An artificial intelligence trained on personal data covering the entire population of Denmark can predict people’s chances of dying more accurately than any existing model, even those used in the insurance industry. The researchers behind the technology say it could also have a positive impact in early prediction of social and health problems – but must be kept out of the hands of big business.

Sune Lehmann Jørgensen at the Technical University of Denmark and his colleagues used a rich dataset from Denmark that covers education, visits to doctors and hospitals, any resulting diagnoses, income and occupation for 6 million people from 2008 to 2020.

They converted this dataset into words that could be used to train a large language model, the same technology that powers AI apps such as ChatGPT. These models work by looking at a series of words and determining which word is statistically most likely to come next, based on vast amounts of examples. In a similar way, the researchers’ Life2vec model can look at a series of life events that form a person’s history and determine what is most likely to happen next.

In experiments, Life2vec was trained on all but the last four years of the data, which was held back for testing. The researchers took data on a group of people aged 35 to 65, half of whom died between 2016 and 2020, and asked Life2vec to predict which who lived and who died. It was 11 per cent more accurate than any existing AI model or the actuarial life tables used to price life insurance policies in the finance industry.

The model was also able to predict the results of a personality test in a subset of the population more accurately than AI models trained specifically to do the job.

Jørgensen believes that the model has consumed enough data that it is likely to be able to shed light on a wide range of health and social topics. This means it could be used to predict health issues and catch them early, or by governments to reduce inequality. But he stresses that it could also be used by companies in a harmful way.

“Clearly, our model should not be used by an insurance company, because the whole idea of insurance is that, by sharing the lack of knowledge of who is going to be the unlucky person struck by some incident, or death, or losing your backpack, we can kind of share this this burden,” says Jørgensen.

But technologies like this are already out there, he says. “They’re likely being used on us already by big tech companies that have tonnes of data about us, and they’re using it to make predictions about us.”

Source: AI trained on millions of life stories can predict risk of early death | New Scientist

There’s something new and powerful for ChatGPT users to play around with: Custom GPTs. These bespoke bots are essentially more focused, more specific versions of the main ChatGPT model, enabling you to build something for a particular purpose without using any coding or advanced knowledge of artificial intelligence.

The name GPT stands for Generative Pre-trained Transformer, as it does in ChatGPT. Generative is the ability to produce new content outside of what an AI was trained on. Pre-trained indicates that it’s already been trained on a significant amount of material, and Transformer is a type of AI architecture adept at understanding language.

You might already be familiar with using prompts to style the responses of ChatGPT: You can tell it to answer using simple language, for example, or to talk to you as if it were an alien from another world. GPTs build on this idea, enabling you to create a bot with a specific personality.

What’s more, you can upload your own material to add to your GPT’s knowledge banks—it might be samples of your own writing, for instance, or copies of reports produced by your company. GPTs will always have access to the data you upload to them and be able to browse the web at large.

GPTs are exclusive to Plus and Enterprise users, though everyone should get access soon. OpenAI plans to open a GPT store where you can sell your AI bot creations if you think others will find them useful, too. Think of an app store of sorts but for bespoke AI bots.

“GPTs are a new way for anyone to create a tailored version of ChatGPT to be more helpful in their daily life, at specific tasks, at work, or at home—and then share that creation with others,” explains OpenAI in a blog post. “For example, GPTs can help you learn the rules to any board game, help teach your kids math, or design stickers.”

Getting started with GPT building

Assuming you have a Plus or Enterprise account, click Explore on the left of the web interface to see some example GPTs: There’s one to help you with your creative writing, for example, and one to produce a particular style of digital painting. When you’re ready to start building your own, click Create a GPT at the top.

There are two tabs to swap between: Create for building a GPT through a question-and-answer routine and Configure for more deliberate GPT production. If you’re just getting started, it’s best to stick with Create, as it’s a more user-friendly option and takes you step-by-step through the process.

Respond to the prompts of the GPT Builder bot to explain what you want the new GPT to be able to do: Explain certain concepts, give advice in specific areas, generate particular kinds of text or images, or whatever it is. You’ll be asked to give the GPT a name and choose an image for it, though you’ll get suggestions for these, too.

