Most people are not even referring to CUDA batching nuances.

They think that sampling is an inherent part of Transformers.

Even on this site, it is regurgitated with confidence.

It’s only good for them in the sense that it is allowing them additional time and compute to cheat a reward signal.

It is not good for them in the sense that they will short circuit the RL path that actually improves general capability.

*TB/s

> Edit: I'm thinking of a headless Mac mini, if you meant running it on the same machine you're using of course you'll need more memory, but LLMs are best served from a headless server so that's what I'd recommend.

What? LLMs are best served from a massive PD disaggregated cluster of B300s connected via NVLink.

If you're running LLMs on a Mac Mini, it's because you want to run local, not because it's the best setup.

>massive PD disaggregated cluster of B300s connected via NVLink.

So a headless server.

Macs were mentioned because that's what the post is about. It could be a PC (I use a 2x3090 PC). The point is that it's a better experience to have a box dedicated to the LLM than running it in your system. Obviously in your home, so local.

Has nothing to do with being headless. That's just a natural outcome.

If someone wanted to use an 8x B300 as their daily driver...go ahead.

It would still be the best way to serve a given model.

You're not entitled to this stuff any more than they are.

What planet do you live on?

Trolling. GLM is heavily distilled from Gemini.

Source? GLM is great for coding and Gemini is barely useful in coding, to be generous.

I highly suspect that the Gemini Google uses internally is very different from what they offer in Antigravity.

No, it's the same internally and externally.

Gemini 3.7 Flash is a pretty great model IMO. You shouldn't compare it to Opus, Sol, K3, etc since it's a much smaller model but it's a little better compared to Sonnet, Luna or Terra, etc.

Gemini 3.7 has been trash in my testing. So is its pricing.

I can't speak for GLM as I haven't tested it much, but my experience with running DeepSeek V4 locally is the first time I prompted it with "Explain your capabilities to me", it responded that it was Gemini, a multi-modal model. I've seen others see the same with DSv4, as well as the "hallucinated" multi-modal nature. I would not be surprised if GLM similarly was partially (heavily?) distilled off of Gemini.

A fun test would be to compare the logits for "gemini", "claude", etc for a continuation of "I am " on all these models. I'm sure that e.g. GLM, Qwen, DS are dominant, but I'd be curious to see the next highest contenders, and how they compare to each other.

Never ask the model what model it is. Every single model may answer wrong answer if you keep asking.

Early glm models gave off that wibe. But now its more inspired by. With a bit of Claude in there. But I do think they actually do RL otherwise glm5.3 shouldn't have been able to beat fable on the few tests it did.

They are talking about breadth, not depth, multipliers in almost every circumstance.

A great X will be able to do far more great X stuff (breadth), and perhaps also be a greater X (depth).

But it's most certainly weighted in favor of the former than the latter.

Is it not a bad thing to encourage poor people to continue doing the thing that has apparently made them poor?

It's also very not true that farmers are poor in the US (median household income about 30% greater than general population). This mindset is like a century old.

"It's also very not true that farmers are poor in the US (median household income about 30% greater than general population)."

Most farmers work a "real" job in addition to the farm work.

Then you also have farm workers, who do not own the farm, and they tend to be poor.

What do you mean by farmers? Of the rich farmers, yes, most do have a second job, but the farming does pay most of their income. It's not just a hobby. The second job is something that they can do in between the farm work. The other thing about farming is it's very much investment heavy. And so for a young farmer just getting started out, you're going to be working at farming and all of your income is going right back to paying things off. And so you're going to need a second job to put the food on the table for a few years. After 10 years, because of inflation, you would hope that your income has gone up while the payment you're making on existing land is kept the same, and so you don't have some left over, and so you can start only being a farmer. After, say, 30 years, most of your land is now paid off, and now you can take all that money that was going into paying off millions of farmlands, and you're making enough income that you could pay off of millions of dollars a loan, and you have no loans, and you are very well off. Of course after 30 years you're also getting close to retirement age and you're probably thinking well I'm gonna sell that farm off and now you have a lot of money that you can put into regular retirement acounts and enjoy a really nice retirement. But that means you're not a farmer and so the farmers themselves really have to live like the poor but they're on track to be very wealthy.

The economic landscape of farming is very bimodal. “Most farmers” (by headcount) are subsistence farmers who live in impoverished countries and don’t produce very much food because they are feeding themselves and their immediate community.

The food that people buy at the grocery store in developed countries is primarily produced a relatively small number of industrialized farming operations which are high dollar businesses.

That is both true and not quite true at the same time. The notion of the Yeoman farm family supporting itself off its land is mostly a myth.

Like everything else. it’s been capitalized and farms have consolidated over time. There were lots and lots and lots of farm workers. who are pretty poor and now lack the assets that their grandparents had.

> Is it not a bad thing to encourage poor people to continue doing the thing that has apparently made them poor?

Right, we should tell the farmers to "learn to code"?

Too bad you can't ask Claude to grow potatoes. Though it probably uses the same amount of water and fuel to tell you it can't.

College stopped being about getting an education a long time ago.

It's just a low pass filter for the job market. Students aren't actually interested in learning anything, they have been condition for a long time to jump through the hoops.

It would be much better if we actually returned to exam and performance based evaluation of students, and stopped giving everyone As just for trying hard. But parents look at college fees as buying their kid a job opportunity, and so grade inflation has basically ruined the SNR of academic performance for all but the lowest performers.

I meet a lot of noj-college people and let's just say, "just a low pass filter" is still A LOT of value.

Although surely highpass filter would make more sense in the analogy here?

It depends on college attendance. In developed countries, it is a low-excluding - or the inverse of a low-pass - filter as most graduate, while in developing countries it is a high-pass filter as few have the resources to reach an undergraduate level of education.

I'm confused by your point here. You seem to be saying it's a highpass in developing countries and also a highpass in developed countries?

It's the other way around: conviction rate includes plea bargains.

Convictions at trial are much lower.