The expected event count from 17,000 people exposed to a month of constant uniform event risk with an annual hazard rate of 20 per 100,000 per year is 0.28. The p-value for observing a count of 5 from a (Poissonian) mean of 0.28 is 1 in 100,000. So it's pretty unlikely by perfect random chance.

The other thing you see in a life spent in count statistics is a lot of statistically extreme events that turn out to have no deep or predictive significance.

Comments URL: https://news.ycombinator.com/item?id=48968701

Points: 76

# Comments: 29

> With little natural gas supplies of the country’s own, and its heavily supported renewable sector unable to fully make up the shortfall, German leaders faced a dilemma. To maintain enough gas reserves to get the country through the winter, they could try to put off the closure of Germany’s last three remaining nuclear reactors, which were scheduled to shutter by the end of 2022 as part of Germany’s post-Fukushima turn against nuclear power, and even restart already closed reactors.

> Or they could try to reactivate mothballed coal-fired power plants, and make up some of the electricity deficit with Germany’s still-ample coal reserves.

[...]

> Last week, the country’s parliament, with the backing of members of the Green Party in the coalition government, passed emergency legislation to reopen coal-powered plants, as well as further measures to boost the production of renewable energy.

I acknowledge that the politics and history of the nuclear decommission is more complicated than I understood - [2] seems to be a useful walkthrough. But it all happened with Green support and is the fruits of decades of Green ideological advocacy.

[1] https://archive.ph/v3iLS

[2] https://www.cjfp.org/for-germanys-green-party-the-50-year-dr...

But what they are saying is, it would be valuable if it was causative wiggles eyebrows

Last sentence of the abstract:

> Sleep regularity may be a simple, effective target for improving general health and survival.

Being talked at by someone's AI copypasta feels like being in the Truman Show.

"Why don't you let me fix you some of this new Mococo drink? All natural cocoa beans from the upper slopes of Mount Nicaragua, no artificial sweeteners!"

"What the hell are you talking about? Who're you talking to?"

RL post-training alters the parameters of the transformer, while your f(manifold) idea seems to suggest that a new layer on top would suffice, no need to alter the transformer itself at all.

It would be extremely handy if that were so, but I'm guessing it isn't, or it would be the prevailing approach.

1. Avoid overparameterization by design. Manually create or choose a space of functions that has limited degrees of freedom by construction.

2. Accept overparameterization and regularize.

The latter tends to be more robust, because of the bitter lesson. It's not practical to manually design an ideal, on-demand, just-right limited-parameter model for every dataset we are presented with. The best way to approach that ideal, it turns out, is really to just let the computer figure it out via regularized optimization over an overparameterized space.

Statisticians started moving in favor of overparameterization long before deep learning got off the ground. This trend dates back at least to the machine learning bible, Elements of Statistical Learning (2001).

[1] https://mlu-explain.github.io/double-descent/

There is no description of what the thing is, no indication of what value it provides its users. The closest it gets is "the product has been used by hundreds of users internally, including daily internal power users".

But the fact that the thing has a million lines of code is repeated twice in the first few hundred words.

[1] https://openai.com/index/harness-engineering/

[2] https://news.ycombinator.com/item?id=48416264