Latest news with #GeoffreyvonMaltzahn
Forbes
24-04-2025
- Science
- Forbes
A Discovery Every Day: What Does Superintelligence Actually Look Like?
As we start heading toward a certain critical mass with artificial intelligence, this word keeps coming up – superintelligence. It's easy to throw the word around, and talk about that point when AI becomes smarter than humans, but what does superintelligence actually look like? To find out, I asked a panel of builders and physicists, and we talked about what types of efforts are happening now to support superintelligent results. These are some of the main themes. One of the biggest ideas in today's tech world is the idea of reinventing the scientific process itself. Geoffrey von Maltzahn is part of the team at Flagship Pioneering, where a project called Lila Sciences is looking to create 'autonomous science systems' that will, as he says, allow AI to take over 'every step of the wheel' when it comes to scientific discovery. '(It's) the ability to call upon tools, to model the way that the world works, to propose a brilliant hypothesis, to autonomously design a decisive experiment, (and) to test that hypothesis in the real world,' he clarified. Geoffrey von Maltzahn, CEO, Lila John Werner Making the analogy from vibe coding to outsourcing, von Maltzahn talked about how even though human level intelligence supports everything, it's possible to bring together autonomous results for aspects like material science, chemistry, and life sciences. That, he suggested, will really have a positive effect, partly because of inherent human limitations in scientific inquiry. 'Neither our bodies, nor our brains, are really optimally suited for science,' he said, 'particularly learning about how the world works from the atoms make the same case for math. You know, try as our brains have … in the reality in life science, in chemistry, materials and more, our brains really struggle to understand what is going on … but machines are much better at matching those patterns … the implications for every single technology domain that we're familiar with are really, really amazing.' Panel on Superintelligence John Werner Here's another major part of what superintelligence is likely to be able to do: it will excel at math, even in more intuitive, abstract ways. Carina Hong, CEO of Axiom, a quantitative superintelligence moonshot company, talked about how pattern matching is not reasoning, and how traditional models don't excel in showing their work. 'Large language models, despite all the amazing post training breakthroughs, are still pretty bad at doing proofs,' she said. 'They will give you a numerical answer. In fact, they can do it really well on the American Invitational math examination. Frontier large language models achieve a 96% score. However, when you ask (the model) to show its proof, the score drops to 5%, so why is different? It's because of the way we train them … what we want to build at Axiom is to use programming language to train the machine to be able to speak the language of formal proof.' This, she says, will enable humans to trust the result of these engines, and make the world 'math-rich'. Another aspect of this is the setup. Riccardo Sabbatini is a numerical modeling specialist who works on drug discovery and more. Setting the stage for full robotic automation, he talked about a system where millions of molecular experiments can happen with no human involvement whatsoever. 'I see a transition moment between now and super intelligence,' he said, calling the interim a time of 'vibe intelligence.' 'When you look at a coder today,' he said, 'instead of going and searching in Stack Overflow every three seconds, and having to copy and paste from (one's) own old code, you have open on the right side of your screen, an LLM: this is going to do boiler plating for you. It's going to like 80% of the boring coding that has been done in the past.' You can watch the video for some additional scientific assessment of things like probabilistic database design, Gaussian curves, and the evolution of AI math. One anecdote from the panel is where Sabbatini talked about image creation models always displaying watches with the same time setting – 10 minutes after 10 o'clock. It's persistent, he said, based on the training set that the LLM gets off of the Internet. 'None of (the generated watches) will show 2:25pm,' he said of an experiment where a user asks for an image of a watch set at this time in the afternoon. 'They will always show 10:10; the reason is that the majority, if not the complete, set, of pictures of watches in the entire world, points at 10:10.' It's an advertising thing, he suggested, based on how people like to see watch faces. 'So any watch in the world has to be at 10:10, stuck there,' he said. 'You can have a pretty analog watch. You can have a classic analog watch - but you can't have a '2:25' analog watch. It is bizarre, if you think about it, such a simple concept learning thing.' That illustrates some of AI's current blind spots that the panel suggested might be solved with superintelligence, eventually. But what von Maltzahn said about the pace of scientific discovery was extremely interesting. With these new tools, he reasoned, we'll be able to speed up science as a human in endeavor: where a ground truth used to take about a year to develop, AI will free us of those time constraints. What if you could have a breakthrough scientific discovery every day? And do it easily? 