This is Science Friday. I'm Ira Flato. And I'm Flora Lichtman. Over the past week the Trump administration has ordered changes at our biggest federal science agencies including the Centers for Disease Control and the National Institutes of Health. The news is coming fast and furious and it can feel hard to keep up.
Our next guest is here to help. Anil Oza is a reporter for Stat and MIT based in Boston, Massachusetts. Anil has been keeping close tabs on what's happening at the agencies and is here to explain. Anil, welcome to Science Friday. Hi, Flora. Good to be here. Okay, so late last week, we started getting news that data on the CDC's website was disappearing. First of all, what disappeared? Yeah, that's right. So...
Late last week on Friday, we were seeing the CDC scrub a lot of the data sets from their website. And so normally anyone can go to the CDC's website and download data sets for a whole slew of different health issues. But to sort of comply with some of Trump's executive orders, by 5 p.m. on Friday, they were trying to scrub a lot of these data sets to comply with two of these executive orders that Trump put in place. One, to sort of remove DEI, and so that would remove mentions of...
race and ethnicity, and another to sort of deal with, I believe, what he calls gender ideology. And so that would remove mentions of sexual orientation, of gender identity. And so one of the data sets was the Youth Behavior Survey, which measures gender and sexual orientation in teenagers. And a lot of data sets, even if they're not sort of expressly about racial inequities, they include data about people's race.
sort of a foul of this executive order. I also saw that there were data sets about things like HIV and sexual health and tuberculosis and reported just yesterday bird flu. What are the consequences of this information disappearing? Yeah, so I think it's going to be hard to see the sort of short-term impacts of this, but I mean...
In the long term, these data are sort of about understanding where the need is the greatest in the U.S. for these different diseases like bird flu or like tuberculosis or other STIs and things like that. And so if you see these data, be offline. long time and future data not collected it would be much harder to sort of know where we need to dedicate resources and time and attention
throughout the country. And I mean, I think it's particularly shocking because normally the sort of US federal agencies are the sort of gold standard data, the sort of envy of the entire world, which is why I think the fact that they were being... tampered with, taken down was such a shock to researchers and journalists. I think I read some of the data sets have reappeared. Have you seen that? Yeah, that's right. So sort of since the outcry about...
these data sets being removed, some of them have been put up. And even sort of journalists and researchers, when they heard that these data sets were about to be taken down, there was sort of a mad dash to sort of...
download these data sets and to sort of archive them so that they were there in case they weren't put back up online. But I think this raises an issue, obviously, about the sort of veracity of the data, that if they were taken down, they may have been tampered with. And so I know that some of these researchers that
downloaded the data are now sort of cross comparing their archival data with what's been put up to see what if anything has changed. This week, the CDC also ordered withdrawals of certain new scientific papers from its researchers. Can you tell me about this? Yeah, that's right. Also, late last week on Friday, the CDC ordered its researchers to withdraw its papers. These are papers that have not been made public but are going through the scientific review process.
have some of these terms that we've been talking about, about diversity, equity, and inclusion, or about gender. The CDC was asking scientists to sort of withdraw papers that they had collaborated on with these terms included in them. It may seem like a small issue, but often any paper that mentions health and wellness will include these data just as a matter of knowing what the sort of makeup of the patient population is.
And I mean, this must be incredibly frustrating, both to the scientists at the CDC and the NIH, but also the researchers that collaborated with them, because notoriously, scientific publishing is a very slow-moving process. very possible that there are papers that have been in the works for months, if not years, that may have to sort of get pulled at the finish line here.
Let's shift gears to the National Institutes of Health. There's been some ups and downs there in the last few weeks with, you know, communication ban and grant making operations paused. What's been happening and what's going on now? Yeah, absolutely. So I think one of the really shocking things that rattled the scientific community when Trump first took office was this communication ban from the federal agencies.
