ChatGPT makes materials research much more efficient
Apr 2023, phys.org
"This isn't programming in the traditional sense; the method of interacting with these bots is through language," Morgan says. "Asking the program to extract data and then asking it to check if it is sure with normal sentences feels closer to how I train my children to get correct answers than how I usually train computers. It's such a different way to ask a computer to do things. It really changes how you think about what your computer can do."
via University of Wisconsin-Madison: Maciej P. Polak et al, Flexible, Model-Agnostic Method for Materials Data Extraction from Text Using General Purpose Language Models, arXiv (2023). DOI: 10.48550/arxiv.2302.04914
Also: Maciej P. Polak et al, Extracting Accurate Materials Data from Research Papers with Conversational Language Models and Prompt Engineering—Example of ChatGPT, arXiv (2023). DOI: 10.48550/arxiv.2303.05352
Image credit: In one of the first instances, phys.org uses a Stable Diffusion-generated image in the thumbnail, along with this statement: "This image was generated using Stable Diffusion, a text-to-image generator, using the prompt "researchers working with huge piles of data." -Dane Morgan and Maciej Polak, University of Wisconsin-Madison, 2023 [link]
New 'AI scientist' combines theory and data to discover scientific equations
Apr 2023, phys.org
The system rediscovered Kepler's third law of planetary motion, and produced a good approximation of Einstein's relativistic time-dilation law.The new AI scientist—dubbed "AI-Descartes" by the researchers—joins the likes of AI Feynman and other recently developed computing tools that aim to speed up scientific discovery. At the core of these systems is a concept called symbolic regression, which finds equations to fit data. Given basic operators, such as addition, multiplication, and division, the systems can generate hundreds to millions of candidate equations, searching for the ones that most accurately describe the relationships in the data.The system works particularly well on noisy, real-world data, which can trip up traditional symbolic regression programs that might overlook the real signal in an effort to find formulas that capture every errant zig and zag of the data. It also handles small data sets well, even finding reliable equations when fed as few as ten data points."In this work, we needed human experts to write down, in formal, computer-readable terms, what the axioms of the background theory are, and if the human missed any or got any of those wrong, the system won't work."
via IBM Research, Samsung AI, and University of Maryland Baltimore County: Combining Data and Theory for Derivable Scientific Discovery with AI-Descartes, Nature Communications (2023). DOI: 10.1038/s41467-023-37236-y
Researchers say AI emergent abilities are just a 'mirage'
Apr 2023, phys.org
"Previously claimed emergent abilities … might likely be a mirage induced by researcher analyses"Researchers contend that when results are reported in non-linear, or discontinuous, metrics, they appear to show sharp, unpredictable changes that are erroneously interpreted as indicators of emergent behavior, however an alternate means of measuring the identical data using linear metrics shows "smooth, continuous" changes that, contrary to the former measure, reveal predictable—non-emergent—behavior.
Large numbers getcha every time:
"The Stanford team added that failure to use large enough samples also contributes to faulty conclusions."
It's one of the easiest to spot when looking at the success of predictive powers, whether it's the weather, a sports bettor, or your financial advisor -- the law of large numbers makes us suck at identifying patterns. If you flip a perfect coin, there's a 50% chance of it landing on either heads or tails, which is what you would call a perfect chance, right down the middle. But if you flip the coin 10 times, you will probably not get 5 heads and 5 tails. You might have to flip it 100 times for that, or maybe even 10,000. And it depends on how many decimals you want to use, and if you start to get into millions and trillions of flips, then you'll have to cancel the variables in your system, like the weight of the respective sides of the coin, or the tendency for you hand to flip a certain way, or the prevailing winds, or patterns of seismic vibrations of the earth.
People who understand very well the law of large numbers, or more likely people who don't udnerstand it and don't want to -- can do a good job convincing others of seeing whatever patterns they want, just by manipulating the metrics, like looking at performance from January to January instead of September to September, or 18 months instead of 12, or:
"The main takeaway," the researchers said, "is for a fixed task and a fixed model family, the researcher can choose a metric to create an emergent ability or choose a metric to ablate an emergent ability."
via Stanford University: Rylan Schaeffer et al, Are Emergent Abilities of Large Language Models a Mirage?, arXiv (2023). DOI: 10.48550/arxiv.2304.15004
 |
| AI Art - A doctor use his stethoscope on a huge mechanical brain pink background 1 - 2023 |
Study finds source validation issues hurt ChatGPT reliability
May 2023, phys.org
Only about half of generated sentences were fully supported by citations, and one quarter of citations failed to support associated sentences.Moreover, the team found citation recall and precision were inversely correlated with fluency and perceived utility. "The responses that seem more helpful are often those with more unsupported statements or inaccurate citations," they observed.As a consequence, they concluded, "This facade of trustworthiness increases the potential for existing generative search engines to mislead users."
