Meh, disagree. I read it as an expression of frustration, as there are enough interesting and reasonable people with similar technical interests to have a good conversation, but they get drowned out by folks who seem to think an interest in applied matrix multiplication is ontologically evil.

Sorry, man. I keep wanting old school AI/ML Twitter to manifest here (or at least somewhere not Xitter), and it’s just not gonna. *sigh*

Everyone thinks there are just gonna be these “insights.”

(“Insight” being a term of art that means “someone else’s work product that will absolve me of all accountability as a decision maker”)

Hoping the solar powered DAC tech Terraform Industries is developing takes off. Would love to bend the Keeling curve, and practical carbon capture that’s initially deployed to produce synthetic fuel gives you an economic rationale to build out capacity to the point you can capture as a CO2 sink.

And the short-sightedness. They are just running up a score for glorified dick-measuring in a way that destroys the stability of the system in which they are running up said score, as if a sufficiently large number will insulate them from the consequences of it all tumbling down.

And this is the thing that really fucking annoys me, like, that’s all these guys care about, money, money, money, they should be taxed into oblivion

further research from a theoretical angle. E.g. some of the neural methods make it very cheap to examine forecast sensitivity. *To the extent you can show this is physical vs numerical* (e.g. see similar things when probing traditional NWP model), can identify important processes for further study.

This is totally fair, and the fact that stuffing your proposal with “AI/ML“ keywords is a funding unlock (seemingly independent of merit) is not ideal. Agree transferability of knowledge/insight is and important axis. There are some cool contexts where the data driven approaches can inform 1/2

Your points from previous threads on making sure conservation and various physical invariances/symmetries are maintained is very valid, but this can be built into the model. AI2’s ACE2 model does this: www.nature.com/articles/s41...

Full disclosure - AI2’s lead for climate was my thesis advisor.

Traditional numerical methods and modern ML can complement each other, I don’t see this as either/or. Critically, NWP is a great application because validation/verification is very well defined. Nobody will switch to these operationally if they suck.

”all models are wrong, some are useful“ applies here. There are an awful lot of assumptions built into the various components of traditional NWP systems (full Earth system models are yet more complex), and tuning those parameterizations is itself prone to overfitting.

Lmao, fantastic: m.youtube.com/watch?v=FwiB...

This is straight up vandalism.

Ouroboros of Shit

Might efficacy of EW lead to renewed focus on precision guidance that's independent of GPS (at least once launched) - high precision IMUs, visual or multi-spectral navigation, etc.? Also, seems like difficulty gaining air supremacy means laser guidance would be useful with UAS-mounted designators.

Nice work!