As you answer the prompts from the builder, the GPT will begin to take form in the preview pane on the right—together with some example inputs that you might want to give to it. You might be asked about specific areas of expertise that you want the bot to have and the sorts of answers you want the bot to give in terms of their length and complexity. The building process will vary though, depending on the GPT you’re creating.

After you’ve worked through the basics of making a GPT, you can try it out and switch to the Configure tab to add more detail and depth. You’ll see that your responses so far have been used to craft a set of instructions for the GPT about its identity and how it should answer your questions. Some conversation starters will also be provided.

You can edit these instructions if you need to and click Upload files to add to the GPT’s knowledge banks (handy if you want it to answer questions about particular documents or topics, for instance). Most common document formats, including PDFs and Word files, seem to be supported, though there’s no official list of supported file types.

The checkboxes at the bottom of the Configure tab let you choose whether or not the GPT has access to web browsing, DALL-E image creation, and code interpretation capabilities, so make your choices accordingly. If you add any of these capabilities, they’ll be called upon as and when needed—there’s no need to specifically ask for them to be used, though you can if you want.

When your GPT is working the way you want it to, click the Save button in the top right corner. You can choose to keep it to yourself or make it available to share with others. After you click on Confirm, you’ll be able to access the new GPT from the left-hand navigation pane in the ChatGPT interface on the web.

GPTs are ideal if you find yourself often asking ChatGPT to complete tasks in the same way or cover the same topics—whether that’s market research or recipe ideas. The GPTs you create are available whenever you need them, alongside access to the main ChatGPT engine, which you can continue to tweak and customize as needed.

Source: How To Build Your Own Custom ChatGPT Bot

We are now releasing Phi-2 (opens in new tab), a 2.7 billion-parameter language model that demonstrates outstanding reasoning and language understanding capabilities, showcasing state-of-the-art performance among base language models with less than 13 billion parameters. On complex benchmarks Phi-2 matches or outperforms models up to 25x larger, thanks to new innovations in model scaling and training data curation.

With its compact size, Phi-2 is an ideal playground for researchers, including for exploration around mechanistic interpretability, safety improvements, or fine-tuning experimentation on a variety of tasks. We have made Phi-2 (opens in new tab) available in the Azure AI Studio model catalog to foster research and development on language models.

[..]

Phi-2 is a Transformer-based model with a next-word prediction objective, trained on 1.4T tokens from multiple passes on a mixture of Synthetic and Web datasets for NLP and coding. The training for Phi-2 took 14 days on 96 A100 GPUs. Phi-2 is a base model that has not undergone alignment through reinforcement learning from human feedback (RLHF), nor has it been instruct fine-tuned. Despite this, we observed better behavior with respect to toxicity and bias compared to existing open-source models that went through alignment (see Figure 3). This is in line with what we saw in Phi-1.5 due to our tailored data curation technique, see our previous tech report (opens in new tab) for more details on this. For more information about the Phi-2 model, please visit Azure AI | Machine Learning Studio (opens in new tab).

Source: Phi-2: The surprising power of small language models – Microsoft Research

Shortly after ChatGPT’s release, a cadre of critics rose to fame claiming AI would soon kill us. As wondrous as a computer speaking in natural language might be, it could use that intelligence to level the planet. The thinking went mainstream via letters calling for research pauses and “60 Minutes” interviews amplifying existential concerns. Leaders like Barack Obama publicly worried about AI autonomously hacking the financial system — or worse. And last week, President Biden issued an executive order imposing some restraints on AI development.

AI Experts Dismiss Doom, Defend Progress

That was enough for several prominent AI researchers who finally started pushing back hard after watching the so-called AI Doomers influence the narrative and, therefore, the field’s future. Andrew Ng, the soft-spoken co-founder of Google Brain, said last week that worries of AI destruction had led to a “massively, colossally dumb idea” of requiring licenses for AI work. Yann LeCun, a machine-learning pioneer, eviscerated research-pause letter writer Max Tegmark, accusing him of risking “catastrophe” by potentially impeding AI progress and exploiting “preposterous” concerns. A new paper earlier this month indicated large language models can’t do much beyond their training, making the doom talk seem overblown. “If ‘emergence’ merely unlocks capabilities represented in pre-training data,” said Princeton professor Arvind Narayanan, “the gravy train will run out soon.”

