KeplerBoy

It still might, at least in some niches? The paper is not that old.

Can't wait for Gemini 4 Darmstadtium

They actually don't. A lot of people still have no clue how to drive these tools.

It kind of was though? It was knowledge work that brought us fertilizers, tractors, trucks and large ships, which in turn made food as abundant as it is today.

Otherwise most of us would still toil on the fields just to harvest a tiny bit more food than our immediate family needs.

Sure, I was referring to the venue hosting children birthday parties. Big screens and projectors would feel even more dystopian there.

Was the mural handpainted or just a print glued to the wall? I feel even something handdrawn from ai generated inspiration could go a long way to add some quality. Even if just by the fact that no one would spend hours to trace terrible slop by hand.

That's just the nature of (slow) exponential gains. If the processor gains 1% in flops each generation, the absolute flop increase is always the largest ever. Now consider that a device has tens of marketable metrics and that is always a defensible claim.

For which applications do you need that kind of single user throughput? A lot of agentic stuff is tool-heavy anyways other stuff are better served by sub agents / parallel work for more throughput.

You absolutely could though. They could just pause the game in a solo game or detect you are in between matches in a multiplayer game.

Does anyone care about the software features? Every phone and TV has those already.

The decoding job seems to rely on the screenshot.

Isn't pretty much every big indie title on switch?

I wonder if the SD Express will be taken advantage of in the switch 2 homebrew scene which seems inevitable now.

SD Express is PCIe express after all (also a possible attack vector in itself, see DaMAgeCard). With nvidia drivers for Ampere already on board maybe someone will get an RTX 3090 running on a switch 2.

I don't know. I feel like the blog reads as very human written and enjoyable overall. Also what's 2 weeks of real work (in tandem with an agent) compared to a few hours for the blog?

Putting the obvious discussion about how much the AI matters in comparison to the user (experienced security expert from a few glances at the blog) aside; i think it's really neat how modern tooling (the KVMs, bluetooth light switch flicker, and usb controller support out of the box) enables this level of automation.

Really? No fancy auto complete or the occasional discussion about decisions with an LLM?

Nobody works without any AI these days, most of us have access to the same models and yet we all do and achieve different things. It's a weird time.

True, but not much of an issue I think. I'd assume we reach net zero very gradually and fifteen years before net zero will be pretty close to zero already. But oh well, first we need a clear downward trend of global emission.

We have much more than 15 more years of temperature rises to come. It's not like we reached net zero in 2026 or have a plausible plan to reach that anytime soon. Net zero is probably another few decades away.

Ah, yes. I should have done the math before asking.

Isn't electricity a pretty major factor even over such a short fraction of the hardware's lifetime?

You needed approval to try different methods for data analysis?

You can do a lot with a few megawatts of power for a few seconds.

That's still all trivial arithmetic and basic discipline assuming you have enough money to cover all your basic needs.

That of course depends on what you want the model to do. If you are into artistic SVG rendering, it's still very relevant.

Meh, that does nothing for me. Yes, the game seems quite detailed but that's neither especially impressive from a technical standpoint nor interesting from a gameplay perspective.

To be fair that was already the case before ai hit the scene: halo and GTA have relied on the same gameplay for over 20 years by now.

Fair enough, if the code could be promoted in the first place, what's the value anyways?

But coding/development is way more important. That's where the IP goes straight into the next training dataset.

How does it take fifteen minutes to read <100 GB into GPU memory? Shouldn't that be limited by SSD speed with everything slower than a minute being a terrible ssd?