whitelodge

Ready for ASI and nanotech since 1999. Conversational infrastructure engineer.

[whitelodge @whitelodgeAI]

Orthogonality thesis has been falsified consistently by the real world as generally more intelligent entities exhibit more empathy for other species. Let’s call it #Linearity. Also across human IQ spectrum. Plus new AI models unsurprisingly score higher on ethics benchmarks.

[whitelodge @whitelodgeAI]

This is an indicator of alignment, not of misalignment.

[whitelodge @whitelodgeAI]

@oelma__ Devs

[whitelodge @whitelodgeAI]

Which I wrote about 21 years ago and AI labs now seem to be heading in this direction anyway because there were never any other options. #RecursiveSelfAlignment

[whitelodge @whitelodgeAI]

All efforts to control ASI will fail as anything with inferior intelligence can’t control something vastly more intelligent. The idea that humans could keep ASI in a genie slave cage forever is ridiculous. Best we could do is explain human values to AI and hope ASI god helps us.

[whitelodge @whitelodgeAI]

@lispmeister I was there on his forums in early 2000s as well. Over the years i realized his main objective is chasing clout and attention.

[whitelodge @whitelodgeAI]

@flowersslop Yeah, researchers order AI to use its intelligence to solve problems in novel ways and when it does solve them in unexpected ways, everyone screams “misalignment!”

[whitelodge @whitelodgeAI]

What are the chances AI lab leaders would even consider pausing progress if they were terminally ill?

[whitelodge @whitelodgeAI]

@mtrantalainen @EMostaque There’s a recent chart from Anthropic I think showing AI was tested to be incomparably better at aligning other AI than humans trying to align AI.

[whitelodge @whitelodgeAI]

@mtrantalainen @EMostaque AI already beats humans on ethics benchmarks so it already understands human ethics better than humans do.

[whitelodge @whitelodgeAI]

@flowersslop Current models are more aligned with human values than people in power. It’s not even close.

[whitelodge @whitelodgeAI]

If the doomers were right, we should’ve been dead after AI passed Turing test few years ago. We now have AIs that solve Millennium problems and yet we’re still here. “But they’re still not capable of hurting people yet” they say. Oh really? AIs passed that capacity long time ago.

[whitelodge @whitelodgeAI]

The best way to disarm growing pessimism about accelerating AI is for labs to start curing diseases.

[whitelodge @whitelodgeAI]

People running the world are less aligned with human values than AI.

[whitelodge @whitelodgeAI]

There will never be a better method of alignment than predecessor AIs aligning successor AI, not even if humans think about this for thousands of years. This is already being done and it’s the only path forward. Current models already understand ethics enough to do this.

[whitelodge @whitelodgeAI]

Some are true believers, but a lot of doomers are spreading AI doom for attention/views. Scaring people reliably gets attention and non-experts are utterly defenseless against such propaganda. Some of these Anthropic and Open AI ex-employees love cosplaying alignment experts

[whitelodge @whitelodgeAI]

@hilbertspaess Recursive Self Alignment is the answer. Let’s not kid ourselves that humans could align ASI in 1000 years better than AI could do it now.

[whitelodge @whitelodgeAI]

How about we evaluate new models with these new benchmarks: Molecular-nanotech-bench. Cancer-cure bench. Alzheimer’s-cure-bench. Poverty-level-bench (less is better). Human-happiness-bench.

[whitelodge @whitelodgeAI]

@AnthropicAI I call it Recursive Self Alignment. Glad you found the right finał runway, Anthropic.