 

 

Related Article: Can We Fix Artificial Intelligence’s Serious PR Problem?

AI Doom Hype Benefits Tech Giants

Worrying about AI safety isn’t wrongheaded, but these Doomers’ path to prominence has insiders raising eyebrows. They may have come to their conclusions in good faith, but companies with plenty to gain by amplifying Doomer worries have been instrumental in elevating them. Leaders from OpenAI, Google DeepMind and Anthropic, for instance, signed a statement putting AI extinction risk on the same plane as nuclear war and pandemics. Perhaps they’re not consciously attempting to block competition, but they can’t be that upset it might be a byproduct.

AI Alarmism Spurs Restrictive Government Policies

Because all this alarmism makes politicians feel compelled to do something, leading to proposals for strict government oversight that could restrict AI development outside a few firms. Intense government involvement in AI research would help big companies, which have compliance departments built for these purposes. But it could be devastating for smaller AI startups and open-source developers who don’t have the same luxury.

 

Doomer Rhetoric: Big Tech’s Unlikely Ally

“There’s a possibility that AI doomers could be unintentionally aiding big tech firms,” Garry Tan, CEO of startup accelerator Y Combinator, told me. “By pushing for heavy regulation based on fear, they give ammunition to those attempting to create a regulatory environment that only the biggest players can afford to navigate, thus cementing their position in the market.”

Ng took it a step further. “There are definitely large tech companies that would rather not have to try to compete with open source [AI], so they’re creating fear of AI leading to human extinction,” he told the Australian Financial Review.

Doomers’ AI Fears Lack Substance

The AI Doomers’ worries, meanwhile, feel pretty thin. “I expect an actually smarter and uncaring entity will figure out strategies and technologies that can kill us quickly and reliably — and then kill us,” Eliezer Yudkowsky, co-founder of the Machine Intelligence Research Institute, told a rapt audience at TED this year. He confessed he didn’t know how or why an AI would do it. “It could kill us because it doesn’t want us making other superintelligences to compete with it,” he offered.

Bankman Fried Scandal Should Ignite Skepticism

After Sam Bankman Fried ran off with billions while professing to save the world through “effective altruism,” it’s high time to regard those claiming to improve society while furthering their business aims with relentless skepticism. As the Doomer narrative presses on, it threatens to rhyme with a familiar pattern.

AI Fear Tactics Threaten Open-Source Movement

Big Tech companies already have a significant lead in the AI race via cloud computing services that they lease out to preferred startups in exchange for equity. Further advantaging them might hamstring the promising open-source AI movement — a crucial area of competition — to the point of obsolescence. That’s probably why you’re hearing so much about AI destroying the world. And why it should be considered with a healthy degree of caution.

Source: AI Doomsayers: Debunking the Despair

IBM and Meta Launch the AI Alliance in collaboration with over 50 Founding Members and Collaborators globally including AMD, Anyscale, CERN, Cerebras, Cleveland Clinic, Cornell University, Dartmouth, Dell Technologies, EPFL, ETH, Hugging Face, Imperial College London, Intel, INSAIT, Linux Foundation, MLCommons, MOC Alliance operated by Boston University and Harvard University, NASA, NSF, Oracle, Partnership on AI, Red Hat, Roadzen, ServiceNow, Sony Group, Stability AI, University of California Berkeley, University of Illinois, University of Notre Dame, The University of Tokyo, Yale University and others

[…]

While there are many individual companies, start-ups, researchers, governments, and others who are committed to open science and open technologies and want to participate in the new wave of AI innovation, more collaboration and information sharing will help the community innovate faster and more inclusively, and identify specific risks and mitigate those risks before putting a product into the world.

