Mind and Iron: A Buzzy Company Is Building an AI...Scientist
Also, time to resurrect the dead for some help in the classroom?
Hi and welcome back to another savory episode of Mind and Iron. I'm Steven Zeitchik, veteran of The Washington Post and Los Angeles Times, senior editor of tech and politics at The Hollywood Reporter and aspiring Dorothy Parker of this journalistic roundtable.
Every Thursday we bring you the science, tech and business news relevant to our coming AI era. Please consider joining our community.
This week we've got two important developments for you. The first involves the idea of science itself — or more specifically, scientists. You might think their job is more important than ever as we race headlong to solve our enormous planetary challenges. And you'd be right. But what if these folks needed help? And could get it in the form of a machine intelligence?
Next, teaching is hard. But resurrecting famous people as avatars is (suddenly) easy. What if we could use the latter to help address the former?
Quick housekeeping note that depending on schedules we may be off next Thursday. Or we may not be. It's all part of the wonder and mystery of the universe, and publishing exigencies. Either way, we'll see you back here in two weeks with all the cutting-edge news your little heart and vast mind can handle.
First, the future-world quote of the week:
“The ultimate litmus test for the success of AI as a technology will be if it can either make research breakthroughs itself or provably massively accelerate scientists in their ability to make breakthroughs.”
— OpenAI/Anthropic veteran and AI blogger Jack Clark, on what all this machine intelligence is actually for
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
Building an AI scientist; Teaching with the AI dead
1. BUILDING THE PERFECT BEAST, DON HENLEY SANG FOUR DECADES AGO — IN A SONG THAT, improbably, was about genetic engineering when CRISPR was still something you stored bread in. But what about building a perfect scientist? With the use of AI?
Yes, what if the inanimate monster of Dr. Frankenstein could actually build him.
That's the basic proposition of a project called FutureHouse, a nonprofit funded by Eric Schmidt that is trying to use AI to enable discoveries — but instead of individual leaps in specific areas like drugs, an entire system that can spin off Salkian and Curie-ish breakthroughs like nothing. Needless to say, if such a creation could even come close to reality we'll have revolutionized our world and understanding and improvement thereof. Electricity, penicillin, satellites, the human genome, the Theory of Relativity — it could do that all, except the next level up, and in a month instead of 125 grueling years.
Even Henley himself, for all his genius, might have trouble getting to the heart of that matter. To do this, after all, a model would need to attack all aspects of scientific discovery, from early research and hypothesizing to grueling experimentation to the formulation of conclusions to ongoing testing. It's the equivalent of going from a machine that can generate a pizza to one that can open an entire Italian restaurant. The jump from a task to an operation isn't small.
Yet the platform that FutureHouse announced a few days ago seeks to at least begin tackling this multitude of challenges. For now they've broken it down into four discrete areas: Crow, Falcon, Owl and Phoenix (no Eagles). Crow is a general search engine; it's what a lead scientist might have their assistant do in the early stages of their research. Falcon is a deep-dive literature synthesizer — think of it like that “Short Circuit” robot that can gobble up every article ever read and make sense of it in new ways.
Owl enables an answer to the simple question of: Has this been tried before? It might be something that a scientist would use as they're getting closer to completing their hypothesis.
And then, finally, comes Phoenix, which is the start of tools to plan experiments, albeit for now only in chemistry.
This sounds sort of, well, elemental, and seems a long way from achieving the next antibiotic or Sputnik. Of course those discoveries began with all of these stages too, and with none of the AI let alone basic digital tools to help the people making them. Imagine Alexander Fleming or Sergei Korolev doing all they did with the techniques and easy access to information we have today, never mind an AI operating at machine-intelligence capacities.
And anyway, it's just a start. FutureHouse is working on a decade-long time horizon — no "we'll have AGI by 2029 hyperbole here." In short, it's an ambitious goal, but one that, like the scientific experiments it seeks to devise, is going about its quest methodically and with no shortcuts.
As a multi-authored post on the FutureHouse site put it: “Science is bottlenecked by data. The 38 million papers on PubMed, 500,000+ clinical trials, and thousands of specialized tools have created an information bottleneck that even the most brilliant scientists can’t navigate. At FutureHouse, our mission is to solve this problem by building an AI Scientist.”
Or as this tidy video from the company explains it:
"We're trying to take all the advances from AI and the resources put into AI and translate it to accelerated scientific discovery."
Or as FutureHouse chief Sam Rodriques said, in a post I can't decide is either boastful or wonky. "We are releasing three superhuman AI Scientist agents today, each with their own specialization.” (That’s everything but Phoenix.) It acknowledged that one is "an experimental project we put together recently just to demonstrate what can happen if you give the agents access to lots of scientific tools. It is not better than humans at planning experiments yet, and it makes a lot more mistakes than Crow, Falcon, or Owl. We want to see all the ways you can break it!" Tl; dr, AI can read, it can't necessarily teach. And it really can't write the curriculum.
Not for nothing has FutureHouse earned the approval of such AI experts like Jack Clark, the OpenAI veteran who co-founded Anthropic and is now one of the more respected authors on AI advancements. "Why this matters? For the AI revolution to truly pay out, it needs to change science," Clark wrote in his newsletter ImportAI this week. "AI has already massively changed and accelerated the work of computer programmers, but I think for AI to have a large effect in the world we need to apply it to science — the ultimate litmus test for the success of AI as a technology will be if it can either make research breakthroughs itself or provably massively accelerate scientists in their ability to make breakthroughs. FutureHouse is building software to help us see if this is the case." This is, indeed, the end of the innocence.
