5 ms·

ChatGPT/GPT3.5 and its weights can fit on a small thumb drive, and copied infinitely and shared. Tech will get better enough in the next decade to make this acc

ChatGPT/GPT3.5 and its weights can fit on a small thumb drive, and copied infinitely and shared. Tech will get better enough in the next decade to make this accessible to normies. The genie cannot be put back in the bottle.

thefaux 4y ago
True. In the long run though, I expect we will either build something dramatically better than these models or lose interest in them. Throw in hardware advances coupled with bitrot and I would go short on any of the gpt-3 code being available in 2123 (except in something like the arctic code vault, which would likely be effectively the same as it being unavailable).
- vidarh 4y ago
  The point isn't that gpt-3 specifically will be available, or the current models, but that gpt-3 level models or better will be available.
Spooky23 4y ago
Sure it can. Missile guidance systems fit on a tiny missile, but you can’t just get one. The controlled parlor game is there to seed acceptance. Once someone is able to train a similar model with something like the leaked State Department cables or classified information we’ll see the risk and the legislation will follow.
- airtonix 4y ago
  [dead]
- Karunamon 4y ago
  They can try. You will note that nobody except government employees and the guy running the website ever got in trouble for reading cables or classified information. We have the Pentagon papers precedent to the effect of it being a freedom of speech issue.
  - soco 4y ago
    The persons from the state dept or army are heavily vetted to get there. A normie with a thumb drive, less so...
- vidarh 4y ago
  Once someone is able to train a similar model on their own, it's too late for legislation to have any meaningful ability to reduce proliferation.
flangola7 4y ago
> ChatGPT/GPT3.5 and its weights can fit on a small thumb drive, and copied infinitely and shared. So can military and nuclear secrets. Anyone with uranium can build a crude gun-type nuke, but the instructions for making a reliable 3 megaton warhead the size of a motorcycle have been successfully kept under wraps for decades. We also make it very hard to obtain uranium in the first place. >Tech will get better enough in the next decade to make this accessible to normies. Not if future AI research is controlled the same way nuclear weapon research is. You want to write AI code? You'll need a TS/SCI clearance just to begin, the mere acting of writing AI software without a license is a federal felony. Need HPC hardware? You'll need to be part of a project authorized to use the tensor facilities at Langley. Nvidia A100 and better TPUs are already export restricted under the dual-use provisions of munition controls, as of late 2022.
- astrange 4y ago
  How are you going to ban the Transformer paper? It's just matrix multiplies.
  - vidarh 4y ago
    It's also a first amendment issue, and already out there. Reminds me that I'm old enough to remember when PGP bypassed export control by being printed on paper and exported as books and scanned/typed back in, though. They can of course restrict publishing of new research, but that won't be enough to stop significant advances just from the ability of private entities worldwide to train larger models and do research on their own.

thefaux 4y ago

True. In the long run though, I expect we will either build something dramatically better than these models or lose interest in them. Throw in hardware advances coupled with bitrot and I would go short on any of the gpt-3 code being available in 2123 (except in something like the arctic code vault, which would likely be effectively the same as it being unavailable).

vidarh 4y ago
The point isn't that gpt-3 specifically will be available, or the current models, but that gpt-3 level models or better will be available.

Spooky23 4y ago

Sure it can. Missile guidance systems fit on a tiny missile, but you can’t just get one. The controlled parlor game is there to seed acceptance. Once someone is able to train a similar model with something like the leaked State Department cables or classified information we’ll see the risk and the legislation will follow.

airtonix 4y ago
[dead]
Karunamon 4y ago
They can try. You will note that nobody except government employees and the guy running the website ever got in trouble for reading cables or classified information. We have the Pentagon papers precedent to the effect of it being a freedom of speech issue.
- soco 4y ago
  The persons from the state dept or army are heavily vetted to get there. A normie with a thumb drive, less so...
vidarh 4y ago
Once someone is able to train a similar model on their own, it's too late for legislation to have any meaningful ability to reduce proliferation.

flangola7 4y ago

> ChatGPT/GPT3.5 and its weights can fit on a small thumb drive, and copied infinitely and shared. So can military and nuclear secrets. Anyone with uranium can build a crude gun-type nuke, but the instructions for making a reliable 3 megaton warhead the size of a motorcycle have been successfully kept under wraps for decades. We also make it very hard to obtain uranium in the first place. >Tech will get better enough in the next decade to make this accessible to normies. Not if future AI research is controlled the same way nuclear weapon research is. You want to write AI code? You'll need a TS/SCI clearance just to begin, the mere acting of writing AI software without a license is a federal felony. Need HPC hardware? You'll need to be part of a project authorized to use the tensor facilities at Langley. Nvidia A100 and better TPUs are already export restricted under the dual-use provisions of munition controls, as of late 2022.

astrange 4y ago
How are you going to ban the Transformer paper? It's just matrix multiplies.
- vidarh 4y ago
  It's also a first amendment issue, and already out there. Reminds me that I'm old enough to remember when PGP bypassed export control by being printed on paper and exported as books and scanned/typed back in, though. They can of course restrict publishing of new research, but that won't be enough to stop significant advances just from the ability of private entities worldwide to train larger models and do research on their own.