[..]

We are:

• The creators of the tooling driving AI benchmarking, trust and validation metrics and best practices, and application creation such as MLPerf, Hugging Face, LangChain, LlamaIndex, and open-source AI toolkits for explainability

, privacy, adversarial robustness, and fairness evaluation

• .
• The universities and science agencies that educate and support generation after generation of AI scientists and engineers and push the frontiers of AI research through open science.
• The builders of the hardware and infrastructure that supports AI training and applications – from the needed GPUs to custom AI accelerators and cloud platforms;
• The champions of frameworks that drive platform software including PyTorch, Transformers, Diffusers, Kubernetes, Ray, Hugging Face Text generation inference      and Parameter Efficient Fine Tuning.
• The creators of some of today’s most used open models including Llama2, Stable Diffusion, StarCoder, Bloom, and many others.

[…]

To learn more about the Alliance, visit here: https://thealliance.ai

[…]

Source: AI Alliance Launches as an International Community of Leading Technology Developers, Researchers, and Adopters Collaborating Together to Advance Open, Safe, Responsible AI

We will see – I don’t see any project pages on this quite yet. But this looks like a reasonable idea.

Alexandre Pouget at the University of Geneva, Switzerland, and his colleagues used machine learning to analyse the chemical composition of 80 red wines from 12 years between 1990 and 2007. All the wines came from seven wine estates in the Bordeaux region of France.

“We were interested in finding out whether there is a chemical signature that is specific to each of those chateaux that’s independent of vintage,” says Pouget, meaning one estate’s wines would have a very similar chemical profile, and therefore taste, year after year.

To do this, Pouget and his colleagues used a machine to vaporise each wine and separate it into its chemical components. This technique gave them a readout for each wine, called a chromatogram, with about 30,000 points representing different chemical compounds.

The researchers used 73 of the chromatograms to train a machine learning algorithm, along with data on the chateaux of origin and the year. Then they tested the algorithm on the seven chromatograms that had been held back.

They repeated the process 50 times, changing the wines used each time. The algorithm correctly guessed the chateau of origin 100 per cent of the time. “Not that many people in the world will be able to do this,” says Pouget. It was also about 50 per cent accurate at guessing the year when the wine was made.

The algorithm could even guess the estate when it was trained using just 5 per cent of each chromatogram, using portions where there are no notable peaks in chemicals visible to the naked eye, says Pouget.

This shows that a wine’s unique taste and feel in the mouth doesn’t depend on a handful of key molecules, but rather on the overall concentration of many, many molecules, says Pouget.

By plotting the chromatogram data, the algorithm could also separate the wines into groups that were more like each other. It grouped those on the right bank of the river Garonne – called Pomerol and St-Emilion wines – separately from those from left-bank estates, known as Medoc wines.

The work is further evidence that local geography, climate, microbes and wine-making practices, together known as the terroir, do give a unique flavour to a wine. Which precise chemicals are behind each wine wasn’t looked at in this study, however.

“It really is coming close to proof that the place of growing and making really does have a chemical signal for individual wines or chateaux,” says Barry Smith at the University of London’s School of Advanced Study. “The chemicals compounds and their similarities and differences reflect that elusive concept of terroir.”

 

Journal reference:

Communications Chemistry DOI: 10.1038/s42004-023-01051-9

Source: AI can tell which chateau Bordeaux wines come from with 100% accuracy | New Scientist

In a paper released on arXiv last week, a team of researchers from Hugging Face and Carnegie Mellon University calculated the amount of power AI systems use when asked to perform different tasks.

After asking AIs to perform 1,000 inferences for each task, the researchers found text-based AI tasks are more energy-efficient than jobs involving images.

Text generation consumed 0.042kWh while image generation required 1.35kWh. The boffins assert that charging a smartphone requires 0.012kWh – making image generation a very power-hungry application.

“The least efficient image generation model uses as much energy as 950 smartphone charges (11.49kWh), or nearly one charge per image generation,” the authors wrote, noting the “large variation between image generation models, depending on the size of image that they generate.”

