Divinity video game from Larian - may use AI trained on own assets and to help development

Gorgon Zee said:

(but do look up the % of the US GDP spent on health care first)

Click to expand...

Not really commenting on the rest of your post (I do agree there may be limited use cases where genAI could provide better cost/benefit), but using the amount of money the USA spends on health care as a metric is... not ideal. Almost literally every other developed country spends far less for greater end results for their citizens.

Gorgon Zee said:

but arguments that it is worse quality than humans, reducing quality of life for anyone, or costing significant energy are just plain wrong

Click to expand...

Nothing in your post supports this generalized claim. In particular, absolutely nothing you've said contests the point re: costing significant energy. It seems like you just have no understanding of how AI works, so believe because it's not using much computing power on your end, it isn't period. Which is hilarious if so.

BrokenTwin said:

Not really commenting on the rest of your post (I do agree there may be limited use cases where genAI could provide better cost/benefit), but using the amount of money the USA spends on health care as a metric is... not ideal. Almost literally every other developed country spends far less for greater end results for their citizens.

Click to expand...

Yeah that's genuinely hilarious. The US has healthcare cost "efficiency" that's like, what 4x worse than other first-world countries?

BrokenTwin said:

Not really commenting on the rest of your post (I do agree there may be limited use cases where genAI could provide better cost/benefit), but using the amount of money the USA spends on health care as a metric is... not ideal. Almost literally every other developed country spends far less for greater end results for their citizens.

Click to expand...

I think you might be missing my point. The point is that if you are looking for areas where improvement due to AI would be significant, healthcare is a big example. Your statement is very important and supports my position: healthcare is an area where there is large potential for cost reduction. And GenAI a viable path to doing so.

Ruin Explorer said:

Nothing in your post supports this generalized claim. In particular, absolutely nothing you've said contests the point re: costing significant energy.

Click to expand...

Ruin, I think you may have been a little careless reading my post. I specifically said that if you wanted details, just ask. Please re-read my post with a little more attention.

I’ll address your implied request on energy separately, as i said I would.

Ruin Explorer said:

It seems like you just have no understanding of how AI works, so believe because it's not using much computing power on your end, it isn't period. Which is hilarious if so.

Click to expand...

Ruin, I’m surprised you have no memory of any of the many previous threads on this topic, but if you wish to persist in your belief I have no understanding, I can re-post a link to my most recent paper in NEJM-AI and you can post your credentials. I’d actually prefer not to engage in this diet of behavior, but if you insist …

On energy use: Training models is super-expensive, no disagreement. But per-token use is not as large as most people think.

Studies on GPT 4o ( the model we use most) shows that expected energy use of a 500 token query is about 0.3 watt-hours. Now to help understand what that means in real terms, the average American home uses about 30 kWh a day. So, doing the math, the average house uses the same energy per day as about 100,000 500-token queries

The studies people are using also assumes the use of the old H100 chips. As you probably know, the newer chips are more efficient. I’m not sure exactly how much. But based on pricing, probably 2x.

So a good estimate of how much energy is used is: 200k queries is about the same as a house.

Is that a lot? Well, it really doesn’t seem like it to me, but YMMV. If each of these queries saves a nurse 5 minutes of boring data entry, a house day worth of energy saves nearly 250 hours - or 30 nurse days.

So it’s really a question to you - do you think saving 30 days of tedious data entry is worth a day’s worth of house energy?

Gorgon Zee said:

So a good estimate of how much energy is used is: 200k queries is about the same as a house.

Is that a lot?

Click to expand...

Objectively yes.

That isnt even that much usage of LLMs to be able to compare to the energy usage of an average home.

And the larger numbers remain. Total energy usage has skyrocketed since LLMs started being pushed.

I was feeling punchy, so I asked AI how much energy it was using:

Gorgon Zee said:

Ruin, I’m surprised you have no memory of any of the many previous threads on this topic, but if you wish to persist in your belief I have no understanding, I can re-post a link to my most recent paper in NEJM-AI and you can post your credentials. I’d actually prefer not to engage in this diet of behavior, but if you insist …

Click to expand...

This only makes your position drastically worse, because it means you must necessarily aware of how you're not being accurate at all re: datacenter energy usage. I don't see how you think this helps your position.