'The human brain understands a really, really small fraction of how the world works,' von Maltzahn explained. 'And in fact, to understand it, we've been dividing it into sub, sub, sub specialties. So I believe something like imagination and taste for novelty is going to hang around as a human contribution for a while.' 'I believe science is going to get way more fun,' von Maltzahn said. 'If you just take the Edisonian 1% inspiration, 99% perspiration, you know, we can put 99% perspiration into a new paradigm … (improving) the quality of life for scientists, and likely the quantity of output.' Panel on Superintelligence John Werner Talking about being able to source rare earth metals in new ways, and perfect a new system of chemistry that's going to change our supply chains and our scientific methods, he suggested that the 'GDP of civilization' is resting on a brand new paradigm. We've had any number of technological revolutions in the past, he argued, but this is the first intelligence revolution. What's going to happen? One such outcome, posited by von Maltzahn as he discussed changes, is that none of our human intellectual contributions to projects will be safe, if AI can do it better. 'None of us really knows in what order the sea level of intelligence will rise and subsume imagination or … logical derivation,' he said. 'But there's probably a rough boundary where, if searching for information within the repository of what is known, then that is underwater now, or will be underwater virtually immediately.' That brings me back to the eternal specter of job displacement, and the question of how we're going to re-order society around these technologies. We seem to have a vague idea that a re-ordering is needed, but not much clarity on what people are going to be doing for jobs in a business world that's dominated by capable AI. In any case, we can anticipate the likelihood of this new era of science, and everything that is going to bring us. This is something every young person should be thinking about as they study and prepare for a career – and something every public planner (or innovator, or entrepreneur) should be thinking about as they try to understand where we're going next.
New York Times
10-03-2025
- Science
- New York Times
The Quest for A.I. ‘Scientific Superintelligence'
Across the spectrum of uses for artificial intelligence, one stands out. The big, inspiring A.I. opportunity on the horizon, experts agree, lies in accelerating and transforming scientific discovery and development. Fed by vast troves of scientific data, A.I. promises to generate new drugs to combat disease, new agriculture to feed the world's population and new materials to unlock green energy — all in a tiny fraction of the time of traditional research. Technology companies like Microsoft and Google are making A.I. tools for science and collaborating with partners in fields like drug discovery. And the Nobel Prize in Chemistry last year went to scientists using A.I. to predict and create proteins. This month, Lila Sciences went public with its own ambitions to revolutionize science through A.I. The start-up, which is based in Cambridge, Mass., had worked in secret for two years 'to build scientific superintelligence to solve humankind's greatest challenges.' Relying on an experienced team of scientists and $200 million in initial funding, Lila has been developing an A.I. program trained on published and experimental data, as well as the scientific process and reasoning. The start-up then lets that A.I. software run experiments in automated, physical labs with a few scientists to assist. Already, in projects demonstrating the technology, Lila's A.I. has generated novel antibodies to fight disease and developed new materials for capturing carbon from the atmosphere. Lila turned those experiments into physical results in its lab within months, a process that most likely would take years with conventional research. Experiments like Lila's have convinced many scientists that A.I. will soon make the hypothesis-experiment-test cycle faster than ever before. In some cases, A.I. could even exceed the human imagination with inventions, turbocharging progress. 