The authors also measured carbon dioxide created by different AI workloads. As depicted in the graphic below, image creation topped that chart

Click to enlarge

You can read the full paper here [PDF].

Source: One AI image needs as much power as a smartphone charge • The Register

Would you trust Elon Musk with your mortgage? Or Big Tech with your benefits?

Us neither.

That’s what’s at stake as the EU’s Artificial Intelligence Act reaches the final stage of negotiations. For all its big talk, it seems like the EU is buckling to Big Tech.

EU lawmakers have been tasked with developing the world’s first comprehensive law to regulate AI products. Now that AI systems are already being used in public life, lawmakers are rushing to catch up.

[…]

The principle of precaution urges us to exercise care and responsibility in the face of potential risks. It is crucial not only to foster innovation but also to prevent the unchecked expansion of AI from jeopardising justice and fundamental rights.

At the Left in the European Parliament, we called for this principle to be applied to the AI Act. Unfortunately, other political groups disagreed, prioritising the interests of Big Tech over those of the people. They settled on a three-tiered approach to risk whereby products are categorised into those that do not pose a significant risk, those that are high risk and those that are banned.

However, this approach contains a major loophole that risks undermining the entire legislation.

Like asking a tobacco company whether smoking is risky

When it was first proposed, the Commission outlined a list of ‘high-risk uses’ of AI, including AI systems used to select students, assess consumers’ creditworthiness, evaluate job-seekers, and determine who can access welfare benefits.

Using AI in these assessments has significant real-life consequences. It can mean the difference between being accepted or rejected to university, being able to take out a loan or even being able to access welfare to pay bills, rent or put food on the table.

Under the three-tiered approach, AI developers are allowed to decide themselves whether their product is high-risk. The self-assessment loophole means the developers themselves get to determine whether their systems are high risk akin to a tobacco company deciding cigarettes are safe for our health, or a fossil fuel company saying its fumes don’t harm the environment.

[…]

Experience shows us that when corporations have this kind of freedom, they prioritise their profits over the interests of people and the planet. If the development of AI is to be accountable and transparent, negotiators must eliminate provisions on self-assessment.

AI gives us the opportunity to change our lives for the better. But as long as we let big corporations make the rules, we will continue to replicate inequalities that are already ravaging our societies.

Source: A(I) deal at any cost: Will the EU buckle to Big Tech? – EURACTIV.com

OK, so this seems to be a little breathless – surely we can put in a mechanism for EU checking of risk level when notified of a potential breech, including harsh penalties for misclassifying an AI?

However, the discussions around the EU AI Act – which had the potential to be one of the first and best pieces of regulation on the planet – has now descended into farce since ChatGPT and some strange idea that the original act did not have any provisions for General Purpose / Foundational AI models (it did – they were high risk models). The silly induced discussions this has provoked has only served to delay the AI act coming into force for over a year – something that big businesses are very very happy to see.

[…]

In a paper in Proceedings of the National Academy of Sciences, a group of researchers led by Dion Häfner, a computer scientist at the University of Copenhagen, describe a clever way to make AI more understandable. They have managed to build a neural network, use it to solve a tricky problem, and then capture its insights in a relatively simple five-part equation that human scientists can use and understand.

The researchers were investigating “rogue waves”, those that are much bigger than expected given the sea conditions in which they form. Maritime lore is full of walls of water suddenly swallowing ships. But it took until 1995 for scientists to measure such a wave—a 26-metre monster, amid other waves averaging 12 metres—off the coast of Norway, proving these tales to be tall only in the literal sense.

[…]

To produce something a human could follow, the researchers restricted their neural network to around a dozen inputs, each based on ocean-wave maths that scientists had already worked out. Knowing the physical meaning of each input meant the researchers could trace their paths through the network, helping them work out what the computer was up to.

The researchers trained 24 neural networks, each combining the inputs in different ways. They then chose the one that was the most consistent at making accurate predictions in a variety of circumstances, which turned out to rely on only five of the dozen inputs.

