diddid

Qwen 3.8 Flash Next is amazing, i did hundreds of turns and billions of prefill and it may not be as smart as sota but then again it does what i tell it to and it does it well.

But I still don’t think that means eat all the cookies in the cookie jar and then apologize after because nobody would have given permission. That’s still about doing what you believe to be right. She even frames the fallout as “where you supposed to do that?” and not “you shouldn’t have done that”.

I always get annoyed when people misinterpret Occam’s razor. It’s not that the simplest is more likely to be correct, it’s that you should prefer it, because it’s simple.

It’s just like the Hopper quote. She said it’s better to ask for forgiveness during the fog of war, doing something you thought was right, not to do something you knew they were going to say no to and now you are trying to get away with something.

I like my Mac mini as much as the next person, but I don’t think an M5 ultra 64gb with 4tb should be as much as a DGX Spark with 128gb and 4tb.

This is where the DGX spark makes up a bit of the ground it loses on llm work, diffusion and cuda go together like peanut butter and jelly.

Coding has been solved in the same way that writing has been solved. Putting the sentences together is now easier, but having something interesting to say? That’s a lot harder. With the barrier to entry lower there is a lot of things that look good at the surface, but when you dig deeper, aren’t quite right. The nouns are in the right place and it has all the things school taught you a paragraph should have, but it’s just not an interesting story. That’s where all this software stuff is right now. It removed a barrier, but just like visual programming doesn’t solve it, AI can’t solve it either. It lowers the barrier but doesn’t solve the essence (is that a silver bullet reference? I’m not sure). AI can extrapolate and interpolate what’s already been done in fascinating ways but I need to see something new to really believe. Asking it to rewrite something it trained on (like SQLite or a browser) into a language that it was also trained on (rust or go) isn’t all that interesting from a software advancement perspective. I want to see imagination and so far I haven’t. But I have my required ai ticket like everyone else and maybe one day I’ll look out the window and be satisfied.

They should do this to GLM 5.2

I doubt you are the average consumer though. Now days I think the biggest average consumer use case is streaming and game downloading, with game downloading being the biggest “I want it now!” impulse. But do you need 1gb service to download that call of duty game every year? I even think most people could get by with 50mb/s and not even know as long as the latency keeps up. If it’s fast enough to stream Bluey the masses are content.

I think the end game is convenience. Nobody really needs anything more than 200mb/s. If the average person can have their entire family stream their favorite Netflix show at the same time then that’s good enough. “Now lil Jimmy can watch it in the minivan too!”

This has been true for as long as war has been a thing. The Army logistics won’t break, it’ll adapt, because that’s what war machines do. What’s good today isn’t good tomorrow. It’s a nice piece but nothing Sun Tzu or a good game of hearts of iron couldn’t teach you.

I feel like this used to be more true, and is getting less true each year. With covid they really did a lot of negative things related to the parks. They cancelled a lot of things ( magical express, free magic bands for resort guests, etc) raised prices and overall began to focus on “shareholder value” over guest experiences. They also really messed up the Star Wars IP and seemed like anything for a quick buck, and became unable to make original stories. In 30 years they should revisit this and they will probably find they took a nose dive. The people who made the nostalgia are long gone, and the new guard is slowly burning it all away.

With the 5090 you need to buy the rest of the computer though, and the Dgx spark will run 1/4th as slow but use 1/5th the electricity. And the spark would be able to run things the 5090 just couldn’t, like the Qwen3.5 122b. Which is all just to say that for llm workflows there is no easy answer. And if you media generation it gets even more complicated.

I’m always a bit confused by this, don’t we want this? If the end goal is to have ai and robots do it all, and there are no jobs left, what would children provide if they aren’t down payments on a future work force? Only people that wanted children would have children as all the old childless people would be taken care of by the states robots. I mean not exactly like that but something like that. Ai isn’t taken anyone’s job is there is nobody there to take it from. Or is that line of think too… something.

This is the only one I can tell that easily works with local modals through hermes or pi.

I wish google would just go away. The thought of another product with googles tentacles in it makes me nauseous.