Wow, this is a peak bad science reporting headline. I hate to be the one to break the news but no, this is deeply misleading. We all want AI to hit it’s downfall, but these issues with recursive training data or training on small datasets have been near enough solved for 5+ years now. The nature paper is interesting because it explains the modality of how specific kinds of recursion impact broadly across model types, this doesn’t mean AI is going to crawl back into pandoras box. The opposite, in fact, since this will let us design even more robust systems.
I’ve read the source nature article (skimmed though the parts that were beyond my understanding) and I did not get the same impression.
I am aware that LLM service providers regularly use AI generated text for additional training (from my understanding this done to “tune” the results to give a certain style). This is not a new development.
From my limited understanding, LLM model degeneracy is still relevant in the medium to long term. If an increasing % of your net new training content is originally LLM generated (and you have difficulties in identifying LLM generated content), it would stand to reason that you would encounter model degeneracy eventually.
I am not saying you’re wrong. Just looking for more information on this issue.
Ah, to clarify: Model Collapse is still an issue - one for which mitigation techniques are already being developed and applied, and have been for a while. While yes currently LLM content is harder to train against, there’s no reason that must always hold true - this paper actually touches on that weird aspect! Right now, we have to be careful to design with model collapse in mind and work to mitigate it manually, but as the technology improves it’s theorized that we’ll hit a point at which models coalesce towards stability, not collapse, even when fed training data that was generated by an LLM. I’ve seen the concept called Generative Bootstrapping or the Bootstrap Ladder (it’s a new enough concept that we haven’t all agreed on a name for it yet. we can only hope someone comes up with something better because wow the current ones suck…). We’re even seeing some models that are starting to do this coalesce-towards-stability thing, though only in some extremely niche applications. Only time will tell if all models are able to do this stable-coalescing or if it’s only possible in some cases.
My original point though was just that this headline is fairly sensationalist, and that people shouldn’t take too much hope from this collapse because we’re both aware of it, and are working to mitigate it (exactly like the paper itself cautions us to do)
I still find it difficult to get my head around how a decrease in novel training data will not eventually cause problems (even with techniques to work around this in the short term, which I am sure work well on a relative basis).
A bit of an aside, I also have zero trust in the people behind current LLM, both the leadership (e.g. Altman) or the rank and file. If it’s in their interests do downplay the scope and impact of model degeneracy, they will not hesitate to lie about it.
Yikes. Well. I’ll be over here, conspiring with the other NASA lizard people on how best to deceive you by politely answering questions on a site where maaaaybe 20 total people will actually read it. Good luck getting your head around it, there’s lots of papers out there that might help (well, assuming I’m not lying to you about those, too).
AI needs human content and a lot of it, someone calculated that to be good it needs like some extreme amount of data impossible to even gather now hence all the hallucinations and effort to optimize and get by on scraps of semi forged data. Semi forged, artificial data isn’t anywhere close to random gibberish of garbage ai output
Depends on what you do with it. Synthetic data seems to be really powerful if it’s human controlled and well built. Stuff like tiny stories (simple llm-generated stories that only use the complexity of a 3-year olds vocabulary) can be used to make tiny language models produce sensible English output. My favourite newer example is the base data for AlphaProof (llm-generated translations of proofs in Math-Papers to the proof-validation system LEAN) to teach an LLM the basic structure of Mathematics proofs. The validation in LEAN itself can be used to only keep high-quality (i.e. correct) proofs. Since AlphaProof is basically a reinforcement learning routine that uses an llm to generate good ideas for proof steps to reduce the size of the space of proof steps, applying it yields new correct proofs that can be used to further improve its internal training data.
Wow, this is a peak bad science reporting headline. I hate to be the one to break the news but no, this is deeply misleading. We all want AI to hit it’s downfall, but these issues with recursive training data or training on small datasets have been near enough solved for 5+ years now. The nature paper is interesting because it explains the modality of how specific kinds of recursion impact broadly across model types, this doesn’t mean AI is going to crawl back into pandoras box. The opposite, in fact, since this will let us design even more robust systems.
I’ve read the source nature article (skimmed though the parts that were beyond my understanding) and I did not get the same impression.
I am aware that LLM service providers regularly use AI generated text for additional training (from my understanding this done to “tune” the results to give a certain style). This is not a new development.
From my limited understanding, LLM model degeneracy is still relevant in the medium to long term. If an increasing % of your net new training content is originally LLM generated (and you have difficulties in identifying LLM generated content), it would stand to reason that you would encounter model degeneracy eventually.
I am not saying you’re wrong. Just looking for more information on this issue.
Ah, to clarify: Model Collapse is still an issue - one for which mitigation techniques are already being developed and applied, and have been for a while. While yes currently LLM content is harder to train against, there’s no reason that must always hold true - this paper actually touches on that weird aspect! Right now, we have to be careful to design with model collapse in mind and work to mitigate it manually, but as the technology improves it’s theorized that we’ll hit a point at which models coalesce towards stability, not collapse, even when fed training data that was generated by an LLM. I’ve seen the concept called Generative Bootstrapping or the Bootstrap Ladder (it’s a new enough concept that we haven’t all agreed on a name for it yet. we can only hope someone comes up with something better because wow the current ones suck…). We’re even seeing some models that are starting to do this coalesce-towards-stability thing, though only in some extremely niche applications. Only time will tell if all models are able to do this stable-coalescing or if it’s only possible in some cases.
My original point though was just that this headline is fairly sensationalist, and that people shouldn’t take too much hope from this collapse because we’re both aware of it, and are working to mitigate it (exactly like the paper itself cautions us to do)
Thanks for the reply.
I guess we’ll see what happens.
I still find it difficult to get my head around how a decrease in novel training data will not eventually cause problems (even with techniques to work around this in the short term, which I am sure work well on a relative basis).
A bit of an aside, I also have zero trust in the people behind current LLM, both the leadership (e.g. Altman) or the rank and file. If it’s in their interests do downplay the scope and impact of model degeneracy, they will not hesitate to lie about it.
Yikes. Well. I’ll be over here, conspiring with the other NASA lizard people on how best to deceive you by politely answering questions on a site where maaaaybe 20 total people will actually read it. Good luck getting your head around it, there’s lots of papers out there that might help (well, assuming I’m not lying to you about those, too).
This was a general comment, not aimed at you. Honestly, it wasn’t my intention to accuse you specifically. Apologies for that.
AI needs human content and a lot of it, someone calculated that to be good it needs like some extreme amount of data impossible to even gather now hence all the hallucinations and effort to optimize and get by on scraps of semi forged data. Semi forged, artificial data isn’t anywhere close to random gibberish of garbage ai output
Depends on what you do with it. Synthetic data seems to be really powerful if it’s human controlled and well built. Stuff like tiny stories (simple llm-generated stories that only use the complexity of a 3-year olds vocabulary) can be used to make tiny language models produce sensible English output. My favourite newer example is the base data for AlphaProof (llm-generated translations of proofs in Math-Papers to the proof-validation system LEAN) to teach an LLM the basic structure of Mathematics proofs. The validation in LEAN itself can be used to only keep high-quality (i.e. correct) proofs. Since AlphaProof is basically a reinforcement learning routine that uses an llm to generate good ideas for proof steps to reduce the size of the space of proof steps, applying it yields new correct proofs that can be used to further improve its internal training data.
deleted by creator