Closet_Gamer
Thinks s/he gets paid by the post
I've spent a lot of time digging into Generative AI, the new AI that came into view with the launch of Chat GPT last year. I'm not an expert but I would say that I am very well informed.
A few thoughts:
Generative AI is fundamentally different than the pattern matching AI we've all been using for a long time. There was a paper written by Google researchers in 2017 that described an new technique for creating and using neural networks that made them capable for creating rather than just discerning content. This new tech has been bubbling away at OpenAI, Google and other places for that last several years.
The second thing to know is that the scale and complexity underlying these models is completely new. ChatGPT 4 is rumored to have 1 Trillion tokens inside of it. This has been built off the back of $10Bs of investment across these companies...perhaps pushing towards $100B collectively invested now. They can leverage all of the existing advances in cloud techniques to scale quite rapidly. OpenSource techniques are also accelerating development.
A third thing is that the silicon now being employed is AI specific. Originally it leveraged GPUs because advanced graphics and AI both rely on matrix math techniques. Nvidia and others are now building Generative AI-specific chips. The largest version of Nvidia's H100 can supposedly push upwards of a trillion tokens in a single pack of their chips (you can bolt 8 of them together). They already have their next gen chip heading to market.
The technology continues to suffer from hallucinations -- the propensity to invent and quite confidently provide incorrect information. They are working to make this better via search augmentation, referencing articles and training on more focused information sets ... but it is still a real problem. Its unclear whether the core tech can be changed to address this issue or if we will see a series of hacks/patches/protections used to minimize the issue. Regardless, Generative AI cannot be trusted right now to be correct on its own. It will confidently invent citations and references to papers that don't even exist. (My daughter saw it take a real author's name but invent a paper that person hadn't written and didn't exist.)
The technology is also super power hungry with no end in sight to that. And hugely expensive. Right now people are doing wildly unprofitable things to try and own the lead in this. Will be interesting to see how long that lasts.
Finally the copyright litigation is a fundamental fight and it will determine where a lot of the value creation resides. If the courts rule that models can be "trained" on whatever they find on the Internet, that will be a real blow to content creators everywhere.
I believe Generative AI is as a fundamental advancement and lots of fits-and-starts but huge impact. The deep fake issue - and the related flood of AI developed content - will be significant and will add lots of clutter. I read a Yahoo! article the other day that quietly disclaimed at the bottom that it had been written by AI ... potentially with halluciations.
My $0.05 on the issue. Hopefully helpful.
A few thoughts:
Generative AI is fundamentally different than the pattern matching AI we've all been using for a long time. There was a paper written by Google researchers in 2017 that described an new technique for creating and using neural networks that made them capable for creating rather than just discerning content. This new tech has been bubbling away at OpenAI, Google and other places for that last several years.
The second thing to know is that the scale and complexity underlying these models is completely new. ChatGPT 4 is rumored to have 1 Trillion tokens inside of it. This has been built off the back of $10Bs of investment across these companies...perhaps pushing towards $100B collectively invested now. They can leverage all of the existing advances in cloud techniques to scale quite rapidly. OpenSource techniques are also accelerating development.
A third thing is that the silicon now being employed is AI specific. Originally it leveraged GPUs because advanced graphics and AI both rely on matrix math techniques. Nvidia and others are now building Generative AI-specific chips. The largest version of Nvidia's H100 can supposedly push upwards of a trillion tokens in a single pack of their chips (you can bolt 8 of them together). They already have their next gen chip heading to market.
The technology continues to suffer from hallucinations -- the propensity to invent and quite confidently provide incorrect information. They are working to make this better via search augmentation, referencing articles and training on more focused information sets ... but it is still a real problem. Its unclear whether the core tech can be changed to address this issue or if we will see a series of hacks/patches/protections used to minimize the issue. Regardless, Generative AI cannot be trusted right now to be correct on its own. It will confidently invent citations and references to papers that don't even exist. (My daughter saw it take a real author's name but invent a paper that person hadn't written and didn't exist.)
The technology is also super power hungry with no end in sight to that. And hugely expensive. Right now people are doing wildly unprofitable things to try and own the lead in this. Will be interesting to see how long that lasts.
Finally the copyright litigation is a fundamental fight and it will determine where a lot of the value creation resides. If the courts rule that models can be "trained" on whatever they find on the Internet, that will be a real blow to content creators everywhere.
I believe Generative AI is as a fundamental advancement and lots of fits-and-starts but huge impact. The deep fake issue - and the related flood of AI developed content - will be significant and will add lots of clutter. I read a Yahoo! article the other day that quietly disclaimed at the bottom that it had been written by AI ... potentially with halluciations.
My $0.05 on the issue. Hopefully helpful.