Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

The AI bullshit singularity

successfulsoftware.net

Post reply on HN
Print view View original

51–60 of 187 posts

Re: The AI bullshit singularity

#51

Post by 29athrowaway »

hermitcrab wrote: post #17
29athrowaway wrote: post #3

When the ELIZA chatbot came out (1964), some users claimed it was intelligent and sentient. Then, they were shown the diagram that describes its behavior. Even after that, they kept insisting it was sentient, etc. Some people just want to believe bullshit.

When you interact with an LLM (or Eliza) your brain is doing a lot of heavy lifting, without you even realizing it. I think we tend to infer more intelligence than is there, in the same way we see faces in clouds.

Yes. You have made an excellent point.

ELIZA replies to almost everything with questions. So ELIZA is constantly prompting you to do the work required to continue the conversation.

Re: The AI bullshit singularity

#52

Post by sebzim4500 »

smartmic wrote: post #2

I broadly agree with that. Repeated training with self-generated data is the technological equivalent of incest and can lead to nothing good.

Training on self generated data is not necessarily such a problem, see how successful alphazero/muzero is when it is only trained on self play.

The key is that you need some kind of external indicator that tells you which generated examples are good and which are bad. In the case of alphazero you get that by simulating games and seeing who wins, in the case of LLMs you will be only taking the generations that are 'successful', e.g. which HN posts get upvoted.

Re: The AI bullshit singularity

#54

Post by coffeebeqn »

horeszko wrote: post #30

Needle in the haystack. Quality content is the needle, SEO spam is the hay. AI is the hay making machine. Besides curation from a trusted source, what other solutions are there to prevent drowning in hay?

That was already the case. Try to find trustworthy consumer reviews on Google in the last 5 years

Re: The AI bullshit singularity

#55

Post by twobitshifter »

smartmic wrote: post #2

I broadly agree with that. Repeated training with self-generated data is the technological equivalent of incest and can lead to nothing good.

With a mixture of experts approach it’s not incest, we intentionally split the training data to add diversity. Incest is bad because the errors compound on themselves, but this isn’t a problem when the AI parents don’t share the same DNA (problem space.)

I see it this way, maybe there’s no new ‘knowledge’ but the AI can apply our collective knowledge better than we can. Within that set of knowledge there are surely discoveries never yet realized based on by the fusion of ideas.

Pre-AI we rely on individuals Einstein, Bohr, and Oppenheimer and their associations and studies of each others’ work. With AI, we can fuse the corpus of scientific discoveries into a single entity that we each can communicate with. Maybe the AI lacks the spark of creativity needed to make discovery, but put today’s Einstein in front of it and what would he ask? How much boost would it give him?

Einstein said - “I have no special talent. I am only passionately curious.”

Re: The AI bullshit singularity

#56

Post by sebzim4500 »

ronald_raygun wrote: post #4

I always found the idea of infinitely self improving AI to be suspect. Let’s say we have a super smart AI with intelligence 1, and it uses all that to improve itself by 0.5. Then that new 1.5 uses itself to improve by 0.25. Then 0.125, etc etc. obviously it’s always increasing, but it’s not going to have the runaway effect people think.

Do you have some reason to think that will happen?

Re: The AI bullshit singularity

#57

Post by rocqua »

echelon wrote: post #18
jsheard wrote: post #5

The same goes for image generation models, AI art already has a tendency to veer into the same clichés and those are only going to get reinforced if newer models are trained on newer scrapes which now include the million hyper-derivative AI images being uploaded to places like DeviantArt, Twitter and Pixiv every day. Those vendors who got in early have a moat in the form of untainted scrapes, but they'll eventually n…

This is such a wrong take. Even if the data is 100% synthetic, you can still hill climb to new mountains. If you don't believe me, look at evolution. It doesn't matter if we no longer have 100% human art as input. This is the worst these systems will ever look and feel, and they're only going to improve. I'd be willing to do a longbets on this one.

Hill-climbing requires some measure to optimize. current technology uses 'how well can you predict real text' as this measure. If you change the text you try to predict, it changes the measure you are optimizing. It is far from obvious that this will still improve the actual performance.

Re: The AI bullshit singularity

#58

Post by GreedClarifies »

Jesus Christ. If all I did was read hackernews comments I would think that these magical algorithms which have the potential for massive positive change don't exist and that the world is coming to and end.

I get it, engineering trains us to look for failure modes, but my god try to have a little amazement at the progress.

FFS.

Re: The AI bullshit singularity

#59

Post by hermitcrab »

echelon wrote: post #18
jsheard wrote: post #5

The same goes for image generation models, AI art already has a tendency to veer into the same clichés and those are only going to get reinforced if newer models are trained on newer scrapes which now include the million hyper-derivative AI images being uploaded to places like DeviantArt, Twitter and Pixiv every day. Those vendors who got in early have a moat in the form of untainted scrapes, but they'll eventually n…

This is such a wrong take. Even if the data is 100% synthetic, you can still hill climb to new mountains. If you don't believe me, look at evolution. It doesn't matter if we no longer have 100% human art as input. This is the worst these systems will ever look and feel, and they're only going to improve. I'd be willing to do a longbets on this one.

Evolution has a clearly defined fitness function (number of offspring that are produced). What is the fitness function for generated art?

Re: The AI bullshit singularity

#60

Post by valine »

I can read my own writings without overfitting the neurons in my brain. The key I think is contextualization, something LLMs are great at already. The open question is how to utilize that contextualization ability during training.

The argument that LLMs can’t possibly scale because of data contamination falls apart the moment we discover a method to incorporate context-learning into the training loop.

Post reply on HN