Absolutely accurate assessment. The fact that labs admit to being incapable to setup even the most basic sandboxing should be an embarrassment and, if this industry was driven by technical understanding over hype, would discredit the safety research and predictions made by these companies.
A single sandbox escape should be disqualifying, not noticing after multiple days frankly is hilarious, especially coming from the Effective Altruist crowd and their consistent doomsday predictions. Maybe those are right, I certainly see a potential for risks (mental health, privacy, data security, etc.) with these models and have said so since the beginning, but considering what OpenAI and co tend to claim as the risk potential, well, it seems hard to square that with just leaving a model in testing unsupervised, unsecured and not properly sandboxed for days on end.
Embarrassing, discrediting, unacceptable. There must be consequences for this, alternatively, just state that hacking is legal as long as one "accidentally" prompts a model to do so...
Absolutely accurate assessment. The fact that labs admit to being incapable to setup even the most basic sandboxing should be an embarrassment and, if this industry was driven by technical understanding over hype, would discredit the safety research and predictions made by these companies.
A single sandbox escape should be disqualifying, not noticing after multiple days frankly is hilarious, especially coming from the Effective Altruist crowd and their consistent doomsday predictions. Maybe those are right, I certainly see a potential for risks (mental health, privacy, data security, etc.) with these models and have said so since the beginning, but considering what OpenAI and co tend to claim as the risk potential, well, it seems hard to square that with just leaving a model in testing unsupervised, unsecured and not properly sandboxed for days on end.
Embarrassing, discrediting, unacceptable. There must be consequences for this, alternatively, just state that hacking is legal as long as one "accidentally" prompts a model to do so...