Pennomi

I like to build puzzles except for the last piece, just to prove to myself that finishing most tasks is a societal pressure, not an existential one.

I think it’s just the nodes unfortunately. The map is too handcrafted to randomize easily.

They’re adding some randomization options (node positions and qualities) in 1.2 which is coming out soon.

1.2 is about to come out and it makes vehicles way better.

I’m aware! I’m not saying they are healthy in any way. I’m just correcting that specific misinformation, because truth is important.

These companies are fucked if they keep operating the way they currently are, and I strongly suspect it’s all going to pop like the dotcom bubble, but worse.

I’m not saying they’re healthy, I’m saying that inference is the one profitable part of their business.

They’re all going to die because training costs dwarf the inference, and training doesn’t generate ANY revenue.

The new supercritical CO2 generators are pretty cool. Pretty much the same thing but no water!

We do have numbers from comparably sized Chinese models.

Yes, every AI company is bleeding money, they’re not healthy in any way. But inference by itself is profitable, based on everything that we know.

Inference + amortizing the training costs is NOT profitable, which is what most people are talking about.

This is easily fixed by not releasing a slightly different version every month.

This is a common myth, inference is not typically run at a loss, despite claims. It’s only a loss if you include staff and ongoing training costs. They could lock in their models now and be profitable if they wanted to.

Edit: I see the comment above has changed (or I misread initially) to say the companies are running at a loss rather than inference running at a loss. Yes, that’s extremely true. Now my comment doesn’t make any sense and is irrelevant so feel free to ignore my pedantry.

Hot damn, look at that RAM!

Me too, lil cartoon guy, me too.

I don’t think I ever used a divide symbol like that beyond elementary school. In practice always use fraction style notation for division because it’s not ambiguous or a gotcha.

Not precisely true. Most LLMs (all frontier LLMs) are in fact designed at a fundamental level to increase engagement, using a technique called RLHF (reinforcement learning by human feedback). Essentially whichever responses cause people to use an LLM more are baked into its weights.

Definitely Syndrome

Difficult to do it in a way that is physically consistent with a camera lens/sensor.

I don’t see any of the expected issues with AI (garbled text, impossible geometry, strange anatomy, etc) in this picture. Of course it’s quite possible to just edit a portion of an existing picture with AI, and it will match the rest. So I may have been overstating the difficulty.

Very hard to get an AI to make an image that consistently blurred. Very likely a real photo.

Am I talking more slowly? No, it must be time that’s wrong.

It’s Hal from Megamind for those who don’t know: