kristjansson

> Been you didn't consume the entire thing in a 2 hour binge uninterrupted by external needs no matter how pressing like everyone else did??

> effortlessly Real question: does your world model suppose super exponential gains in intelligence and efficiency?

I mean ASI could just generate the pixels of a webpage at 60hz. And who needs cookies if ASI knows who you are after seeing half a frame of the dust on your wall, and then knows all the pixels and all…

How do you determine flawlessness? How you even approximate a guarantee of it? To what specification is flawlessness judged, and can you precisely and completely relay that spec to your friendly local…

The value of a library is not just that it does a thing you want, but that it doesn’t do all the things you’d prefer it didn’t. It’s easy to write a cookie parser for a simple case; clearly your robot…

Wh … why would we want that?

A stack of 38x 1U and a switch does not a cloud make. Slightly less pithy: they're selling rack-scale systems, with power, hardware, network, and control plane software all integrated. Something…

> 5 kW to work with per cabinet have my expectations been shot by reading too much about Nvidia's latest and greatest, or does this seem quite low?

Tesla Roadster took a bunch of preorders at $50-250k down almost a decade ago, More recently, Taycan did reasonable-ish volume at $100-200k/unit. There (at least once was) a market for such thi…

Set your parameters to make that point then. “Yeah just run a 1T+ model on CPU”

An example or two would go a long way here.

Perhaps we'll go the way of the Space Shuttle? One group writes a highly-structured, highly-granular, branch-by-branch 2500 page spec, and another group (LLM) writes 25000 lines of code, then th…

Code written in a HLL is a sufficient[1] description of the resulting program/behavior. The code, in combination with the runtime, define constraints on the behavior of the resulting program. A …

I always appreciate the near Literate Programming completeness of Mark’s posts, and his consistency in that over years. Every post feels like one could spin up a clean VM and end up with exactly his r…

200 GB is an unfathomable amount of main memory for a CPU (with apologies for snark,) give gpt-oss-120b a try. It’s not fast at all, but it can generate on CPU.

Your original question was a bit of a challenge to accept earnestly. It’s a pretty charged situation; joking about ETs led me to discount it and may have led others to downvote. That said, I described…

> Responsible actors in civil society have a mandate to aggressively detect and interdict fraud. If they do not, they cede the field to irresponsible demagogues. They will not be careful in their c…

On Shirely, and the reaction: there's a markedly different valence to a fraud 'investigation' seeking to arrest, try, convict, and imprison _fraudsters_ vs one seeking (through a thin v…

Oh you're right that is a misread on my part, the appendix charts don't say that. I think they're just useless then though? Since they're reporting absolute error (on a log10 sca…

> Right next door is Hawthorne 30 minutes drive in no traffic, crossing half a dozen cities and the 405. There's reasons to inveigh against the YIMBYs (why are they celebrating densifying a c…

> approximately the same magnitude and they really do mean that, their results show +/- 1 on log10 plots.

> self-attention is efficiently computable to arbitrary precision with constant cost per token This paper at least aspires to reproduce 'true' attention, which distinguishes it from many …

50MW might be one aisle of a really dense DC. A single rack might draw 120kW.

I think OP was saying hotter part -> hotter radiator attached to the part, not that the part itself will radiate significantly.

The physical constraints aren’t insane for black body radiators. The engineering to run radiators at 90C in space OTOH…

I see what you’re saying - higher design temp radiates better despite more energy overall to dissipate.

Space is empty, not cold.

1kW TDP chips need LESS cooling?

Same reason humans write in higher-level languages instead of machine code? Each additional unit of program text costs energy at write time, so there's a bias toward more compact _representations…