Comment by alexpotato
1 day ago
I was in college in the late 1990s/early 2000s and I distinctly remember an econometrics professor state the following:
"As cable TV and Pay Per View came out, there were studies done about how many movies people would watch if given unlimited access to films. The results were bandied about as proof that we should build out all this infrastructure to support this line of business. When the data was further analyzed by statisticians etc, it turned out that people claimed they were going to watch films 10-12 hours a day, every day of the week. Impossible."
I feel like we are in a similar boat here where some people are assuming:
- EVERYONE is going to be using max tokens
- tokens will NEVER get cheaper due to improvements in hardware, software, design, market forces etc etc
>I feel like we are in a similar boat here where some people are assuming: >- EVERYONE is going to be using max tokens >- tokens will NEVER get cheaper due to improvements in hardware, software, design, market forces etc etc
I feel like the reverse assumption is being made, that the current model looks like IBM doubling down on Mainframes soon to become cheap enough to deploy everywhere, when the real action is that the costs coming down represents cheaper hardware or more efficient software, and that a big chunk of "cheaper" AI will be eaten by smaller products deployed by individuals. Whatever the Personal Computer of AI looks like is going to be more disruptive than just an API endpoint you can fling tokens at.
We already see this with things like chrome auto installing an LLM.
You cant tell me with complete certainty that theres a moat here for the people spending 1 trillion + on this infra.
>When the data was further analyzed by statisticians etc, it turned out that people claimed they were going to watch films 10-12 hours a day, every day of the week. Impossible.
I also think this applies to people suggesting that companies will sack workers for AI, when the costs of replacing everything someone does in a day is more expensive in terms of tokens (likely even at a reduced price) than just hiring a bloke.
> it turned out that people claimed they were going to watch films 10-12 hours a day, every day of the week. Impossible."
I realized it long ago: one needs output to make meaning. Input can only be the cherry on a cake in one's life. That, actually, makes FIRE or Fat FIRE not so sustainable unless one has other hobbies.
> it turned out that people claimed they were going to watch films 10-12 hours a day, every day of the week. Impossible.
And what happened? How many hours per day/week are people spending watching now?
What people: I'm sure some people are watching 10-12 hours per day - in places like nursing homes or hospitals. I know a reasonable number of people who watch a film nearly every day: 2-3 hours. Most people watch something every few days - often a tuesday night movie night for the family (or something like that). There are some who never watch anything. I don't know what the statistics are on this.
My friends in day care tell me the kids hate "movie day" because movies are all the get at home and they are sick of them - they want to play all day. (but I'm not sure if this is representative of anything other than the types of people who put their kids in that particular daycare)
> they were going to watch films 10-12 hours a day, every day of the week. Impossible.
A lot of these LLM demand scaling scenarios make broad "up and to the right" assumptions about things which in practice have finite limits. Only some percentage of knowledge work benefits from acceleration, optimization or other improvements, and even then the amount of economic gain is capped.
But isn't it wonderful that they did?
It's vaguely disturbing that people "watch" films 10-12 hours a day. Many of them are using it as a radio, for background noise, without really caring what the program is beyond vague genre, tuning in and out without particular regard to the plot… and yet we have all the cost of transmitting high-resolution video point-to-point.
Surely we could just put better stuff on the radio, and accomplish most of the same goals for a far lower price?
My Dad was in the hospital, and just wanted to watch the Pirates play. The TV was filled with apps, some of them free to watch, others demanding a subscription and log in once you selected something.
None of them had the Pirates game.
I was thinking how the transistor radio was a far superior experience for this use case. Just tune to the channel broadcasting the game.
1 reply →
Radio has not gone anywhere you know? There is of course podcasts, but for instance Radio France has amazing music services like FIP: https://www.radiofrance.fr/fip
Then there’s NTS, BBC… Ypu can listen to them from online service, but at least in Europe there’s amazing national FM broadcastimg services.
TV is just bad radio with flickerimg lights.
Who has the time to watch films 10-12 hours a day?
I think the comment put forward that as an incorrect assumption that was made prior to the cable build-out.
2 replies →
> - EVERYONE is going to be using max tokens
anthropic already hunts down OpenClaw users for using too much on their plan.
I'll give different example: When LED lights started to be more popular, the power usage didn't drop by the amount of power saved
>- tokens will NEVER get cheaper due to improvements in hardware, software, design, market forces etc etc
Well, first, improvements in computing stalled or even rolled back just purely because price of everything compute shot up cos of AI and that will NOT be fixed for a while and ESPECIALLY if AI usage will continue to increase
Second, the token per model might go down in time but better models have more expensive tokens, so we quickly get into spot when:
* price increase in token might not be worth marginal improvement next, better model brings
* more and more models are passing "good enough for the task" threshold so for less and less companies there is any economic sense to pay for the "best" instead of paying deepseek or some other company to run "previous gen" models