tomsyouruncle

I’m not filled with confidence when the author’s first paper takes an RL approach but then doesn’t use it to change the action taken in the next turn. Seems like simple classification would achieve the same end. And this quote from the paper isn’t overly reassuring:

“I personally found that this sequential approach captured sales dynamics much more effectively than traditional classification models.”

https://arxiv.org/pdf/2503.23303

Here’s the academic paper behind it: https://arxiv.org/abs/2602.04101

I think this article overlooks the act of engaging in problem solving with an agent.

Personally I find it super helpful to discuss stuff back and forth: It takes a view, explores the code and brings some insight. I take a view and steer the analysis. And together we arrive at a conclusion.

By that point the AI’s got so much context it typically does a great job summarising the thought process for wider discussion so I can tweak and polish and share.

Thanks a lot for adding these details about scoping[0]. I’ve definitely learned something there.

So my scope 1 emissions can be your scope 3 emissions if I emit carbon to make something that I sell to you… but the “real” emissions are always _somebody’s_ scope 1. Interesting stuff!

[0]: https://www.carbontrust.com/resources/briefing-what-are-scop...

> When you switched from burning lignin to solar, you reduced your carbon usage. However, what happened to the lignin that you stopped using?

In my scenario, CarbonCruncher bought the lignin I stopped using and made a road out of it. Crucially, in doing so they claim to have a negative carbon impact because they'd trapped that carbon in the ground. But I already claimed that impact when I stopped buying and burning it myself and switched to a zero-emission energy source.

So my (genuine) question remains: we can't _both_ claim the benefit, so who's right?

Maybe the concern others are expressing here can be thought of as the risk of double counting of carbon.

If I run an industrial plant that's currently fuelled by burning waste lignin from the paper industry, and I decide to stop that and install some solar panels instead, it seems reasonable for me to claim that change is carbon negative (i.e. I've reduced carbon emissions).

Now if CarbonCrusher comes along and buys the lignin I no longer need, uses it to build a road, and claims the same carbon saving as I did, we end up double counting.

Which of us is wrong?

This statistic actually says nothing at all about 90% of the people who took the test so the headline is pretty misleading.

You can’t really talk about “Australian developers” (or indeed any other nation’s) and ignore 90% of the distribution !

It’s a bit absurd to say you’re not using the mean because you’re worried about the effect of outliers and then instead choose a statistic that largely focuses on outliers! Why not use the median?

Thanks yeah 7zip looks useful although the standard macOS unzipping utility can’t decrypt it when it’s encrypted using AES256 (not sure if windows can natively). Hence I’m concerned that the third party will also need a copy of 7zip...

Adding a bit more detail to the question:

I'm keen to understand people's solutions for sending sensitive data (e.g. CSVs, PDF docs) to third party organisations that can only be assumed to have standard (MS Windows / MS Office) tools at their disposal for decryption.

I'm keen for encryption to be of a modern, robust standard, and preferably available via mac / unix tools.

Thanks HN!