Some caveats. Science is exact, rigorous, demanding. AI...is not. That would seem to make the idea of an all-encompassing scientific being a little foolhardy to say the least; how much could we trust a machine to discover the next vaccine when it doesn't even know what to put on pizza. Not to mention that simply scouring and synthesizing scientific research is a long way from coming up with a quarkian hypothesis, which, as many scientists will tell you, is not (just) a matter of banging your head against the wall a million times but locating that elusive spark of creativity that seems to visit us humans and us humans alone; science may be more art than science. And even the most advanced microprocessor and reasoning-seeming machines may be unable to welcome the muse all humans are open and ready to receive.
We'll have scientists and folks who study the epistemology of science drop by down the road to discuss how realistic this all is. But a point that strikes me as notable is that at a time when traditional scientific research is under such assault in the U.S., this just might be our ticket out. The ability of humans to conduct experiments with the kind of resources and time horizons needed is being drastically reduced. So the universe has (maybe) given us a gift — a means of shortcutting ourselves around all the time-consuming research.
We’ll see what this yields. But the march of progress is slow and laborious. It's nice to think that, every once in a while, something can come along and make strides in a New York minute.
2. BACK IN 2021 I WROTE ABOUT A WHITNEY HOUSTON HOLOGRAM show that had come to Vegas. At the time it was a pretty cool, spooky thing — in fact come to think of it it's still a cool, spooky thing. A sophisticated projection that could move around, sing and interact with the audience had been mounted at a Sin City theater, and in the process given people a night they could never forget, as if it was the heyday of The Voice and we were jamming out to "How Will I Know" and "I Will Always Love You" like it was the early-90's and you were at middle-school prom all over again.
At the time such projections were thought of as primarily entertainment-use cases — fun, even nifty distractions, but little more than that. Increasingly in the past few years though, they've become different kinds of tools. In late 2022 Wal Mart mounted a Sam Walton hologram for visitors at its Bentonville museum, giving it a corporate promotion tint, while the next year a university president in Oman gave a conference speech as a hologram.
You see where this is going — these representations are getting a little less frilly and a little more serious. The latest shift is into education, via an Agatha Christie avatar that aims to teach you about writing mysteries. As the NY Times reported this week, the online-learning platform BBC Maestro has introduced a new course featuring the late legendary author offering instructions on how to plant just the right clue and how to make the murderer just credible enough.
The idea of wanting to resurrect the dead goes at least back to the early necromancy of the ancient Greeks and their dancing around fire pits in nocturnal hours; the modern era is really just continuing the pattern with more technology. In fact the tech here is actually not especially advanced — the idea of using AI expressions projected over an actor's body, as BBC Maestro does, was deployed by the Houston show nearly four years ago. Also, the avatar is reading a script, not interacting ChatGPT-style.
But if the tech is standard, the application is forward-looking. And provocative. And promising.
Because education might seem like a frivolous use case. But given how badly American kids are doing in so many academic departments, some new techniques might be needed. Reading scores on the benchmark National Assessment of Educational Progress fell for both elementary and middle-school students in 2024, dropping even below the levels they were at during the very dismal pandemic era.
The question is whether a Christie-like teaching device is a gimmick or a real educational tool. Is there reason to think that we learn better when taught this way, with lifelike avatars and holograms? The full studies have yet to come in, but it wouldn't be crazy to assume that seeing a beloved person in front of us, let alone someone who we can't otherwise see in real life, will up our level of interest and engagement.
Education has already been looking pretty hard for ways to lean on AI to improve matters in the classroom. Teachers have used ChatGPT to ease their burdens and make them more effective, while companies like TeachMateAI offer ways for AI to design lesson plans, write student reports and take on other tasks that divert from the, you know, teaching.
Very few of these programs seek to become teachers themselves, and understandably so. For one thing, the tech probably isn't good enough (a script-reading online-avatar is a far cry from a live, interacting teacher in the classroom). For another, the idea of computers teaching kids may not go over so well with either the UFT or the PTA. And the notion of students already bombarded with digital distractions now getting a fresh dose of one might feel a little like treating scurvy with another voyage on the Santa Maria.
And yet for all these challenges I can't but wonder if there's a glimmer of something here. We may not want kids’ minds to be so fragmented yet they are; we may wish teachers have the mental space to become the next Edward James Olmos yet here they are. Maybe a famous-AI teacher is not in every school, and it's certainly not for every kid. But when a crisis is this deep, measures have to be this desperate. The Times piece is dotted with all sorts of skeptical and defensive statements, but I'm not sure they're needed.
Technology used to be a scourge in the classroom; now it's often thought helpful. Interactive digital tools were long considered verboten by teachers; now games like Kahoot! are often believed a godsend. In fact a study last year by researchers from King Juan Carlos University in Madrid found all kinds of encouraging evidence for these digital tools as teaching aids. "Our students belong to a highly digitized generation with easy and rapid access to information. They are dependent on technology and tend to become bored quickly," the authors wrote. 'The use of interactive applications helps to maintain attention...the employ of several tools increases the student’s participation and improves their academic marks."
And really how different is this from the showing of Shakespearean movies or historical epics to hook students on the subject matter, as idea-thirsty teachers did when I and I suspect many of you were in school? It's a starting point.
The concept of getting AI resurrections to teach or re-enact might seem kind of crazy to us now, but I suspect it won't soon enough. These tools could be helpful. And no one needs to dance around any fire pits to employ them.
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. This year started on kind of a good note. But it’s been pretty rough since. This week? Arrow pointing up.
AI MIGHT BE ABLE TO BUILD A FULL-ON SCIENTIST: Ambitious but promising. +3.5
DEAD ICONS GET STUDENTS’ ATTENTION. Why not? +2.5