'A.I. will power the next revolution of this most valuable thing humans ever stumbled across — the scientific method,' said Geoffrey von Maltzahn, Lila's chief executive, who has a Ph.D. in biomedical engineering and medical physics from the Massachusetts Institute of Technology. The push to reinvent the scientific discovery process builds on the power of generative A.I., which burst into public awareness with the introduction of OpenAI's ChatGPT just over two years ago. The new technology is trained on data across the internet and can answer questions, write reports and compose email with humanlike fluency. The new breed of A.I. set off a commercial arms race and seemingly limitless spending by tech companies including OpenAI, Microsoft and Google. (The New York Times has sued OpenAI and Microsoft, which formed a partnership, accusing them of copyright infringement regarding news content related to A.I. systems. OpenAI and Microsoft have denied those claims.) Lila has taken a science-focused approach to training its generative A.I., feeding it research papers, documented experiments and data from its fast-growing life science and materials science lab. That, the Lila team believes, will give the technology both depth in science and wide-ranging abilities, mirroring the way chatbots can write poetry and computer code. Still, Lila and any company working to crack 'scientific superintelligence' will face major challenges, scientists say. While A.I. is already revolutionizing certain fields, including drug discovery, it's unclear whether the technology is just a powerful tool or on a path to matching or surpassing all human abilities. Since Lila has been operating in secret, outside scientists have not been able to evaluate its work and, they add, early progress in science does not guarantee success, as unforeseen obstacles often surface later. 'More power to them, if they can do it,' said David Baker, a biochemist and director of the Institute for Protein Design at the University of Washington. 'It seems beyond anything I'm familiar with in scientific discovery.' Dr. Baker, who shared the Nobel Prize in Chemistry last year, said he viewed A.I. more as a tool. Lila was conceived inside Flagship Pioneering, an investor in and prolific creator of biotechnology companies, including the Covid-19 vaccine maker Moderna. Flagship conducts scientific research, focusing on where breakthroughs are likely within a few years and could prove commercially valuable, said Noubar Afeyan, Flagship's founder. 'So not only do we care about the idea, we care about the timeliness of the idea,' Dr. Afeyan said. Lila resulted from the merger of two early A.I. company projects at Flagship, one focused on new materials and the other on biology. The two groups were trying to solve similar problems and recruit the same people, so they combined forces, said Molly Gibson, a computational biologist and a Lila co-founder. The Lila team has completed five projects to demonstrate the abilities of its A.I., a powerful version of one of a growing number of sophisticated assistants known as agents. In each case, scientists — who typically had no specialty in the subject matter — typed in a request for what they wanted the A.I. program to accomplish. After refining the request, the scientists, working with A.I. as a partner, ran experiments and tested the results — again and again, steadily homing in on the desired target. One of those projects found a new catalyst for green hydrogen production, which involves using electricity to split water into hydrogen and oxygen. The A.I. was instructed that the catalyst had to be abundant or easy to produce, unlike iridium, the current commercial standard. With A.I.'s help, the two scientists found a novel catalyst in four months — a process that more typically might take years. That success helped persuade John Gregoire, a prominent researcher in new materials for clean energy, to leave the California Institute of Technology last year to join Lila as head of physical sciences research. George Church, a Harvard geneticist known for his pioneering research in genome sequencing and DNA synthesis who has co-founded dozens of companies, also joined recently as Lila's chief scientist. 'I think science is a really good topic for A.I.,' Dr. Church said. Science is focused on specific fields of knowledge, where truth and accuracy can be tested and measured, he added. That makes A.I. in science less prone to the errant and erroneous answers, known as hallucinations, sometimes created by chatbots. The early projects are still a long way from market-ready products. Lila will now work with partners to commercialize the ideas emerging from its lab. Lila is expanding its lab space in a six-floor Flagship building in Cambridge, alongside the Charles River. Over the next two years, Lila says, it plans to move into a separate building, add tens of thousands of square feet of lab space and open offices in San Francisco and London. On a recent day, trays carrying 96 wells of DNA samples rode on magnetic tracks, shifting directions quickly for delivery to different lab stations, depending partly on what the A.I. suggested. The technology appeared to improvise as it executed experimental steps in pursuit of novel proteins, gene editors or metabolic pathways. In another part of the lab, scientists monitored high-tech machines used to create, measure and analyze custom nanoparticles of new materials. The activity on the lab floor was guided by a collaboration of white-coated scientists, automated equipment and unseen software. Every measurement, every experiment, every incremental success and failure was captured digitally and fed into Lila's A.I. So it continuously learns, gets smarter and does more on its own. 'Our goal is really to give A.I. access to run the scientific method — to come up with new ideas and actually go into the lab and test those ideas,' Dr. Gibson said.
