Tenobrus

what basis do you have for assuming an LLM is fundamentally incapable of doing this?

o1 has a METR time horizon of around 40 minutes, opus 4.7 has an implied horizon of 18 hours based on its ECI score. this study is on a model that's several generations behind wrt the kind of tasks it can complete. it would be shocking if this number were anywhere near as low with GPT 5.5, to the point it seems nearly totally irrelevant to talk about these results

"Japanese soldier who kept fighting 29 years after World War 2"

those two stipulations were always their only ones, and they were included explicitly in their original contract with the DoW.

some ai detectors work now. pangram detects this as 57% AI written, and the parts it thinks are human are.... the ascii diagrams / screenshots. all the actual text it detects as generated.

A lot of people forget how whimsical and strange and beautiful the old smaller GPT models and the original 3 base model pre-RHLF could be. Nowadays hundreds of millions of people have talked to heavily assistant tuned 4 or 5, but comparatively very few people have ever even seen GPT-1 outputs. It's cheap to run so I threw up a simple interface + single server hosting it.

it is very clear to me that humans do in fact have a recursive self-improvement ability, and i'm confused why you think otherwise

all this capability was already easily available in GPT-3 with some prompt engineering, this mostly just changed it from few-shot to zero-shot.

You're talking the US first amendment and how freedom of speech relates to governments, but that's only a limited subset of the general concept of free speech. Obviously discussions around platforms involve more complexity, but for example if AT&T started refusing to send my texts based on what kinds of ideas I was arguing for then this is a serious attempt to limit speech and discussion which is morally incorrect, regardless of legalities.

The main interesting thing about GPT-3 is that prompt engineering turns it into an incredibly general tool. This summary was likely generated with a prompt something like:

Here's an article about x: <text of sample article>

A quick summary of the article, focusing on the main relevant points and keeping critical detail: <sample summary>

as an example, then duplicating that with the real article and text.

That's the "few shot learning" from the original paper. It could also be that it's good enough at summarizing specifically that you don't need an example, just the right framing prompt around the article text. Either way, that kind of prompt engineering is how you get "text completion" to perform basically any text processing task, generate code or play tic tac toe, so on