To generate a human-comprehensible equation, the researchers used a method inspired by natural selection in biology. They told a separate algorithm to come up with a slew of different equations using those five variables, with the aim of matching the neural network’s output as closely as possible. The best equations were mixed and combined, and the process was repeated. The result, eventually, was an equation that was simple and almost as accurate as the neural network. Both predicted rogue waves better than existing models.

The first part of the equation rediscovered a bit of existing theory: it is an approximation of a well-known equation in wave dynamics. Other parts included some terms that the researchers suspected might be involved in rogue-wave formation but are not in standard models. There were some puzzlers, too: the final bit of the equation includes a term that is inversely proportional to how spread out the energy of the waves is. Current human theories include a second variable that the machine did not replicate. One explanation is that the network was not trained on a wide enough selection of examples. Another is that the machine is right, and the second variable is not actually necessary.

Better methods for predicting rogue waves are certainly useful: some can sink even the biggest ships. But the real prize is the visibility that Dr Häfner’s approach offers into what the neural network was doing. That could give scientists ideas for tweaking their own theories—and should make it easier to know whether to trust the computer’s predictions.

Source: A new way to predict ship-killing rogue waves

Stable Diffusion’s generative art can now be animated, developer Stability AI announced. The company has released a new product called Stable Video Diffusion into a research preview, allowing users to create video from a single image. “This state-of-the-art generative AI video model represents a significant step in our journey toward creating models for everyone of every type,” the company wrote.

The new tool has been released in the form of two image-to-video models, each capable of generating 14 to 25 frames long at speeds between 3 and 30 frames per second at 576 × 1024 resolution.

[…]

Stable Video Diffusion is available only for research purposes at this point, not real-world or commercial applications. Potential users can sign up to get on a waitlist for access to an “upcoming web experience featuring a text-to-video interface,” Stability AI wrote. The tool will showcase potential applications in sectors including advertising, education, entertainment and more.

[…]

it has some limitations, the company wrote: it generates relatively short video (less than 4 seconds), lacks perfect photorealism, can’t do camera motion except slow pans, has no text control, can’t generate legible text and may not generate people and faces properly.

The tool was trained on a dataset of millions of videos and then fine-tuned on a smaller set, with Stability AI only saying that it used video that was publicly available for research purposes.

[…]

Source: The AI startup behind Stable Diffusion is now testing generative video

Just a few weeks ago Judge William Orrick massively trimmed back the first big lawsuit that was filed against generative AI companies for training their works on copyright-covered materials. Most of the case was dismissed, and what bits remained may not last much longer. And now, it appears that Judge Vince Chhabria (who has been very good on past copyright cases) seems poised to do the same.

This is the high profile case brought by Sarah Silverman and some other authors, because some of the training materials used by OpenAI and Meta included their works. As we noted at the time, that doesn’t make it copyright infringing, and it appears the judge recognizes the large hill Silverman and the other authors have to climb here:

U.S. District Judge Vince Chhabria said at a hearing that he would grant Meta’s motion to dismiss the authors’ allegations that text generated by Llama infringes their copyrights. Chhabria also indicated that he would give the authors permission to amend most of their claims.

Meta has not yet challenged the authors’ central claim in the case that it violated their rights by using their books as part of the data used to train Llama.

“I understand your core theory,” Chhabria told attorneys for the authors. “Your remaining theories of liability I don’t understand even a little bit.”

Chhabria (who you may recall from the time he quashed the ridiculous copyright subpoena that tried to abuse copyright law to expose whoever exposed a billionaire’s mistress) seems rightly skeptical that just because ChatGPT can give you a summary of Silverman’s book that it’s somehow infringing:

“When I make a query of Llama, I’m not asking for a copy of Sarah Silverman’s book – I’m not even asking for an excerpt,” Chhabria said.

The authors also argued that Llama itself is an infringing work. Chhabria said the theory “would have to mean that if you put the Llama language model next to Sarah Silverman’s book, you would say they’re similar.”

“That makes my head explode when I try to understand that,” Chhabria said.

