The answer is that it's obviously regressive, but that its training data is recent enough that there isn't much difference with the test data... yet. I'm also very skeptical of this lol. The...
The answer is that it's obviously regressive, but that its training data is recent enough that there isn't much difference with the test data... yet.
I'm also very skeptical of this lol. The headline seems worded specifically to scare people as much as possible, but this is just guessing how long you'll live based on various statistical factors. A normal doctor or statistician working in medicine would probably be able to do that if you asked them. And even the researchers basically say it's overfit on its training data in as many words.
The article describes researchers who trained an AI model on statistical data from the entire country of Denmark. The model has made predictions that are accurate enough that they are growing...
The article describes researchers who trained an AI model on statistical data from the entire country of Denmark. The model has made predictions that are accurate enough that they are growing concerned about the future policy impact of such systems. Their article in Nature is here:
Not a ton to add, but I'd love to know what their data inputs are. I'm guessing it's a combination of demographic, financial, health, genetic, and social information but I'd really like to know...
Not a ton to add, but I'd love to know what their data inputs are. I'm guessing it's a combination of demographic, financial, health, genetic, and social information but I'd really like to know what databases they created or harvested.
I don't have access to Nature, any students of faculty with a subscription care to elaborate? If so could you please post the methods section?
I don't have access either, but it says in the article that they made use of a database that the government of Denmark keeps on their citizens.
I don't have access either, but it says in the article that they made use of a database that the government of Denmark keeps on their citizens.
Built using transformer models, which power large language models (LLMs) like ChatGPT, the new tool, life2vec, is trained on a data set pulled from the entire population of Denmark — 6 million people. The data set was made available to the researchers only by the Danish government.
At the heart of life2vec is the massive data set that the researchers used to train their model. The data is held by Statistics Denmark, the central authority on Danish statistics, and, although tightly regulated, can be accessed by some members of the public, including researchers. The reason it’s so tightly controlled is it includes a detailed registry of every Danish citizen.
Traditionally, neural networks suffer from overfitting on tabular data, and the data they used to train sure sounds like tabular data. I'm really doubtful their model is any better in actuality...
Traditionally, neural networks suffer from overfitting on tabular data, and the data they used to train sure sounds like tabular data. I'm really doubtful their model is any better in actuality than boosted trees, which tends to be what wins in kaggle competitions for tabular data.
This sounds like they just got 100% training accuracy because their parameter space was larger than the input space. It's not like any statistical model is doing magic - it can't project what's not there. It is simply not possible to predict human lifespan with any kind of accuracy from the data the government has about you.
The answer is that it's obviously regressive, but that its training data is recent enough that there isn't much difference with the test data... yet.
I'm also very skeptical of this lol. The headline seems worded specifically to scare people as much as possible, but this is just guessing how long you'll live based on various statistical factors. A normal doctor or statistician working in medicine would probably be able to do that if you asked them. And even the researchers basically say it's overfit on its training data in as many words.
I'm quite confident this model will be just as skilled as my human doctors at telling me to lose weight whenever I have any medical problem 🥲
And definitely won’t conflate the difficulty some folks find in accessing medical care from things like this with other health outcomes, I’m sure 🙄
The article describes researchers who trained an AI model on statistical data from the entire country of Denmark. The model has made predictions that are accurate enough that they are growing concerned about the future policy impact of such systems. Their article in Nature is here:
https://www.nature.com/articles/s43588-023-00586-0
Not a ton to add, but I'd love to know what their data inputs are. I'm guessing it's a combination of demographic, financial, health, genetic, and social information but I'd really like to know what databases they created or harvested.
I don't have access to Nature, any students of faculty with a subscription care to elaborate? If so could you please post the methods section?
I don't have access either, but it says in the article that they made use of a database that the government of Denmark keeps on their citizens.
Not sure how many (if any) changes were made for the Nature version, but the preprint is here: https://arxiv.org/abs/2306.03009
Traditionally, neural networks suffer from overfitting on tabular data, and the data they used to train sure sounds like tabular data. I'm really doubtful their model is any better in actuality than boosted trees, which tends to be what wins in kaggle competitions for tabular data.
This sounds like they just got 100% training accuracy because their parameter space was larger than the input space. It's not like any statistical model is doing magic - it can't project what's not there. It is simply not possible to predict human lifespan with any kind of accuracy from the data the government has about you.
"Ai can do extremely exploitable thing. Scientist hope it isn't exploited."
Every. AI. Article. Ever.
This how we end up with rehoboam from West World. All it will need is the accompanying app, RICO.