via Stanford University's Human-Centered AI research group: Nelson F. Liu et al, Evaluating Verifiability in Generative Search Engines, arXiv (2023). DOI: 10.48550/arxiv.2304.09848
Online consumers at risk from 'intelligent' price manipulation, say experts
May 2023, phys.org
"Widespread use of intelligent algorithmics and dynamic pricing by online retailers, puts the public at risk of 'adversarial collusion"More sophisticated algorithms can manipulate weaker algorithms and therefore collude together to increase prices for everyone, subtly undermine the competitiveness of online markets and harm consumers.
via University of Oxford: Luc Rocher, Adversarial competition and collusion in algorithmic markets, Nature Machine Intelligence (2023). DOI: 10.1038/s42256-023-00646-0
Ethical, legal issues raised by ChatGPT training literature
May 2023, phs.org
"Our work here has shown that OpenAI models know about books in proportion to their popularity on the web and the accuracy of such models is strongly dependent on the frequency with which a model has seen information in the training data."Few if any details about data used to train the models are known to the public.Also, science fiction and fantasy books dominate the list of memorized books, presenting a built-in bias on the nature of responses ChatGPT may provide.We should be thinking about whose narrative experiences are encoded in these models.
via University of California, Berkeley: Kent K. Chang et al, Speak, Memory: An Archaeology of Books Known to ChatGPT/GPT-4, arXiv (2023). DOI: 10.48550/arxiv.2305.00118
Post Script: I'm more interested in the simple statistical reality of algorithms trained on a completely "wild" dataset. The internet as a dataset is not curated, it's not designed, it's neither tamed nor maintained in any way; it is completely wild. When you apply the current state of the art in machine learning to a wild dataset, you multiply the wild part.
 |
| AI Art - A doctor use his stethoscope on a huge mechanical brain pink background 2 - 2023 |
AI: War crimes evidence erased by social media platforms
Jun 2023, BBC News
AI-powered church service in Germany draws a large crowd
Jun 2023, Ars Technica
New tool explains how AI 'sees' images and why it might mistake an astronaut for a shovel
Jun 2023, phys.org
CRAFT -- for Concept Recursive Activation FacTorization for Explainability -- was a joint project with the Artificial and Natural Intelligence Toulouse Institute.One of the concepts associated with the tench (a type of fish) is the face of a white male, because there are many photos online of white male sports fishermen holding fish that look like tench. In another example, the predominant concept associated with a soccer ball in neural networks is the presence of soccer players on the field.One way to explain AI vision is through what's called attribution methods, which employ heatmaps to identify the most influential regions of an image that impact AI decisions. However, these methods mainly focus on the most prominent regions of an image—revealing "where" the model looks, but failing to explain "what" the model sees in those areas.But with CRAFT we can see how the system is ranking the concepts.With the 'image of an astronaut was incorrectly classified as a shovel' problem, CRAFT showed that the neural network identified the concept of "dirt" commonly found in members of the image class "shovel" and the concept of "ski pants" typically worn by people clearing snow from their driveway with a shovel.
via Brown University's Carney Institute for Brain Science: Thomas Fel et al, CRAFT: Concept Recursive Activation FacTorization for Explainability (2023)
 |
| AI Art - A doctor use his stethoscope on a huge mechanical brain pink background 3 - 2023 |
Study says AI data contaminates vital human input
Jun 2023, phys.org
They dubbed this phenomenon "artificial artificial artificial intelligence."
(But you should know this is because of the already-in-use term for Mechanical Turks, dubbed "artificial artificial intelligence"; and although they used to provide human input are now relying on AI-generated content, thus the "artificial" hole of recursion.)
"It is tempting to rely on crowdsourcing to validate large language model outputs or to create human gold-standard data for comparison," Veselovsky said. "But what if crowd workers themselves are using LLMs … in order to increase their productivity, and thus their income, on crowdsourcing platforms?"Based on a limited study of the use of large language models by workers at MTurk, Amazon's crowd sourcing operation, the EPFL researchers estimated that 33% to 46% of worker assignments were completed with the aid of large language models.
via École polytechnique fédérale de Lausanne (EPFL), Lausanne, Switzerland: Veniamin Veselovsky et al, Artificial Artificial Artificial Intelligence: Crowd Workers Widely Use Large Language Models for Text Production Tasks, arXiv (2023). DOI: 10.48550/arxiv.2306.07899
Post Script: Figure this one out flesh engine of the future! "fringe benefits, french benefits, and friends with benefits" boy is that a good one.
Is it growing pains or is ChatGPT just becoming dumber?
Jul 2023, phys.org
"We don't fully understand what causes these changes in ChatGPT's responses because these models are opaque.""Any results on closed-source models are not reproducible and not verifiable, and therefore, from a scientific perspective, we are comparing raccoons and squirrels." -Sasha Luccioni of the AI company Hugging Face
via Stanford and UC Berkeley: Lingjiao Chen et al, How is ChatGPT's behavior changing over time?, arXiv (2023). DOI: 10.48550/arxiv.2307.09009
No comments:
Post a Comment