natemurthy

There's been a lot talk of time series foundation models in the last year, but how well do they in fact perform on real-world operational tasks where decision makers face actual stakes?

Researchers led by Tao Han at The Hong Kong University of Science and Technology collaborated with AI/ML weather scientists in both China and Australia to answer this in a rare, one-of-its-kind study comparing the ECMWF-HRES Numerical Weather Prediction (NWP) model against a panel of nearly a dozen time series foundation models (TSFMs), and found that, while recent TSFMs perform well on benchmarks, they still lag behind operational NWP systems in capturing complex weather dynamics and extreme events.

In other words, modeling the physics of these meteorological phenomena outperforms simply generating confidence-bounded trend lines from statistical covariates and correlations.

Their work was accepted at Int'l Conference on Machine Learning (ICML) this July 2026.

Takeaway: TSFMs are still useful for quick back-of-the-envelope planning estimation at low cost, they require just a fraction of the inference compute of running NWP models; but true accuracy and skill comes from modeling the underlying causal dynamics of complex systems — especially when lives and livelihood are at risk.

An update to https://news.ycombinator.com/item?id=49279890 (originally published Aug 2026) on the state of world models since then. Lots of progress!

It's so strange how much skepticism technologists on HN have of our court system. Take a look at the evidence already presented and read the complaint sheets that are publicly available (4:23-cv-05448-YGR - People of the State of California v. Meta Platforms, Inc. et al). The state AGs have a very strong body of evidence. If I were betting person, I would expect that Meta loses this case from the arguments I've seen presented. The reason why they're in court in the first place is because the AGs have provable claims.

P.s. I'm literally typing this from inside the courthouse in Oakland, public audio live stream is available from the court website

Pretty damning remarks from the California AG about Meta: the company knew their products were harmful, they ignored the harms, and they intentionally misled and lied to the public.

There's been such a massive Cambrian explosion of world models in development that I decided to create a survey of all the world models I've encountered to show just how active this space has become.

From the models that train Waymo vehicles and industrial robotic systems, weather forecasting models, and Nobel-prize winning protein folding classifiers (arguably), to even contested generative video models — I've grouped these into roughly a dozen model families and made an infographic to illustrate the activity in this research and engineering domain over the last 8 years or so

By no means is this list comprehensive, and I'm sure I'll discover many new things over time that are not in this diagram. So I'll do my best to keep this regularly updated. If anyone is looking for more detailed notes on these families and lineages, and follow any updates I make to this survey, I created a GitHub repo for public study and examination of world models in applied AI research and engineering.

I hope folks find this useful!

Remember that time when Apple's Global Head of Security bribed the Santa Clara County Sheriff's office with tens of thousands of dollars worth of iPads to expedite concealed hand-gun permits for its company agents? Welp, federal appeals court put that criminal case back on the docket. Paywall removed, plain text available: https://pastebin.com/Fzgs5t2X

Hi everyone, so I'm the guy who wrote this article. I'm sharing a video recording of George Hotz writing some Python code. Note his reaction at 5:55

https://www.youtube.com/watch?v=7Hlb8YX2-W8&t=355s

His outburst is an emotional manifestation of the different sources of frustration I've had working with Python that I've tried to document.

A common theme I've seen in responses is that many claim that I've mistaken Python as a weakly-typed language countering that it is indeed a strongly-typed language -- obviously an argument of semantics. I'll define a "strongly-typed" language as one where it is trivial to identify the type signature and definition of any function or variable in a program. In Python, this is non-trivial.

FYI, this article is also cross-posted on:

https://news.ycombinator.com/item?id=16337435

https://news.ycombinator.com/item?id=16371338

https://news.ycombinator.com/item?id=16411765

Let me ask you all this: Why was Hadoop written in Java and not Python? Why was Spark written in Scala and not Python? Why was Kubernetes written in Go and not Python? Why was Mesos written in C/C++ and not Python?

This looks like an awesome project, and I hope this receives significant attention from city planners. I imagine as cars become more connected, we'll have better interfaces for integrating mobile traffic sensors with stationary traffic sensors for optimal management of vehicle flow along congested streets. But I do worry about the pace at which city planners adopt such technology at a time when local government budgets are tight.

I would think that the chip manufacturers (Intel, AMD, etc) should be reaching out to OS vendors for correct patch implementations.. unless they're committed to selling us shit forever.