4 ms·

Will this still happen when humans are in the loop? Isn't this kind of like running an evolutionary algorithm on the models? - AI generates 20 images - A hum

Will this still happen when humans are in the loop? Isn't this kind of like running an evolutionary algorithm on the models? - AI generates 20 images - A human picks out the best 2 images to publish to the internet. - Eventually the 2 best images end up in the training data of a new model, while the 18 bad images get discarded. - Successive generations of models bias more towards the better images. This might get rid of some of the defects that plague the worse images. Humans end up being the fitness function and even the results from different models ends up mixed together! I'm not sure that this would be a big problem for images, because there usually isn't an emphasis on correctness and validity of information. It seems more of a problem for LLMs.

ben_w 3y ago
Are humans still in the loop? 1. AI (say ChatGPT or whatever LLM you like) makes blogspam 2. AI (Google Pagerank etc.) fooled into high-ranking it 3. AI (unnamed scammer) searches for a suitable comment form to post spam on, pretends to be real human as it does so 4. AI (Google Ad network or similar) selects best bid for an advert to show to the fake-user in #3 4. a. AI (Amazon, eBay, etc) has meanwhile been placing bids: dictionary.filter($0.isNoun)).map { "\($0): Buy it cheap on Amazon!" } 5. AI (products scammers) look at "best" opportunities on Amazon, create fake listings to match each noun in 4. a. 6. AI (same scammers) upvote the fake reviews for the fake products for the fake nouns advertised to fake users on fake blogs, but the metrics… which are A/B tests… which is GOFAI (AKA applied statistics)… say it's good, so everyone keeps doing it.
- 6510 3y ago
  still some humans in the click farm
- birdyrooster 3y ago
  I like how you are thinking and I wondered if the company operates as a monopoly or is in a very small market, people won’t have alternatives to pick from and so their demand inelasticity combined with AI will lead to bad choices about A/B tests. I assume this is already problematic without AI, as it seems many company’s don’t know what made them successful as they ascend and A/B tests lie to them. When their product is in hot demand they think every stinker is a rose by misinterpreting the noise.
bhouston 3y ago
> Will this still happen when humans are in the loop? Yes it will. Unless the humans are generating new high fidelity content.
- SketchySeaBeast 3y ago
  Even if humans do generate high fidelity content would it help when it constitutes a miniscule sliver of all new content?
  - stOneskull 3y ago
    in many fields, it'd be the largest sliver. perhaps, in general, it'd be fiction collapses and non-fiction grows?
    - slowmovintarget 3y ago
      In some senses that will be the greatest damage. We live our lives in the fictions we share, using them as touchstones for how we talk to each other. Generative AI is the genie we've let out of the bottle, and we keep making wishes.
svieira 3y ago
The invalid presumption is "A human picks out the best 2 images". People will absolutely stop looking at these in any meaningful way (or even produce them in a direct-to-the-internet pipeline)

ben_w 3y ago

Are humans still in the loop? 1. AI (say ChatGPT or whatever LLM you like) makes blogspam 2. AI (Google Pagerank etc.) fooled into high-ranking it 3. AI (unnamed scammer) searches for a suitable comment form to post spam on, pretends to be real human as it does so 4. AI (Google Ad network or similar) selects best bid for an advert to show to the fake-user in #3 4. a. AI (Amazon, eBay, etc) has meanwhile been placing bids: dictionary.filter($0.isNoun)).map { "\($0): Buy it cheap on Amazon!" } 5. AI (products scammers) look at "best" opportunities on Amazon, create fake listings to match each noun in 4. a. 6. AI (same scammers) upvote the fake reviews for the fake products for the fake nouns advertised to fake users on fake blogs, but the metrics… which are A/B tests… which is GOFAI (AKA applied statistics)… say it's good, so everyone keeps doing it.

6510 3y ago
still some humans in the click farm
birdyrooster 3y ago
I like how you are thinking and I wondered if the company operates as a monopoly or is in a very small market, people won’t have alternatives to pick from and so their demand inelasticity combined with AI will lead to bad choices about A/B tests. I assume this is already problematic without AI, as it seems many company’s don’t know what made them successful as they ascend and A/B tests lie to them. When their product is in hot demand they think every stinker is a rose by misinterpreting the noise.

bhouston 3y ago

> Will this still happen when humans are in the loop? Yes it will. Unless the humans are generating new high fidelity content.

SketchySeaBeast 3y ago
Even if humans do generate high fidelity content would it help when it constitutes a miniscule sliver of all new content?
- stOneskull 3y ago
  in many fields, it'd be the largest sliver. perhaps, in general, it'd be fiction collapses and non-fiction grows?
  - slowmovintarget 3y ago
    In some senses that will be the greatest damage. We live our lives in the fictions we share, using them as touchstones for how we talk to each other. Generative AI is the genie we've let out of the bottle, and we keep making wishes.

svieira 3y ago

The invalid presumption is "A human picks out the best 2 images". People will absolutely stop looking at these in any meaningful way (or even produce them in a direct-to-the-internet pipeline)