Anthropic, which was accused of illegally training its A.I. on books, just settled the major class action brought by those books’ authors, rather than go to trial in December. We’ll have to wait until next week to learn the details, but as my colleague Eriq Gardner notes, “The deal flips the math behind class certification.” Previously, Eriq explains, “Anthropic wanted the smallest class possible to limit exposure. Now? The bigger the better as broad coverage reduces the risk of future suits from other authors.” Watch this space…

For some time now, IBM has been working on creating a “hybrid cloud”—something that would combine the benefits of classical computing (which encompasses A.I.) with quantum computing and supercomputing. The company took a step toward that ambitious goal today by announcing a partnership with AMD, the chip manufacturer and Nvidia competitor, to pursue the development of next-gen computing architecture that could power a hybrid, quantum-centric computer.

The plan, according to a press release, is to develop “scalable, open-source platforms that could redefine the future of computing,” to “explore solutions to complex problems… including in fields such as drug discovery, materials discovery, optimization, and logistics.” Buzzword soup, maybe, but shares of both companies rose on the news.

One thousand days ago, a still little-known startup called OpenAI quietly released a web app called ChatGPT. You know what happened next: It quickly became one of the fastest-growing apps in history, launched a multitrillion-dollar industry, and catalyzed a relentless global hype cycle of investment, pearl-clutching, excitement, yada yada. Several weeks ago, in the lead-up to the anniversary, a reader asked me how the generative technology that ChatGPT popularized has tangibly contributed to humanity. So I made a few calls to researchers and scientists to get their perspectives.

For what it’s worth, generative algorithms keep getting more advanced, enabling faster, cheaper, and sometimes better weather prediction, drug discovery, and medical diagnostics. Similarly, they are paving the way to enhanced and empowered conservation research and bringing us closer to the early detection of wildfires, among other life-changing developments. This is all great, and advancements in these areas are objectively a good thing.

But I think it’s clear now, and will continue to become more clear, that the impact of chatbots—and the industry’s push for their rapid and uncritical adoption—is hardly an unalloyed good. It is perhaps worth noting that today, another lawsuit was filed, this time against OpenAI, alleging that ChatGPT “validated, encouraged and assisted” a 16-year-old California boy in the planning and taking of his own life earlier this year.

Anyway, here’s a smattering of the most interesting responses I received—lightly edited, of course.

And now for the main event…

If you spent the past week watching Hurricane Erin whipping up huge waves along the East Coast before mercifully heading out to sea, you probably noticed how frequently forecasters mentioned the various models used to predict the storm’s path. Historically, this has been the domain of the National Hurricane Center (N.H.C.)—a division of the National Oceanic and Atmospheric Administration (NOAA), which has been targeted by DOGE and the Trump administration for potentially crippling budget cuts. (Fortunately, Congress has since pushed back.) The N.H.C. uses dozens of models to track storms and alert the public to danger. These dynamic models, which are fed on a vast buffet of data gathered from satellites, weather balloons, buoys, radar, airplane and ship sensors, etcetera, and are powered by banks of supercomputers, have long been the indispensable tool at the heart of weather forecasting.

And yet, despite how far numerical weather prediction (N.W.P.) has come, the process tends to be slow, enormously costly, and remarkably energy-intensive. There are also reliability issues: Atmospheric conditions change rapidly, which means models need to be updated regularly with real-time data to maintain accuracy. And that’s where new A.I. methods have entered the picture.

As with all things A.I., there are a few kinks to work out. But industry experts are encouraged that these models are much, much faster—and therefore cheaper—than traditional methods. For example, training a machine-learning model essentially “front-loads the main computational expense into the training,” explained Dr. Istvan Szunyogh, a professor of atmospheric sciences at Texas A&M. “In some sense, a weather forecast is like a 3D movie, and a [machine-learning] model is good at guessing the continuation of the movie based on the last few frames.” As a result, Szunyogh said, A.I. forecasting can be “orders of magnitude faster than a comparable-quality N.W.P. model.”

The big difference between A.I. models and physics-based forecasts, though, is that A.I. doesn’t actually use math to make predictions. Instead, these models rely on massive swaths of historical data to make probabilistic predictions based on new inputs. This has its pros and, of course, its cons. Dr. Ruoying He, the director of the Ocean Observing and Modeling Group at North Carolina State University, told me that these A.I. forecasts are now “mostly being tested in research settings and as supplements, helping refine physics-based forecasts, filling data gaps, or downscaling to local detail.”

As powerful as these new A.I. models have become, the numerical approach probably won’t ever go away. A.I.-based methods, for instance, struggle with rare, extreme weather events, especially in cases where there’s not much training data available. “A.I. is powerful at pattern recognition and bias correction,” He said. “Rather than one replacing the other, I believe the future is hybrid: A.I.-powered, physics-based models that combine strengths of both to deliver more accurate, timely forecasts.”

Szunyogh agreed that hybrid systems—which have been the focus of his research for the past seven years—are the likely denouement, but added the caveat that “the purely [machine-learning]-based models have surprised us before, so I do not find it unimaginable that we may have highly capable, purely M.L.-based models in the very near future.” But Szunyogh also explained that A.I. models are unlikely to dramatically improve our current weather-prediction capabilities, given that state-of-the-art N.W.P. models are already close to hitting what he called the “predictability limit”—i.e., the most accurate a forecasting model can be given the “chaotic nature of atmospheric dynamics.”

Szunyogh does, however, expect A.I. models to transform the enterprise of weather prediction. After all, machine learning forecasting models are much more attractive to corporations because they cost less to run and don’t require as much expertise as physics-based models to construct. As a result, Szunyogh expects the private sector to take on a much more significant role in weather prediction than it has historically.

Google DeepMind, for instance, recently launched Weather Lab, a publicly accessible website that features a handful of different A.I. weather-prediction models, with one based on stochastic neural networks specifically designed for hurricanes. (You can check out Google’s A.I. forecasts of Hurricane Erin here.) That model, according to internal tests run by Google, is about on par with—and in some aspects, more capable than—the dynamic models that are still most common today.

In July, Google announced a partnership with the N.H.C. to provide its experimental models for hurricane forecasting. But so far, the dynamic models aren’t out of a job. An N.H.C. spokesperson told me on Tuesday that the team is currently testing Google’s models to assess their efficacy, but confirmed that the Erin forecasts we’ve seen have come from the center’s preexisting N.W.P. models. The spokesperson explained that it’s still too early to make a thorough assessment of the A.I. models’ performance, though the team plans to evaluate the models after hurricane season wraps up. (DeepMind did not respond to a request for comment.)

In any case, as weather prediction agencies around the world begin to incorporate A.I. models into their forecasting process, Szunyogh observed that “A.I./M.L.-based models could not exist without all that has been learned from [N.W.P.] models and the data that they have produced. It is not clear whether a fully autonomous system can be built in which M.L. itself produces the training data without any physics-based knowledge.” When that day comes, we’ll let you know.

That’s all for today. I’ll see you Thursday.

Ian

Need help? Review our FAQ page or contact us for assistance. For brand partnerships, email ads@puck.news.

You received this email because you signed up to receive emails from Puck, or as part of your Puck account associated with {{customer.email}}. To stop receiving this newsletter and/or manage all your email preferences, click here.

Puck is published by Heat Media LLC. 107 Greenwich St, New York, NY 10006