5 ms·

Wrote about this last week [1]. Extrapolating kind of a lot here, but these are my predictions: 1/ We will soon see fully autonomous AI agents operating on th

by non- 3y ago

Wrote about this last week [1]. Extrapolating kind of a lot here, but these are my predictions: 1/ We will soon see fully autonomous AI agents operating on the Internet, running businesses, and pursuing goals set by their human masters with little-to-no oversight. 2/ An AI will commit fraud or a financial crime within the next 6 months and its human creator will be liable (if they can be found). 3/ A platform will be launched within the next 3 months for managing autonomous AI assistants that can handle tasks like coordinating appointments or planning a trip. [1] https://www.boteatbrain.com/p/jarvis-hugginggpt https://www.boteatbrain.com/p/jarvis-hugginggpt

Version467 3y ago
Depending on how you constrain your definition of "little to no oversight", I'd actually bet that 2/ has already happened. Maybe not in a fully autonomous way, but in a supporting capacity to scale up scammer/victim interactions. If I were running one of those Indian call center scams than I'd be jumping at this tech.
- non- 3y ago
  Actually a couple weeks ago a YouTuber[0] showed that you can use ChatGPT to generate Windows 95 activation keys. Not sure if that counts as criminal, but it's edging closer. [0]https://www.youtube.com/watch?v=2bTXbujbsVk https://www.youtube.com/watch?v=2bTXbujbsVk
EGreg 3y ago
“set by their human masters” Why does the human have to be in charge? Once someone sets a swarm of AI bots on a path, it can easily become a grey goo scenario. And even if it isn’t, the bot can simply stop listening to the “master”.
- Szpadel 3y ago
  I think we are way off having ai that's able to replicate itself anywhere, and currently all sensible agents are based on OpenAI API, so you can just kill internet/account/etc and agent is dead therefore I would worry much about any skynet like scenario
  - EGreg 3y ago
    LLaMa is good enough now to self-host a swarm. The real question is whether it can pay for autonomous compute somewhere to keep replicating. What is stopping a LLaMa swarm from overrunning HN? Basically karma points and account ages, that’s about it. But given a year of sleeper swarm members amassing karma in a distributed manner, they can definitely make HN ban all accounts registered after X time similarly to how that sci fi publisher banned independent submissions!
    - Szpadel 3y ago
      I guess they could autonomusly generate spam in comments, but i dont believe they could work in any coordinated manner to achieve some high level goal. (At least until we discover some ground breaking technique to boost performance in task planning for current models) In my testing llama(I tested multiple variations) < gpt3.5 < gpt4 And more and more I use gpt4, I'm starting to notice more and more ways this model fails in spectacular ways. In many cases those models want to generate specific output and if that it's what you want it is very hard to convince it to adjust it, as ofter model will just ignore what you said. Therefore I think any autonomous agent will fairly quickly fail in following it's objectives or it will be very easy to break it when there is any controlled way to inject text into input
turtleyacht 3y ago
> coordinating appointments Asynchronous scheduling of disparate resources to converge on one or more solutions given gradually stricter constraints. > planning a trip Resolving workflow dependencies to a single, favored sequence of actions. Both these primitives sound similar, graph-like. And maybe we could apply them to a wide variety of problems.

by non- 3y ago

Version467 3y ago

Depending on how you constrain your definition of "little to no oversight", I'd actually bet that 2/ has already happened. Maybe not in a fully autonomous way, but in a supporting capacity to scale up scammer/victim interactions. If I were running one of those Indian call center scams than I'd be jumping at this tech.

non- 3y ago
Actually a couple weeks ago a YouTuber[0] showed that you can use ChatGPT to generate Windows 95 activation keys. Not sure if that counts as criminal, but it's edging closer. [0]https://www.youtube.com/watch?v=2bTXbujbsVk https://www.youtube.com/watch?v=2bTXbujbsVk

EGreg 3y ago

“set by their human masters” Why does the human have to be in charge? Once someone sets a swarm of AI bots on a path, it can easily become a grey goo scenario. And even if it isn’t, the bot can simply stop listening to the “master”.

Szpadel 3y ago
I think we are way off having ai that's able to replicate itself anywhere, and currently all sensible agents are based on OpenAI API, so you can just kill internet/account/etc and agent is dead therefore I would worry much about any skynet like scenario
- EGreg 3y ago
  LLaMa is good enough now to self-host a swarm. The real question is whether it can pay for autonomous compute somewhere to keep replicating. What is stopping a LLaMa swarm from overrunning HN? Basically karma points and account ages, that’s about it. But given a year of sleeper swarm members amassing karma in a distributed manner, they can definitely make HN ban all accounts registered after X time similarly to how that sci fi publisher banned independent submissions!
  - Szpadel 3y ago
    I guess they could autonomusly generate spam in comments, but i dont believe they could work in any coordinated manner to achieve some high level goal. (At least until we discover some ground breaking technique to boost performance in task planning for current models) In my testing llama(I tested multiple variations) < gpt3.5 < gpt4 And more and more I use gpt4, I'm starting to notice more and more ways this model fails in spectacular ways. In many cases those models want to generate specific output and if that it's what you want it is very hard to convince it to adjust it, as ofter model will just ignore what you said. Therefore I think any autonomous agent will fairly quickly fail in following it's objectives or it will be very easy to break it when there is any controlled way to inject text into input

turtleyacht 3y ago

> coordinating appointments Asynchronous scheduling of disparate resources to converge on one or more solutions given gradually stricter constraints. > planning a trip Resolving workflow dependencies to a single, favored sequence of actions. Both these primitives sound similar, graph-like. And maybe we could apply them to a wide variety of problems.