selectodude

Yeah, they all do it though. And now that models are being natively trained at FP8/NVFP4, I’m not sure it matters.

Only way you can really know you’re getting the full model is to host it yourself. Every inference provider has every reason to lie and it’s impossible to find out the degree to which they are.

DeepInfra does and it's the same price. That's what I use.

I look at it as like being able to start a fire in the woods without a lighter. It’s a skill, a helpful one even on the extremely rare occasions that it’s necessary. But realistically with even the bare minimum amount of prep, not a useful one.

How many people out there can do hand lithography to make an 8086 to bootstrap computing? One?

I really do not know when I will replace my five year old M1 Pro. I'm not even starting to begin to get to the point where I start thinking about replacing this thing which I'm frankly astonished by.

I think the interesting part of that sort of success is that if/when this thing will start to lose a step (local inference would be cool, 16GB of RAM is starting to get a little tight, especially for rust builds and agentic programming), I really have no serious qualms about spending $5000+ on its replacement.

I would guess he supports those who do.

Yeah all that time not working is spent in the slack chats. Sounds like place just sucks.

They have to cover their fixed costs somehow. Collusion and cartels only get you so far. OPEC and Mexican drug cartels are still exposed to market pressures.

Large tech companies like cloudflare are spending developer hours lowering memory usage because they feel the impact too.

The idea that companies should have priority price controlled access to memory puts us even deeper into a situation where we can never afford local compute ever again.

It’s a fork of Roo Code due to Roo no longer being developed.

There aren’t many primary and secondary colors. If you plan to have a more comprehensive system you’ll run out pretty quick. Chicago comes to mind. The newest one is pink and I’m not sure what they’d do if they built a new one.

Because the hyper-tailored vibe coded browser front end that you like might be ever so slightly different from the hyper-tailored vibe coded browser front end that I want to use.

You don’t. LLMs make vibe coding highly customized software that operates exactly as the user wants it super easy, which is awesome. However humans still haven’t caught up to the idea that it doesn’t need to go on GitHub because if it was useful to somebody, they’d have already vibe coded it themselves.

Push comes to shove, OpenAI could go out of business tomorrow and I could pick up roughly where I left off for $25k, which is the cost to serve GLM 5.3 Flash on four Nvidia GB10s. Granted, if OpenAI et al go kaput all at the same time, I could probably get a whole lot more compute for a whole lot less money.

Okay? I didn’t sign a 10 year contract. We’re month to month and I use my own harness.

If they’re subsidizing my usage, that’s great.

My M1 Pro MBP is 6 years old and continues to be the best computer I own, so if that’s Apple not trying, god help everybody else once they do.

Once I get some kind of settlement after getting beaten up by a cop my first purchase will be some RTX Pro 6000s.

Dogs use the same x-ray machines that people do.

The Gulf countries paid handsomely to get Trump in there to keep the United States as the buyer of last resort once the rest of the world goes renewable.

Problem is that installing a bunch of sick sons of bitches to do your bidding can backfire when those some morons lose a war to Iran in a matter of months.

My only concern is that they'll sell out so quickly for so long that I'll have to wait for the iPhone Duo Two-o because they won't be easy to get until like May 2027.

150 tokens per second on a ternary model implies that it’s GPU bound, I’d bet a Q6 model is even faster because it’s existed longer and seen more optimization. You’d have to be insane to not run an NVFP4 quant over a ternary quant on Blackwell if they both fit.

Elon fucking Musk has literally every single piece of personal information of every person in the country. It’s so far past too late for any of this to matter.

I’m not sure how we start over but this data plus LLMs is gonna make it a full time job to keep your parents from sending every penny to a scammer.

Local frontier costs a half million dollars to run locally in anything higher than basically ternary.

I can’t run a corruption racket in my basement.

Hermes still makes saddles that are somewhat reasonably priced compared to their handbags.

It's still profitable, it's their original raison d'etre, and there's no real reason for them to stop even if it its rounding error on their regular business.

Probably will never, ever see an Nvidia card with >32GB of VRAM though unless they start making dies that lack LLM performance like the gimped ethereum mining cards.

It's still a 15B market for Nvidia, it's not nothing. But I'm curious how they're going to turn Rubin into a gaming GPU. I think at this point consumer GPU upgrades are going to be AI-related upgrades that happen to also help raster capabilities. Blackwell was already kind of a dud on performance uplift from Ada beyond the new LLM features.

Zero impact on iOS 27.

The cursor acquisition just shows that there’s always a dumber shithead out there. Though vaporizing Elon Musks money is about as pure of a good as there is out there these days.

I’m pretty sure the efficiency cores in the A20 Pro have more juice than the performance cores in my M1 Pro laptop.

NVFP4 would buy them a huge increase in capacity but I think it would be noticeable.

The passport card is almost exclusively for people who go back and forth to Canada and want to keep it in their wallet. Nobody is complaining about the size of a passport.