Researchers at these federal institutions couldn't collaborate with people that they've collaborated one of the things that was really shocking and abrupt was they canceled study sections at the NIH, which is these groups that are formed to review grants for applications to sort of fund research. And so these groups of professors and researchers will meet and sort of determine whether or not research is worth funding and they completely just stop those kind of on a dime and so
My colleague Megan Moulteni and I this week reported that some of them are restarting. The first study section to meet in over three weeks was on Tuesday, which initially seemed like a good sign. But I think there are still concerns about whether or not the Trump administration.
is sort of tampering with these processes. So not all study sections have been restarted. Some researchers have said that some have been rescheduled. And other sort of advisory meetings for the NIH have been paused to sort of with some of Trump's other executive orders. So things are not sort of completely as normal at the NIH, but they are slowly starting to come back online. Anil, thanks for your reporting on this. Of course, it was a pleasure to be here.
Anil Oza is a reporter for STAT and MIT based in Boston. Next up, we are leaving the Beltway for a look at other science news of the week. As you know, we have been tracking the bird flu situation closely here at Science Friday. This week, we got news that dairy cows in Nevada have been infected with a different version of H5N1 bird flu, a form that may put people who work closely with cows at higher risk. Here to give us an update on this and other science stories.
from the week is Kathleen Davis, Sci-Fi producer based in New York. Welcome, Kathleen. Hi, Flora. Okay, what do we know about this form of bird flu? So on Wednesday, the U.S. Department of Agriculture announced that As you said, a new form of the bird flu virus was found in Nevada dairy cows. This is still H5N1 bird flu, but a different form. And this is called D1.1. And this is actually the predominant form of bird flu that is found in migrating birds.
But the thing with this form is that it wasn't previously known to be able to spread to cows. So the infection that we have seen in dairy cows up until now was another form of bird flu called B3.13. I mean, we know dozens of dairy workers have been infected with B3.13, and they've had relatively mild symptoms. Have any people been infected with this form, D1.1?
Yes, and that is a big part of the concern. So we know that this form of bird flu has killed one person. That was a Louisiana resident over 65 years old. This person died back in January. Also, a 13-year-old Canadian girl was infected by this virus back in November. She had a pre-existing condition and was on life support, but she did recover. So there are new concerns for dairy workers and people who work with cows.
And the more that this virus moves around and replicates, the more chances there are for mutations that might make it easier for it to spread to and among humans. How's the government responding to this development? Yeah, so the USDA is testing milk for the bird flu virus.
This is how the strain was detected in the first place. This testing is done before milk is pasteurized. So we know that the virus can be present in unpasteurized raw milk. But if you're worried about whether or not drinking milk is safe... If you're drinking good old pasteurized milk, which is what most of us do drink, this heats the milk for a short amount of time and that virus is inactivated. So you shouldn't have anything to worry about.
Let's move on to some good news about AI detecting breast cancer. Tell me about this. So AI is everywhere now, obviously. But one of the things that years ago scientists were talking about and were really excited about were the prospects of AI in the medical field. So this is doing things like detecting cancer. x-rays. The idea here is
For example, you can feed an AI model a ton of images of what breast cancer looks like in a mammogram. The AI then knows what breast cancer looks like, and in theory, it can pick out cancer more accurately than a person could. It's also a lot faster than what one radiologist can do. So now this has been put into practice. We recently got the results of the largest randomized trial for this.
More than 100,000 women were part of the study, where in some cases, AI took a look at mammograms, and in other cases, a human radiologist did. And the group that used AI had a 29% higher detection of cancer with no false positives. The AI beat the doctors. It did. That's amazing. It is. And a lot of experts are looking at the future of this as like...
AI and radiologists work together to detect breast cancer. And, you know, for breast cancer in particular, this can be huge. Women here in the U.S. have a 13 percent average risk of getting breast cancer sometime in their life. And so better early detection. could be a lifesaver. You've got one more story for us featuring a charismatic mega fossil, one of our favorites, Lucy.
