Recent content by Baughn

Or, with a little bit of work, longer videos. Veo 3 may not do it yet, but the likes of Framepack makes it extremely obvious that that's only a matter of time. I can generate 1-minute video clips on my desktop PC. It takes forever, but it works.

That's one possibility. Another I've heard—outside of work, or I couldn't mention it—is that the lawyers wanted to retrain the AI such that the CoT would never contain anything that might be considered legally dubious. Here's the problem: Sometimes AIs attempt to lie or otherwise mislead the...

Yes; as does every AI available. The price is often different -- prefill is more parallelisable than decode, hence typically cheaper -- but this makes conversations in general an N^2 operation. That is, the cost for any conversation is proportional to the square of their length. (Prompt...

"Limited access to 2.5 Pro" means "Two queries per day". Not conversations. Two messages. Which is fair, considering how expensive it is to run, and works fine for a demo... 2.5 Pro really is that smart. It's very much worth the sub. As for Deep Think, that's 2.5 Pro tuned for longer reasoning...

One of their free models does do reasoning, but importantly, it's not the default. Even then it's just not nearly as good as the alternatives. NotebookLM is an interface, not a model; pretty sure it still uses Gemini under the hood. Gemini itself has a context size of one million tokens...

Yes; all the time. It's not going to do it very discriminately.

I'm torn between 'insightful' and 'funny', so here's a light bulb. That's not the only reason, nor is it even strictly a reason, but there definitely is a valid causal chain you're pointing at.

ChatGPT is also designed to be... there's no better word for it. Obsequious. If you want an LLM that will actually push back, get Gemini 2.5. Or if you want one that can handle longer documents. Or a smarter one. In fact, just use 2.5 unconditionally—the only way it's 'worse' is that it by...

Friday, October 23, 2009 "I speak unto the gemstone, resilient, unyielding. I speak unto the lightning trapped inside. Raise thy gaze to the world. Mirror the barest fragments of my mind, releasing thy power only once I—ouch!" A torrent of static tore Lulu's attempt at divination into little...

This is huge point, really. ChatGPT is what people use. ChatGPT is what almost everybody is familiar with, and that usually means ChatGPT-4o, because that's the free one and even if you pay they aggressively steer you towards 4o to keep their costs down. (ChatGPT runs exclusively on nvidia...

They also may be funding bots on Reddit to talk shit about AI and generally inflame tensions. That’s harder to prove — I heard it from someone who works there, and it was a rumour to him as well — but doesn’t feel surprising. They want an apparent consensus on “regulating AI”.

High-quailty NVMe/SATA-connected flash storage is a lot more durable than HDDs, because it has no moving parts. The downside is that it needs to be regularly (ideally constantly) connected to power, as the data is stored in the form of precise voltage gradients that need regular maintenance...

Why wouldn't a spaceship be steam-powered? Because it sounds primitive? But there really aren't any alternatives. Steam is just much too good at what it does. Gas turbines are the only exception I can think of, and those depend on combustible fuel; you're right back to steam once you suggest...

The best part is the explosions were rarely instant. Plasma would cook anything in the room, of course, and skyrocket the pressure until airlocks failed, but we usually had time to investigate and discover just how screwed we were. Eventually it would always find the power system.

Hah... yes... when it works. Its internal architecture is a bit perpendicular to everything else, so we had a couple outages. Still, it's worth a try! IIRC the free tier gives you 5 uses per month, so for everyone else here, there's no real excuse not to try it. Just try not to do it while I'm...