This is bananas!

One thing he doesn't mention is memory. What happens when the context runs out? How do agents preserve state and what are they continuously being prompted with?

So, one of the rules I came up with and put into all of my Claude.md files is that after discovering any piece of information, and before doing anything else, that information has to be written to local file storage. That local file storage is then re-read any time the session is compacted or the main process starts up again. Agents may not need this information, but the top-level process does.

And this is on a per-task basis, so we are not polluting the memory of one top-level Claude process with the data from another.

This method should survive all compaction, and allow me to go back and revisit any previous task that I’ve worked on with Claude.

I also have a rule that any important lessons learned also get recorded into the Claude.md, for the benefit of future tasks.

Why only 20%? Is there a theoretical maximum?

The theoretical maximum for the mentioned cell is around 32-33%, just a touch higher than the theoretical limit for silicon.

The Shockley–Queisser limit still applies i.e. with the current understanding of physics the limit for single cell is ~33%

I said ~20% because that's the minimum required to bring it to commercial market while being competitive. If they go more, great, bonus.

For those dizzied by the quick unit changes:

Forest loss in Finland - 243,000 hectares = 600,000 acres

Finland total - 20.3-26.3 million hectares = 50-65 million acres

US forest loss - 1.2M hectares = 3M acres US forest conversions - 1.05M hectares = 2.6M acres

Brazil - 985k hectares = 2.43M acres India - 26,858 hectares = 66,368 acres

Thanks. It wasn't clear to me what the actual issue was from the headline.

I wonder if there'll be an episode of Hello Internet on it.

The study of arcane procedures and immigration of all sorts comes up here very often.

Now put six of them in a field and throw in a huge soccer ball.

The headline is "Online ads weaponised to deliver ‘zero-click’ spyware, researchers say"

You're right. Never in history and never anywhere else in the world have Jews been persecuted.

It's harder in the exact way that AI is good at, though. You need to slog through very large amounts of boilerplate until you find what you're looking for.

It's not really more complex once you're practiced at it, just tedious.