaling argument.”
Or AI pioneer and Stanford university collaborator Ray Perrault, who told Mind and Iron in April that, “There seem to be in the mill maybe one more generation of models out of OpenAI and Google and maybe Meta that will cost an astronomical amount of money and will do somewhat better. But my guess is it still won't do hard math and planning problems. How do you get to the next stage? That’s the $64,000 dollar technical question.”
New approaches, that’s how, he said. (Possibly by iterating on an older approaches, like something called the symbolic approach. Or a new notion entirely.) Of course, that’s a lot harder than just trying to scale the current LLMs.
The AI blogger Alberto Romero, for his part, has fired back, saying that “Scaling is far from over, only this particular way of scaling this specific thing is falling short of expectations.” And on it goes.
The best evidence that the anti-scaling camp may be right is that OpenAI, despite continuously releasing new models over the past year (like the recent o1 nee Strawberry) have essentially just been offering more bells-and-whistles-y variations on what they’ve already had out there. The so-called “GPT-5” — which is supposed to be a game-changer — has not arrived and changed no games.
None of this is to imply that large language models as they exist now can’t be deployed in all kinds of new ways to our lives, whether it’s serving as our assistants, predicting health risks (see above), matching us romantically or even helping us with decisions. Even with the LLM approach, it seems pretty clear we’ve only begun to scratch the surface of how this powerful capability to peer into words and data will be integrated into our worlds. We may not have scientific breakthroughs, but we’ll have commercial ones.
It would be a little like if you were standing in an empty orange grove with a fresh fruit in your hand. There may be no new trees. But you can still squeeze a lot out of the orange in front of you.
The debate over scaling will go on, fights between those who think AI is on the verge of more breakthroughs and those who roll their eyes at the grandiosity. But don’t discount the middle approach. Our lives can change a lot sans a revolution too. Even without new drugs, there’s still a lot of potency in Vitamin C.
The Mind and Iron Totally Scientific Apocalypse Score
Every week we bring you the TSAS — the TOTALLY SCIENTIFIC APOCALYPSE SCORE (tm). It’s a barometer of the biggest future-world news of the week, from a sink-to-our-doom -5 or -6 to a life-is-great +5 or +6 the other way. Last year ended with a score of -21.5 — gulp. Can 2024 do better? The year started off strong but the summer and fall haven’t been great. But this week the outlook’s a little better.
A CLOCK THAT FORECASTS OUR DEATH DATE? Color us intrigued +2.0
TO SCALE OR NOT TO SCALE: Hard to know just yet who’s right on this one. 0