3 ms·

Let's actually be generous and assume that all parties involved did the math and some due diligence and are not just idiots. If we try that approach, what could

Let's actually be generous and assume that all parties involved did the math and some due diligence and are not just idiots. If we try that approach, what could that plausibly tell us about a situation where OpenAI has struck deals with not one, but basically all the major chip/infra providers?

dontlikeyoueith 1y ago
> Let's actually be generous and assume that all parties involved did the math and some due diligence and are not just idiots Economic history strongly suggests this would be a bad assumption.
- jstummbillig 1y ago
  How do you mean? Western economic history is, on average, one of success. So on average, that's a pretty good assumption.
  - ben_w 1y ago
    Western economic history is 75% of businesses failing in the first 15 years, and the market still growing because the last 25% has outsized rewards. More pertinently, we have a long history of people buying into bubbles only for them to crash hard, no matter how often people tell them "past performance is not a guarantee of future growth" or whatever the legally mandated phrase is for the supply of investment opportunities to the public where you live. Sometimes the bubbles do useful things before they burst, like the railways. Sometimes the response to the burst creates a bunch of social safety nets, sometimes it leads to wars, sometimes both (e.g. Great Depression).
cmiles8 1y ago
The history of bubbles strongly suggests this is precisely evidence of a bad decision, not a good one. For a bubble to exist and be sustained everyone needs to get on board with things that wouldn’t normally make any sense.
- jstummbillig 1y ago
  See, here the trick is that you assume a bubble and reason from there. But what if, maybe, it ain't so? Of course, lots of AI things are going to fail, and nobody is exactly sure of the future. But what if, after in depth inspection, the overall thing is actually looking pretty good and OpenAI like a winner?
  - ben_w 1y ago
    A trillion dollar valuation for a company losing money does naturally lead to the belief "this is a bubble", that's not really what most call an "assumption" as evidence led to the belief. May be incorrect, but it's not writing down the answer first and working backwards. > But what if, maybe, it ain't so? https://www.youtube.com/watch?v=9z70BKwfSUA https://www.youtube.com/watch?v=9z70BKwfSUA Comedic take from last time, but the point at the conclusion remains. "Just this once, we think we might". > Of course, lots of AI things are going to fail, and nobody is exactly sure of the future. But what if, after in depth inspection, the overall thing is actually looking pretty good and OpenAI like a winner? Much as I like what LLMs and VLMs can do, much as I think they can provide value to the tune of trillions of USD, I have no confidence that any of this would return to the shareholders. The big players are all in a Red Queen's race, moving as fast as they can just to stay at the same (relative) ranking for the SOTA models; at the same time, once those SOTA models are made, there are ways to compress them effectively with minimal losses of performance, and if you combine that with the current rate of phone hardware improvements it's plausible we'll get {state of the art for 2025} models running on-device sometime between 2027 and 2030, with no money going to any model provider.
  - Libidinalecon 1y ago
    In other words, "It's different this time!"

dontlikeyoueith 1y ago

> Let's actually be generous and assume that all parties involved did the math and some due diligence and are not just idiots Economic history strongly suggests this would be a bad assumption.

jstummbillig 1y ago
How do you mean? Western economic history is, on average, one of success. So on average, that's a pretty good assumption.
- ben_w 1y ago
  Western economic history is 75% of businesses failing in the first 15 years, and the market still growing because the last 25% has outsized rewards. More pertinently, we have a long history of people buying into bubbles only for them to crash hard, no matter how often people tell them "past performance is not a guarantee of future growth" or whatever the legally mandated phrase is for the supply of investment opportunities to the public where you live. Sometimes the bubbles do useful things before they burst, like the railways. Sometimes the response to the burst creates a bunch of social safety nets, sometimes it leads to wars, sometimes both (e.g. Great Depression).

cmiles8 1y ago

The history of bubbles strongly suggests this is precisely evidence of a bad decision, not a good one. For a bubble to exist and be sustained everyone needs to get on board with things that wouldn’t normally make any sense.

jstummbillig 1y ago
See, here the trick is that you assume a bubble and reason from there. But what if, maybe, it ain't so? Of course, lots of AI things are going to fail, and nobody is exactly sure of the future. But what if, after in depth inspection, the overall thing is actually looking pretty good and OpenAI like a winner?
- ben_w 1y ago
  A trillion dollar valuation for a company losing money does naturally lead to the belief "this is a bubble", that's not really what most call an "assumption" as evidence led to the belief. May be incorrect, but it's not writing down the answer first and working backwards. > But what if, maybe, it ain't so? https://www.youtube.com/watch?v=9z70BKwfSUA https://www.youtube.com/watch?v=9z70BKwfSUA Comedic take from last time, but the point at the conclusion remains. "Just this once, we think we might". > Of course, lots of AI things are going to fail, and nobody is exactly sure of the future. But what if, after in depth inspection, the overall thing is actually looking pretty good and OpenAI like a winner? Much as I like what LLMs and VLMs can do, much as I think they can provide value to the tune of trillions of USD, I have no confidence that any of this would return to the shareholders. The big players are all in a Red Queen's race, moving as fast as they can just to stay at the same (relative) ranking for the SOTA models; at the same time, once those SOTA models are made, there are ways to compress them effectively with minimal losses of performance, and if you combine that with the current rate of phone hardware improvements it's plausible we'll get {state of the art for 2025} models running on-device sometime between 2027 and 2030, with no money going to any model provider.
- Libidinalecon 1y ago
  In other words, "It's different this time!"