The iPad at launch was just a giant iPhone. This looks like it's just a giant Apple Watch face attached to a speaker.

Blind?

So I guess "being into AI" is a signal for being older in the same sense as "not being broke" is a signal of being older. They are intertwined with one another.

Most of the people who are into AI are into using it for productivity and accomplishing new things. Average kids in their late teens and early 20s generally aren't really accomplishing anything, and don't really have much professional pressure to do so. They're just playing.

That's basically what you're seeing here: M3.1 Flash's road output looks like a 240p downsampled YouTube video (with road antialiasing/dithering, etc.), while DeepSeek's output looks like an upsampled Google Maps navigation screen.

DeepSeek doesn't have a video model, while MinMax has H3. Would not be surprised if this "video model lineage" made its way into M3.1 Flash as training data somehow, causing it to "visualize" scenes like this top-down and translate to code, rather than building bottom-up from code to a visualization.

You must be confused. The chart shows the model/CC pair climbing with the new checkpoint, and shows it climbing again when switching from CC to DSH. My point is that this doesn't say "DSH good", it just says "CC bad". There are lots of good non-CC harnesses.

Your prompt is incredibly contradictory in multiple places, but this line right here means that MiniMax absolutely wins the benchmark according to the stated primary criteria. You explicitly state that it should "not be like a video game", but you imply throughout that it should be an exaggerated cinematic style (e.g. "like a video game"). But despite your mixed messages, explicit instructions should always override conflicting implicit undertones. So again, MiniMax did a better job with instruction following on this adversarial prompt.

The one good thing these benchmarks are for isn't evaluating models, but instead drumming up social media engagement bait. One of the best ways to get social media replies is to get people to debate something that doesn't have an objectively-correct answer, after all.

Neither is better in any general sense, which is why these arbitrary vibe benchmarks are essentially worthless. One of them has "cozy video game lighting" but a weird looking fire. The other is more realistic but less romanticized. And neither model will do the same twice.

Genuine question: Here in the US, in Silicon Valley, what do you suppose the average age is of the the average engineer working for an AI startup? Do you think it's closer to 20, or closer to 50?

It's interesting that you're unable to distinguish your bubble from the greater world around it. Your children seem to be biased by your beliefs and self-selected their friends accordingly. You're in Europe. If you told me "18 year old European girls and their friends tend to hate AI", I'd ask you whether or not water was wet as well.

The problem is that one man's "reward hacking" is another man's "efficiency". I suspect Google researchers were being clever doing something similar to what you're suggesting, which is probably why they ended up with a model that repeatedly re-reads the same unmodified file.

If this isn't the textbook definition of "mutually-exclusive", then nothing is lol. I can't spend time with my family at the office. When I've been forced to be in onsites for a week at a time in SF/NYC, my woman constantly complains I'm impossible to reach.

They are literally mutually-exclusive, since the time you spend in office takes you away from these things lol.

Apparently because it showed up on my For You feed and I took the bait.

Single Spark is horrible, not recommended. Nvidia hinted at this by making the expensive ConnectX port on the back non-optional. You should think of them as being sold in pairs by default, like shoes. If you only have 1, it is owning "half a set" rather than a full unit.

Of course Nvidia has since hiked prices and Apple hasn't yet so that changes the economics of this. But for people who were buying 2 to 3 months ago, the more appropriate comparison would have been a single M5 Max against two Asus GX10.

This is also a little bit of a weird comparison because it at least used to be the case (couple of months ago) that you could buy roughly two Asus GX10 for around the price of a single M5 Max. These had aggregate bandwidth that was in the ballpark of the bandwidth of the Max.

Basically, the more you're using an agent like how people used them in late 2024, the more decode-heavy it is. And the more you're using them like how people use them in late 2026, the more prefill starts to increasingly become important in the mix.