Machine learning is neither good or evil | viznut

🌈 Abstract

The article discusses the author's complex relationship with "artificial intelligence" (AI) and their perspective on the current state of AI and machine learning (ML) technologies. The author explores their fascination with systems that can do "impossible" things, their exploration of neural networks and language models, and their concerns about the corporate adoption and application of these technologies.

🙋 Q&A

[01] The author's relationship with AI

1. What is the author's perspective on AI and ML technologies?

• The author has a strong scientific and hacker-like curiosity towards ML models, but they dislike how these technologies are being adopted by corporations and society.
• The author feels that both the skeptics and advocates of AI are misguided, and that simplistic judgement of these technologies is impossible due to the vast territory of possibilities they represent.
• The author is often obsessed with systems that can do "impossible" things, such as computer programs that can generate music or neural networks that can produce psychedelic imagery.

2. How does the author's background influence their perspective on AI?

• The author generally prefers systems that are relatively simple on the surface but have a lot of emergent complexity, such as elegant algorithms, simple formulas, and 8-bit microchips.
• The author finds the "black box" nature of large neural models to be somewhat alien and difficult to reverse-engineer, which they understand is a source of anxiety for "oldschool hackers".

[02] The history and evolution of AI and IA

1. What is the difference between AI and IA (intelligence amplification)?

• AI researchers aimed to create machines that would simulate human intelligence and compete with humans, while IA researchers wanted to create machines that would assist and "amplify" people's natural intellects.
• The author notes that the current framing of machine learning models as "AI" has brought back the old separation of AI and IA, and that we seem to have forgotten the decades of research in human-computer interaction.

2. How has the perception of AI and IA evolved over time?

• The author suggests that the "prompt'n'pray" interface of many current AI systems is particularly alienating to those who want to affect and control the output, in contrast to the "boss types" who prefer to give orders.
• The author believes that corporate generative ML will not get closer to IA ideals, as intelligent and knowledgeable users are "bad for business".

[03] The limitations of current generative ML models

1. What are the limitations of current generative ML models?

• Generative ML models are based on statistical probabilities and tend to output content that is popular or considered "safe", rather than unique or unconventional.
• The author suggests that to get something truly unique from a generative model, the seeds of uniqueness need to come from outside the "probability-and-popularity engine" of the model.

2. How do the author's views on "content" relate to the limitations of generative ML models?

• The author sees "content" as an industrial-capitalist concept that prioritizes the corporate platform over the creator, and believes that once all "content" becomes as cheap as "tap water", the current hierarchies and games of "content production" will be demolished.
• The author suggests that the culture will need to grow beyond "content" altogether, and that the most essential characteristic of computers - their flexibility - has been suppressed by oppressive ideologies and corporate interests.

[04] The author's vision for the future of AI and technology

1. What is the author's perspective on the polarization around AI and technology?

• The author believes that power dynamics of technology often create polarization, which weakens the ability to see the issues properly and eliminates the potential for "third options" or alternative visions.
• The author suggests that it is important to work on creating a "third option" that can gain traction once the current "AI bubble" bursts, rather than simply fighting a "strawman".

2. What kind of AI and technology is the author interested in?

• The author is mostly interested in ML and AI that stay away from the hype, such as classical hand-coded heuristics, small neural nets that do small things, and the unexplored gaps between AI and IA.
• The author believes that practicality may not matter as much as creating a cultural alternative to how the establishment is using technology, similar to the early days of hobbyist microcomputers.