FSFLA/ blogs/ lxo/ stochastic parrots vs intelligence

FSFLA

Myself

Blonging for Freedom

This is Alexandre Oliva's personal blog, that does not necessarily represent the opinions held by FSFLA, LibrePlanet-BR, AdaCore or any other organization the author may be or may have been associated with, or even by myself, as of now or ever.

See also my home page, and follow me on the distributed P2P networks Twister (@lxoliva) and Secure Scuttlebutt (lxoliva), or on the federated and decentralized ones Diaspora, GNU social or Pump.io.

I'm also lxo in GNU Jami, and I have a Tox ID.

This blog has different entries in Portuguese, Spanish, and perhaps in other languages.

Legal matters

The following catch-all copyright notice and permission apply to all documents and postings in this blog that don't contain a copyright notice of their own, and whose copyright can be reasonably assumed to be held by Alexandre Oliva.

Copyright 2007-2024 Alexandre Oliva

Permission is granted to make and distribute verbatim copies of this entire document worldwide without royalty, provided the copyright notice, the document's official URL, and this permission notice are preserved.

The following licensing terms also apply to all documents and postings in this blog that don't contain a copyright notice of their own, or that contain a notice equivalent to the one above, and whose copyright can be reasonably assumed to be held by Alexandre Oliva.

This work is licensed under the Creative Commons License BY-SA (Attribution ShareAlike) 3.0 Unported. To see a copy of this license, visit http://creativecommons.org/licenses/by-sa/3.0/ or send a letter to Creative Commons, 444 Castro Street, Suite 900, Mountain View, California, 94041, USA.

Anti-piracy policy

Please do not attack ships, especially while visiting this site!

Machine learning is often framed as artificial intelligence.

Based on large statistical models, it can learn what is common.

That is only vaguely correlated with what is intelligent, let alone with what is good.

To figure out what is good, we need to reflect, to weight consequences, to identify benefits and risks to ourselves and to others.

The machine learning systems better understood as stochastic parrots don't do that.

They may seem to, when they're trained on datasets that embody our own collective intelligence.

But it's a lossy process. It's not just that it doesn't advance wisdom, it rather dilutes it in noise.

As training datasets encompass more and more data produced by such intelligence-diluting processes, the pearls that would advance our collective intelligence will become ever harder for us, the reflecting beings, to identify, select and promote in the sea of stochastic noise generated by the artificial parrots.

This vulnerability doesn't affect only these brute-force algorithms. Humans too are vulnerable to it. Social media filter bubbles, much older propaganda techniques and cults exploit just that: if we're surrounded by people (or bots) voicing a similar position, we tend to take that position for granted, as a reflection of collectively-selected wisdom.

It takes huge effort to overcome peer pressure and maintain and advance the unusual, surprising, even unreasonable positions that progress depends on.

Stochastic parrots, with all their computing power, can't crack us out of these serpent eggs: being unable to reason, they can only further entrap us, reinforcing already common sedimented positions.

They won't figure out that we've been fooled and warn us. They won't find solutions for climate catastrophe, for war, for famine, for inequity, for prejudice, for exploitation.

They are artificial parrots made to identify prevalent patterns and to boost them.

The prevalent systems of beliefs that they are made to reinforce are the ones that brought us these problems.

The pearls we wish and hope for are too rare to catch the stochastic parrots' algorithms' attention, and the extra noise the parrots output won't make the pearls any easier for us intelligent beings to find.

So blong...