Seems like a case of Simpson’s Paradox https://en.wikipedia.org/wiki/Simpson%27s_paradox

This is not a case of Simpson's Parodox, at least the analogy about accuracy vs speed from the article isn't. You're not comparing global vs subgroup correlations. On the one hand you're measuring how speed and accuracy are correlated across the population when you ask subjects to solve a problem. On the other hand, (rather than measuring subgroup correlations) you're measuring how accuracy is affected when you ask an individual to speed up or slow down.

The actual paper (linked in another comment here) is literally titled “Nonergodicity and Simpson’s paradox in neurocognitive dynamics of cognitive control”. Why they omit it from the blog post version, I’ve no idea.

never heard of this before, very cool

Not to be confused with Flanderization.

https://tvtropes.org/pmwiki/pmwiki.php/Main/Flanderization

If you are running on 4090 and get 5 t/s, then you exceeded your VRAM and are offloading to the CPU (or there is some other serious perf. issue)

You're almost certainly (definitely, in fact) confusing the 120b and 20b models.

I'm most certainly not doing so.

   seg@seg-epyc:~/models$ du -sh * /llmzoo/models/* | sort -n
   4.0K metrics.txt
   4.0K opus
   4.0K start_llama
   8.2G nvidia_Orchestrator-8B-Q8_0.gguf
   12K  config.ini
   34G  Qwen3.5-27B
   47G  Qwen3.5-35B
   51G  Qwen3.5-27B-BF16
   61G  gpt-oss-120b-F16.gguf
   65G  Qwen3.5-35B-BF16
   106G Qwen3.5-122B-Q6
   117G GLM4.6V
   175G MiniMax-M2.5
   232G /llmzoo/models/small_models
   240G Ernie4.5-300B
   377G DeepSeekv3.2-nolight
   380G /llmzoo/models/DeepSeek-V3.2-UD
   400G /llmzoo/models/Qwen3.5-397B-Q8
   424G /llmzoo/models/KimiK2Thinking
   443G DeepSeek-Math-v2
   443G DeepSeek-V3-0324-Q5
   500G /llmzoo/models/GLM5-Q5
   546G /llmzoo/models/KimiK2.5

Oh I missed the "quad" before 3090.

Yes. Even 90 Hz is a noticeable improvement over 60 Hz. I wouldn’t pick it over high-DPI, though.

No, the connector is longer than it is tall.

A solid refutation to the first point but not the second suggestion.

Maybe, but Apple doesn’t make them thinner anyway so the argument is invalid. iPhone 6S with headphone jack: 7.1mm thick. iPhone 17 is 7.95mm thick.

It’s easier than that. Replace shorts with watch in the URL.

Non-Shorts can be vertical as well. It works fine, and without the swipe-encouraging UX. If they were “just making long/medium-form videos in portrait now” they don’t need YT Shorts to do that.

DisplayPort 2.1 (which the monitor supports) provides sufficient bandwidth for 7680x4320@60 Hz 10-bit without DSC when using UHBR20. The press release unfortunately doesn’t clarify whether the monitor supports UHBR20 or only the lower UHBR10 or UHBR13.5 speeds. Of course, the GPU must also support that (Nvidia RTX 5000 only at the moment, as I believe AMD RX 9000 is only UHBR13.5).

I believe you're right regarding AMD's lack of UHBR20 on its cards. Fingers crossed for their next gen!

AMDs current workstation cards do support UHBR20, just not their consumer cards, even though it's the same silicon. Artificial segmentation on GPUs is nothing new but segmenting on display bandwidth is a strange move, especially when the market leader isn't doing that.

Redownloading everything isn’t a risk when the lock file contains a hash of the download on first update.