Yes, Lucy is one of the most famous fossils. She was a human ancestor, an Australopithecus afarensis. She lived about 3.2 million years ago. And we do know a lot about Lucy because her skeleton is pretty complete. We have about 40 percent of it. And now researchers have put her on a virtual treadmill to answer the age old question.
How the heck would Lucy run? And how would she run? Well, these researchers reconstructed what they think her leg and pelvic muscles would look like. Obviously, those aren't present in the fossils that we do have of Lucy, but they did their best estimate. And they found that Lucy standing upright was about three and a half feet tall. She weighed somewhere between 29 and 93 pounds, which is obviously a huge spread.
And she was capable of standing and walking upright. So Lucy does have more ape-like characteristics than we do. her upper body is proportionately larger than ours, for example. She's got longer arms and shorter legs. So if she did have to run, it was probably just for short bursts. The generous explanation from the researchers is that Lucy was not a natural runner. They always do these reconstructions. Did they reconstruct her running style and what did it look like?
Oh, they sure did. I mean, running's a pretty generous term. She's kind of got both legs at like 90 degree angles. She's got this forward lean. They say that she had a max speed of 11 miles per hour. She doesn't have that springy running gait that modern humans have. So, I mean, no offense to Lucy, but...
I think you and I, Flora, could really beat her pretty easily on the track if we wanted to. Speak for yourself, Kathleen. When you were describing max speed of 11 miles per hour, I was like, that's me. That's my running style. It's all relative, Flora. That's all the time we have. Kathleen, thanks so much. Thank you. Kathleen Davis, sci-fi producer based in New York.
After the break, a deep dive into the AI model from DeepSeek. It's really, really good. I think the thing that has got people really shocked is that it is as good as the best that the U.S. has made. Stay with us. Support for Science Friday comes from the Alfred P. Sloan Foundation, working to enhance public understanding of science, technology, and economics in the modern world.
Last week, if you recall, we briefly talked about new advances in AI, especially this offering from a Chinese company called DeepSeek, which supposedly needs a lot less computing power to run. than many of the other AI models on the market, and it costs lots less money to use. It's been described as so revolutionary that I really wanted to take a deeper dive into DeepSake. What's all the buzz? makes one model smarter than another, less power hungry.
Joining me to help dive into that is Will Douglas Heaven, Senior Editor for AI Coverage at MIT Technology Review. He's based in the UK. Welcome back to the program, Will. Hi. Thanks a lot for having me. Happy to have you. Is it really as good as people are saying? Yeah. The thing is, I think it's really, really good. I think the thing that has got people really...
is that it is as good as the best that the US has made. We're in a stage now where, you know, the margins between the best new models are pretty slim. You know, is OpenAI's best better than... Google's best? Is that better than Anthropic's best? Just the fact that a Chinese company has matched what the best US labs can do is itself a shocking thing. I don't think people thought that China had caught up.
So what is its competitive advantage here? From what I've been reading, it seems that DeepSeek computer geeks figured out a much simpler way to program the less powerful. cheaper NVIDIA chips that the U.S. government allowed to be exported to China, basically. They've done a lot of interesting things. There's also a lot of things that aren't quite clear. So we don't know exactly what computer chips.
DeepSeq has. And it's also unclear how much of this work they did before the export controls kicked in. But from the several papers that they've released, and the very cool thing about them is that they are sharing all their information, which we're not seeing from the US companies. It looks like they have squeezed a lot more juice out of the...
Nvidia chips they do have. They've done some very clever engineering work to sort of reprogram them down at very low levels to kind of get more power out of the box than Nvidia gives you.
by default so that's one cool thing they've done these are also sort of um got innovative techniques in you know how they gather data to train the models but one key thing in their approach is they've found ways to sidestep the use of human data labelers, which, you know, if you think about how you have to build one of these large language models, the first stage is you basically scrape as much.
information you can from the internet and millions of books etc and you know we're probably familiar with that part of the story stealing other people's data Yeah, pretty much. And as an aside, you've got to laugh when OpenAI is upset. It's claiming now that DeepSeek maybe sold some of the output from its models. I mean, the Schadenfreude is sweet.