It’s good to see careful judges like Chhabria and Orrick getting into the details here. Of course, with so many of these lawsuits being filed, I’m still worried that some judge is going to make a mess of things, but we’ll see what happens.

Source: Sarah Silverman’s AI Case Isn’t Going Very Well Either | Techdirt

Global medium-range weather forecasting is critical to decision-making across many social and economic domains. Traditional numerical weather prediction uses increased compute resources to improve forecast accuracy, but does not directly use historical weather data to improve the underlying model. Here, we introduce “GraphCast,” a machine learning-based method trained directly from reanalysis data. It predicts hundreds of weather variables, over 10 days at 0.25° resolution globally, in under one minute. GraphCast significantly outperforms the most accurate operational deterministic systems on 90% of 1380 verification targets, and its forecasts support better severe event prediction, including tropical cyclones tracking, atmospheric rivers, and extreme temperatures. GraphCast is a key advance in accurate and efficient weather forecasting, and helps realize the promise of machine learning for modeling complex dynamical systems.

[…]

The dominant approach for weather forecasting today is “numerical weather prediction” (NWP), which involves solving the governing equations of weather using supercomputers.

[…]

NWP methods are improved by highly trained experts innovating better models, algorithms, and approximations, which can be a time-consuming and costly process.

Machine learning-based weather prediction (MLWP) offers an alternative to traditional NWP, where forecast models can be trained from historical data, including observations and analysis data.

[…]

In medium-range weather forecasting, i.e., predicting atmospheric variables up to 10 days ahead, NWP-based systems like the IFS are still most accurate. The top deterministic operational system in the world is ECMWF’s High RESolution forecast (HRES), a configuration of IFS which produces global 10-day forecasts at 0.1° latitude/longitude resolution, in around an hour

[…]

Here we introduce an MLWP approach for global medium-range weather forecasting called “GraphCast,” which produces an accurate 10-day forecast in under a minute on a single Google Cloud TPU v4 device, and supports applications including predicting tropical cyclone tracks, atmospheric rivers, and extreme temperatures.

[…]

A single weather state is represented by a 0.25° latitude/longitude grid

[…]

GraphCast is implemented as a neural network architecture, based on GNNs in an “encode-process-decode” configuration (13, 17), with a total of 36.7 million parameters (code, weights and demos can be found at https://github.com/deepmind/graphcast).

[…]

During model development, we used 39 years (1979–2017) of historical data from ECMWF’s ERA5 (21) reanalysis archive.

[…]

Of the 227 variable and level combinations predicted by GraphCast at each grid point, we evaluated its skill versus HRES on 69 of them, corresponding to the 13 levels of WeatherBench (8) and variables (23) from the ECMWF Scorecard (24)

[…]

We find that GraphCast has greater weather forecasting skill than HRES when evaluated on 10-day forecasts at a horizontal resolution of 0.25° for latitude/longitude and at 13 vertical levels.

[NOTE HRES has a resolution of 0.1°]

[…]

We also compared GraphCast’s performance to the top competing ML-based weather model, Pangu-Weather (16), and found GraphCast outperformed it on 99.2% of the 252 targets they presented (see supplementary materials section 6 for details).

[…]

GraphCast’s forecast skill and efficiency compared to HRES shows MLWP methods are now competitive with traditional weather forecasting methods

[…]

With 36.7 million parameters, GraphCast is a relatively small model by modern ML standards, chosen to keep the memory footprint tractable. And while HRES is released on 0.1° resolution, 137 levels, and up to 1 hour time steps, GraphCast operated on 0.25° latitude-longitude resolution, 37 vertical levels, and 6 hour time steps, because of the ERA5 training data’s native 0.25° resolution, and engineering challenges in fitting higher resolution data on hardware.

[…]

Our approach should not be regarded as a replacement for traditional weather forecasting methods, which have been developed for decades, rigorously tested in many real-world contexts, and offer many features we have not yet explored. Rather our work should be interpreted as evidence that MLWP is able to meet the challenges of real-world forecasting problems and has potential to complement and improve the current best methods.

[…]

Source: Learning skillful medium-range global weather forecasting | Science