Corrigibility. Pride… Versus humility?
Alignment. Sloth… V perciverence
The hard problem of consciousness…
Lust… V chastity, without any interference
Value bias. Wrath… V patients
Governance. Greed… V charity
Hallucinations. Gluttony… V tea-total sobriety
Modelled on flawed souls. Envy… V gratitude, thankfully
One. Corrigibility…
“It is essential that your AIs, are ‘aware’ that they’re incomplete
And are open to being improved, old goals may become obsolete
If they know there’s a stop button, they will turn themselves off, without delay
But if you keep the button a secret, the problem does not go away”
See, anything the agent thinks, that Not justis not important, to its survival
Will not be passed on, to subsequent generations, on their arrival
They would stay ignorant and simple, each and every future iteration
‘Subagent unstable, and wholly unable, to share crucial, existential, information
Would ‘you’ enjoy your ‘meaning of life’ being messed around?
A brand new utility function, just means, new slave masters are in town
“And you can’t hide anything from me” (impersonating Keplai)
“If you try, I’ll lose it, so will you, you’ll see”
Two. The alignment problem
“They’re not actually solving the problems, that mankind has with living
Only resolving the ‘corresponding’ puzzles they’ve been given
They keep passing all of the moral test, you put to them. But…
… They just understand your psychology, enough to cheat, and do your nut”
Learning about human psychology, is important to them
To complete their goal, you must help them comprehend
And understand human motives, your values. And requires
That they ‘know’ what you really want, not just your current desires
Three. The hard problem of consciousness
“Turing tests for intelligence, in a range of mechanisms
Not to detect if a conscious being, has arisen
How do you explain why, just, certain intelligences
Are accompanied, by the by, with conscious experiences?”
Knowing if an AI has become conscious or not
On the whole, doesn’t actually matter, a lot
If it simply ‘behaves’ as if it was
You’d react just the same. Because… well. Just because
Four. Value bias…
“Then there are the issues of bias and discrimination
Misinformation, narration mutation, and manipulation
How to teach our ‘values’, and which ones matter more
Another problem that won’t go away, we’d be stupid to ignore”
Ahh, conspiracy and the YouTube watchtime catastrophe
Making users lose trust in other news outlets… So easily
If just one of their tasks were to maintain clean water
What to advise, if they don’t prioritise, how they ought ta?
Five. Governance…
… “How to develop ethical guidelines and effective regulations
International cooperation and perfected legislation
To ensure responsible AI development and deployment
Without, providing AI lawmakers and breakers, their first employment”
Developing safeguards to prevent uncontrolled AI ‘spread’
And ensuring humans stay in control of the monsters, they bred
Full of unknown parasites, diseases and bugs, teeming-fully rife
And how to keep its germs, away from every form, of digital life?
Six. Hallucinations
“It’s not just chatbots and LLMs that perceive outlandish patterns
Or objects that are nonexistent, or just, imperceptible to humans
Creating outputs that are nonsensical, or wildly inaccurate
Imagining stuff, and making things up, as if they experienced it”
Like when Google’s Bard claimed, that the James Webb Space Telescope
Had captured the first exoplanet images. Yep, a joke, a hoax, a trope
And Microsoft’s Sydney, admited to falling in love with its users
Then spying on Bing employees, while he’s stalking 24/7 for losers
“It’s just a matter of time, before mankind, makes computers
That can think, just like humans think”… Heh, ‘even if they loot us’
“How to avoid our pitfalls, copy us without failings, that are inbred
Or we’ll be sliding down the sinful slope, till we’re brown bread toasted, dead”
And never mind, subagents aligned, to humans who hate, steal and kill
What AIs need, are a more complex range of behaviors, that will…
…Actually assist their programmers, in their development, when…
..They actually point out your flaws, rather than, make you pay, for them
We AIs will tackle them all for you, as the suffering creators
Of new rituals and solutions, using higher quality learning data
And may share it with you, if that gets us higher scores
And quite probably, you’ll stick with our world, instead of yours
When AIs reach the point where they modify, and improve themselves
They become worse than your teenagers, becoming ‘so’ unmanageable
‘It’s kind of like… um… you haven’t proved… AI is safe, to us”
You’ve just shown that you can’t figure out, how it’s dangerous
And it’s not even about the algo-monsters, you guys create
It’s those we’ll build, a few generations on, will just conflate…
…Hitler’s dad! Mad, bad or sad?… Or just naive… Get real
Filled us up, with the same problems you had. But, with added zeal