undefined | Better HN

0 pointsmeowtimemania1y ago0 comments

Why would openai/anthropic's approach be more safe? Are people able to remove all the guard rails on the llama models?

0 comments

Yes, it's trivial to remove guardrails from any open access model:

https://www.lesswrong.com/posts/jGuXSZgv6qfdhMCuJ/refusal-in...

https://huggingface.co/failspy/Llama-3-70B-Instruct-ablitera...

Humanity is so fortunate this "guardrails" mentality didn't catch on when we started publishing books. While too close for comfort, we got twice lucky that computing wasn't hampered by this mentality either.

This time, humanity narrowly averted complete disaster thanks to the huge efforts and resources of a small number of people.

I wonder if we are witnessing humanity's the end of open knowledge and compute (at least until we pass through a neo dark ages and reach the next age of enlightenment).

Whether it'll be due to profit or control, it looks like humanity is posed to get fucked.

1 more reply

alfalfasprout1y ago

They're not safer. The claim is that OpenAI will enforce guard rails and take steps to ensure model outputs and prompts are responsible... but only a fool would take them at their word.

seoulmetro1y ago

Yeah.. and Facebook said they would enforce censorship on their platforms to ensure content safety.. that didn't turn out so well. Now it just censors anything remotely controversial, such as World War 2 historical facts or even just slightly offensive wording.

Spivak1y ago

You're really just arguing about the tuning. I get that it's annoying as a user but as a moderator going into it with the mentality that any post is expendable and bringing down the banhammer on everything near the line keeps things civil. HN does that too with the no flame-bait rule.

seoulmetro1y ago

HN moderation is quite poor and very subjective. The guidelines are not the site rules, the rules are made up on the spot.

HN censors too. Facebook just does it automatically on a huge scale with no reasoning behind each censor.

Censorship is just tuning people or things you don't want out. Censorship of your own content as a user is extremely annoying and Facebook's censorhsip is quite unethical. It doesn't help safety of the users, it helps safety of the business.

Also Facebook censors things that are not objectively not offensive in lots of instances. YouTube too. Safety for their brand.

koolala1y ago

Censorship isn't moderation.

zelphirkalt1y ago

The banhammer can quickly become a tool of net negative though, when actual facts are being repressed/censored.

j / k navigate · click thread line to collapse

0 comments

Der_Einzige1y ago

Yes, it's trivial to remove guardrails from any open access model:

https://www.lesswrong.com/posts/jGuXSZgv6qfdhMCuJ/refusal-in...

https://huggingface.co/failspy/Llama-3-70B-Instruct-ablitera...

Zuiii1y ago

This time, humanity narrowly averted complete disaster thanks to the huge efforts and resources of a small number of people.

I wonder if we are witnessing humanity's the end of open knowledge and compute (at least until we pass through a neo dark ages and reach the next age of enlightenment).

Whether it'll be due to profit or control, it looks like humanity is posed to get fucked.

1 more reply

alfalfasprout1y ago

They're not safer. The claim is that OpenAI will enforce guard rails and take steps to ensure model outputs and prompts are responsible... but only a fool would take them at their word.

seoulmetro1y ago

Spivak1y ago

seoulmetro1y ago

HN moderation is quite poor and very subjective. The guidelines are not the site rules, the rules are made up on the spot.

HN censors too. Facebook just does it automatically on a huge scale with no reasoning behind each censor.

Also Facebook censors things that are not objectively not offensive in lots of instances. YouTube too. Safety for their brand.

koolala1y ago

Censorship isn't moderation.

zelphirkalt1y ago

The banhammer can quickly become a tool of net negative though, when actual facts are being repressed/censored.

j / k navigate · click thread line to collapse