EleutherAI executive director Stella Biderman singled out CSAM generation as the clearest case where closed-source models hold a safety edge over open ones, since the latter can be more readily adapted for that misuse; she framed it as the sole unambiguous example she knows of in current debates.
@xeophon The answer to your question is CSAM generation, by the way. CSAM is an unambiguous win for closed source being safer. It’s the only example I know.
I have notes on why the WormGPT stuff was a total fraud, if that’s helpful to include too.
yeah, looking further into this and it seems like open models are safer. blog about this later
Biderman also pointed to compiled documentation that services like WormGPT amounted to outright scams rather than functional malicious models.
The claim arrives from a leading open-source researcher amid broader discussion, where evidence for unique open-model risks remains narrower outside this specific domain.
Some users agree open source AI enables CSAM generation that closed models prevent, while others dismiss the safety claims as exaggerated.
@BlancheMinerva The counter-example is Grok
@xeophon The answer to your question is CSAM generation, by the way. CSAM is an unambiguous win for closed source being safer. It’s the only example I know.
I have notes on why the WormGPT stuff was a total fraud, if that’s helpful to include too.
@BlancheMinerva But I agree
@BlancheMinerva @xeophon Really, really stretching the meaning of "safe" with that one.
EleutherAI executive director Stella Biderman singled out CSAM generation as the clearest case where closed-source models hold a safety edge over open ones, since the latter can be more readily adapted for that misuse; she framed it as the sole unambiguous example she knows of in current debates.
@xeophon The answer to your question is CSAM generation, by the way. CSAM is an unambiguous win for closed source being safer. It’s the only example I know.
I have notes on why the WormGPT stuff was a total fraud, if that’s helpful to include too.
yeah, looking further into this and it seems like open models are safer. blog about this later