3 ms·

> The economic value of mediocre and derivative work is plummeting. But the value of truly excellent and original work will become even more unique, more highly

by chewz 4y ago

> The economic value of mediocre and derivative work is plummeting. But the value of truly excellent and original work will become even more unique, more highly prized. > A chasm between the third best and fourth best script writer, graphic designer, lawyer, etc. is just starting to open. ChatGPT might write the next Marvel movie but will never write the next Everything Everywhere All at Once. Not my words. By Eric Peters CIO of One River Asset Management

nightski 4y ago
So a race to the bottom. I wouldn't rely on the "truly excellent and original work" crutch for very long.
Al-Khwarizmi 4y ago
I think it may turn out to be like that but I don't see it as guaranteed. How can we be so sure that GPT-8 will not write things that we consider to be really original and unique? I know, it's trained with human data and its output is based on that. The same can be said of human authors, though. Arguably we never create an idea from scratch, originality is just putting ideas together in really non-standard ways.
- lanstin 4y ago
  Useful non standard ways. I think trying to make a context free intelligent system is making it excessively hard for the AGI folks. Our brain definitely figures out what to do for pre-sentient reasons, or using pre-sentient hardware.
Retric 4y ago
Except they can’t actually write the next Marvel movie today. I suspect AI getting good enough to create a Marvel movie is basically indistinguishable from it being able to create a good movie. The problem is LLM’s are basically just creating gibberish which mimics meaningful text well enough you don’t notice it’s meaningless as long as it’s a short enough sample. Modern AI techniques are approaching human level capacity by combining a lifetime worth of information in their training sets with nearly human levels of computation. It’s not at the level of professionals, but being roughly as good as a teenager means crossing that threshold could happen well before people are ready for it.
- lanstin 4y ago
  Are you sure about the nearly human levels of computational? Even at a trillion or so, I thought we still had several orders of magnitude before the free parameters approached anything like the (currently known) information density of an adult brain?
  - Retric 4y ago
    Yea, I mean close on a logarithmic scale though it’s not easy to make direct comparisons. A low end estimate of say 10 cycles per neuron / second * 90 billion neurons * 10,000 connections per neuron is already 90 peta operations per second. Of course we also expect AI’s to be faster but conversely we aren’t trying to simulate a person that needs to walk and talk, just a rough equivalent of whatever is needed for some task.
- aezart 4y ago
  "Short enough sample" is definitely key. With a small model like llama 13B, you can watch the output degrade in real time as it gets longer. When the only thing in its context is 1000 words of human prompt, the next LLM-predicted sentence will probably be acceptable. But by the time it's added a few hundred words of its own output to the context, it's going downhill fast. Bigger models slow the rate of degradation, but that's it.

by chewz 4y ago

nightski 4y ago

So a race to the bottom. I wouldn't rely on the "truly excellent and original work" crutch for very long.

Al-Khwarizmi 4y ago

I think it may turn out to be like that but I don't see it as guaranteed. How can we be so sure that GPT-8 will not write things that we consider to be really original and unique? I know, it's trained with human data and its output is based on that. The same can be said of human authors, though. Arguably we never create an idea from scratch, originality is just putting ideas together in really non-standard ways.

lanstin 4y ago
Useful non standard ways. I think trying to make a context free intelligent system is making it excessively hard for the AGI folks. Our brain definitely figures out what to do for pre-sentient reasons, or using pre-sentient hardware.

Retric 4y ago

Except they can’t actually write the next Marvel movie today. I suspect AI getting good enough to create a Marvel movie is basically indistinguishable from it being able to create a good movie. The problem is LLM’s are basically just creating gibberish which mimics meaningful text well enough you don’t notice it’s meaningless as long as it’s a short enough sample. Modern AI techniques are approaching human level capacity by combining a lifetime worth of information in their training sets with nearly human levels of computation. It’s not at the level of professionals, but being roughly as good as a teenager means crossing that threshold could happen well before people are ready for it.

lanstin 4y ago
Are you sure about the nearly human levels of computational? Even at a trillion or so, I thought we still had several orders of magnitude before the free parameters approached anything like the (currently known) information density of an adult brain?
- Retric 4y ago
  Yea, I mean close on a logarithmic scale though it’s not easy to make direct comparisons. A low end estimate of say 10 cycles per neuron / second * 90 billion neurons * 10,000 connections per neuron is already 90 peta operations per second. Of course we also expect AI’s to be faster but conversely we aren’t trying to simulate a person that needs to walk and talk, just a rough equivalent of whatever is needed for some task.
aezart 4y ago
"Short enough sample" is definitely key. With a small model like llama 13B, you can watch the output degrade in real time as it gets longer. When the only thing in its context is 1000 words of human prompt, the next LLM-predicted sentence will probably be acceptable. But by the time it's added a few hundred words of its own output to the context, it's going downhill fast. Bigger models slow the rate of degradation, but that's it.