I looked into this last week, and it seemed like if I wanted to get something in the Opus 5.5 range I'd be talking about $500k in hardware. Is this actually all much cheaper than I was seeing?

The weights of Kimi K3.1 Pro are 1.56 TB. I guess with four of the upcoming MaixBlade cards from Sipeed you'll be able to run it without quantization but that setup would still be >$50k.

Isn't 1.56TB the 4-bit quantization?

But yes, I was thinking about the hardware that's already available, and the MaxiBlades are exciting.

That's the native size (MXFP4 weights / MXFP8 activations).

I'm considering getting the MaixStation Pro, which is the same as one Blade (4x AX8860P) and the absolute maximum my budget would allow.

Pedantically: it's not a windfall if they specifically brought it about.

What do you mean "lost the majority vote"? More people voted for him than anyone else, and he got a majority of the electoral college.

Look up plurality vs majority.

In American English "majority" can mean "plurality", it's a bit weird! [1] Saying "lost the majority vote" when you mean "won the election with under 50%" is bound to confuse people.

[1] https://www.merriam-webster.com/dictionary/majority definition 1.c

I am not making this up, majority vote and popular vote are literal phrases used to describe the results. Please look it up. I agree it’s confusing.

I know you're not making it up, "plurality" is a real word and in a technical context "majority" can be contrasted with "plurality". But even then you really wouldn't say "lost the majority vote" for someone who won an election (try searching for that phrase: none of the hits are people using it like you are).

The majority vote is being used often to indicate whether or not majority actually voted in a given President. In the USA however, majority vote win happens infrequently nowadays, only 4 out of last 9 elections with winner won a majority vote. so perhaps unfair to bring that up, the 2024 was an absolute blowout by USA's standards

They slept with foreign spies and helped them evade export controls?

If we had a pathogen-agnostic early warning system (what I'm working on: https://securebio.org/detection ) we would know. But we need a larger and more sensitive system than anything deployed today.

This is mostly best practices for making websites work better, not for pretty HTML.

Firefox is not a skin of Chrome; it has the only browser engine in common use that's not descended from KHTML.

That's true, but in that case just say "Firefox and Chrome", Edge isn't contributing anything.

Yes. RAM got expensive, because demand went up a lot.

Is there a lot of new demand for the kind of RAM on a RPi?

This is not the same stuff an H100 uses, and it’s not even made in the same facilities, is it?

I think it’s more likely that the volume is not large enough to make up for the previously crappy margins on cheap RAM products.

It was great while it lasted.

A price increase in one market can displace demand to other markets.

A workstation user or gamer might want a desktop PC loaded with DDR5. But they'll also consider a desktop with DDR4 or even a laptop with LPDDR4X if they get more for their money.

So a rise in demand/prices for DDR5 can trigger a lesser rise in demand/prices for LPDDR4X

That's a good point that I had not considered, thanks.

It's a surprise bonanza for the low-end RAM producers. Raise prices substantially with no new investment needed. Some of these factories must have paid for themselves years ago.

The key is where the RAM comes from. The sources are the same companies/factories that are making memory for H100s....

> Is there a lot of new demand for the kind of RAM on a RPi?

Yes. Due to the way the RPi board is laid out, they can only support one DRAM chip. And 16 Gb modules are the most expensive ones available, they are the ones that go onto 128Gb and 256Gb sticks.

Right, and I learned in this thread that the same wafer processes are used to make these as the globally bottlenecked HBM chips in NVIDIA servers. It's a shame that hobbyists are competing with Big AI for these factory lines.

All of the world's RAM manufacturing was switched to making the kind of RAM that's on graphics cards. Hence the shortage.

The kind of RAM on an RPi uses the same wafers and fab equipment as the kind of RAM in a $40k GPU

Up to the finished wafer, right? Assembly is much more exotic for HBM, with 3D stacks, etc.

If the wafers are the bottleneck, e.g. ASML's god-machines, then yeah, it's annoyingly the same thing.

> RAM got expensive, because demand went up a lot.

Somewhat, but not all of that demand was real; some of it was promised and not followed through on.

Also, supply isn't growing to match, in part because some companies exited the business.

That is literally demand

I wish that (regardless of iOS vs Android) when a swype hit their "this is risky, probably don't do it automatically" list their best guess showed up as one of the options in the "did you mean" list above the keyboard, instead of excluding it entirely.

As a Claude Code user I'm very glad to see Boris and other devs asking for our feedback and advice. That's how you make good products, and should be welcome on HN.

I've always been into programming as a way to solve problems, and that CC allows us to get more done faster is a good thing.

(Up to the point that we get to RSI and extremely powerful models that we don't know how to align or control. Which I think is uncomfortably soon, and so there's a part of me that says "don't make CC better, it's on the critical path to reaching RSI before we're ready")