A bit ironic considering that Yamamoto himself never got a chance to play in Koshien.

It reminds me of the NASA Commercial Crew/Cargo contract model. The government footed the initial development cost and bought a supply agreement to use the service itself, a “subscription” if you will. Cygnus and Dragon were developed with taxpayer funds and yet the government doesn’t own the IP.

Though President Jimmy Carter tried to this with the F-16/79 variant.

Not really.

Their low compute version satisfies the <10k cost cap for the public leaderboard. The public leaderboard is separate from the main prize and allows closed models.

So their “official” public leaderboard score is 75.7 % which is still much higher than any other models.

The public leaderboard which allows closed models had an API cost cap of <10k.

OpenAI’s o3 got 75.7 % with this requirement so this is the official score on the leaderboard.

OpenAI then tried the high compute version to target the 85% goal and they spent millions to get 87.5 % on that.

There are several tracks. To be eligible for the 1 million dollar grand prize fund you had to satisfy the open source and compute requirement.

So the open source requirement is only relevant for the prize money.

OpenAI targeted the public leaderboard which allows closed models to participate.

Modern armor would also need to deal with shrapnel as in peer-to-peer conflicts shrapnel from artillery generally kills more people that firearms.

Interesting bifurcation seems to be SOF more focusing on firearms and mobility while regular infantry also needs to protect against shrapnel as well.

I also believe that another example is that when using LPIPS loss for something like VAE you need adversarial loss as well if you don’t want artifacts.

x.com/madebyollin/...

Concept of adversarial loss as well as VQ-GAN seems makes GAN a pretty important paper.

bsky.app/profile/sedi...