@RandomRanger's banner p

RandomRanger

Just build nuclear plants!

5 followers   follows 1 user  
joined 2022 September 05 00:46:54 UTC

				

User ID: 317

RandomRanger

Just build nuclear plants!

5 followers   follows 1 user   joined 2022 September 05 00:46:54 UTC

					

No bio...


					

User ID: 317

I'm reminded of a prediction we had just 4 months ago: https://www.themotte.org/post/3758/culture-war-roundup-for-the-week/446278?context=8#context

The TLDR of that post was of AI being mostly fake tricks and hype, costs rising faster than intelligence (Opus 5 drops in 2029)... Remember the discourse we had about the AI plateau? Don't hear so much about that anymore.

And why now? AI was good for several months up to now.

Kind of? It gets better over time though, whips out new tricks easier and cheaper than before.

We had AI images that were kind of good since the end of 2022, if you did a lot of tricks with them and good prompting. Back in 2023 you could say images were good, albeit text was impossible and certain detailed direction impossible. Then there was that Ghibli phase when GPT image rolled out. And then everyone learned to accept what images could be made and everyone understood that AI could make cartoons with text and images in them, learned to recognize and resent that specific artstyle of GPT-image and nano-banana.

And so AI images are still 'kind of good' in 2026 even though they're much better than 2023.

It's the same with coding I think, except replace 2023 with 2024. Trying to code by copy pasting snippets back and forth was possible. I did it. It was a huge pain though and stuff would just not work. Then we had coding agents. Then the agents get better and better. They start being able to review and check if what they've done works. They get better and better at that. They can see, they can operate all kinds of other software for sound and images.

Progress is constant I think but we suck at gauging steady progress as humans.

Yeah, I remember this discourse about Buffy the Vampire Slayer being 'performatively feminist' as a piece of media and how there were all these secret ways it was actually demeaning to women... and no that wasn't true. It was just feminist. There was a girl power speech in it that Buffy gives and the central premise is basically girl power.

I think you Mr '@deepseek_ai stan #1, 2023–Deep Time' are a bit too emotionally attached to Deepseek to perceive what's going on

You're ritalin addled, obsessed with penises and write like an Indian!

That might work on twitter for ratios. Wrong website!

45 on Artificial Analysis, what is that, a model for ants? What's Deepseek's excuse, they been selling their compute to Anthropic? Making image and video models? Actually doing their own inference rather than opensourcing it?

One (saturated) cyber benchmark: sub-Astra, cheaper? In the real world the Americans nerf their released model's cyber capabilities. They've been public about this. Cyber naturally follows programming ability and general intelligence.

Quantity works. I see it working today and throughout history, so I say it works and expect the trend to continue. I say that's the most important thing. And you don't like it, you don't want to see it, you want something else to happen and so you imagine these worlds where Huawei is going to be anywhere near TPUs, or where 10:1 is a somehow tolerable disparity, or where Deepseek 'auditing the stack' and finding vulnerabilities will save them from hardware level problems that need those extra faculties to detect, let alone fix.

And that might happen. But you seem so certain that it will, when the natural balance of power is against it. It's cope.