But all you get from training a large language model on the internet is a model that's really good at mimicking internet documents. It's not something that's very useful. The chatbots that we've come to know... where you can ask them questions and make them do all sorts of different tasks. To make them do those things, you need to do this extra layer of training. And that's typically being done by getting a lot of people to come up with ideal.
question-answer scenarios and training the model to sort of act more like that. So you need a lot of people involved, is basically what you're saying. Yeah, exactly. That's time-consuming and costly. found a way to do without that. Probably the coolest trick that DeepSeek used is this thing called reinforcement learning, which essentially an AI model sort of learned by trial and error. Listeners might recall DeepMind.
Back in 2016, they built this board game playing AI called AlphaGo. And so the amazing thing that they showed was if you get an AI to start just trying things at random and then... If it gets it slightly right, you nudge it more in that direction. If you do that many, many, many, many times, then you end up incrementally getting better and better and better. So you ended up in DeepMind's case with an AI that could...
starting from scratch, went on to beat a human grandmaster at Go. What Deep Seek has done is applied that technique to language models. You know, they didn't want to play a game. Obviously, they wanted it to... get better at giving thought through answers to questions that you asked the language model. And again, to start off with, it did a pretty poor job, but they nudged it bit by bit in the right direction.
And you let that run enough times and it sort of figures out itself how to get better, sort of improving bit by bit as it goes. So what you're basically saying is that it's teaching itself how to get better. Yeah, I hesitate sort of phrases like that because it always gives the AI some sense of agency and it's, you know, going to do its own thing. Yeah.
There is a term called self-play. It sort of learns to play itself and get better as it goes. So you can think of it in that way. You know, aside from the human involvement, one of the problems with AI, as we know, is that the computer computers use a tremendous amount of energy, even more than crypto mining, which is shockingly high. I mean, is DeepSeek less energy hungry then for all its advantages across the board?
Yet again, this is something that we've heard a lot about in the last week or so. And the answer to that as well is not as clear as it was initially made out. It does seem that they trained the model. They built the model using... less energy and more cheaply. Running it may be cheaper as well. But the thing is, with the latest type of model that they've built, they're known as some chain of thought models.
rather than if you're familiar with using something like chat gpt and you ask it a question and it pretty much gives the the first response it it comes up with back at you but there's a brand new sort of paradigm in chat box now where You ask it a question and it sort of takes its time and steps through, sort of shows its answers, shows its reasoning as it steps through its response. And each one of those steps is like a whole separate call to the language model.
Although DeepSeek's new model R1 may be more efficient, the fact that it is one of these sort of chain of thought reasoning models. may end up using more energy than the vanilla type of language models we've actually seen. And another complicating factor is that now they've shown everybody how they did it and essentially given away the model for free.
I think we can expect so many other companies and startups and research groups sort of picking it up and rolling their own based on group six techniques. One of the criticisms... of AIs that sometimes it's going to make up the answers if it doesn't know it, right? Right. There are two layers here. One, how does it stack up on reliability or this issue, as they call it, hallucinations? And second, because it's...
It's a Chinese model. Is there censorship going on here? Yeah. So a lot of stuff happening there as well. All models hallucinate and they will continue to do so as long as they're sort of built in this way. You know, there's statistical slot machines. You can. polish them up as much as you like but you're still going to have the chance that it'll make stuff up and i've seen examples that deep six model actually isn't great uh in in this respect um if it's
can't answer a question, it will still have a go at answering it and give you a bunch of nonsense. Anecdotally, based on a bunch of examples of people posting online having played around with it, it looks like it makes some howlers. But yeah, the question of censorship is interesting. I mean, I guess it's not surprising at all that a model built in China, it can't tell you anything about Tiananmen Square. It won't answer questions about Chinese politics.
at all. It just says, you know, I'm sorry, let's talk about something else. If they're innovating like this, but making their code available in open source, as you say, are we likely to see the other competitors saying we're going to use this? Because why not? Yeah, I mean, you can download the DeepSig app from the App Store or Google Play and have a go with it right now. For free?
