The compute cost for DeepSeek V3 was $5.6M. Training costs for traditional US frontier labs model is about $70 to $200M+, with estimates of up to $1B+ to train the next generation models. It's expensive, but it's not even in the ballpark of "There is no viable economic model to do so." Google revenue in Q2 was $1.32 billion *per day*. $1B is trivially less than a day; $5.6M is just 6 minutes of revenue.

The cost problem is in the compute buildout and the subsidized compute to capture users.

> It remains to be seen if even using their models can be profitable,

I presume we can stipulate there exists some tasks that are cheaper to do using an LLM than doing it "by hand". As a such, using an LLM model can be profitable. To me, it is also obvious that the business of serving an LLM model can be profitable. There's large benefits to being able to get tokens quickly, and you can expect people to pay for that, even if we presume they could run the model locally with no maintenance overhead. There's even cases where using the cloud is cheaper than you can serve locally at all - I saw one guy that measured his added cost of electricity when his computer ran a specific LLM, and it was more expensive to run it locally than to buy tokens from the cheapest inference provider for the same model, presumably because they had better hardware and cheaper electricity.

I love a term a friend of mine used to use when talking about AI: "Narrowly superhuman" (before "superintelligence" got to be a thing).

AI is narrowly superhuman. It is able to do specific things that used to be defined as a "hallmark of intelligence" much better than a human. The easiest one? Summing numbers. It's both faster and more precise than a human.

AI gets more and more such areas it can do better than a human.

If we define "Superintelligent" to be "A computer system able to do absolutely everything we can somehow call intelligence better than any human", we will *never* have superintelligence. The things we define as intelligence will be in conflict; we'll land at the system being better at one at the cost of another.

I'm in favor with being careful about this development. But "prohibiting the development of superintelligence" isn't an option. It's already here.

Dividing the yearly US spend on Internet advertising by the number of people in the US that has Internet access gives $900-$970 spend per person per year, or $75-$81 per month.

Please choose the option you want:

1. I want to pay about $75 per month per member in my family into the ad auctions to not see any ads. Ie, keep the infrastructure, but let you opt to pay your share through money rather than attention.
2. I want to prohibit ads, and accept that everybody will have to pay for the sites they want access to, that there will be much fewer services, that there will be much higher transaction costs, and that the poor will have much less ability to access services. (Microtransactions won't solve this, due to the need for trust evaluation of same.)
3. I want some combination of 1 and 2, where some sides of ads are prohibited so it is possible to buy my way out of them at a lower price,. I accept that this will result in the loss of some services and lower access to others for the poor.
4. I want legislation that makes Internet ad blockers work more consistently. I accept that this just shifts the attention cost to other people, and have me getting the benefit without paying the cost.
5. I want to remove Internet ads and subsidize the sites through taxes.
6. I want to increase the requirements of Trust and Safety checks and format restrictions for ads, potentially accepting higher overhead (in what goes to content vs infrastructure) for less annoying/bad ads. I accept that doing this creates significant advantages of scale for ad networks.
7. Other, please describe, including where costs gets shifted with your proposal.

The above are my best attempt at listing possible options and their correct caveats. In most situations, I find ads annoying (who doesn't). Unfortunately, they are the main money highway of Internet content.

DISCLOSURE: I used to work for Google, where I rang infrastructure for stopping bad ads/advertisers and for removing fake clicks. I've also been interested at the use of sub-cent microtransactions to pay for content since the 1990s, though I've become disillusioned as I've learned more about the attention economics involved. Approving individual microtransactions has a higher attention cost than dealing with the ads, and bulking them per-site (through micro-subscriptions) will limit the number of sites that can have content.

There's quite a bit of hyperbole and assumption in what you're saying - even from enthusiasts getting Qwen model running locally "90%" of frontier is a bold statement given the capabilities of actual frontier models. I don't think you're pricing in demand - would demand decrease from cheaper models or increase? You seem to imply that we're at a saturation point from demand, but I feel this couldn't be further from the truth. There's a lot of integration points and a lot of huge benefits a 100x model could produce in terms of software use cases - we aren't really seeing peak usage or anywhere near peak usage there. The tools are too expensive right now for everyone to maximize use. If the price crashes I think you just see more usage taking up the slack.

I'm right there with you though that there are going to be winners and losers, but certainly not on the timescale you're predicting and certainly not with some of the big HW players.

Ok, lets break this down for you. The previous poster was complaining that they went to California twice and said it was a third world slum. I told them to stay out of areas like Tahoe, Mendicino, Carmel and Yosemite which are not generally known as third world slums and directed them to places that obviously could be slummy like an overpass or the afore mentioned walk of fame, which IS shitty and THAT is the joke. And now that I've explained it, its that much funnier! /S

Sounds like a feature not a bug for some.

God damn people really do need the "/s". Am I in the wrong? No...no...they're the ones that need to change!

Let me guess? Mendicino and Lake Tahoe right? Or Carmel and Yosemite? See there's your problem, you went to the shittiest places. I suggest one of the many interstate interchanges or perhaps the Hollywood walk of fame. That's the "real" California

Waymo operates just fine in San Francisco which has plenty of very narrow tight streets (2 way with 1 lane) with bad visibility and lots of extra shitty drivers, double parkers etc. It goes around these just fine, and when it gets legit stuck on something (road closed off by cops or something) it can dial home to have an operator give instructions on what it should do next.

Games haven't run off disc for a while, they install to the drive. Even HDD loading speeds are intolerable in a lot of games, forget about optical.

Discs have only been a distribution mechanism now for 2+ generations. You don't run the game off the disc it just installs it to the HD. (Then layers in a 12 GB patch anyways). Two sets of people benefit: those that want to resell the title and those that don't want the gigantic download. The no download crew is pretty much lost as you generally need "day 0" patch to make the game anything but a buggy mess, so this is 100% about resale.

I can think of a few uses, such as being able to read a document or something on the go. My phone is too small, and I don't want to bring a laptop or tablet with me. But I think a six inch screen is too small still, so yeah, not sure what people are using this version for.

It doesn't matter what they say. Whenever the Apple version launches, that will be when a company like Apple was brave enough to launch such an innovative device that no one else ever thought to do. Truly they are the inventive minds of our time, and we should be so lucky to get an iFold of our own, on sale now for $3500.

OK, which particular aspect of the LLM learning do you think doesn't exist in the human brain? Backprop exists (and the other training methods the brain uses has similar effects), embeddings more or less exists and seems to be very similar to the LLM embeddings (shown by brain activation in various areas linearly matching with embeddings) - see e.g. High-level visual representations in the human brain are aligned with large language models (Adrien Doerig et al, 2025), and speech/text generation by humans has similar error patterns as for LLMs.

They obviously don't learn in the EXACT same way, but there's also clearly a lot of bits that are the same, and I'm fairly sure nobody knows how much is similar yet, particularly not weighted by importance.
 

"Its legal when the president gets his taste."