zlacker

>>JimDab+(OP)
I don’t quite buy your Cyberpunk utopia where the Megacorp finally rids us of those pesky ethics qualms (or ”shackles“, as you phrased it.) Microsoft can now proceed without the guidance of a council that actually has humanities interests in mind, not only those of Microsoft shareholders. I don’t know whether all that caution will turn out to have been necessary, but I guess we’re just gleefully heading into whatever lies ahead without any concern whatsoever, and learn it the hard way.

It’s a bit tragic that Ilya and company achieved the exact opposite of what they intended apparently, by driving those they attempted to slow down into the arms of people with more money and less morals. Well.

>>9dev+w9
Ilya should just go to Anthropic AI at this point. They have better momentum at this point after all this, and share his ideals. But it would be funny because they broke off of OpenAI because of their Microsoft ventures already in 2019, haha. He'd be welcomed with a big "We told you so!"

>>jug+ve
I don't consider Anthropic's approach to safety fantastic. They train the model to lie, play cat and mouse with jailbreakers, run moderation on generations with delay etc. This makes the model appear safer, as it's harder to jailbreak, but this approach solves nothing fundamentally.

If Ilya is concerned about safety and alignment, he probably has a better chance to get there with OpenAI, now the he has more control over it.

>>Athari+1l
I haven't paid a lot of attention to Anthropic. Are you able to summarize, or link anything about, those events for those who missed it? Particularly the "training to lie" bit

>>didntc+vv
David Shapiro complained about Anthropic's approach to alignment. In his video https://www.youtube.com/watch?v=PgwpqjiKkoY he discusses ableism, moralism, lying.

As to cat-and-mouse with jailbreakers, I don't remember any thorough articles or videos. It's mostly based on discussions on LLM forums. Claude is widely regarded as one of the best models for NSFW roleplay, which completely invalidates Antropic's claims about safety and alignment being "solved."