For free? Yeah, I've been playing with it. I mean, I quite like it. In many ways, it's more friendly than ChatGPT or Google's Gemini. But there are also lots and lots of companies that sort of offer services that sort of... provide a wrapper to all these different chatbots that are now on the market and you sort of just, you go to those companies and you can pick and choose whichever one you want.
Within days of it being released, many of these companies were offering DeepSeek as one of the alternatives. Of course, not just companies offering DeepSeek's model as is to people, but... Because it's open source, you can adapt it. So there's a company called Huggy Face that sort of reverse engineered it and made their own version called Open R1. There's also a technique called distillation where you can take a really powerful language model.
and sort of use it to teach a smaller, less powerful one, but give it most of the abilities that the better one has. DeepSeq's one has already been distilled into a bunch of different models. Yeah, so I think we're going to see adaptations of it and people copying it for some time to come. So what's your take on...
Artificial general intelligence. We've talked about this before on the show. Will we be getting there and when, in your opinion? I am super skeptical about everything about AGI. You are. Partly is just a term that means very little. It means different things to different people who use it. I think it's become a marketing term more than anything else. I mean, I roll my eyes when people like Sam Altman. tell us that AGI is coming. He was telling us that two or three years ago.
And when I spoke to him then, he'd say, you know, the reason OpenAI is releasing these models is to show people what's possible. This society needs to know what's coming and there's going to be such a big societal adjustment.
to this new technology that we all need to sort of educate ourselves and get prepared. Now he's talking about AGI still coming, but he means something, you know, I don't know, like a sort of a workplace productivity tool that we're all going to use. And that's now what he means.
And I'm picking soundbar that has the example here, but like most of the big tech CEOs all write blog posts talking about, you know, this is what they're building. I think AGI has been this term that essentially means, you know. AI, but better than what we have today. The definition that's most usually used is, you know, an AI that can match humans on a wide range of cognitive tasks. But how would you really test that? And how would you know when we've got that?
And it's not clear at all that we'll get there on the current path, even with these large language models. Maybe they'll plateau soon. Maybe they'll just be very, very good language mimics. And, you know, we'll stop there. And would there have to be a whole nother...
breakthrough and a different type of AI technology to take us further. But I just, the AGI is my least favorite term. Well, Will, I want to thank you for taking us really into the weeds on this. I hope we still have a few listeners left. We appreciate how deeply we've taken a dive here, but I really enjoyed it. Thank you for taking time to be with us today. Thanks a lot. Will Douglas Heaven, Senior Editor for AI Coverage at MIT Technology Review.
Before we go, next Friday is Valentine's Day, and we want to hear from you. We are looking for your geekiest love stories. Maybe your meet-cute bubbled up in the lab. Yeah, maybe you dropped your beaker and as you bent to pick it up, you locked eyes on the grad student across the table who beat you to it.
Or you found your life partner traipsing through leech-infested rainforests during a scientific field expedition. Yeah, or stuck on a line of code in Starbucks and the latte next to you showed you a shortcut. And you thought of how really smart your kids are going to be. Was it love at first science? Did you fall for someone across the fume hood or get butterflies over that bacterial culture? Call us and tell us about it.
Do we have a phone number we're going to give them? Oh yeah, maybe that would be a good idea. Okay, that number. is 646-767-6532. Grab a pencil, 646-767-6532. Or... If you want to stay digital, send us a voice memo. That's a, you know, you'll record a little file and you'll upload that to SciFry at ScienceFriday.com. And that is about all we have time for. Lots of folks helped make this show happen, including... I'm Flora Lichtman. Thanks for listening.